BCa U-Net3D BraTS 2020 Segmentation¶
FileCopyrightText: Copyright (C) 2022 Ebtihal Alwadee AlwadeeEJ@cardiff.ac.uk, PhD student at Cardiff University
FileCopyrightText: Copyright (C) 2022-2023 Frank C Langbein frank@langbein.org, Cardiff UniversityLicense-Identifier: AGPL-3.0-or-later
We consider the BraTS 20202 dataset to explore the performance of a simple 3D UNet architecture for cancer segmentation. We test the same architecture across a range of input/output combinations, for whole tumour segmentation and segmenting various tumour regions. The purpose is to establish a baseline for performance evaluations and acquire some basic information about which MRI modality are best suited for the different tasks.
In [1]:
%load_ext autoreload
%autoreload 2
%matplotlib inline
import os
In [2]:
%%html
<style>
.dataframe th { font-size: 11px; }
.dataframe td { font-size: 11px; }
</style>
BraTS2020 Dataset¶
This dataset contains t1, t1ce, t2, and flair MRI data. We use a version where the images are cropped to the bounding box across all modalities (later scaled to 128x128x128) and one cropped to a fixed 128x128x128 cube (for some limited test).
In [3]:
# We assume https://qyber.black/ca/brainca-dev/code-bca is cloned into bca in the top folder
# This has been registered as a sub-module.
from bca.bca.dataset import Dataset
# Setup dataset
set="TrainingData"
##set="Small" # Smaller subset of the BraTs2020 dataset; for testing only
# Bounding-box cropped dataset - we use this for most experiments later
ds_bb = Dataset(os.path.join("data","brats2020","MICCAI_BraTS2020_"+set), # Dataset folder - brought in via a git submodule (not public, so replace as needed for repeats)
os.path.join("results","brats2020","MICCAI_BraTS2020_"+set)) # Results folder
ds_bb.crop_to_bb()
# Fixed crop dataset - only used to compare, briefly
ds_fi = Dataset(os.path.join("data","brats2020","MICCAI_BraTS2020_"+set), # Dataset folder (see above)
os.path.join("results","brats2020","MICCAI_BraTS2020_"+set)) # Cache folder
ds_fi.crop((56,184), (56,184), (13,141))
# Show bounding-box cropped dataset info and slices.
print(ds_bb)
ds_bb.browse()
# Dataset: data/brats2020/MICCAI_BraTS2020_TrainingData [369 patients] Cache: results/brats2020/MICCAI_BraTS2020_TrainingData Channels (patient BraTS20_Training_001): flair: (240, 240, 155) (int16) seg: (240, 240, 155) (uint8) t1: (240, 240, 155) (int16) t1ce: (240, 240, 155) (int16) t2: (240, 240, 155) (int16) Crop: bounding-box
interactive(children=(IntSlider(value=185, description='idx', max=369, min=1), IntSlider(value=78, description…
UNet3D with Dice Loss¶
We use a standard 3D UNet with few filters in encoder and decoder (due to memory) across all input/output and datset options.
In [4]:
from bca.bca.unet import UNet3D
from bca.bca.trainer import Trainer, dsc_loss, dsc, iou
import tensorflow.keras as keras
model = UNet3D(name="UNet3D_dice",
enc=[{"filters": 16},
{"filters": 32},
{"filters": 64},
{"filters":128, "kernel_regularizer":keras.regularizers.l2(0.02)},
{"filters":256, "kernel_regularizer":keras.regularizers.l2(0.02), "max_pooling":None}],
dec=[{"filters":128},
{"filters": 64},
{"filters": 32, "kernel_regularizer":keras.regularizers.l2(0.02)},
{"filters": 16, "kernel_regularizer":keras.regularizers.l2(0.02)}],
loss=dsc_loss,
metrics=[dsc,iou])
# Plot model structure for single output map; we need to fix an input shape to be able to do this.
# First part is shape of input, last entry in tuple is number of classes.
model.plot((128,128,128,4,1))
Segmentation Models: using `tf.keras` framework.
Experiments¶
We use a 5-fold cross-validation to explore model performance over a range of inputs and segmentation targets defined below. Each dataset is split identical using a seed.
Note, the preprocessed data (cropped according to the dataset version, scaled to 128x128x128, normalised to $[0,1]$) is not stored in the repo and will be recreated if the notebook is executed.
In [5]:
def setup(ds, inp_chs, out_chs):
# Setup training run for input/output channels inp_chs, out_chs for the dataset ds.
# See specification in the runs dict below.
# Parameters fixed across all runs
K = 5 # 5-fold cross validation
EPOCHS = 500 # 500 epochs
SEED = 8120341116777169704 # Seed for dataset 5-fold split
BATCH_SIZE = 4 # Batch size for training
# Generate a list of K keras (train,test)-pair sequences for dataset splits
seqs = ds.sequences(K, (128,128,128), inp_chs, out_chs,
batch_size=BATCH_SIZE, pre_proc=Dataset.norm_minmax, seed=SEED)
# Setup trainer for the model/data
trainer = Trainer(model, epochs=EPOCHS)
# We run this with a slurm scheduler later, so this just sets up the training task, and
# does not execute training (it would run the training if remote=False here immediately).
results = trainer.train(seqs, jit_compile=True, remote=True)
return trainer, seqs
import numpy as np
# Functions mapping output of network to standardised output
# Default function acting as simple threshold only; we set the name via lambda below
def std_eval(name, P, Y):
p = np.where(P[0] >= 0.5, 1.0, 0.0).astype(np.float32)
return { name: [p, Y[0]] }
# Specifically for the nested architectures below to map nested regions to individual regions
def std_eval_124_14_1(P,Y):
P[0] = np.where(P[0] >= 0.5, 1.0, 0.0).astype(np.float32)
r = {
'whl': [P[0][0,...,0], Y[0][0,...,0]], # Labels 1,2,4
'ede': [P[0][0,...,0]-P[0][0,...,1], Y[0][0,...,0]-Y[0][0,...,1]], # Label 2
'enh': [P[0][0,...,1]-P[0][0,...,2], Y[0][0,...,1]-Y[0][0,...,2]], # Label 4
'nec': [P[0][0,...,2], Y[0][0,...,2]], # Label 1
'cor': [P[0][0,...,1], Y[0][0,...,1]] # Labels 1, 4
}
for k in ['ede', 'enh']:
r[k] = [np.expand_dims(np.where(r[k][0] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4)),
np.expand_dims(np.where(r[k][1] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4))]
for k in ['whl', 'nec', 'cor']:
r[k] = [np.expand_dims(r[k][0], axis=(0,4)),
np.expand_dims(r[k][1], axis=(0,4))]
return r
def std_eval_124_14(P,Y):
P[0] = np.where(P[0] >= 0.5, 1.0, 0.0).astype(np.float32)
r = {
'whl': [P[0][0,...,0], Y[0][0,...,0]], # Labels 1,2,4
'ede': [P[0][0,...,0]-P[0][0,...,1], Y[0][0,...,0]-Y[0][0,...,1]], # Label 2
'cor': [P[0][0,...,1], Y[0][0,...,1]] # Labels 1, 4
}
for k in ['ede']:
r[k] = [np.expand_dims(np.where(r[k][0] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4)),
np.expand_dims(np.where(r[k][1] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4))]
for k in ['whl', 'cor']:
r[k] = [np.expand_dims(r[k][0], axis=(0,4)),
np.expand_dims(r[k][1], axis=(0,4))]
return r
def std_eval_124_2(P,Y):
P[0] = np.where(P[0] >= 0.5, 1.0, 0.0).astype(np.float32)
r = {
'whl': [P[0][0,...,0], Y[0][0,...,0]], # Labels 1,2,4
'ede': [P[0][0,...,1], Y[0][0,...,1]], # Label 2
'cor': [P[0][0,...,0]-P[0][0,...,1], Y[0][0,...,0]-Y[0][0,...,1]] # Labels 1, 4
}
for k in ['cor']:
r[k] = [np.expand_dims(np.where(r[k][0] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4)),
np.expand_dims(np.where(r[k][1] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4))]
for k in ['whl', 'ede']:
r[k] = [np.expand_dims(r[k][0], axis=(0,4)),
np.expand_dims(r[k][1], axis=(0,4))]
return r
def std_eval_1_4_2(P,Y):
P[0] = np.where(P[0] >= 0.5, 1.0, 0.0).astype(np.float32)
r = {
'whl': [P[0][0,...,0]+P[0][0,...,1]+P[0][0,...,2], Y[0][0,...,0]+Y[0][0,...,1]]+Y[0][0,...,2], # Labels 1,2,4
'ede': [P[0][0,...,2], Y[0][0,...,2]], # Label 2
'enh': [P[0][0,...,1], Y[0][0,...,1]], # Label 4
'nec': [P[0][0,...,0], Y[0][0,...,0]], # Label 1
'cor': [P[0][0,...,0]+P[0][0,...,1], Y[0][0,...,0]]+Y[0][0,...,1] # Labels 1, 4
}
for k in ['whl', 'cor']:
r[k] = [np.expand_dims(np.where(r[k][0] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4)),
np.expand_dims(np.where(r[k][1] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4))]
for k in ['ede', 'enh', 'nec']:
r[k] = [np.expand_dims(r[k][0], axis=(0,4)),
np.expand_dims(r[k][1], axis=(0,4))]
return r
def std_eval_14_2(P,Y):
P[0] = np.where(P[0] >= 0.5, 1.0, 0.0).astype(np.float32)
r = {
'whl': [P[0][0,...,0]+P[0][0,...,1], Y[0][0,...,0]+Y[0][0,...,1]], # Labels 1,2,4
'ede': [P[0][0,...,1], Y[0][0,...,1]], # Label 2
'cor': [P[0][0,...,0], Y[0][0,...,0]] # Labels 1, 4
}
for k in ['whl']:
r[k] = [np.expand_dims(np.where(r[k][0] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4)),
np.expand_dims(np.where(r[k][1] >= 0.5, 1.0, 0.0).astype(np.float32), axis=(0,4))]
for k in ['ede', 'cor']:
r[k] = [np.expand_dims(r[k][0], axis=(0,4)),
np.expand_dims(r[k][1], axis=(0,4))]
return r
# Dataset input/output options to create keras sequences for training.
# These are all options analysed in groups after training / evaluation.
# Dict: "NAME": DATASET, INPUT, OUTPUT, STD_EVAL (standardise output for evaluation, if needed)
runs = {
# Whole tumour with single modality
"bb_whole_t1": [ds_bb, ["t1"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
"bb_whole_t1ce": [ds_bb, ["t1ce"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
"bb_whole_t2": [ds_bb, ["t2"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
"bb_whole_flair": [ds_bb, ["flair"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
# Whole tumour with two modalities
"bb_whole_flair+t2": [ds_bb, ["flair","t2"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
# Whole tumour with three modalities
"bb_whole_flair+t1+t2": [ds_bb, ["flair","t1","t2"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
"bb_whole_flair+t1ce+t2": [ds_bb, ["flair","t1ce","t2"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
# Whole tumour with four modalities
"bb_whole_flair+t1+t1ce+t2": [ds_bb, ["flair","t1","t1ce","t2"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
# Whole tumour with fixed crop instead of bounding box
"fi_whole_flair+t1+t1ce+t2": [ds_fi, ["flair","t1","t1ce","t2"], ["seg+1+2+4"], lambda P, Y : std_eval("whl", P, Y)],
# Necrotic/non-enhancing tumour with single modality
"bb_necrotic_t1": [ds_bb, ["t1"], ["seg+1"], lambda P, Y : std_eval("nec", P, Y)],
"bb_necrotic_t1ce": [ds_bb, ["t1ce"], ["seg+1"], lambda P, Y : std_eval("nec", P, Y)],
"bb_necrotic_t2": [ds_bb, ["t2"], ["seg+1"], lambda P, Y : std_eval("nec", P, Y)],
"bb_necrotic_flair": [ds_bb, ["flair"], ["seg+1"], lambda P, Y : std_eval("nec", P, Y)],
# Enhancing tumour with single modality
"bb_enhancing_t1": [ds_bb, ["t1"], ["seg+4"], lambda P, Y : std_eval("enh", P, Y)],
"bb_enhancing_t1ce": [ds_bb, ["t1ce"], ["seg+4"], lambda P, Y : std_eval("enh", P, Y)],
"bb_enhancing_t2": [ds_bb, ["t2"], ["seg+4"], lambda P, Y : std_eval("enh", P, Y)],
"bb_enhancing_flair": [ds_bb, ["flair"], ["seg+4"], lambda P, Y : std_eval("enh", P, Y)],
# Peritumoural edema with single modality
"bb_edema_t1": [ds_bb, ["t1"], ["seg+2"], lambda P, Y : std_eval("ede", P, Y)],
"bb_edema_t1ce": [ds_bb, ["t1ce"], ["seg+2"], lambda P, Y : std_eval("ede", P, Y)],
"bb_edema_t2": [ds_bb, ["t2"], ["seg+2"], lambda P, Y : std_eval("ede", P, Y)],
"bb_edema_flair": [ds_bb, ["flair"], ["seg+2"], lambda P, Y : std_eval("ede", P, Y)],
# Core (Necrotic/non-enhancing + enhancing tumour) with single modality
"bb_core_t1": [ds_bb, ["t1"], ["seg+1+4"], lambda P, Y : std_eval("cor", P, Y)],
"bb_core_t1ce": [ds_bb, ["t1ce"], ["seg+1+4"], lambda P, Y : std_eval("cor", P, Y)],
"bb_core_t2": [ds_bb, ["t2"], ["seg+1+4"], lambda P, Y : std_eval("cor", P, Y)],
"bb_core_flair": [ds_bb, ["flair"], ["seg+1+4"], lambda P, Y : std_eval("cor", P, Y)],
# Nested/separated prediction of all regions with four modalities
"bb_nested_flair+t1+t1ce+t2": [ds_bb, ["flair","t1","t1ce","t2"], ["seg+1+2+4","seg+1+4","seg+1"], std_eval_124_14_1],
"bb_nestedCore_flair+t1+t1ce+t2": [ds_bb, ["flair","t1","t1ce","t2"], ["seg+1+2+4","seg+1+4"], std_eval_124_14],
"bb_nestedEdema_flair+t1+t1ce+t2": [ds_bb, ["flair","t1","t1ce","t2"], ["seg+1+2+4","seg+2"], std_eval_124_2],
"bb_separate_flair+t1+t1ce+t2": [ds_bb, ["flair","t1","t1ce","t2"], ["seg+1","seg+4","seg+2"], std_eval_1_4_2],
"bb_separateCore_flair+t1+t1ce+t2": [ds_bb, ["flair","t1","t1ce","t2"], ["seg+1+4","seg+2"], std_eval_14_2],
# Nested prediction of all regions with two/three modalities
"bb_nested_flair+t1ce": [ds_bb, ["flair","t1ce"], ["seg+1+2+4","seg+1+4","seg+1"], std_eval_124_14_1],
"bb_nested_flair+t1ce+t2": [ds_bb, ["flair","t1ce","t2"], ["seg+1+2+4","seg+1+4","seg+1"], std_eval_124_14_1]
