def run(self):
##### Create identifiers
myconfig = self.config
##### Set network specifics
patchsize = myconfig["Patch size"]
batchsize = myconfig["Batch size"]
output_type = myconfig["Output type"]
only_first = myconfig["Only first"]
#### Data specifics
splits = myconfig["Splits"]
##### load datamanager
loc = myconfig["Nifti Source"]
if self.config['Dataset'] == 'Brats18':
dataset = Brats18.fromFile(loc)
elif self.config['Dataset'] == 'BTD':
dataset = BTD.fromFile(loc)
elif self.config['Dataset'] == 'Generic':
dataset = GenericData.fromFile(loc)
if "Splits from File" in myconfig:
dataset.loadSplits(myconfig["Splits from File"])
##### Load model from source step
sourcestep = myconfig["Model Source"]
loc_model = os.path.join(self.datadir, sourcestep)
config_model = self.loadConfig(os.path.join(loc_model, 'config.json'))
sequences = config_model["Sequences"]
transform = ToTensor()
testset = dataset.getDataset(splits, sequences, transform=transform)
dataloader = DataLoader(testset, batch_size=batchsize, num_workers=0, shuffle=True)
evaluator = StandardEvaluator.loadFromCheckpoint(os.path.join(loc_model, 'model.pt'))
self.logger.info('Dataloader has {n} images.'.format(n=len(testset)))
for i_batch, sample_batched in enumerate(dataloader):
images = sample_batched["data"]
subjects = sample_batched["subject"]
header_sources = sample_batched["header_source"]
t1_sources = sample_batched["t1_source"]
resultpaths = [os.path.join(self.tmpdir, s + '_segmented.nii.gz') for s in subjects]
evaluator.segmentNifti(images, header_sources, patchsize, resultpaths)
for i in range(0, len(subjects)):
plotResultImageWithoutGT(t1_sources[i], resultpaths[i], self.tmpdir,
subjects[i], output_type=output_type)
self.logger.info('Nifti image segmented for {}.'.format(subjects[i]))
if only_first: break
self.tearDown()
def getDataloader(self):
n_workers = self.config['Num Workers']
batchsize = self.config['Batch size']
sequences = self.config["Sequences"]
loc = os.path.join(self.datadir, self.config["Nifti Source"])
# locate, organize and split the dataset
dataset = Brats18.fromFile(loc)
#Data specifics
splits = self.config['Splits']
testsplits = self.config['Testsplits']
dataset.saveSplits(self.tmpdir)
targetsize = tuple(self.config["Patch size"])
imgsize = targetsize
# initialize input transforms
transforms = [RandomCrop(output_size=imgsize), ToTensor()]
transform = Compose(transforms)
# prepare the training set loader
trainset = dataset.getDataset(splits, sequences, transform=transform)
trainset.saveListOfPatients(os.path.join(self.tmpdir, 'trainset.json'))
self.logger.info('Generating patches with input size ' + str(imgsize) +
' and outputsize ' + str(targetsize))
trainloader = DataLoader(trainset,
batch_size=batchsize,
num_workers=n_workers,
shuffle=True)
# prepare the testing set loader
if len(testsplits) > 0:
testset = dataset.getDataset(testsplits,
sequences,
transform=transform)
testloader = DataLoader(testset,
batch_size=batchsize,
num_workers=n_workers,
shuffle=True)
else:
testloader = None
# plot and save samples of a mini-batch
logDataLoader(trainloader, self.tmpdir)
return trainloader, testloader
def evaluate(self):
# Create identifiers
myconfig = self.config
# Set network specifics
patchsize = myconfig["Patch size"]
batchsize = myconfig["Batch size"]
output_type = myconfig["Output type"]
only_first = myconfig["Only first"]
sequences = myconfig["Sequences"]
# Data specifics
splits = myconfig["Testsplits"]
# load data-manager
loc = os.path.join(self.datadir, self.config["Nifti Source"])