}
# Generate the dataset and trainers for each dataset option.
# This generates the cropped/scaled/normalised cached samples for training,
# and sets the trainers up for running with the scheduler later.
trainers = {}
seqs = {}
for r in runs:
print(f"# {r}: in {runs[r][1]} - out {runs[r][2]}")
trainers[r], seqs[r] = setup(runs[r][0], runs[r][1], runs[r][2])
# bb_whole_t1: in ['t1'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_whole_t1ce: in ['t1ce'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_whole_t2: in ['t2'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_whole_flair: in ['flair'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_whole_flair+t2: in ['flair', 't2'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t2-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t2-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t2-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t2-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t2-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_whole_flair+t1+t2: in ['flair', 't1', 't2'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t2-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t2-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t2-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t2-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t2-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_whole_flair+t1ce+t2: in ['flair', 't1ce', 't2'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_whole_flair+t1+t1ce+t2: in ['flair', 't1', 't1ce', 't2'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # fi_whole_flair+t1+t1ce+t2: in ['flair', 't1', 't1ce', 't2'] - out ['seg+1+2+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/f56_184x56_184x13_141-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/f56_184x56_184x13_141-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/f56_184x56_184x13_141-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/f56_184x56_184x13_141-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/f56_184x56_184x13_141-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_necrotic_t1: in ['t1'] - out ['seg+1'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_necrotic_t1ce: in ['t1ce'] - out ['seg+1'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_necrotic_t2: in ['t2'] - out ['seg+1'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_necrotic_flair: in ['flair'] - out ['seg+1'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_enhancing_t1: in ['t1'] - out ['seg+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_enhancing_t1ce: in ['t1ce'] - out ['seg+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_enhancing_t2: in ['t2'] - out ['seg+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_enhancing_flair: in ['flair'] - out ['seg+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_edema_t1: in ['t1'] - out ['seg+2'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+2/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+2/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+2/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+2/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+2/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_edema_t1ce: in ['t1ce'] - out ['seg+2'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+2/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+2/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+2/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+2/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+2/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_edema_t2: in ['t2'] - out ['seg+2'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+2/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+2/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+2/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+2/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+2/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_edema_flair: in ['flair'] - out ['seg+2'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+2/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+2/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+2/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+2/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+2/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_core_t1: in ['t1'] - out ['seg+1+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1-seg+1+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_core_t1ce: in ['t1ce'] - out ['seg+1+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t1ce-seg+1+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_core_t2: in ['t2'] - out ['seg+1+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-t2-seg+1+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_core_flair: in ['flair'] - out ['seg+1+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair-seg+1+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_nested_flair+t1+t1ce+t2: in ['flair', 't1', 't1ce', 't2'] - out ['seg+1+2+4', 'seg+1+4', 'seg+1'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_nestedCore_flair+t1+t1ce+t2: in ['flair', 't1', 't1ce', 't2'] - out ['seg+1+2+4', 'seg+1+4'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+1+4/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_nestedEdema_flair+t1+t1ce+t2: in ['flair', 't1', 't1ce', 't2'] - out ['seg+1+2+4', 'seg+2'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+2/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+2/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+2/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+2/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+2+4_seg+2/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_separate_flair+t1+t1ce+t2: in ['flair', 't1', 't1ce', 't2'] - out ['seg+1', 'seg+4', 'seg+2'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1_seg+4_seg+2/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1_seg+4_seg+2/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1_seg+4_seg+2/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1_seg+4_seg+2/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1_seg+4_seg+2/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_separateCore_flair+t1+t1ce+t2: in ['flair', 't1', 't1ce', 't2'] - out ['seg+1+4', 'seg+2'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+4_seg+2/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+4_seg+2/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+4_seg+2/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+4_seg+2/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1_t1ce_t2-seg+1+4_seg+2/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_nested_flair+t1ce: in ['flair', 't1ce'] - out ['seg+1+2+4', 'seg+1+4', 'seg+1'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0 # bb_nested_flair+t1ce+t2: in ['flair', 't1ce', 't2'] - out ['seg+1+2+4', 'seg+1+4', 'seg+1'] * Fold 1: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-0 * Fold 2: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-1 * Fold 3: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-2 * Fold 4: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-3 * Fold 5: done - results/brats2020/MICCAI_BraTS2020_TrainingData/bb-128_128_128-norm_minmax/UNet3D_dice-flair_t1ce_t2-seg+1+2+4_seg+1+4_seg+1/500-4/8120341116777169704-5-4 => Done: 5; Failed: 0; Training: 0; Start: 0
Run¶
This runs all the training tasks not completed yet using whatever scheduler is defined via the bca code.
The schedulers are defined via the cfg.json file of bca.
Instead of running this in a jupyter notebook, it may be sipler to run schedule.py on the terminal. This
runs exactly the same commands. Repeatedly running the previous block will give a short summary of the
state of all relevant tasks, until everything is done. The ealuation below can be run anytime and will
update the evaluation results for the completed models.
In [6]:
from bca.bca.scheduler import schedule, schedule_clean
task_folder="results"
schedule_clean(task_folder=task_folder) # Cleanup failed task (assuming issues have been fixed)
schedule(task_folder=task_folder) # Schedule tasks
All tasks complete.
Evaluation¶
This runs the evaluation of the trained models locally, for any completely trained model not yet evaluated. Only models trained completly are evaluataed; everything else is ignored.
In [7]:
for t in trainers:
print(f"# {t}")
trainers[t].eval(seqs[t], std_eval=runs[t][3])
trainers[t].plot_model(seqs[t], save_only=True)
print("Evaluation complete.")
Evaluation complete.
Results¶
We present the results from the training runs, separated mainly over using different modalities and their combincations, and either nested segmentation regions or separate segmentation regions for segmenting multiple/all regions. This is mostly done on the bounding-box cropped, scaled to 128x128x128, and $[0,1]$ normalised dataset.
Note that below the "prime" (DSC', val_DSC', etc) values are per-sample metrics (and then averaged), while the others are the metrics used for training (per batch values, and then averaged). The STD values are the results from the standardised metrics using the standardised outputs with the std_eval functions above, making the values better comparable (for the dataset they were used on). They are averages over the values per sample (and then further averaged across the folds).
In [8]:
import numpy as np
import pandas as pd
from IPython import display
pd.set_option('display.max_columns', None)
def show_results(id):
# Show the results for the specified trainers and the evaluation summary
results={}
for t in id:
print(f"# {t} results")
r = trainers[t].plot_results(seqs[t])
if r is not None:
results[t] = r
if len(results) > 0:
print("Results Summary")
# Get mean/std performance results into a single pandas frame to report mean and std
# of performance metrics across faults per option analysed.
keys = []
for res in results:
for k in results[res].keys():
if 'loss' not in k and k not in keys:
keys.append(k)
data = np.zeros((2*len(results),len(keys)),dtype=np.float64())
idx = []
for l, res in enumerate(results):
data[2*l,:] = [results[res].at['Mean', k] if k in results[res].keys() else np.nan for k in keys]
data[2*l+1,:] = [results[res].at['Std', k] if k in results[res].keys() else np.nan for k in keys]
idx.append(f"{res} Mean")
idx.append(f"{res} Std")
df = pd.DataFrame(data, columns=keys, index=idx)
# Find minimum/maximum indices of mean performance results
idx_max = np.nanargmax(data[0::2], axis=0)*2
idx_min = np.nanargmin(data[0::2], axis=0)*2
def style_cells(idx_min,idx_max,x):
df1 = pd.DataFrame('', index=x.index, columns=x.columns)
for c,r in enumerate(idx_min):
if not '_c' in x.columns[c]: # skip as channels, as these can be inconsistent (careful, simple test, works with our metric names)
df1.iloc[r,c] = 'background-color: teal' # Color minimum teal per column
for c,r in enumerate(idx_max):
if not '_c' in x.columns[c]: # skip as channels, as these can be inconsistent (careful, simple test, works with our metric names)
df1.iloc[r,c] = 'background-color: green' # Color maximum green per column
return df1
return df.style.apply(lambda x : style_cells(idx_min,idx_max,x), axis=None)
Whole Tumour Segmentation with a Single Modality¶
We evaluate the 3D UNet model for all single modality inputs to segment the whole tumour (all non-background regions).
In [9]:
show_results(["bb_whole_t1", "bb_whole_t1ce", "bb_whole_t2", "bb_whole_flair"])
# bb_whole_t1 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.844994 | 0.701447 | 0.733085 | 0.545981 | 0.176450 | 0.320833 | 0.825442 | 0.635273 | 0.713646 | 0.500122 | 0.827472 | 0.636230 | 0.716633 | 0.501297 |
| Fold 2 | 0.861948 | 0.763849 | 0.758147 | 0.620386 | 0.158092 | 0.256538 | 0.835058 | 0.692753 | 0.731262 | 0.559868 | 0.835976 | 0.693437 | 0.732654 | 0.560726 |
| Fold 3 | 0.831730 | 0.701723 | 0.713381 | 0.544505 | 0.189938 | 0.320530 | 0.796173 | 0.637578 | 0.679533 | 0.498045 | 0.798243 | 0.639177 | 0.682437 | 0.499940 |
| Fold 4 | 0.858229 | 0.735982 | 0.752532 | 0.586641 | 0.164292 | 0.287428 | 0.837187 | 0.684954 | 0.729660 | 0.544998 | 0.838144 | 0.685671 | 0.731104 | 0.545897 |
| Fold 5 | 0.819657 | 0.648832 | 0.697093 | 0.489759 | 0.199703 | 0.374852 | 0.798066 | 0.620408 | 0.679824 | 0.477091 | 0.800250 | 0.621813 | 0.682805 | 0.478583 |
| Mean | 0.843311 | 0.710366 | 0.730848 | 0.557455 | 0.177695 | 0.312036 | 0.818385 | 0.654193 | 0.706785 | 0.516025 | 0.820017 | 0.655265 | 0.709127 | 0.517289 |
| Std | 0.015914 | 0.038620 | 0.023086 | 0.044032 | 0.015490 | 0.039449 | 0.017819 | 0.029012 | 0.022974 | 0.031157 | 0.017342 | 0.028712 | 0.022352 | 0.030855 |
# bb_whole_t1ce results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.769252 | 0.674515 | 0.629160 | 0.517854 | 0.253355 | 0.348021 | 0.727048 | 0.594008 | 0.597070 | 0.464608 | 0.728119 | 0.594841 | 0.598474 | 0.465583 |
| Fold 2 | 0.850417 | 0.742702 | 0.742337 | 0.593804 | 0.172336 | 0.279547 | 0.826566 | 0.687648 | 0.718360 | 0.548128 | 0.827516 | 0.688288 | 0.719756 | 0.548901 |
| Fold 3 | 0.774188 | 0.642441 | 0.635410 | 0.480888 | 0.249829 | 0.386385 | 0.712031 | 0.557335 | 0.583791 | 0.425678 | 0.713475 | 0.558428 | 0.585612 | 0.426799 |
| Fold 4 | 0.626999 | 0.605725 | 0.465587 | 0.441069 | 0.394487 | 0.420883 | 0.564739 | 0.513881 | 0.423422 | 0.380521 | 0.565591 | 0.514527 | 0.424375 | 0.381235 |
| Fold 5 | 0.851767 | 0.689648 | 0.742638 | 0.536444 | 0.173020 | 0.341186 | 0.812083 | 0.609954 | 0.705200 | 0.482556 | 0.813129 | 0.610625 | 0.706733 | 0.483343 |
| Mean | 0.774525 | 0.671006 | 0.643026 | 0.514012 | 0.248605 | 0.355204 | 0.728494 | 0.592565 | 0.605568 | 0.460298 | 0.729566 | 0.593342 | 0.606990 | 0.461172 |
| Std | 0.081876 | 0.045995 | 0.101509 | 0.051580 | 0.081038 | 0.047451 | 0.093482 | 0.057927 | 0.106181 | 0.056198 | 0.093516 | 0.057873 | 0.106341 | 0.056176 |
# bb_whole_t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.888529 | 0.842338 | 0.799982 | 0.729086 | 0.127617 | 0.173332 | 0.875042 | 0.809227 | 0.781414 | 0.699077 | 0.877122 | 0.811018 | 0.784613 | 0.701661 |
| Fold 2 | 0.864223 | 0.818430 | 0.761792 | 0.696882 | 0.150758 | 0.196543 | 0.848444 | 0.775628 | 0.742017 | 0.656075 | 0.853183 | 0.778969 | 0.748980 | 0.660703 |
| Fold 3 | 0.890825 | 0.824137 | 0.804027 | 0.706477 | 0.124453 | 0.186320 | 0.875746 | 0.793690 | 0.783711 | 0.679503 | 0.877753 | 0.795140 | 0.786830 | 0.681524 |
| Fold 4 | 0.887255 | 0.819155 | 0.797792 | 0.697430 | 0.129942 | 0.197681 | 0.874518 | 0.792395 | 0.780389 | 0.669990 | 0.876312 | 0.793980 | 0.783141 | 0.672159 |
| Fold 5 | 0.865103 | 0.805727 | 0.763092 | 0.677922 | 0.157177 | 0.218991 | 0.841924 | 0.770373 | 0.736209 | 0.646919 | 0.847537 | 0.775640 | 0.744372 | 0.653826 |
| Mean | 0.879187 | 0.821957 | 0.785337 | 0.701559 | 0.137990 | 0.194573 | 0.863135 | 0.788263 | 0.764748 | 0.670313 | 0.866381 | 0.790950 | 0.769587 | 0.673975 |
| Std | 0.011917 | 0.011865 | 0.018805 | 0.016611 | 0.013318 | 0.015027 | 0.014806 | 0.013898 | 0.021039 | 0.018231 | 0.013211 | 0.012707 | 0.018800 | 0.016787 |
# bb_whole_flair results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.908639 | 0.854074 | 0.833142 | 0.749718 | 0.108853 | 0.163835 | 0.897663 | 0.844842 | 0.816780 | 0.740067 | 0.899366 | 0.846165 | 0.819503 | 0.742042 |
| Fold 2 | 0.921982 | 0.882493 | 0.855577 | 0.792038 | 0.099596 | 0.140066 | 0.912832 | 0.858621 | 0.841692 | 0.766646 | 0.915015 | 0.860219 | 0.845347 | 0.769180 |
| Fold 3 | 0.919085 | 0.868868 | 0.850800 | 0.770502 | 0.105148 | 0.156459 | 0.908827 | 0.835862 | 0.835674 | 0.741926 | 0.910642 | 0.837301 | 0.838659 | 0.744107 |
| Fold 4 | 0.920667 | 0.871569 | 0.854144 | 0.774205 | 0.097128 | 0.147161 | 0.910485 | 0.846704 | 0.840607 | 0.746011 | 0.912352 | 0.848035 | 0.843662 | 0.748068 |
| Fold 5 | 0.904472 | 0.851423 | 0.827183 | 0.744515 | 0.112857 | 0.167316 | 0.885706 | 0.832944 | 0.806740 | 0.729748 | 0.887164 | 0.834158 | 0.809102 | 0.731542 |
| Mean | 0.914969 | 0.865685 | 0.844169 | 0.766196 | 0.104716 | 0.154967 | 0.903103 | 0.843795 | 0.828298 | 0.744879 | 0.904908 | 0.845176 | 0.831254 | 0.746988 |
| Std | 0.007055 | 0.011537 | 0.011694 | 0.017275 | 0.005786 | 0.010162 | 0.010141 | 0.009056 | 0.014019 | 0.012132 | 0.010354 | 0.009153 | 0.014396 | 0.012369 |
Results Summary
Out[9]:
| DSC | val_DSC | IoU | val_IoU | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bb_whole_t1 Mean | 0.843311 | 0.710366 | 0.730848 | 0.557455 | 0.818385 | 0.654193 | 0.706785 | 0.516025 | 0.820017 | 0.655265 | 0.709127 | 0.517289 |
| bb_whole_t1 Std | 0.015914 | 0.038620 | 0.023086 | 0.044032 | 0.017819 | 0.029012 | 0.022974 | 0.031157 | 0.017342 | 0.028712 | 0.022352 | 0.030855 |
| bb_whole_t1ce Mean | 0.774525 | 0.671006 | 0.643026 | 0.514012 | 0.728494 | 0.592565 | 0.605568 | 0.460298 | 0.729566 | 0.593342 | 0.606990 | 0.461172 |
| bb_whole_t1ce Std | 0.081876 | 0.045995 | 0.101509 | 0.051580 | 0.093482 | 0.057927 | 0.106181 | 0.056198 | 0.093516 | 0.057873 | 0.106341 | 0.056176 |
| bb_whole_t2 Mean | 0.879187 | 0.821957 | 0.785337 | 0.701559 | 0.863135 | 0.788263 | 0.764748 | 0.670313 | 0.866381 | 0.790950 | 0.769587 | 0.673975 |
| bb_whole_t2 Std | 0.011917 | 0.011865 | 0.018805 | 0.016611 | 0.014806 | 0.013898 | 0.021039 | 0.018231 | 0.013211 | 0.012707 | 0.018800 | 0.016787 |
| bb_whole_flair Mean | 0.914969 | 0.865685 | 0.844169 | 0.766196 | 0.903103 | 0.843795 | 0.828298 | 0.744879 | 0.904908 | 0.845176 | 0.831254 | 0.746988 |
| bb_whole_flair Std | 0.007055 | 0.011537 | 0.011694 | 0.017275 | 0.010141 | 0.009056 | 0.014019 | 0.012132 | 0.010354 | 0.009153 | 0.014396 | 0.012369 |
The flair and t2 modalities perform best, with flair being ideal here; t1 is better than t1ce and still reasonable; t1ce performs poorly. Validation indicates some overfitting in all cases.
Whole Tumour Segmentation with Multiple Modalities¶
We combine the best performing single modalities input multi-modality inputs to predict the whole tumour again. Here we also look at the fixed crop results, not just the bounding-box crop.
In [10]:
show_results(["bb_whole_flair+t2", "bb_whole_flair+t1+t2", "bb_whole_flair+t1ce+t2", "bb_whole_flair+t1+t1ce+t2", "fi_whole_flair+t1+t1ce+t2"])
# bb_whole_flair+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.898459 | 0.859028 | 0.816151 | 0.755824 | 0.117680 | 0.156766 | 0.886184 | 0.849789 | 0.798863 | 0.746465 | 0.889391 | 0.852540 | 0.803905 | 0.750614 |
| Fold 2 | 0.914740 | 0.880557 | 0.843411 | 0.788730 | 0.103867 | 0.138834 | 0.905095 | 0.860127 | 0.829258 | 0.767778 | 0.906724 | 0.861394 | 0.831940 | 0.769739 |
| Fold 3 | 0.900961 | 0.856266 | 0.820459 | 0.750712 | 0.113035 | 0.157297 | 0.887732 | 0.815074 | 0.802367 | 0.715804 | 0.889442 | 0.816324 | 0.805070 | 0.717694 |
| Fold 4 | 0.900158 | 0.861559 | 0.820178 | 0.758812 | 0.117435 | 0.156726 | 0.885655 | 0.835452 | 0.803398 | 0.732234 | 0.887550 | 0.837057 | 0.806483 | 0.734662 |
| Fold 5 | 0.900360 | 0.852539 | 0.819546 | 0.745148 | 0.113273 | 0.160773 | 0.880116 | 0.832431 | 0.796063 | 0.728228 | 0.882745 | 0.834575 | 0.800255 | 0.731426 |
| Mean | 0.902936 | 0.861990 | 0.823949 | 0.759845 | 0.113058 | 0.154079 | 0.888956 | 0.838574 | 0.805990 | 0.738102 | 0.891171 | 0.840378 | 0.809531 | 0.740827 |
| Std | 0.005960 | 0.009755 | 0.009853 | 0.015172 | 0.005001 | 0.007769 | 0.008470 | 0.015427 | 0.011921 | 0.017775 | 0.008151 | 0.015575 | 0.011393 | 0.017851 |
# bb_whole_flair+t1+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.903092 | 0.852108 | 0.823817 | 0.747126 | 0.112505 | 0.163396 | 0.889394 | 0.844461 | 0.804462 | 0.743562 | 0.891651 | 0.846478 | 0.808043 | 0.746596 |
| Fold 2 | 0.884718 | 0.873791 | 0.794463 | 0.777715 | 0.131552 | 0.143866 | 0.872836 | 0.848729 | 0.779657 | 0.750987 | 0.874910 | 0.850453 | 0.782892 | 0.753611 |
| Fold 3 | 0.904121 | 0.857734 | 0.825436 | 0.752736 | 0.109213 | 0.156184 | 0.893493 | 0.822031 | 0.809720 | 0.719382 | 0.894143 | 0.822515 | 0.810761 | 0.720089 |
| Fold 4 | 0.911084 | 0.865858 | 0.837678 | 0.765300 | 0.100996 | 0.146762 | 0.896722 | 0.836716 | 0.819007 | 0.735003 | 0.897995 | 0.837648 | 0.821057 | 0.736418 |
| Fold 5 | 0.900525 | 0.861180 | 0.819854 | 0.757833 | 0.111423 | 0.152548 | 0.887667 | 0.837163 | 0.802322 | 0.735624 | 0.888306 | 0.837674 | 0.803338 | 0.736392 |
| Mean | 0.900708 | 0.862134 | 0.820250 | 0.760142 | 0.113138 | 0.152551 | 0.888023 | 0.837820 | 0.803034 | 0.736912 | 0.889401 | 0.838954 | 0.805218 | 0.738621 |
| Std | 0.008726 | 0.007356 | 0.014202 | 0.010630 | 0.010052 | 0.006925 | 0.008225 | 0.009100 | 0.013030 | 0.010535 | 0.007907 | 0.009617 | 0.012582 | 0.011327 |
# bb_whole_flair+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.910921 | 0.864465 | 0.836917 | 0.765053 | 0.103250 | 0.150114 | 0.899958 | 0.846724 | 0.821377 | 0.749295 | 0.901209 | 0.847788 | 0.823397 | 0.750927 |
| Fold 2 | 0.904908 | 0.869597 | 0.827269 | 0.771888 | 0.108985 | 0.145399 | 0.893663 | 0.848846 | 0.812668 | 0.752461 | 0.894836 | 0.849774 | 0.814558 | 0.753867 |
| Fold 3 | 0.887926 | 0.853008 | 0.800192 | 0.746552 | 0.125755 | 0.162412 | 0.873293 | 0.797078 | 0.784181 | 0.699720 | 0.874781 | 0.798239 | 0.786528 | 0.701481 |
| Fold 4 | 0.900134 | 0.860106 | 0.819615 | 0.756300 | 0.116953 | 0.157848 | 0.885573 | 0.838832 | 0.801724 | 0.732233 | 0.887570 | 0.840558 | 0.804878 | 0.734785 |
| Fold 5 | 0.900217 | 0.861888 | 0.819742 | 0.759333 | 0.113226 | 0.152569 | 0.883112 | 0.839144 | 0.800102 | 0.738527 | 0.884856 | 0.840623 | 0.802901 | 0.740777 |
| Mean | 0.900821 | 0.861813 | 0.820747 | 0.759825 | 0.113634 | 0.153668 | 0.887120 | 0.834125 | 0.804010 | 0.734447 | 0.888650 | 0.835396 | 0.806452 | 0.736367 |
| Std | 0.007561 | 0.005442 | 0.012073 | 0.008505 | 0.007583 | 0.005936 | 0.009135 | 0.018949 | 0.012572 | 0.018828 | 0.008988 | 0.018947 | 0.012368 | 0.018747 |
# bb_whole_flair+t1+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.890522 | 0.836671 | 0.804013 | 0.725487 | 0.125974 | 0.180345 | 0.869114 | 0.816980 | 0.782389 | 0.712822 | 0.871242 | 0.818932 | 0.785720 | 0.715697 |
| Fold 2 | 0.903852 | 0.876937 | 0.825133 | 0.782948 | 0.108185 | 0.135024 | 0.891301 | 0.856593 | 0.807690 | 0.763701 | 0.892281 | 0.857453 | 0.809233 | 0.765008 |
| Fold 3 | 0.905919 | 0.862193 | 0.828580 | 0.759250 | 0.105038 | 0.148670 | 0.895848 | 0.830421 | 0.814213 | 0.726626 | 0.897115 | 0.831363 | 0.816253 | 0.728007 |
| Fold 4 | 0.907746 | 0.865860 | 0.831610 | 0.764747 | 0.105372 | 0.147789 | 0.897316 | 0.848064 | 0.816316 | 0.744422 | 0.898211 | 0.848830 | 0.817760 | 0.745543 |
| Fold 5 | 0.904384 | 0.872875 | 0.826116 | 0.776039 | 0.111543 | 0.145425 | 0.892695 | 0.859228 | 0.809488 | 0.761015 | 0.894656 | 0.860942 | 0.812627 | 0.763604 |
| Mean | 0.902485 | 0.862907 | 0.823091 | 0.761694 | 0.111222 | 0.151451 | 0.889255 | 0.842257 | 0.806019 | 0.741717 | 0.890701 | 0.843504 | 0.808319 | 0.743572 |
| Std | 0.006133 | 0.014098 | 0.009798 | 0.019918 | 0.007738 | 0.015246 | 0.010297 | 0.016164 | 0.012218 | 0.019625 | 0.009943 | 0.015989 | 0.011681 | 0.019411 |
# fi_whole_flair+t1+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.902794 | 0.886608 | 0.823414 | 0.797397 | 0.109691 | 0.125740 | 0.891512 | 0.867874 | 0.807883 | 0.776553 | 0.892610 | 0.868868 | 0.809610 | 0.778042 |
| Fold 2 | 0.905444 | 0.886931 | 0.827864 | 0.798481 | 0.107504 | 0.126486 | 0.890724 | 0.865905 | 0.808204 | 0.776193 | 0.891463 | 0.866458 | 0.809395 | 0.777080 |
| Fold 3 | 0.918697 | 0.875000 | 0.850062 | 0.780081 | 0.091895 | 0.136571 | 0.907763 | 0.836112 | 0.833953 | 0.741232 | 0.908772 | 0.836819 | 0.835596 | 0.742305 |
| Fold 4 | 0.919316 | 0.877490 | 0.851154 | 0.783293 | 0.097718 | 0.140323 | 0.909556 | 0.852361 | 0.836634 | 0.752647 | 0.911749 | 0.854080 | 0.840252 | 0.755282 |
| Fold 5 | 0.921537 | 0.879444 | 0.854884 | 0.786241 | 0.090318 | 0.130836 | 0.911124 | 0.859446 | 0.839115 | 0.766710 | 0.912198 | 0.860294 | 0.840868 | 0.768019 |
| Mean | 0.913557 | 0.881094 | 0.841476 | 0.789099 | 0.099425 | 0.131991 | 0.902136 | 0.856339 | 0.825158 | 0.762667 | 0.903359 | 0.857304 | 0.827144 | 0.764146 |
| Std | 0.007809 | 0.004844 | 0.013105 | 0.007484 | 0.007915 | 0.005677 | 0.009062 | 0.011482 | 0.014069 | 0.013798 | 0.009326 | 0.011458 | 0.014519 | 0.013645 |
Results Summary
Out[10]:
| DSC | val_DSC | IoU | val_IoU | DSC' | val_DSC' | IoU' | val_IoU' | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bb_whole_flair+t2 Mean | 0.902936 | 0.861990 | 0.823949 | 0.759845 | 0.888956 | 0.838574 | 0.805990 | 0.738102 | 0.891171 | 0.840378 | 0.809531 | 0.740827 |
| bb_whole_flair+t2 Std | 0.005960 | 0.009755 | 0.009853 | 0.015172 | 0.008470 | 0.015427 | 0.011921 | 0.017775 | 0.008151 | 0.015575 | 0.011393 | 0.017851 |
| bb_whole_flair+t1+t2 Mean | 0.900708 | 0.862134 | 0.820250 | 0.760142 | 0.888023 | 0.837820 | 0.803034 | 0.736912 | 0.889401 | 0.838954 | 0.805218 | 0.738621 |
| bb_whole_flair+t1+t2 Std | 0.008726 | 0.007356 | 0.014202 | 0.010630 | 0.008225 | 0.009100 | 0.013030 | 0.010535 | 0.007907 | 0.009617 | 0.012582 | 0.011327 |
| bb_whole_flair+t1ce+t2 Mean | 0.900821 | 0.861813 | 0.820747 | 0.759825 | 0.887120 | 0.834125 | 0.804010 | 0.734447 | 0.888650 | 0.835396 | 0.806452 | 0.736367 |
| bb_whole_flair+t1ce+t2 Std | 0.007561 | 0.005442 | 0.012073 | 0.008505 | 0.009135 | 0.018949 | 0.012572 | 0.018828 | 0.008988 | 0.018947 | 0.012368 | 0.018747 |
| bb_whole_flair+t1+t1ce+t2 Mean | 0.902485 | 0.862907 | 0.823091 | 0.761694 | 0.889255 | 0.842257 | 0.806019 | 0.741717 | 0.890701 | 0.843504 | 0.808319 | 0.743572 |
| bb_whole_flair+t1+t1ce+t2 Std | 0.006133 | 0.014098 | 0.009798 | 0.019918 | 0.010297 | 0.016164 | 0.012218 | 0.019625 | 0.009943 | 0.015989 | 0.011681 | 0.019411 |
| fi_whole_flair+t1+t1ce+t2 Mean | 0.913557 | 0.881094 | 0.841476 | 0.789099 | 0.902136 | 0.856339 | 0.825158 | 0.762667 | 0.903359 | 0.857304 | 0.827144 | 0.764146 |
| fi_whole_flair+t1+t1ce+t2 Std | 0.007809 | 0.004844 | 0.013105 | 0.007484 | 0.009062 | 0.011482 | 0.014069 | 0.013798 | 0.009326 | 0.011458 | 0.014519 | 0.013645 |
Interestingly, combining the modalities does not improve the prediction; flair seems sufficient for whole tumour segmentation and maybe even better. There is a similar indication of overfitting as for the single modalities.
We see a slightly higher score for the fixed crop, but as this runs on a different dataset it is not easy to directly compare this and the values are not significantly larger.
Potentially, relations between modalities are more complex and convolutions across the 3D modalities, then averaged across the modalities, cannot find these, if they are there at all. 4D convolutions or differently aligned convolutions or completely different arrangements may produce different results.
Necrotic/Non-Enhancing Tumour Segmentation with a Single Modality¶
These are the results using single modalities for the necrotic/non-enhancing region.