dataset = Brats18.fromFile(loc)
transforms = [ToTensor()]
transform = Compose(transforms)
testset = dataset.getDataset(splits, sequences, transform=transform)
dataloader = DataLoader(testset,
batch_size=batchsize,
num_workers=0,
shuffle=True)
self.logger.info('Dataloader has {n} images.'.format(n=len(testset)))
# Load model from source step
sourcestep = self.name
loc_model = os.path.join(self.datadir, sourcestep)
evaluator = StandardEvaluator.loadFromCheckpoint(
os.path.join(loc_model, 'model.pt'))
all_dice = []
all_dice_core = []
all_dice_enhancing = []
results = pd.DataFrame(columns=[
'sample', 'subject', 'class', 'TP', 'FP', 'FN', 'TN', 'dice'
])
for i_batch, sample_batched in enumerate(dataloader):
images = sample_batched['data']
segfiles = sample_batched['seg_file']
subjects = sample_batched['subject']
segs = sample_batched['seg']
resultpaths = [
os.path.join(self.tmpdir, s + '_segmented.nii.gz')
for s in subjects
]
uncertaintypaths = [
os.path.join(self.tmpdir, s + '_epistimci.nii.gz')
for s in subjects
]
classifications, epistemicUncertainty = evaluator.segmentNifti(
images, segfiles, patchsize, resultpaths, uncertaintypaths)
for i in range(0, len(subjects)):
seg = segs[i].numpy()
plotResultImage(dataset,
resultpaths[i],
uncertaintypaths[i],
self.tmpdir,
subjects[i],
output_type=output_type)
for c in range(0, 5):
truth = seg == c
positive = classifications[i] == c
(dice, TT, FP, FN,
TN) = getPerformanceMeasures(positive, truth)
results = results.append(
{
'sample': i_batch,
'class': c,
'subject': subjects[i],
'TP': TT,
'FP': FP,
'FN': FN,
'TN': TN,
'dice': dice
},
ignore_index=True)
if c == 4:
all_dice_enhancing.append(dice)
class_whole = classifications[i] > 0
result_core = (classifications[i] == 1) | (classifications[i]
== 4)
truth_whole = seg > 0
truth_core = (seg == 1) | (seg == 4)
(dice, TT, FP, FN,
TN) = getPerformanceMeasures(class_whole, truth_whole)
(dice_core, TT_core, FP_core, FN_core,
TN_core) = getPerformanceMeasures(result_core, truth_core)
all_dice.append(dice)
all_dice_core.append(dice_core)
self.logger.info('Nifti image segmented for ' + subjects[i] +
'. Dice: ' + str(dice))
results = results.append(
{
'sample': i_batch,
'class': 'whole',
'subject': subjects[i],
'TP': TT,
'FP': FP,
'FN': FN,
'TN': TN,
'dice': dice
},
ignore_index=True)
results = results.append(
{
'sample': i_batch,
'class': 'core',
'subject': subjects[i],
'TP': TT_core,
'FP': FP_core,
'FN': FN_core,
'TN': TN_core,
'dice': dice_core
},
ignore_index=True)
if only_first: break
dice_mean = sum(all_dice) / len(all_dice)
dice_core = sum(all_dice_core) / len(all_dice_core)
dice_enhancing = sum(all_dice_enhancing) / len(all_dice_enhancing)
plt.boxplot(all_dice)
plt.savefig(os.path.join(self.tmpdir, 'boxplot_dice.png'))
plt.close()
results.to_csv(os.path.join(self.tmpdir, 'results_eval.csv'))
dataset.saveSplits(self.tmpdir)
self.logger.info(
'evaluation finished. Dice coefficient: whole: {}, core: {}, enhancing: {}'
.format(dice_mean, dice_core, dice_enhancing))
def getDataloader(self):
batchsize = self.config['Batch size']
loc = os.path.join(self.datadir, self.config["Nifti Source"])
sequences = self.config[
"Sequences"] if "Sequences" in self.config else None
if self.config['Dataset'] == 'Brats18':
dataset = Brats18.fromFile(loc)
elif self.config['Dataset'] == 'BTD':
if self.config['Brainmask'] == False:
dataset = BTD.fromFile(loc, brainmask=False)
else:
dataset = BTD.fromFile(loc, brainmask=True)
elif self.config['Dataset'] == 'Hippocampus':