In [11]:
show_results(["bb_necrotic_t1", "bb_necrotic_t1ce", "bb_necrotic_t2", "bb_necrotic_flair"])
# bb_necrotic_t1 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.656041 | 0.405809 | 0.503492 | 0.265205 | 0.381665 | 0.635162 | 0.498993 | 0.268924 | 0.385142 | 0.188863 | 0.503214 | 0.268946 | 0.389484 | 0.188959 |
| Fold 2 | 0.674565 | 0.393212 | 0.519846 | 0.265471 | 0.363115 | 0.639089 | 0.523817 | 0.306877 | 0.404105 | 0.214586 | 0.528672 | 0.306901 | 0.409163 | 0.214764 |
| Fold 3 | 0.601155 | 0.313405 | 0.446056 | 0.198437 | 0.429707 | 0.725726 | 0.456805 | 0.201798 | 0.344510 | 0.134233 | 0.457952 | 0.202879 | 0.345767 | 0.135184 |
| Fold 4 | 0.456596 | 0.347459 | 0.316807 | 0.223600 | 0.567172 | 0.675457 | 0.332410 | 0.249500 | 0.238824 | 0.171755 | 0.333911 | 0.250589 | 0.240335 | 0.172850 |
| Fold 5 | 0.525799 | 0.364369 | 0.374755 | 0.233629 | 0.504129 | 0.673196 | 0.401104 | 0.313567 | 0.292938 | 0.213573 | 0.403590 | 0.328568 | 0.295445 | 0.228731 |
| Mean | 0.582831 | 0.364851 | 0.432191 | 0.237268 | 0.449157 | 0.669726 | 0.442626 | 0.268133 | 0.333104 | 0.184602 | 0.445468 | 0.271577 | 0.336039 | 0.188098 |
| Std | 0.081607 | 0.032958 | 0.076905 | 0.025627 | 0.076532 | 0.032601 | 0.069054 | 0.040776 | 0.060632 | 0.029861 | 0.070160 | 0.043960 | 0.061815 | 0.032849 |
# bb_necrotic_t1ce results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.751874 | 0.545931 | 0.614698 | 0.403112 | 0.286226 | 0.497017 | 0.637632 | 0.454941 | 0.522781 | 0.347884 | 0.645263 | 0.457697 | 0.531686 | 0.351188 |
| Fold 2 | 0.754268 | 0.627476 | 0.612624 | 0.470202 | 0.282785 | 0.409050 | 0.636048 | 0.508094 | 0.516765 | 0.390531 | 0.646106 | 0.512814 | 0.528213 | 0.396130 |
| Fold 3 | 0.710248 | 0.479992 | 0.559527 | 0.342923 | 0.326347 | 0.565098 | 0.591932 | 0.379037 | 0.469703 | 0.277540 | 0.598468 | 0.381492 | 0.476969 | 0.280149 |
| Fold 4 | 0.783534 | 0.591127 | 0.651220 | 0.455104 | 0.263196 | 0.462113 | 0.673991 | 0.486389 | 0.554942 | 0.384402 | 0.681421 | 0.491049 | 0.564479 | 0.391144 |
| Fold 5 | 0.746503 | 0.637842 | 0.606202 | 0.485826 | 0.292610 | 0.406889 | 0.619582 | 0.512216 | 0.500317 | 0.403340 | 0.624973 | 0.514827 | 0.506919 | 0.407099 |
| Mean | 0.749285 | 0.576474 | 0.608854 | 0.431433 | 0.290233 | 0.468033 | 0.631837 | 0.468135 | 0.512902 | 0.360739 | 0.639246 | 0.471576 | 0.521653 | 0.365142 |
| Std | 0.023393 | 0.058007 | 0.029269 | 0.052265 | 0.020554 | 0.059185 | 0.026724 | 0.048960 | 0.027944 | 0.045497 | 0.027308 | 0.049514 | 0.028953 | 0.046509 |
# bb_necrotic_t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.679164 | 0.398174 | 0.523368 | 0.258413 | 0.362546 | 0.647488 | 0.511967 | 0.299043 | 0.391157 | 0.206366 | 0.517140 | 0.299676 | 0.396572 | 0.206992 |
| Fold 2 | 0.606156 | 0.425600 | 0.448369 | 0.290319 | 0.424524 | 0.597514 | 0.475054 | 0.331540 | 0.351963 | 0.234276 | 0.480815 | 0.332023 | 0.357628 | 0.234918 |
| Fold 3 | 0.722747 | 0.365344 | 0.575934 | 0.240643 | 0.311993 | 0.664525 | 0.583961 | 0.266561 | 0.457839 | 0.182863 | 0.588913 | 0.267195 | 0.463034 | 0.183372 |
| Fold 4 | 0.649624 | 0.409539 | 0.495876 | 0.272264 | 0.387470 | 0.621778 | 0.529053 | 0.305748 | 0.400360 | 0.208174 | 0.533934 | 0.305966 | 0.405254 | 0.208486 |
| Fold 5 | 0.481049 | 0.296111 | 0.331487 | 0.179532 | 0.562038 | 0.751826 | 0.370610 | 0.279871 | 0.261877 | 0.185876 | 0.373361 | 0.283580 | 0.264294 | 0.189416 |
| Mean | 0.627748 | 0.378954 | 0.475007 | 0.248234 | 0.409714 | 0.656626 | 0.494129 | 0.296553 | 0.372639 | 0.203511 | 0.498833 | 0.297688 | 0.377356 | 0.204637 |
| Std | 0.082624 | 0.045889 | 0.082787 | 0.038026 | 0.084488 | 0.052759 | 0.071033 | 0.022328 | 0.064910 | 0.018513 | 0.071763 | 0.021819 | 0.065819 | 0.018006 |
# bb_necrotic_flair results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.623752 | 0.388735 | 0.467427 | 0.251803 | 0.410989 | 0.643000 | 0.504066 | 0.309698 | 0.380714 | 0.212399 | 0.505811 | 0.310110 | 0.382517 | 0.212902 |
| Fold 2 | 0.640693 | 0.398043 | 0.479623 | 0.255825 | 0.394670 | 0.634910 | 0.508813 | 0.312456 | 0.380426 | 0.213095 | 0.510552 | 0.313156 | 0.382221 | 0.213770 |
| Fold 3 | 0.540229 | 0.334970 | 0.390134 | 0.209123 | 0.504799 | 0.712760 | 0.442778 | 0.247390 | 0.323324 | 0.163843 | 0.447664 | 0.248037 | 0.328268 | 0.164415 |
| Fold 4 | 0.670525 | 0.330933 | 0.521723 | 0.206980 | 0.370739 | 0.705997 | 0.551475 | 0.272476 | 0.424786 | 0.180614 | 0.552709 | 0.272747 | 0.426254 | 0.180893 |
| Fold 5 | 0.670454 | 0.361964 | 0.512994 | 0.227043 | 0.370204 | 0.680094 | 0.529100 | 0.317084 | 0.397376 | 0.213381 | 0.529732 | 0.317444 | 0.398011 | 0.213766 |
| Mean | 0.629131 | 0.362929 | 0.474380 | 0.230155 | 0.410280 | 0.675352 | 0.507246 | 0.291821 | 0.381325 | 0.196666 | 0.509294 | 0.292299 | 0.383454 | 0.197149 |
| Std | 0.047932 | 0.027223 | 0.046702 | 0.020575 | 0.049694 | 0.031759 | 0.036339 | 0.027320 | 0.033202 | 0.020649 | 0.034970 | 0.027298 | 0.031911 | 0.020670 |
Results Summary
Out[11]:
| DSC | val_DSC | IoU | val_IoU | DSC' | val_DSC' | IoU' | val_IoU' | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bb_necrotic_t1 Mean | 0.582831 | 0.364851 | 0.432191 | 0.237268 | 0.442626 | 0.268133 | 0.333104 | 0.184602 | 0.445468 | 0.271577 | 0.336039 | 0.188098 |
| bb_necrotic_t1 Std | 0.081607 | 0.032958 | 0.076905 | 0.025627 | 0.069054 | 0.040776 | 0.060632 | 0.029861 | 0.070160 | 0.043960 | 0.061815 | 0.032849 |
| bb_necrotic_t1ce Mean | 0.749285 | 0.576474 | 0.608854 | 0.431433 | 0.631837 | 0.468135 | 0.512902 | 0.360739 | 0.639246 | 0.471576 | 0.521653 | 0.365142 |
| bb_necrotic_t1ce Std | 0.023393 | 0.058007 | 0.029269 | 0.052265 | 0.026724 | 0.048960 | 0.027944 | 0.045497 | 0.027308 | 0.049514 | 0.028953 | 0.046509 |
| bb_necrotic_t2 Mean | 0.627748 | 0.378954 | 0.475007 | 0.248234 | 0.494129 | 0.296553 | 0.372639 | 0.203511 | 0.498833 | 0.297688 | 0.377356 | 0.204637 |
| bb_necrotic_t2 Std | 0.082624 | 0.045889 | 0.082787 | 0.038026 | 0.071033 | 0.022328 | 0.064910 | 0.018513 | 0.071763 | 0.021819 | 0.065819 | 0.018006 |
| bb_necrotic_flair Mean | 0.629131 | 0.362929 | 0.474380 | 0.230155 | 0.507246 | 0.291821 | 0.381325 | 0.196666 | 0.509294 | 0.292299 | 0.383454 | 0.197149 |
| bb_necrotic_flair Std | 0.047932 | 0.027223 | 0.046702 | 0.020575 | 0.036339 | 0.027320 | 0.033202 | 0.020649 | 0.034970 | 0.027298 | 0.031911 | 0.020670 |
No option performs well here. T1ce outperforms all other options noticeably. There is some overfitting.
Enhancing Tumour Segmentation with a Single Modality¶
These are the single modality results for the enhacning region.
In [12]:
show_results(["bb_enhancing_t1", "bb_enhancing_t1ce", "bb_enhancing_t2", "bb_enhancing_flair"])
# bb_enhancing_t1 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.553917 | 0.219573 | 0.388752 | 0.131892 | 0.489504 | 0.831517 | 0.428809 | 0.157115 | 0.307746 | 0.097866 | 0.475405 | 0.170600 | 0.354635 | 0.111408 |
| Fold 2 | 0.559332 | 0.254285 | 0.393864 | 0.152825 | 0.485932 | 0.789677 | 0.437451 | 0.186792 | 0.311983 | 0.117897 | 0.457135 | 0.214009 | 0.331652 | 0.145204 |
| Fold 3 | 0.488949 | 0.230685 | 0.330127 | 0.135551 | 0.550279 | 0.811768 | 0.374664 | 0.167843 | 0.258816 | 0.103228 | 0.393409 | 0.181547 | 0.277526 | 0.116944 |
| Fold 4 | 0.616552 | 0.218351 | 0.451089 | 0.125420 | 0.426240 | 0.829386 | 0.479263 | 0.173017 | 0.354406 | 0.103373 | 0.521246 | 0.172897 | 0.396498 | 0.103393 |
| Fold 5 | 0.573644 | 0.249950 | 0.407695 | 0.150042 | 0.464240 | 0.779258 | 0.439365 | 0.198965 | 0.315952 | 0.125632 | 0.478074 | 0.212621 | 0.354741 | 0.139380 |
| Mean | 0.558479 | 0.234569 | 0.394305 | 0.139146 | 0.483239 | 0.808321 | 0.431911 | 0.176747 | 0.309781 | 0.109599 | 0.465054 | 0.190335 | 0.343010 | 0.123266 |
| Std | 0.041130 | 0.015022 | 0.038868 | 0.010581 | 0.040375 | 0.020909 | 0.033512 | 0.014650 | 0.030450 | 0.010420 | 0.041545 | 0.019120 | 0.038861 | 0.016226 |
# bb_enhancing_t1ce results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.804738 | 0.700303 | 0.675877 | 0.562849 | 0.235871 | 0.343928 | 0.695454 | 0.587474 | 0.579104 | 0.477925 | 0.721974 | 0.604975 | 0.607522 | 0.496623 |
| Fold 2 | 0.836366 | 0.688629 | 0.721355 | 0.553872 | 0.198630 | 0.348329 | 0.731807 | 0.599209 | 0.625278 | 0.486377 | 0.760376 | 0.642847 | 0.655643 | 0.530950 |
| Fold 3 | 0.789248 | 0.705611 | 0.656010 | 0.561762 | 0.241711 | 0.329688 | 0.659586 | 0.612269 | 0.544511 | 0.497640 | 0.683645 | 0.642134 | 0.569750 | 0.528318 |
| Fold 4 | 0.818727 | 0.738416 | 0.697004 | 0.595116 | 0.215155 | 0.298358 | 0.697019 | 0.666306 | 0.588303 | 0.543371 | 0.777902 | 0.668531 | 0.670238 | 0.546230 |
| Fold 5 | 0.839091 | 0.754406 | 0.724886 | 0.622601 | 0.196856 | 0.287083 | 0.714084 | 0.683063 | 0.610279 | 0.560003 | 0.743723 | 0.701969 | 0.642211 | 0.580200 |
| Mean | 0.817634 | 0.717473 | 0.695027 | 0.579240 | 0.217644 | 0.321477 | 0.699590 | 0.629664 | 0.589495 | 0.513063 | 0.737524 | 0.652091 | 0.629073 | 0.536464 |
| Std | 0.018892 | 0.024780 | 0.026373 | 0.025884 | 0.018500 | 0.024535 | 0.023967 | 0.037958 | 0.027732 | 0.032578 | 0.032667 | 0.032133 | 0.036213 | 0.027175 |
# bb_enhancing_t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.563666 | 0.287766 | 0.399343 | 0.177885 | 0.477191 | 0.758798 | 0.466333 | 0.226635 | 0.336541 | 0.143677 | 0.501219 | 0.226916 | 0.371599 | 0.143917 |
| Fold 2 | 0.368155 | 0.248420 | 0.232924 | 0.147915 | 0.677491 | 0.796391 | 0.287177 | 0.207954 | 0.188004 | 0.129373 | 0.301104 | 0.248764 | 0.201945 | 0.170136 |
| Fold 3 | 0.515590 | 0.281677 | 0.352833 | 0.169468 | 0.521733 | 0.749821 | 0.385684 | 0.195806 | 0.269000 | 0.122517 | 0.451011 | 0.238031 | 0.334028 | 0.164386 |
| Fold 4 | 0.203035 | 0.175204 | 0.116822 | 0.100704 | 0.849164 | 0.872818 | 0.152317 | 0.139934 | 0.091983 | 0.084834 | 0.152363 | 0.139907 | 0.092021 | 0.084827 |
| Fold 5 | 0.622182 | 0.282950 | 0.458129 | 0.175227 | 0.418240 | 0.763186 | 0.498302 | 0.232518 | 0.369435 | 0.147884 | 0.560378 | 0.246248 | 0.431614 | 0.161700 |
| Mean | 0.454526 | 0.255203 | 0.312010 | 0.154240 | 0.588764 | 0.788203 | 0.357963 | 0.200569 | 0.250993 | 0.125657 | 0.393215 | 0.219973 | 0.286241 | 0.144993 |
| Std | 0.151299 | 0.042372 | 0.122472 | 0.028766 | 0.156013 | 0.045145 | 0.126102 | 0.033027 | 0.100919 | 0.022401 | 0.147948 | 0.040753 | 0.122907 | 0.031332 |
# bb_enhancing_flair results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.571206 | 0.303068 | 0.405230 | 0.186871 | 0.474024 | 0.744970 | 0.458524 | 0.254953 | 0.327705 | 0.165160 | 0.486919 | 0.255190 | 0.356146 | 0.165471 |
| Fold 2 | 0.434619 | 0.265831 | 0.286585 | 0.161864 | 0.592477 | 0.765359 | 0.346333 | 0.232330 | 0.235509 | 0.149164 | 0.364341 | 0.272918 | 0.253512 | 0.189762 |
| Fold 3 | 0.407635 | 0.308152 | 0.261740 | 0.191345 | 0.624979 | 0.721293 | 0.321348 | 0.264357 | 0.214607 | 0.171978 | 0.342025 | 0.278091 | 0.235253 | 0.185683 |
| Fold 4 | 0.563186 | 0.290972 | 0.400207 | 0.179221 | 0.485234 | 0.751334 | 0.441012 | 0.245042 | 0.319138 | 0.156364 | 0.493006 | 0.245295 | 0.371254 | 0.156654 |
| Fold 5 | 0.517836 | 0.323708 | 0.356734 | 0.199918 | 0.516741 | 0.717052 | 0.420288 | 0.272213 | 0.294034 | 0.172731 | 0.438845 | 0.272739 | 0.312641 | 0.173135 |
| Mean | 0.498896 | 0.298346 | 0.342099 | 0.183844 | 0.538691 | 0.740002 | 0.397501 | 0.253779 | 0.278199 | 0.163079 | 0.425027 | 0.264847 | 0.305761 | 0.174141 |
| Std | 0.066605 | 0.019356 | 0.058507 | 0.012871 | 0.059766 | 0.018291 | 0.053951 | 0.014067 | 0.045264 | 0.009115 | 0.061992 | 0.012477 | 0.053994 | 0.012322 |
Results Summary
Out[12]:
| DSC | val_DSC | IoU | val_IoU | DSC' | val_DSC' | IoU' | val_IoU' | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bb_enhancing_t1 Mean | 0.558479 | 0.234569 | 0.394305 | 0.139146 | 0.431911 | 0.176747 | 0.309781 | 0.109599 | 0.465054 | 0.190335 | 0.343010 | 0.123266 |
| bb_enhancing_t1 Std | 0.041130 | 0.015022 | 0.038868 | 0.010581 | 0.033512 | 0.014650 | 0.030450 | 0.010420 | 0.041545 | 0.019120 | 0.038861 | 0.016226 |
| bb_enhancing_t1ce Mean | 0.817634 | 0.717473 | 0.695027 | 0.579240 | 0.699590 | 0.629664 | 0.589495 | 0.513063 | 0.737524 | 0.652091 | 0.629073 | 0.536464 |
| bb_enhancing_t1ce Std | 0.018892 | 0.024780 | 0.026373 | 0.025884 | 0.023967 | 0.037958 | 0.027732 | 0.032578 | 0.032667 | 0.032133 | 0.036213 | 0.027175 |
| bb_enhancing_t2 Mean | 0.454526 | 0.255203 | 0.312010 | 0.154240 | 0.357963 | 0.200569 | 0.250993 | 0.125657 | 0.393215 | 0.219973 | 0.286241 | 0.144993 |
| bb_enhancing_t2 Std | 0.151299 | 0.042372 | 0.122472 | 0.028766 | 0.126102 | 0.033027 | 0.100919 | 0.022401 | 0.147948 | 0.040753 | 0.122907 | 0.031332 |
| bb_enhancing_flair Mean | 0.498896 | 0.298346 | 0.342099 | 0.183844 | 0.397501 | 0.253779 | 0.278199 | 0.163079 | 0.425027 | 0.264847 | 0.305761 | 0.174141 |
| bb_enhancing_flair Std | 0.066605 | 0.019356 | 0.058507 | 0.012871 | 0.053951 | 0.014067 | 0.045264 | 0.009115 | 0.061992 | 0.012477 | 0.053994 | 0.012322 |
No option performs well here, but T1ce has a noticeable advantage. There is some overfitting.
Peritumoural Edema Segmentation with a Single Modality¶
These are the single modality results for the peritumoural edema segmentation.
In [13]:
show_results(["bb_edema_t1", "bb_edema_t1ce", "bb_edema_t2", "bb_edema_flair"])
# bb_edema_t1 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.735397 | 0.425970 | 0.583658 | 0.276597 | 0.297229 | 0.606085 | 0.683449 | 0.360839 | 0.538121 | 0.237833 | 0.684691 | 0.361287 | 0.539589 | 0.238196 |
| Fold 2 | 0.622729 | 0.437944 | 0.455472 | 0.286135 | 0.418131 | 0.605321 | 0.564083 | 0.364590 | 0.413932 | 0.244388 | 0.565059 | 0.364956 | 0.414942 | 0.244713 |
| Fold 3 | 0.687117 | 0.458557 | 0.527151 | 0.301687 | 0.348139 | 0.575080 | 0.633085 | 0.395925 | 0.483277 | 0.264577 | 0.634965 | 0.397027 | 0.485338 | 0.265535 |
| Fold 4 | 0.735582 | 0.452848 | 0.583690 | 0.300017 | 0.299842 | 0.586842 | 0.673394 | 0.367320 | 0.529006 | 0.248630 | 0.679469 | 0.368605 | 0.535583 | 0.249867 |
| Fold 5 | 0.751556 | 0.427769 | 0.604052 | 0.277857 | 0.281724 | 0.602238 | 0.695579 | 0.365381 | 0.554269 | 0.245148 | 0.699995 | 0.365837 | 0.558913 | 0.245576 |
| Mean | 0.706476 | 0.440618 | 0.550805 | 0.288458 | 0.329013 | 0.595113 | 0.649918 | 0.370811 | 0.503721 | 0.248115 | 0.652836 | 0.371542 | 0.506873 | 0.248777 |
| Std | 0.047125 | 0.013101 | 0.054118 | 0.010650 | 0.049810 | 0.012207 | 0.047773 | 0.012732 | 0.050711 | 0.008941 | 0.048933 | 0.012955 | 0.051996 | 0.009173 |
# bb_edema_t1ce results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.750149 | 0.519686 | 0.603346 | 0.362854 | 0.283250 | 0.512320 | 0.691226 | 0.423787 | 0.551444 | 0.298158 | 0.693598 | 0.424504 | 0.554366 | 0.298979 |
| Fold 2 | 0.751663 | 0.539595 | 0.605752 | 0.378095 | 0.278808 | 0.495224 | 0.704772 | 0.452340 | 0.563000 | 0.322897 | 0.709829 | 0.452975 | 0.568468 | 0.323606 |
| Fold 3 | 0.711482 | 0.553556 | 0.554846 | 0.390227 | 0.320089 | 0.481822 | 0.643768 | 0.466678 | 0.497865 | 0.334170 | 0.645312 | 0.467432 | 0.499589 | 0.334938 |
| Fold 4 | 0.666340 | 0.478009 | 0.504645 | 0.326041 | 0.363108 | 0.556442 | 0.605161 | 0.431262 | 0.457685 | 0.300829 | 0.606104 | 0.431740 | 0.458708 | 0.301301 |
| Fold 5 | 0.727456 | 0.499256 | 0.575176 | 0.343618 | 0.303591 | 0.535935 | 0.671922 | 0.442434 | 0.527623 | 0.307459 | 0.674416 | 0.443769 | 0.530615 | 0.308813 |
| Mean | 0.721418 | 0.518020 | 0.568753 | 0.360167 | 0.309769 | 0.516349 | 0.663370 | 0.443300 | 0.519523 | 0.312703 | 0.665852 | 0.444084 | 0.522349 | 0.313527 |
| Std | 0.031334 | 0.027144 | 0.037175 | 0.023111 | 0.030501 | 0.027002 | 0.035597 | 0.015196 | 0.038126 | 0.013746 | 0.036794 | 0.015232 | 0.039468 | 0.013736 |
# bb_edema_t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.770666 | 0.576108 | 0.629762 | 0.409798 | 0.260583 | 0.453865 | 0.728988 | 0.508375 | 0.589684 | 0.364190 | 0.731090 | 0.509631 | 0.592356 | 0.365476 |
| Fold 2 | 0.723519 | 0.554561 | 0.569679 | 0.389190 | 0.306878 | 0.478960 | 0.679425 | 0.485744 | 0.532722 | 0.345740 | 0.681222 | 0.487120 | 0.534861 | 0.347096 |
| Fold 3 | 0.740972 | 0.587819 | 0.591329 | 0.423049 | 0.291722 | 0.441812 | 0.693501 | 0.549651 | 0.547512 | 0.400693 | 0.695884 | 0.551062 | 0.550322 | 0.402199 |
| Fold 4 | 0.607657 | 0.541142 | 0.442191 | 0.379878 | 0.428730 | 0.498978 | 0.536580 | 0.473622 | 0.393360 | 0.334119 | 0.539172 | 0.475822 | 0.396177 | 0.336329 |
| Fold 5 | 0.727982 | 0.546231 | 0.576275 | 0.380342 | 0.300208 | 0.477972 | 0.672468 | 0.498786 | 0.527274 | 0.352569 | 0.674020 | 0.499793 | 0.529080 | 0.353614 |
| Mean | 0.714159 | 0.561172 | 0.561847 | 0.396451 | 0.317624 | 0.470318 | 0.662192 | 0.503236 | 0.518110 | 0.359462 | 0.664277 | 0.504686 | 0.520559 | 0.360943 |
| Std | 0.055737 | 0.017896 | 0.063354 | 0.017161 | 0.057772 | 0.020189 | 0.065762 | 0.026014 | 0.066102 | 0.022805 | 0.065569 | 0.025849 | 0.066013 | 0.022689 |
# bb_edema_flair results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.825641 | 0.638905 | 0.705096 | 0.475039 | 0.204660 | 0.392478 | 0.791154 | 0.600829 | 0.666011 | 0.448425 | 0.793715 | 0.602446 | 0.669492 | 0.450173 |
| Fold 2 | 0.818635 | 0.673130 | 0.695196 | 0.513094 | 0.211913 | 0.361456 | 0.783663 | 0.620900 | 0.657424 | 0.471061 | 0.786242 | 0.622581 | 0.660907 | 0.472882 |
| Fold 3 | 0.719951 | 0.653348 | 0.565983 | 0.492013 | 0.313139 | 0.380480 | 0.677110 | 0.611123 | 0.531215 | 0.464482 | 0.680101 | 0.613095 | 0.534705 | 0.466743 |
| Fold 4 | 0.701419 | 0.559068 | 0.543037 | 0.393665 | 0.331400 | 0.475766 | 0.646239 | 0.515038 | 0.498678 | 0.366877 | 0.648777 | 0.516898 | 0.501533 | 0.368702 |
| Fold 5 | 0.781819 | 0.606341 | 0.644113 | 0.439043 | 0.252864 | 0.423815 | 0.744796 | 0.570372 | 0.608855 | 0.417176 | 0.747145 | 0.571846 | 0.611872 | 0.418739 |
| Mean | 0.769493 | 0.626158 | 0.630685 | 0.462571 | 0.262795 | 0.406799 | 0.728592 | 0.583653 | 0.592437 | 0.433604 | 0.731196 | 0.585373 | 0.595702 | 0.435448 |
| Std | 0.050612 | 0.040005 | 0.065951 | 0.042134 | 0.051591 | 0.039999 | 0.057692 | 0.038266 | 0.067000 | 0.038206 | 0.057614 | 0.038252 | 0.067151 | 0.038293 |
Results Summary
Out[13]:
| DSC | val_DSC | IoU | val_IoU | DSC' | val_DSC' | IoU' | val_IoU' | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bb_edema_t1 Mean | 0.706476 | 0.440618 | 0.550805 | 0.288458 | 0.649918 | 0.370811 | 0.503721 | 0.248115 | 0.652836 | 0.371542 | 0.506873 | 0.248777 |
| bb_edema_t1 Std | 0.047125 | 0.013101 | 0.054118 | 0.010650 | 0.047773 | 0.012732 | 0.050711 | 0.008941 | 0.048933 | 0.012955 | 0.051996 | 0.009173 |
| bb_edema_t1ce Mean | 0.721418 | 0.518020 | 0.568753 | 0.360167 | 0.663370 | 0.443300 | 0.519523 | 0.312703 | 0.665852 | 0.444084 | 0.522349 | 0.313527 |
| bb_edema_t1ce Std | 0.031334 | 0.027144 | 0.037175 | 0.023111 | 0.035597 | 0.015196 | 0.038126 | 0.013746 | 0.036794 | 0.015232 | 0.039468 | 0.013736 |
| bb_edema_t2 Mean | 0.714159 | 0.561172 | 0.561847 | 0.396451 | 0.662192 | 0.503236 | 0.518110 | 0.359462 | 0.664277 | 0.504686 | 0.520559 | 0.360943 |
| bb_edema_t2 Std | 0.055737 | 0.017896 | 0.063354 | 0.017161 | 0.065762 | 0.026014 | 0.066102 | 0.022805 | 0.065569 | 0.025849 | 0.066013 | 0.022689 |
| bb_edema_flair Mean | 0.769493 | 0.626158 | 0.630685 | 0.462571 | 0.728592 | 0.583653 | 0.592437 | 0.433604 | 0.731196 | 0.585373 | 0.595702 | 0.435448 |
| bb_edema_flair Std | 0.050612 | 0.040005 | 0.065951 | 0.042134 | 0.057692 | 0.038266 | 0.067000 | 0.038206 | 0.057614 | 0.038252 | 0.067151 | 0.038293 |
No option performs well, but flair has a noticeable advantage. There is some overfitting.
Core Segmentation with a Single Modality¶
We call the union of the necrotic/non-enhancing and enhancing tumour region core here. These are the single modality segmentation results.
In [14]:
show_results(["bb_core_t1", "bb_core_t1ce", "bb_core_t2", "bb_core_flair"])
# bb_core_t1 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.825789 | 0.508800 | 0.706331 | 0.349408 | 0.213181 | 0.532416 | 0.741521 | 0.418551 | 0.629886 | 0.303641 | 0.742936 | 0.418567 | 0.631901 | 0.303811 |
| Fold 2 | 0.793605 | 0.472918 | 0.662508 | 0.322296 | 0.244925 | 0.559562 | 0.705907 | 0.396274 | 0.590697 | 0.278442 | 0.708993 | 0.396851 | 0.594737 | 0.279113 |
| Fold 3 | 0.840970 | 0.493199 | 0.727581 | 0.343434 | 0.202095 | 0.553859 | 0.743572 | 0.398755 | 0.640144 | 0.290417 | 0.745019 | 0.401850 | 0.642247 | 0.293442 |
| Fold 4 | 0.711935 | 0.459857 | 0.558842 | 0.308692 | 0.328230 | 0.585055 | 0.626113 | 0.389516 | 0.494864 | 0.272613 | 0.627334 | 0.389882 | 0.496401 | 0.273008 |
| Fold 5 | 0.844421 | 0.518607 | 0.732894 | 0.357716 | 0.197385 | 0.526060 | 0.763030 | 0.445922 | 0.659637 | 0.323149 | 0.764393 | 0.446322 | 0.661546 | 0.323562 |
| Mean | 0.803344 | 0.490676 | 0.677631 | 0.336309 | 0.237163 | 0.551390 | 0.716029 | 0.409804 | 0.603046 | 0.293652 | 0.717735 | 0.410694 | 0.605366 | 0.294587 |
| Std | 0.049107 | 0.021818 | 0.064362 | 0.018109 | 0.048458 | 0.021014 | 0.048590 | 0.020483 | 0.058575 | 0.018190 | 0.048591 | 0.020171 | 0.058663 | 0.018061 |
# bb_core_t1ce results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.893375 | 0.801990 | 0.808876 | 0.683727 | 0.128805 | 0.222339 | 0.852363 | 0.728726 | 0.767917 | 0.627022 | 0.854314 | 0.730116 | 0.770948 | 0.629057 |
| Fold 2 | 0.762362 | 0.656528 | 0.636446 | 0.514988 | 0.255759 | 0.362675 | 0.717072 | 0.585020 | 0.619576 | 0.476147 | 0.718503 | 0.585872 | 0.621574 | 0.477302 |
| Fold 3 | 0.902816 | 0.788767 | 0.824806 | 0.667253 | 0.115962 | 0.233083 | 0.870529 | 0.729637 | 0.788197 | 0.625564 | 0.872145 | 0.730943 | 0.790732 | 0.627377 |
| Fold 4 | 0.898264 | 0.774697 | 0.816502 | 0.663295 | 0.120830 | 0.248394 | 0.858267 | 0.727310 | 0.771835 | 0.620876 | 0.859761 | 0.728288 | 0.774089 | 0.622205 |
| Fold 5 | 0.868458 | 0.804497 | 0.776209 | 0.691415 | 0.152812 | 0.222146 | 0.826144 | 0.701957 | 0.742363 | 0.609985 | 0.828466 | 0.703319 | 0.745994 | 0.612180 |
| Mean | 0.865055 | 0.765296 | 0.772568 | 0.644136 | 0.154833 | 0.257727 | 0.824875 | 0.694530 | 0.737978 | 0.591919 | 0.826638 | 0.695708 | 0.740667 | 0.593624 |
| Std | 0.052703 | 0.055413 | 0.070028 | 0.065396 | 0.052028 | 0.053342 | 0.055818 | 0.055721 | 0.060996 | 0.058193 | 0.055912 | 0.055873 | 0.061241 | 0.058458 |
# bb_core_t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.765589 | 0.538829 | 0.627759 | 0.376198 | 0.279664 | 0.508232 | 0.686423 | 0.450171 | 0.564248 | 0.329659 | 0.688730 | 0.451160 | 0.567300 | 0.330879 |
| Fold 2 | 0.814146 | 0.531804 | 0.690142 | 0.372132 | 0.226853 | 0.505927 | 0.740046 | 0.444630 | 0.622654 | 0.320063 | 0.742578 | 0.445966 | 0.626223 | 0.321384 |
| Fold 3 | 0.695219 | 0.469635 | 0.539094 | 0.320979 | 0.348695 | 0.570116 | 0.598538 | 0.370426 | 0.468650 | 0.265451 | 0.600209 | 0.370929 | 0.470809 | 0.266251 |
| Fold 4 | 0.759782 | 0.575197 | 0.618226 | 0.411514 | 0.279255 | 0.461246 | 0.681009 | 0.489793 | 0.551172 | 0.356954 | 0.686352 | 0.491826 | 0.557780 | 0.359323 |
| Fold 5 | 0.699065 | 0.514724 | 0.549997 | 0.354288 | 0.336394 | 0.527423 | 0.603510 | 0.438371 | 0.480249 | 0.319622 | 0.604511 | 0.438925 | 0.481538 | 0.320246 |
| Mean | 0.746760 | 0.526038 | 0.605044 | 0.367022 | 0.294172 | 0.514589 | 0.661905 | 0.438678 | 0.537394 | 0.318350 | 0.664476 | 0.439761 | 0.540730 | 0.319617 |
| Std | 0.044713 | 0.034417 | 0.055337 | 0.029570 | 0.044091 | 0.035239 | 0.053847 | 0.038571 | 0.056871 | 0.029736 | 0.054576 | 0.039033 | 0.057790 | 0.030183 |
# bb_core_flair results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | IoU' | val_IoU' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.762900 | 0.522564 | 0.624926 | 0.368759 | 0.269591 | 0.511220 | 0.676501 | 0.483391 | 0.558735 | 0.354057 | 0.677989 | 0.484082 | 0.560765 | 0.354877 |
| Fold 2 | 0.767437 | 0.514076 | 0.633839 | 0.353254 | 0.271744 | 0.523148 | 0.703372 | 0.474699 | 0.583634 | 0.343240 | 0.704687 | 0.475438 | 0.585312 | 0.344066 |
| Fold 3 | 0.841017 | 0.589142 | 0.728014 | 0.427585 | 0.203620 | 0.458676 | 0.792805 | 0.488876 | 0.679276 | 0.361714 | 0.794365 | 0.489529 | 0.681392 | 0.362363 |
| Fold 4 | 0.821062 | 0.581418 | 0.703402 | 0.414962 | 0.219410 | 0.457132 | 0.771257 | 0.498097 | 0.655367 | 0.364366 | 0.772398 | 0.498550 | 0.656858 | 0.364824 |
| Fold 5 | 0.783634 | 0.520505 | 0.650022 | 0.358578 | 0.254830 | 0.524741 | 0.729278 | 0.488426 | 0.605239 | 0.356151 | 0.730960 | 0.489529 | 0.607494 | 0.357414 |
| Mean | 0.795210 | 0.545541 | 0.668041 | 0.384628 | 0.243839 | 0.494983 | 0.734642 | 0.486698 | 0.616450 | 0.355906 | 0.736080 | 0.487426 | 0.618364 | 0.356709 |
| Std | 0.030711 | 0.032659 | 0.040498 | 0.030595 | 0.027484 | 0.030638 | 0.042686 | 0.007650 | 0.044711 | 0.007338 | 0.042667 | 0.007582 | 0.044682 | 0.007232 |
Results Summary
Out[14]:
| DSC | val_DSC | IoU | val_IoU | DSC' | val_DSC' | IoU' | val_IoU' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bb_core_t1 Mean | 0.803344 | 0.490676 | 0.677631 | 0.336309 | 0.716029 | 0.409804 | 0.603046 | 0.293652 | 0.717735 | 0.410694 | 0.605366 | 0.294587 |
| bb_core_t1 Std | 0.049107 | 0.021818 | 0.064362 | 0.018109 | 0.048590 | 0.020483 | 0.058575 | 0.018190 | 0.048591 | 0.020171 | 0.058663 | 0.018061 |
| bb_core_t1ce Mean | 0.865055 | 0.765296 | 0.772568 | 0.644136 | 0.824875 | 0.694530 | 0.737978 | 0.591919 | 0.826638 | 0.695708 | 0.740667 | 0.593624 |
| bb_core_t1ce Std | 0.052703 | 0.055413 | 0.070028 | 0.065396 | 0.055818 | 0.055721 | 0.060996 | 0.058193 | 0.055912 | 0.055873 | 0.061241 | 0.058458 |
| bb_core_t2 Mean | 0.746760 | 0.526038 | 0.605044 | 0.367022 | 0.661905 | 0.438678 | 0.537394 | 0.318350 | 0.664476 | 0.439761 | 0.540730 | 0.319617 |
| bb_core_t2 Std | 0.044713 | 0.034417 | 0.055337 | 0.029570 | 0.053847 | 0.038571 | 0.056871 | 0.029736 | 0.054576 | 0.039033 | 0.057790 | 0.030183 |
| bb_core_flair Mean | 0.795210 | 0.545541 | 0.668041 | 0.384628 | 0.734642 | 0.486698 | 0.616450 | 0.355906 | 0.736080 | 0.487426 | 0.618364 | 0.356709 |
| bb_core_flair Std | 0.030711 | 0.032659 | 0.040498 | 0.030595 | 0.042686 | 0.007650 | 0.044711 | 0.007338 | 0.042667 | 0.007582 | 0.044682 | 0.007232 |
T1ce performs considerably better than the other options, even if all of them produce some limited, useful results. This matches the t1ce results for necrotic and enhancing regions. Overfitting is also visible here.