dataset = Hippocampus.fromFile(loc)
if "Splits from File" in self.config:
dataset.setSplits(self.config["Splits from File"])
elif "Crossval Splits" in self.config:
dataset.createCVSplits(self.config["Crossval Splits"])
#### Data specifics
splits = self.config['Splits']
testsplits = self.config['Testsplits']
dataset.saveSplits(self.tmpdir)
# check if patch size isn't too large. If so make it smaller
# targetsize = tuple(self.config["Patch size"])
targetsize = tuple(set_patch_size(self.config))
imgsize = targetsize
transforms = [RandomCrop(output_size=imgsize), ToTensor()]
if 'Whole Tumor' in self.config and self.config["Whole Tumor"]:
transforms = [BinarySegmentation()] + transforms
transform = Compose(transforms)
if 'Target' in self.config and self.config['Target'] == 'Brainmask':
trainset = dataset.getBrainmaskDataset(splits,
sequences,
transform=transform)
else:
trainset = dataset.getDataset(splits,
sequences,
transform=transform)
trainset.saveListOfPatients(os.path.join(self.tmpdir, 'trainset.json'))
self.logger.info('Generating patches with input size ' + str(imgsize) +
' and outputsize ' + str(targetsize))
n_workers = self.config['Num Workers']
trainloader = DataLoader(trainset,
batch_size=batchsize,
num_workers=n_workers,
shuffle=True)
if len(testsplits) > 0:
if 'Target' in self.config and self.config['Target'] == 'Brainmask':
testset = dataset.getBrainmaskDataset(testsplits,
sequences,
transform=transform)
else:
testset = dataset.getDataset(testsplits,
sequences,
transform=transform)
testloader = DataLoader(testset,
batch_size=batchsize,
num_workers=n_workers,
shuffle=True)
else:
testloader = None
logDataLoader(trainloader, self.tmpdir)
return trainloader, testloader
def getDataloader(self):
batchsize = self.config['Batch size']
loc = os.path.join(self.datadir, self.config["Nifti Source"])
sequences = self.config[
"Sequences"] if "Sequences" in self.config else None
# Loading the specialized config settings from file
print("normalization technique is:", self.config["technique"])
print("using otsu threshold for normalization:",
self.config["using_otsu_ROI"])
print("resampling factor is:", self.config["resampling_factor"])
if self.config['Dataset'] == 'Brats18':
dataset = Brats18.fromFile(loc)
elif self.config['Dataset'] == 'BTD':
print(self.config['Brainmask'])
if self.config['Brainmask'] == False:
dataset = BTD.fromFile(loc, brainmask=False)
else:
dataset = BTD.fromFile(loc, brainmask=True)
elif self.config['Dataset'] == 'Hippocampus':
dataset = Hippocampus.fromFile(loc)
elif self.config['Dataset'] == 'LipoData':
dataset = LipoData.fromFile(loc)
elif self.config['Dataset'] == 'LitsData':
dataset = LitsData.fromFile(loc)
elif self.config['Dataset'] == 'ErgoData':
dataset = ErgoData.fromFile(loc)
if "Splits from File" in self.config:
dataset.setSplits(self.config["Splits from File"])
elif "Crossval Splits" in self.config:
dataset.createCVSplits(self.config["Crossval Splits"])
#### Data specifics
splits = self.config['Splits']
testsplits = self.config['Testsplits']
dataset.saveSplits(self.tmpdir)
# check if patch size isn't too large. If so make it smaller
# targetsize = tuple(self.config["Patch size"])
targetsize = tuple(set_patch_size(self.config))
imgsize = targetsize
transforms = [RandomCrop(output_size=imgsize), ToTensor()]
if 'Whole Tumor' in self.config and self.config["Whole Tumor"]:
transforms = [BinarySegmentation()] + transforms
transform = Compose(transforms)
if 'Target' in self.config and self.config['Target'] == 'Brainmask':