Segmenting Multiple Nested or Separate Regions using all Modalities¶
We now look at the nested segmentation region and separate segmentation regions using all modalities. We could have combined some of them first, but the results on the whole tumour are not encouraging here, as they indicate that there is mainly one modality that enables segmentation (at least for this UNet architecture).
Nested options:
- nested: seg+1+2+4 > seg+1+4 > seg+1
- nested core: seg+1+2+4 > seg+1+4
- nested edema: seg+1+2+4 > seg+2
- separte: seg+1, seg+4, seg+2
- separate core: seg+1+4, seg+2
In [15]:
show_results(["bb_nested_flair+t1+t1ce+t2", "bb_nestedCore_flair+t1+t1ce+t2", "bb_nestedEdema_flair+t1+t1ce+t2", "bb_separate_flair+t1+t1ce+t2", "bb_separateCore_flair+t1+t1ce+t2"])
# bb_nested_flair+t1+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | DSC_c2' | val_DSC_c2' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | IoU_c2' | val_IoU_c2' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.816706 | 0.715437 | 0.699311 | 0.578984 | 0.213033 | 0.315034 | 0.749345 | 0.650715 | 0.840338 | 0.791750 | 0.820972 | 0.707337 | 0.586726 | 0.453059 | 0.635936 | 0.533474 | 0.735226 | 0.679532 | 0.715587 | 0.589334 | 0.456996 | 0.331556 | 0.822776 | 0.708716 | 0.718374 | 0.591252 | 0.633905 | 0.533981 | 0.488552 | 0.385836 | 0.514377 | 0.447359 | 0.386177 | 0.321188 | 0.592416 | 0.453833 | 0.463235 | 0.332900 | 0.841802 | 0.792945 | 0.737410 | 0.681243 |
| Fold 2 | 0.806320 | 0.698856 | 0.685952 | 0.565504 | 0.226737 | 0.329728 | 0.743742 | 0.646551 | 0.830538 | 0.815961 | 0.834737 | 0.689881 | 0.565951 | 0.433811 | 0.628411 | 0.528056 | 0.718110 | 0.703992 | 0.732946 | 0.567041 | 0.434178 | 0.313135 | 0.836538 | 0.690630 | 0.735559 | 0.568158 | 0.649773 | 0.570859 | 0.499431 | 0.425378 | 0.488249 | 0.371624 | 0.356372 | 0.256956 | 0.570663 | 0.434684 | 0.439017 | 0.314017 | 0.831502 | 0.817108 | 0.719523 | 0.705668 |
| Fold 3 | 0.769543 | 0.648300 | 0.635967 | 0.506661 | 0.259669 | 0.384770 | 0.704338 | 0.565283 | 0.789100 | 0.728552 | 0.777255 | 0.612607 | 0.546659 | 0.354690 | 0.579604 | 0.451231 | 0.666982 | 0.604272 | 0.660224 | 0.498851 | 0.411606 | 0.250570 | 0.778664 | 0.613466 | 0.662223 | 0.500164 | 0.552905 | 0.509215 | 0.405977 | 0.367828 | 0.446908 | 0.390108 | 0.322429 | 0.276076 | 0.551263 | 0.355458 | 0.416269 | 0.251304 | 0.790019 | 0.729221 | 0.668269 | 0.605172 |
| Fold 4 | 0.791505 | 0.693635 | 0.665586 | 0.549282 | 0.242801 | 0.343124 | 0.721607 | 0.617795 | 0.807556 | 0.780490 | 0.796652 | 0.664598 | 0.560613 | 0.408296 | 0.602126 | 0.494465 | 0.695300 | 0.659620 | 0.684254 | 0.535616 | 0.426822 | 0.288158 | 0.798500 | 0.665727 | 0.686923 | 0.537348 | 0.574157 | 0.521755 | 0.430701 | 0.379249 | 0.456438 | 0.418339 | 0.333109 | 0.294852 | 0.563884 | 0.410688 | 0.430461 | 0.290600 | 0.810011 | 0.782899 | 0.698829 | 0.662910 |
| Fold 5 | 0.789307 | 0.671594 | 0.666555 | 0.530805 | 0.237186 | 0.359300 | 0.733619 | 0.627964 | 0.842366 | 0.790336 | 0.799114 | 0.685630 | 0.559378 | 0.407926 | 0.616860 | 0.506040 | 0.736611 | 0.668241 | 0.687176 | 0.563189 | 0.426792 | 0.286689 | 0.800508 | 0.686552 | 0.689195 | 0.564466 | 0.620223 | 0.524990 | 0.470071 | 0.374918 | 0.423063 | 0.377843 | 0.300939 | 0.257724 | 0.561399 | 0.408619 | 0.429224 | 0.287685 | 0.843260 | 0.790821 | 0.737938 | 0.668940 |
| Mean | 0.794676 | 0.685564 | 0.670674 | 0.546247 | 0.235885 | 0.346391 | 0.730530 | 0.621662 | 0.821979 | 0.781418 | 0.805746 | 0.672011 | 0.563865 | 0.411556 | 0.612587 | 0.502653 | 0.710446 | 0.663131 | 0.696037 | 0.550806 | 0.431279 | 0.294022 | 0.807397 | 0.673018 | 0.698455 | 0.552278 | 0.606192 | 0.532160 | 0.458946 | 0.386642 | 0.465807 | 0.401055 | 0.339805 | 0.281359 | 0.567925 | 0.412657 | 0.435641 | 0.295301 | 0.823319 | 0.782599 | 0.712394 | 0.664787 |
| Std | 0.016071 | 0.023315 | 0.021452 | 0.025524 | 0.015628 | 0.024130 | 0.016148 | 0.030649 | 0.020567 | 0.028897 | 0.020054 | 0.032670 | 0.013067 | 0.033092 | 0.020056 | 0.029387 | 0.026366 | 0.032989 | 0.025473 | 0.031085 | 0.014811 | 0.027387 | 0.020189 | 0.032764 | 0.025710 | 0.031172 | 0.036688 | 0.020915 | 0.035335 | 0.020234 | 0.032046 | 0.028170 | 0.029265 | 0.024292 | 0.013739 | 0.033083 | 0.015596 | 0.027495 | 0.020456 | 0.029034 | 0.026307 | 0.033217 |
# bb_nestedCore_flair+t1+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.880785 | 0.817964 | 0.788362 | 0.700932 | 0.142165 | 0.205572 | 0.854977 | 0.780387 | 0.866610 | 0.825614 | 0.843345 | 0.735161 | 0.757762 | 0.669891 | 0.771316 | 0.721554 | 0.744207 | 0.618228 | 0.845668 | 0.737063 | 0.747607 | 0.620828 | 0.697076 | 0.605532 | 0.556000 | 0.455384 | 0.868165 | 0.827092 | 0.773707 | 0.723715 |
| Fold 2 | 0.887225 | 0.791334 | 0.798026 | 0.668335 | 0.136205 | 0.230675 | 0.863601 | 0.756213 | 0.872608 | 0.829706 | 0.854594 | 0.682719 | 0.767705 | 0.644127 | 0.779169 | 0.725308 | 0.756240 | 0.562945 | 0.856835 | 0.683997 | 0.759590 | 0.564772 | 0.707303 | 0.582121 | 0.566339 | 0.438602 | 0.874323 | 0.831363 | 0.781843 | 0.727767 |
| Fold 3 | 0.889687 | 0.810888 | 0.802216 | 0.689144 | 0.134017 | 0.214106 | 0.865895 | 0.755121 | 0.876393 | 0.806668 | 0.855397 | 0.703574 | 0.772683 | 0.643219 | 0.785182 | 0.702368 | 0.760185 | 0.584070 | 0.857764 | 0.705372 | 0.763715 | 0.586393 | 0.706119 | 0.633557 | 0.564479 | 0.495239 | 0.878077 | 0.808087 | 0.787804 | 0.704468 |
| Fold 4 | 0.881363 | 0.813692 | 0.790112 | 0.693638 | 0.142645 | 0.211995 | 0.853712 | 0.779290 | 0.869564 | 0.837173 | 0.837860 | 0.721407 | 0.758731 | 0.662563 | 0.777020 | 0.729950 | 0.740443 | 0.595176 | 0.840099 | 0.723186 | 0.743926 | 0.597552 | 0.701901 | 0.629089 | 0.563584 | 0.482854 | 0.871076 | 0.838803 | 0.779373 | 0.732341 |
| Fold 5 | 0.863996 | 0.804016 | 0.763320 | 0.682666 | 0.160394 | 0.224977 | 0.831291 | 0.765224 | 0.857736 | 0.812653 | 0.804846 | 0.717794 | 0.730504 | 0.653592 | 0.760371 | 0.704733 | 0.700636 | 0.602451 | 0.807593 | 0.720079 | 0.704999 | 0.605603 | 0.684095 | 0.599293 | 0.543862 | 0.453291 | 0.860159 | 0.814585 | 0.764083 | 0.707681 |
| Mean | 0.880611 | 0.807579 | 0.788407 | 0.686943 | 0.143085 | 0.217465 | 0.853895 | 0.767247 | 0.868582 | 0.822363 | 0.839208 | 0.712131 | 0.757477 | 0.654678 | 0.774612 | 0.716783 | 0.740342 | 0.592574 | 0.841592 | 0.713939 | 0.743967 | 0.595029 | 0.699299 | 0.609918 | 0.558853 | 0.465074 | 0.870360 | 0.823986 | 0.777362 | 0.719194 |
| Std | 0.008975 | 0.009303 | 0.013531 | 0.011045 | 0.009276 | 0.009095 | 0.012250 | 0.010868 | 0.006320 | 0.011175 | 0.018431 | 0.017814 | 0.014595 | 0.010367 | 0.008390 | 0.011152 | 0.021157 | 0.018507 | 0.018268 | 0.018044 | 0.020813 | 0.018835 | 0.008408 | 0.019137 | 0.008282 | 0.020783 | 0.006077 | 0.011178 | 0.008035 | 0.011101 |
# bb_nestedEdema_flair+t1+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.768011 | 0.716529 | 0.641441 | 0.578613 | 0.246613 | 0.298551 | 0.745783 | 0.694385 | 0.865678 | 0.820839 | 0.625888 | 0.567931 | 0.626131 | 0.566766 | 0.772998 | 0.716913 | 0.479265 | 0.416620 | 1.297326e-09 | 8.953387e-10 | 1.297326e-09 | 8.953387e-10 | 0.627098 | 0.568969 | 0.480646 | 0.417709 | 0.866986 | 0.822008 | 0.775048 | 0.718678 |
| Fold 2 | 0.768008 | 0.753593 | 0.643613 | 0.623868 | 0.245591 | 0.261306 | 0.751832 | 0.718400 | 0.880239 | 0.850218 | 0.623425 | 0.586582 | 0.632087 | 0.597517 | 0.790680 | 0.753228 | 0.473493 | 0.441807 | 1.015554e-03 | 1.201093e-03 | 5.083242e-04 | 6.014264e-04 | 0.624421 | 0.587361 | 0.474614 | 0.442630 | 0.881372 | 0.851162 | 0.792460 | 0.754674 |
| Fold 3 | 0.768151 | 0.755943 | 0.644339 | 0.621449 | 0.247219 | 0.260085 | 0.750360 | 0.717226 | 0.882213 | 0.814405 | 0.618508 | 0.620048 | 0.630303 | 0.594543 | 0.792339 | 0.713340 | 0.468268 | 0.475746 | 3.472502e-03 | 3.318186e-03 | 1.742028e-03 | 1.667203e-03 | 0.620037 | 0.621378 | 0.469949 | 0.477290 | 0.883859 | 0.815649 | 0.794929 | 0.715228 |
| Fold 4 | 0.777564 | 0.738544 | 0.654294 | 0.603654 | 0.234757 | 0.275113 | 0.759121 | 0.709810 | 0.885590 | 0.828241 | 0.632653 | 0.591380 | 0.641660 | 0.582034 | 0.798324 | 0.720936 | 0.484997 | 0.443133 | 7.472297e-05 | 7.255922e-05 | 3.736725e-05 | 3.628495e-05 | 0.633761 | 0.592262 | 0.486251 | 0.444096 | 0.886739 | 0.829173 | 0.800120 | 0.722309 |
| Fold 5 | 0.775685 | 0.736807 | 0.653942 | 0.600617 | 0.236717 | 0.272171 | 0.758905 | 0.709721 | 0.888000 | 0.835985 | 0.629811 | 0.583458 | 0.641492 | 0.581026 | 0.801601 | 0.729118 | 0.481383 | 0.432935 | 6.787886e-04 | 6.949784e-04 | 3.396530e-04 | 3.477419e-04 | 0.630436 | 0.584023 | 0.482093 | 0.433521 | 0.888669 | 0.836584 | 0.802663 | 0.730000 |
| Mean | 0.771484 | 0.740283 | 0.647526 | 0.605640 | 0.242179 | 0.273445 | 0.753200 | 0.709909 | 0.880344 | 0.829938 | 0.626057 | 0.589879 | 0.634335 | 0.584377 | 0.791188 | 0.726707 | 0.477481 | 0.442048 | 1.048314e-03 | 1.057363e-03 | 5.254747e-04 | 5.305314e-04 | 0.627151 | 0.590799 | 0.478710 | 0.443049 | 0.881525 | 0.830915 | 0.793044 | 0.728178 |
| Std | 0.004239 | 0.014154 | 0.005468 | 0.016382 | 0.005322 | 0.013860 | 0.005149 | 0.008564 | 0.007807 | 0.012448 | 0.004930 | 0.017007 | 0.006221 | 0.010981 | 0.009918 | 0.014262 | 0.005928 | 0.019321 | 1.269736e-03 | 1.212531e-03 | 6.370551e-04 | 6.092951e-04 | 0.004744 | 0.017155 | 0.005755 | 0.019527 | 0.007682 | 0.012309 | 0.009699 | 0.014127 |
# bb_separate_flair+t1+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | DSC_c2' | val_DSC_c2' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | IoU_c2' | val_IoU_c2' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.737905 | 0.607213 | 0.590359 | 0.451657 | 0.295094 | 0.426982 | 0.635994 | 0.527295 | 0.573972 | 0.420168 | 0.616519 | 0.547404 | 0.717491 | 0.614313 | 0.503764 | 0.395664 | 0.449088 | 0.305099 | 0.485087 | 0.416312 | 0.577118 | 0.465580 | 0.882940 | 0.766335 | 0.803468 | 0.675041 | 0.719452 | 0.615794 | 0.579425 | 0.467200 | 0.636306 | 0.550843 | 0.505102 | 0.419437 | 0.578926 | 0.420819 | 0.454486 | 0.306014 | 0.926074 | 0.896277 | 0.864761 | 0.816937 |
| Fold 2 | 0.754270 | 0.605833 | 0.612413 | 0.454307 | 0.280202 | 0.425562 | 0.662973 | 0.537770 | 0.606427 | 0.472497 | 0.665750 | 0.526766 | 0.716740 | 0.614048 | 0.531422 | 0.410480 | 0.481234 | 0.353688 | 0.537779 | 0.403133 | 0.575251 | 0.474620 | 0.888923 | 0.754135 | 0.815163 | 0.660480 | 0.719158 | 0.615243 | 0.578135 | 0.476008 | 0.692121 | 0.582495 | 0.564611 | 0.459035 | 0.612019 | 0.474122 | 0.487419 | 0.355427 | 0.928621 | 0.915937 | 0.869127 | 0.851383 |
| Fold 3 | 0.670508 | 0.555071 | 0.520251 | 0.405275 | 0.359822 | 0.478861 | 0.605617 | 0.464427 | 0.581213 | 0.348180 | 0.540194 | 0.451308 | 0.695445 | 0.593794 | 0.470984 | 0.346769 | 0.452764 | 0.254703 | 0.411386 | 0.332955 | 0.548803 | 0.452649 | 0.899630 | 0.790013 | 0.833620 | 0.693817 | 0.697674 | 0.594978 | 0.551373 | 0.454022 | 0.548601 | 0.479228 | 0.420075 | 0.361159 | 0.587311 | 0.348834 | 0.459363 | 0.255727 | 0.907194 | 0.883571 | 0.833034 | 0.803751 |
| Fold 4 | 0.510702 | 0.360023 | 0.414564 | 0.255669 | 0.513239 | 0.665719 | 0.441636 | 0.314072 | 0.596482 | 0.338287 | 0.012117 | 0.012621 | 0.716309 | 0.591307 | 0.352393 | 0.228499 | 0.474138 | 0.234460 | 0.006133 | 0.006383 | 0.576909 | 0.444654 | 0.025947 | 0.026609 | 0.013260 | 0.013628 | 0.718610 | 0.592457 | 0.579662 | 0.446061 | 0.012391 | 0.012868 | 0.006273 | 0.006509 | 0.598610 | 0.338857 | 0.476557 | 0.235070 | 0.060852 | 0.059983 | 0.031714 | 0.031275 |
| Fold 5 | 0.642484 | 0.557565 | 0.487046 | 0.403335 | 0.383881 | 0.473660 | 0.570485 | 0.495976 | 0.531709 | 0.431172 | 0.531474 | 0.509798 | 0.648274 | 0.546960 | 0.437881 | 0.368428 | 0.401814 | 0.320459 | 0.404613 | 0.377543 | 0.507217 | 0.407282 | 0.876220 | 0.799987 | 0.805607 | 0.707720 | 0.651505 | 0.548224 | 0.510812 | 0.408792 | 0.536499 | 0.525090 | 0.410026 | 0.392919 | 0.534272 | 0.445487 | 0.404649 | 0.335309 | 0.932563 | 0.928086 | 0.878570 | 0.869074 |
| Mean | 0.663174 | 0.537141 | 0.524927 | 0.394048 | 0.366447 | 0.494157 | 0.583341 | 0.467908 | 0.577961 | 0.402061 | 0.473211 | 0.409579 | 0.698852 | 0.592084 | 0.459289 | 0.349968 | 0.451808 | 0.293682 | 0.369000 | 0.307265 | 0.557059 | 0.448957 | 0.714732 | 0.627416 | 0.654224 | 0.550137 | 0.701280 | 0.593339 | 0.559882 | 0.450416 | 0.485184 | 0.430105 | 0.381218 | 0.327812 | 0.582227 | 0.405624 | 0.456495 | 0.297509 | 0.751061 | 0.736771 | 0.695441 | 0.674484 |
| Std | 0.086733 | 0.091365 | 0.071515 | 0.072541 | 0.082976 | 0.088663 | 0.077273 | 0.081084 | 0.025765 | 0.051199 | 0.235838 | 0.201042 | 0.026615 | 0.024561 | 0.061972 | 0.064584 | 0.027829 | 0.043531 | 0.187989 | 0.153108 | 0.027128 | 0.023254 | 0.344478 | 0.300847 | 0.320659 | 0.268736 | 0.026232 | 0.024588 | 0.026783 | 0.023243 | 0.243248 | 0.211340 | 0.195910 | 0.163830 | 0.026428 | 0.053282 | 0.028492 | 0.045826 | 0.345215 | 0.338744 | 0.332216 | 0.322453 |
# bb_separateCore_flair+t1+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.409572 | 0.340202 | 0.337150 | 0.252379 | 0.603352 | 0.673179 | 0.385417 | 0.316878 | 0.025318 | 0.027712 | 0.745515 | 0.606043 | 0.313887 | 0.233810 | 0.012942 | 0.014168 | 0.614832 | 0.453452 | 0.025544 | 0.028214 | 0.013059 | 0.014430 | 0.747480 | 0.607479 | 0.617396 | 0.454992 | 0.061327 | 0.058090 | 0.031970 | 0.030255 |
| Fold 2 | 0.828139 | 0.682577 | 0.711618 | 0.527865 | 0.212529 | 0.356913 | 0.791394 | 0.640644 | 0.843605 | 0.675937 | 0.739182 | 0.605350 | 0.671131 | 0.503066 | 0.742586 | 0.545608 | 0.599676 | 0.460524 | 0.846362 | 0.677798 | 0.746593 | 0.548128 | 0.742171 | 0.607198 | 0.603399 | 0.462540 | 0.874113 | 0.829887 | 0.780457 | 0.724545 |
| Fold 3 | 0.825418 | 0.724857 | 0.708227 | 0.580347 | 0.202475 | 0.305830 | 0.780865 | 0.643991 | 0.831553 | 0.657443 | 0.730178 | 0.630538 | 0.661715 | 0.522159 | 0.730955 | 0.548449 | 0.592474 | 0.495870 | 0.833862 | 0.658775 | 0.734262 | 0.550295 | 0.732203 | 0.631332 | 0.594891 | 0.496906 | 0.875756 | 0.788141 | 0.785538 | 0.685212 |
| Fold 4 | 0.670886 | 0.625123 | 0.519849 | 0.468406 | 0.358222 | 0.405222 | 0.610339 | 0.571703 | 0.688479 | 0.620834 | 0.532198 | 0.522573 | 0.480181 | 0.438667 | 0.572098 | 0.495577 | 0.388265 | 0.381756 | 0.690032 | 0.622517 | 0.574323 | 0.497684 | 0.533235 | 0.523533 | 0.389386 | 0.382800 | 0.808648 | 0.789889 | 0.706279 | 0.677013 |
| Fold 5 | 0.843655 | 0.722984 | 0.733076 | 0.578923 | 0.190251 | 0.316285 | 0.800332 | 0.677412 | 0.847923 | 0.716320 | 0.752742 | 0.638503 | 0.685408 | 0.545508 | 0.749846 | 0.599748 | 0.620970 | 0.491267 | 0.850731 | 0.718385 | 0.754017 | 0.602537 | 0.756178 | 0.640920 | 0.625344 | 0.494041 | 0.887514 | 0.829823 | 0.801364 | 0.722768 |
| Mean | 0.715534 | 0.619149 | 0.601984 | 0.481584 | 0.313366 | 0.411486 | 0.673669 | 0.570125 | 0.647376 | 0.539649 | 0.699963 | 0.600601 | 0.562464 | 0.448642 | 0.561685 | 0.440710 | 0.563243 | 0.456574 | 0.649306 | 0.541138 | 0.564451 | 0.442615 | 0.702253 | 0.602093 | 0.566083 | 0.458256 | 0.701471 | 0.659166 | 0.621122 | 0.567959 |
| Std | 0.165394 | 0.144101 | 0.153216 | 0.121720 | 0.157306 | 0.135445 | 0.160308 | 0.131207 | 0.316636 | 0.257811 | 0.084210 | 0.041166 | 0.145145 | 0.113139 | 0.282138 | 0.215802 | 0.088083 | 0.040916 | 0.317554 | 0.258319 | 0.283576 | 0.216646 | 0.084866 | 0.041447 | 0.088982 | 0.041224 | 0.321268 | 0.301093 | 0.296398 | 0.269538 |
Results Summary
Out[15]:
| DSC | val_DSC | IoU | val_IoU | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | DSC_c2' | val_DSC_c2' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | IoU_c2' | val_IoU_c2' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bb_nested_flair+t1+t1ce+t2 Mean | 0.794676 | 0.685564 | 0.670674 | 0.546247 | 0.730530 | 0.621662 | 0.821979 | 0.781418 | 0.805746 | 0.672011 | 0.563865 | 0.411556 | 0.612587 | 0.502653 | 0.710446 | 0.663131 | 0.696037 | 0.550806 | 0.431279 | 0.294022 | 0.807397 | 0.673018 | 0.698455 | 0.552278 | 0.606192 | 0.532160 | 0.458946 | 0.386642 | 0.465807 | 0.401055 | 0.339805 | 0.281359 | 0.567925 | 0.412657 | 0.435641 | 0.295301 | 0.823319 | 0.782599 | 0.712394 | 0.664787 |
| bb_nested_flair+t1+t1ce+t2 Std | 0.016071 | 0.023315 | 0.021452 | 0.025524 | 0.016148 | 0.030649 | 0.020567 | 0.028897 | 0.020054 | 0.032670 | 0.013067 | 0.033092 | 0.020056 | 0.029387 | 0.026366 | 0.032989 | 0.025473 | 0.031085 | 0.014811 | 0.027387 | 0.020189 | 0.032764 | 0.025710 | 0.031172 | 0.036688 | 0.020915 | 0.035335 | 0.020234 | 0.032046 | 0.028170 | 0.029265 | 0.024292 | 0.013739 | 0.033083 | 0.015596 | 0.027495 | 0.020456 | 0.029034 | 0.026307 | 0.033217 |
| bb_nestedCore_flair+t1+t1ce+t2 Mean | 0.880611 | 0.807579 | 0.788407 | 0.686943 | 0.853895 | 0.767247 | 0.868582 | 0.822363 | 0.839208 | 0.712131 | nan | nan | 0.757477 | 0.654678 | 0.774612 | 0.716783 | 0.740342 | 0.592574 | nan | nan | 0.841592 | 0.713939 | 0.743967 | 0.595029 | 0.699299 | 0.609918 | 0.558853 | 0.465074 | nan | nan | nan | nan | nan | nan | nan | nan | 0.870360 | 0.823986 | 0.777362 | 0.719194 |
| bb_nestedCore_flair+t1+t1ce+t2 Std | 0.008975 | 0.009303 | 0.013531 | 0.011045 | 0.012250 | 0.010868 | 0.006320 | 0.011175 | 0.018431 | 0.017814 | nan | nan | 0.014595 | 0.010367 | 0.008390 | 0.011152 | 0.021157 | 0.018507 | nan | nan | 0.018268 | 0.018044 | 0.020813 | 0.018835 | 0.008408 | 0.019137 | 0.008282 | 0.020783 | nan | nan | nan | nan | nan | nan | nan | nan | 0.006077 | 0.011178 | 0.008035 | 0.011101 |
| bb_nestedEdema_flair+t1+t1ce+t2 Mean | 0.771484 | 0.740283 | 0.647526 | 0.605640 | 0.753200 | 0.709909 | 0.880344 | 0.829938 | 0.626057 | 0.589879 | nan | nan | 0.634335 | 0.584377 | 0.791188 | 0.726707 | 0.477481 | 0.442048 | nan | nan | 0.001048 | 0.001057 | 0.000525 | 0.000531 | 0.627151 | 0.590799 | 0.478710 | 0.443049 | nan | nan | nan | nan | nan | nan | nan | nan | 0.881525 | 0.830915 | 0.793044 | 0.728178 |
| bb_nestedEdema_flair+t1+t1ce+t2 Std | 0.004239 | 0.014154 | 0.005468 | 0.016382 | 0.005149 | 0.008564 | 0.007807 | 0.012448 | 0.004930 | 0.017007 | nan | nan | 0.006221 | 0.010981 | 0.009918 | 0.014262 | 0.005928 | 0.019321 | nan | nan | 0.001270 | 0.001213 | 0.000637 | 0.000609 | 0.004744 | 0.017155 | 0.005755 | 0.019527 | nan | nan | nan | nan | nan | nan | nan | nan | 0.007682 | 0.012309 | 0.009699 | 0.014127 |
| bb_separate_flair+t1+t1ce+t2 Mean | 0.663174 | 0.537141 | 0.524927 | 0.394048 | 0.583341 | 0.467908 | 0.577961 | 0.402061 | 0.473211 | 0.409579 | 0.698852 | 0.592084 | 0.459289 | 0.349968 | 0.451808 | 0.293682 | 0.369000 | 0.307265 | 0.557059 | 0.448957 | 0.714732 | 0.627416 | 0.654224 | 0.550137 | 0.701280 | 0.593339 | 0.559882 | 0.450416 | 0.485184 | 0.430105 | 0.381218 | 0.327812 | 0.582227 | 0.405624 | 0.456495 | 0.297509 | 0.751061 | 0.736771 | 0.695441 | 0.674484 |
| bb_separate_flair+t1+t1ce+t2 Std | 0.086733 | 0.091365 | 0.071515 | 0.072541 | 0.077273 | 0.081084 | 0.025765 | 0.051199 | 0.235838 | 0.201042 | 0.026615 | 0.024561 | 0.061972 | 0.064584 | 0.027829 | 0.043531 | 0.187989 | 0.153108 | 0.027128 | 0.023254 | 0.344478 | 0.300847 | 0.320659 | 0.268736 | 0.026232 | 0.024588 | 0.026783 | 0.023243 | 0.243248 | 0.211340 | 0.195910 | 0.163830 | 0.026428 | 0.053282 | 0.028492 | 0.045826 | 0.345215 | 0.338744 | 0.332216 | 0.322453 |
| bb_separateCore_flair+t1+t1ce+t2 Mean | 0.715534 | 0.619149 | 0.601984 | 0.481584 | 0.673669 | 0.570125 | 0.647376 | 0.539649 | 0.699963 | 0.600601 | nan | nan | 0.562464 | 0.448642 | 0.561685 | 0.440710 | 0.563243 | 0.456574 | nan | nan | 0.649306 | 0.541138 | 0.564451 | 0.442615 | 0.702253 | 0.602093 | 0.566083 | 0.458256 | nan | nan | nan | nan | nan | nan | nan | nan | 0.701471 | 0.659166 | 0.621122 | 0.567959 |
| bb_separateCore_flair+t1+t1ce+t2 Std | 0.165394 | 0.144101 | 0.153216 | 0.121720 | 0.160308 | 0.131207 | 0.316636 | 0.257811 | 0.084210 | 0.041166 | nan | nan | 0.145145 | 0.113139 | 0.282138 | 0.215802 | 0.088083 | 0.040916 | nan | nan | 0.317554 | 0.258319 | 0.283576 | 0.216646 | 0.084866 | 0.041447 | 0.088982 | 0.041224 | nan | nan | nan | nan | nan | nan | nan | nan | 0.321268 | 0.301093 | 0.296398 | 0.269538 |
Nested approaches perform better than separate regions if we try to identify multiple regions and look at the raw metrics over all regions. On the individual regions with standardised metrics, the results are quite mixed between the two options.
The whole tumour is best identified by nestedEdema, but nestedCore is close, both effectively only identfiying the core. For nested, separating the core into enahcning and necrotic, we still get a good result, better than the separata options.
Expectedly, enhancing and necrotic regions alone are difficult to identify with no noticeable difference between the nested and separate option.
Combining enhancing and necrotic region into core works noticeably better and the nested options work better than the separate options, with the exception of nestedEdema. nestedEdema does not explicitly identify the core, but the edema, which may be the reason (it is an inverted nesting in that sense and so a whole would have to be identified).
Edema is similarly identified across all options, except for the inverted nestedEdema option, which fails completely. The separate options work better than the nested options.
All show some overfitting.
None of the approaches give better result than the networks identifying individual regions and are at best nearly as good. The exception here is the core where nestedCore gives slightly better results. This indicates features across modalities, if they exist, are not well exploited.
Segmenting Nested Regions using Two/Three Modalities¶
Finally, we reduce the number of modalities for segmenting the full nested regions.
In [16]:
show_results(["bb_nested_flair+t1ce","bb_nested_flair+t1ce+t2"])
# bb_nested_flair+t1ce results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | DSC_c2' | val_DSC_c2' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | IoU_c2' | val_IoU_c2' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.761034 | 0.620482 | 0.628323 | 0.473546 | 0.272361 | 0.411384 | 0.689902 | 0.584355 | 0.817110 | 0.785168 | 0.726786 | 0.590752 | 0.525811 | 0.377143 | 0.568251 | 0.458021 | 0.714258 | 0.668036 | 0.596774 | 0.446067 | 0.393722 | 0.259959 | 0.729781 | 0.591839 | 0.600604 | 0.447253 | 0.554336 | 0.432603 | 0.417217 | 0.290797 | 0.312543 | 0.229077 | 0.204968 | 0.141095 | 0.528084 | 0.378040 | 0.396280 | 0.260938 | 0.819193 | 0.787155 | 0.717354 | 0.670824 |
| Fold 2 | 0.817127 | 0.668141 | 0.699529 | 0.531734 | 0.218437 | 0.366766 | 0.749479 | 0.605683 | 0.834816 | 0.784852 | 0.825477 | 0.632109 | 0.588145 | 0.400088 | 0.637927 | 0.492346 | 0.730989 | 0.674024 | 0.722738 | 0.514458 | 0.460055 | 0.288554 | 0.828362 | 0.633318 | 0.727106 | 0.516357 | 0.648827 | 0.558088 | 0.503822 | 0.420886 | 0.531306 | 0.456189 | 0.404166 | 0.339748 | 0.594056 | 0.401565 | 0.466556 | 0.290147 | 0.836641 | 0.786577 | 0.733726 | 0.676383 |
| Fold 3 | 0.819805 | 0.682323 | 0.703424 | 0.543186 | 0.210927 | 0.351262 | 0.760029 | 0.594672 | 0.842901 | 0.763917 | 0.831662 | 0.646155 | 0.605525 | 0.373945 | 0.644825 | 0.479316 | 0.737975 | 0.652613 | 0.726833 | 0.522305 | 0.469668 | 0.263030 | 0.833956 | 0.647534 | 0.730276 | 0.524193 | 0.628745 | 0.550483 | 0.480798 | 0.408124 | 0.465728 | 0.387184 | 0.336756 | 0.268410 | 0.608250 | 0.374260 | 0.472611 | 0.263703 | 0.844081 | 0.764901 | 0.739760 | 0.653993 |
| Fold 4 | 0.803490 | 0.706285 | 0.682281 | 0.566393 | 0.226897 | 0.324089 | 0.744643 | 0.636235 | 0.837509 | 0.812719 | 0.808214 | 0.669528 | 0.588205 | 0.426459 | 0.628446 | 0.513883 | 0.732889 | 0.698256 | 0.695710 | 0.535440 | 0.456739 | 0.307954 | 0.810989 | 0.671704 | 0.699697 | 0.538178 | 0.625201 | 0.569792 | 0.479761 | 0.417777 | 0.500611 | 0.460635 | 0.376219 | 0.322345 | 0.590516 | 0.426731 | 0.459781 | 0.309132 | 0.839197 | 0.814609 | 0.735446 | 0.700976 |
| Fold 5 | 0.776107 | 0.646162 | 0.643881 | 0.500513 | 0.257524 | 0.394200 | 0.691503 | 0.583712 | 0.793543 | 0.763808 | 0.740145 | 0.588347 | 0.540821 | 0.398981 | 0.571389 | 0.465584 | 0.683540 | 0.642932 | 0.620796 | 0.463343 | 0.409831 | 0.290477 | 0.741721 | 0.589441 | 0.623250 | 0.464923 | 0.576271 | 0.499650 | 0.439870 | 0.362012 | 0.368320 | 0.355108 | 0.255138 | 0.244137 | 0.542822 | 0.414307 | 0.411969 | 0.305955 | 0.797050 | 0.768963 | 0.688648 | 0.650170 |
| Mean | 0.795513 | 0.664679 | 0.671488 | 0.523074 | 0.237229 | 0.369540 | 0.727111 | 0.600931 | 0.825176 | 0.782093 | 0.786457 | 0.625378 | 0.569701 | 0.395323 | 0.610168 | 0.481830 | 0.719930 | 0.667172 | 0.672570 | 0.496323 | 0.438003 | 0.281995 | 0.788962 | 0.626767 | 0.676187 | 0.498181 | 0.606676 | 0.522123 | 0.464294 | 0.379919 | 0.435702 | 0.377638 | 0.315450 | 0.263147 | 0.572746 | 0.398980 | 0.441440 | 0.285975 | 0.827232 | 0.784441 | 0.722987 | 0.670469 |
| Std | 0.023179 | 0.029498 | 0.030160 | 0.032613 | 0.023655 | 0.030876 | 0.030145 | 0.019388 | 0.018027 | 0.017999 | 0.044147 | 0.031611 | 0.030745 | 0.018937 | 0.033366 | 0.019875 | 0.019861 | 0.019044 | 0.053707 | 0.035064 | 0.030313 | 0.018078 | 0.044263 | 0.031959 | 0.054012 | 0.035515 | 0.035414 | 0.050772 | 0.031263 | 0.049336 | 0.082418 | 0.084531 | 0.074648 | 0.070225 | 0.031371 | 0.020304 | 0.031135 | 0.020375 | 0.017269 | 0.017569 | 0.018777 | 0.018162 |
# bb_nested_flair+t1ce+t2 results
| DSC | val_DSC | IoU | val_IoU | loss | val_loss | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | DSC_c2' | val_DSC_c2' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | IoU_c2' | val_IoU_c2' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Fold 1 | 0.779002 | 0.672003 | 0.648688 | 0.529349 | 0.255551 | 0.360792 | 0.719423 | 0.604914 | 0.808188 | 0.767718 | 0.803298 | 0.676052 | 0.546784 | 0.370971 | 0.598273 | 0.482279 | 0.686752 | 0.635360 | 0.689695 | 0.546439 | 0.418371 | 0.265039 | 0.806052 | 0.677486 | 0.693538 | 0.548413 | 0.600253 | 0.512683 | 0.450773 | 0.362851 | 0.531608 | 0.448417 | 0.402653 | 0.322077 | 0.551861 | 0.371505 | 0.423819 | 0.266024 | 0.809950 | 0.769915 | 0.689209 | 0.638156 |
| Fold 2 | 0.827764 | 0.683941 | 0.713086 | 0.546527 | 0.203324 | 0.342268 | 0.763364 | 0.632032 | 0.848093 | 0.816181 | 0.840611 | 0.652884 | 0.601390 | 0.427031 | 0.650383 | 0.514511 | 0.742711 | 0.704199 | 0.736764 | 0.527665 | 0.471673 | 0.311670 | 0.844272 | 0.654541 | 0.742204 | 0.529872 | 0.662827 | 0.563397 | 0.516705 | 0.418426 | 0.538133 | 0.425310 | 0.409423 | 0.308033 | 0.607624 | 0.429368 | 0.478810 | 0.314289 | 0.850091 | 0.818122 | 0.745717 | 0.706958 |
| Fold 3 | 0.764279 | 0.567504 | 0.631895 | 0.431506 | 0.264601 | 0.462027 | 0.693899 | 0.479870 | 0.797439 | 0.720458 | 0.730472 | 0.463688 | 0.553785 | 0.255465 | 0.572566 | 0.376563 | 0.686771 | 0.614727 | 0.608108 | 0.342516 | 0.422820 | 0.172446 | 0.731825 | 0.464218 | 0.610626 | 0.343346 | 0.549546 | 0.473581 | 0.410567 | 0.335781 | 0.344579 | 0.217857 | 0.228485 | 0.141517 | 0.558475 | 0.255673 | 0.427723 | 0.172674 | 0.798667 | 0.721513 | 0.688591 | 0.616397 |
| Fold 4 | 0.782030 | 0.734676 | 0.654056 | 0.599196 | 0.245270 | 0.294846 | 0.715339 | 0.660821 | 0.797784 | 0.787334 | 0.801597 | 0.738290 | 0.546634 | 0.456837 | 0.596886 | 0.540839 | 0.677857 | 0.663773 | 0.693306 | 0.618804 | 0.419495 | 0.339942 | 0.803453 | 0.740217 | 0.696065 | 0.621359 | 0.581990 | 0.558323 | 0.434213 | 0.407843 | 0.551878 | 0.550425 | 0.426098 | 0.413404 | 0.551178 | 0.457601 | 0.424665 | 0.341141 | 0.799103 | 0.788612 | 0.679712 | 0.665513 |
| Fold 5 | 0.828476 | 0.697183 | 0.714185 | 0.556555 | 0.206962 | 0.341428 | 0.768859 | 0.645558 | 0.860668 | 0.811314 | 0.827481 | 0.680249 | 0.618426 | 0.445111 | 0.655753 | 0.527372 | 0.763309 | 0.698642 | 0.716730 | 0.550474 | 0.487219 | 0.333000 | 0.830531 | 0.682551 | 0.721258 | 0.553452 | 0.655564 | 0.538722 | 0.511035 | 0.391749 | 0.520485 | 0.494635 | 0.393330 | 0.358391 | 0.621848 | 0.460075 | 0.491048 | 0.348356 | 0.864061 | 0.814496 | 0.768557 | 0.703372 |
| Mean | 0.796310 | 0.671061 | 0.672382 | 0.532627 | 0.235141 | 0.360272 | 0.732177 | 0.604639 | 0.822434 | 0.780601 | 0.800692 | 0.642233 | 0.573404 | 0.391083 | 0.614772 | 0.488313 | 0.711480 | 0.663340 | 0.688920 | 0.517179 | 0.443916 | 0.284419 | 0.803227 | 0.643803 | 0.692738 | 0.519288 | 0.610036 | 0.529341 | 0.464658 | 0.383330 | 0.497337 | 0.427329 | 0.371998 | 0.308684 | 0.578197 | 0.394844 | 0.449213 | 0.288497 | 0.824374 | 0.782532 | 0.714357 | 0.666079 |
| Std | 0.026659 | 0.055893 | 0.034469 | 0.055558 | 0.025272 | 0.055343 | 0.029085 | 0.065036 | 0.026666 | 0.034750 | 0.038078 | 0.093603 | 0.030399 | 0.073935 | 0.032622 | 0.059153 | 0.034682 | 0.034819 | 0.043847 | 0.092639 | 0.029460 | 0.061804 | 0.038824 | 0.094125 | 0.044753 | 0.093275 | 0.043361 | 0.033089 | 0.042202 | 0.030278 | 0.077051 | 0.113146 | 0.072551 | 0.091172 | 0.030278 | 0.076549 | 0.029446 | 0.064714 | 0.027364 | 0.035236 | 0.035826 | 0.035525 |
Results Summary
Out[16]:
| DSC | val_DSC | IoU | val_IoU | DSC' | val_DSC' | DSC_c0' | val_DSC_c0' | DSC_c1' | val_DSC_c1' | DSC_c2' | val_DSC_c2' | IoU' | val_IoU' | IoU_c0' | val_IoU_c0' | IoU_c1' | val_IoU_c1' | IoU_c2' | val_IoU_c2' | STD-cor-DSC | val_STD-cor-DSC | STD-cor-IoU | val_STD-cor-IoU | STD-ede-DSC | val_STD-ede-DSC | STD-ede-IoU | val_STD-ede-IoU | STD-enh-DSC | val_STD-enh-DSC | STD-enh-IoU | val_STD-enh-IoU | STD-nec-DSC | val_STD-nec-DSC | STD-nec-IoU | val_STD-nec-IoU | STD-whl-DSC | val_STD-whl-DSC | STD-whl-IoU | val_STD-whl-IoU | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bb_nested_flair+t1ce Mean | 0.795513 | 0.664679 | 0.671488 | 0.523074 | 0.727111 | 0.600931 | 0.825176 | 0.782093 | 0.786457 | 0.625378 | 0.569701 | 0.395323 | 0.610168 | 0.481830 | 0.719930 | 0.667172 | 0.672570 | 0.496323 | 0.438003 | 0.281995 | 0.788962 | 0.626767 | 0.676187 | 0.498181 | 0.606676 | 0.522123 | 0.464294 | 0.379919 | 0.435702 | 0.377638 | 0.315450 | 0.263147 | 0.572746 | 0.398980 | 0.441440 | 0.285975 | 0.827232 | 0.784441 | 0.722987 | 0.670469 |
| bb_nested_flair+t1ce Std | 0.023179 | 0.029498 | 0.030160 | 0.032613 | 0.030145 | 0.019388 | 0.018027 | 0.017999 | 0.044147 | 0.031611 | 0.030745 | 0.018937 | 0.033366 | 0.019875 | 0.019861 | 0.019044 | 0.053707 | 0.035064 | 0.030313 | 0.018078 | 0.044263 | 0.031959 | 0.054012 | 0.035515 | 0.035414 | 0.050772 | 0.031263 | 0.049336 | 0.082418 | 0.084531 | 0.074648 | 0.070225 | 0.031371 | 0.020304 | 0.031135 | 0.020375 | 0.017269 | 0.017569 | 0.018777 | 0.018162 |
| bb_nested_flair+t1ce+t2 Mean | 0.796310 | 0.671061 | 0.672382 | 0.532627 | 0.732177 | 0.604639 | 0.822434 | 0.780601 | 0.800692 | 0.642233 | 0.573404 | 0.391083 | 0.614772 | 0.488313 | 0.711480 | 0.663340 | 0.688920 | 0.517179 | 0.443916 | 0.284419 | 0.803227 | 0.643803 | 0.692738 | 0.519288 | 0.610036 | 0.529341 | 0.464658 | 0.383330 | 0.497337 | 0.427329 | 0.371998 | 0.308684 | 0.578197 | 0.394844 | 0.449213 | 0.288497 | 0.824374 | 0.782532 | 0.714357 | 0.666079 |
| bb_nested_flair+t1ce+t2 Std | 0.026659 | 0.055893 | 0.034469 | 0.055558 | 0.029085 | 0.065036 | 0.026666 | 0.034750 | 0.038078 | 0.093603 | 0.030399 | 0.073935 | 0.032622 | 0.059153 | 0.034682 | 0.034819 | 0.043847 | 0.092639 | 0.029460 | 0.061804 | 0.038824 | 0.094125 | 0.044753 | 0.093275 | 0.043361 | 0.033089 | 0.042202 | 0.030278 | 0.077051 | 0.113146 | 0.072551 | 0.091172 | 0.030278 | 0.076549 | 0.029446 | 0.064714 | 0.027364 | 0.035236 | 0.035826 | 0.035525 |
Reducing tahe number of input channels reduced the performance slightl, but noticebly. There is no significant differene between the two and three channel options, though.
This still hints towards limited ability of the networks to exploit relations between the channels.
Summary¶
Overall we see the following:
- Single Modality:
- Flair (and to some small extend t2) dominant:
- Whole tumour: Especially flair and also t2 perform very well; t1/t1ce perform reasonable, but not as good.
- Peritumoural edema: flair performs well; t2 and t1ce have some useful information, but are not as good.
- T1ce dominant:
- Necrotic/non-enhancing tumour: t1ce performs reasonable, but overall performance is poor.
- Enhancing tumour: t1ce performs reasonable, but all others perform poorly.
- Core: t1ce performs well, while t2, flair and t1ce show some good, but generally mixed results across the folds. This makes sense given their performance on the individual regions making up the core.
- Flair (and to some small extend t2) dominant:
- Combine multiple input modalities
- Only for whole tumour: no noticeable improvement when compared to single modality results, sometimes maybe worse; so did not test for others.
- Multi-region segmentation:
- Nested performs better than separated regions overall, but quite mixed results on individual regions.
- Using fewer regions, expectedly, performs better in both cases, as in particular the core is easier to identify.
- For nested approaches, necrotic is in particular hard to identify, but also core performs poorly if edema is last in nested modality (needs to identify a hole).
- For separate approaches, we generally get poorer results, but there are better results for the enhancing and necrotic regions individually compared to the nested approaches (even if they are overall poor).
- Using fewer modalities reduces the performance slightly, but makes not major difference to the results; there is no difference between two and three modalities.
Different modalities are useful for different regions, in particular flair and t1ce are useful and t2 to some extend; t1 does not seem to add much.
All networks seem to be struggling to exploit relations between the modalities.
We see some overfitting.