trainset = dataset.getBrainmaskDataset(splits,
sequences,
transform=transform)
else:
trainset = dataset.getDataset(splits,
sequences,
transform=transform,
preprocess_config=self.config)
trainset.saveListOfPatients(os.path.join(self.tmpdir, 'trainset.json'))
self.logger.info('Generating patches with input size ' + str(imgsize) +
' and outputsize ' + str(targetsize))
n_workers = self.config['Num Workers']
now = datetime.now()
print('train loader is initializing', now)
trainloader = DataLoader(trainset,
batch_size=batchsize,
num_workers=n_workers,
shuffle=True)
later = datetime.now()
print('train loader is done initializing', later - now)
if len(testsplits) > 0:
if 'Target' in self.config and self.config['Target'] == 'Brainmask':
testset = dataset.getBrainmaskDataset(testsplits,
sequences,
transform=transform)
else:
testset = dataset.getDataset(testsplits,
sequences,
transform=transform,
preprocess_config=self.config)
testloader = DataLoader(testset,
batch_size=batchsize,
num_workers=n_workers,
shuffle=True)
else:
testloader = None
logDataLoader(trainloader, self.tmpdir)
return trainloader, testloader
def createNewDatamanager(location, dataset, nsplits):
if dataset == 'Brats':
return Brats18.fromFile(location, nsplits=nsplits)
def run(self):
##### Create identifiers
myconfig = self.config
##### Set network specifics
patchsize = myconfig["Patch size"]
batchsize = myconfig["Batch size"]
output_type = myconfig["Output type"]
only_first = myconfig["Only first"]
#### Data specifics
splits = myconfig["Splits"]
##### load datamanager
loc = os.path.join(self.datadir, self.config["Nifti Source"])
if self.config['Dataset'] == 'Brats18':
dataset = Brats18.fromFile(loc)
elif self.config['Dataset'] == 'BTD':
dataset = BTD.fromFile(loc)
elif self.config['Dataset'] == 'Hippocampus':
dataset = Hippocampus.fromFile(loc)
if "Splits from File" in myconfig:
dataset.loadSplits(myconfig["Splits from File"])
##### Load model from source step
sourcestep = myconfig["Model Source"]
loc_model = os.path.join(self.datadir, sourcestep)
config_model = self.loadConfig(os.path.join(loc_model, 'config.json'))
if "Sequences" in self.config:
sequences = self.config["Sequences"]
else:
sequences = config_model["Sequences"]
transforms = [ToTensor()]
if 'Whole Tumor' in self.config and self.config["Whole Tumor"]:
transforms = [BinarySegmentation()] + transforms
transform = Compose(transforms)
if 'Target' in self.config and self.config['Target'] == 'Brainmask':
testset = dataset.getBrainmaskDataset(splits, sequences, transform=transform)
else:
testset = dataset.getDataset(splits, sequences, transform=transform)
dataloader = DataLoader(testset, batch_size=batchsize, num_workers=0, shuffle=True)
evaluator = StandardEvaluator.loadFromCheckpoint(os.path.join(loc_model, 'model.pt'))
self.logger.info('Dataloader has {n} images.'.format(n=len(testset)))
all_dice = []
all_dice_core = []
all_dice_enhancing = []
results = pd.DataFrame(columns=['sample', 'subject', 'class', 'TP', 'FP', 'FN', 'TN', 'dice'])
for i_batch, sample_batched in enumerate(dataloader):
######### TODO: work with original dataset
images = sample_batched['data']
segfiles = sample_batched['seg_file']
subjects = sample_batched['subject']
segs = sample_batched['seg']
resultpaths = [os.path.join(self.tmpdir, s+'_segmented.nii.gz') for s in subjects]
uncertaintypaths = [os.path.join(self.tmpdir, s + '_epistimci.nii.gz') for s in subjects]
classifications, epistemicUncertainty = evaluator.segmentNifti(images, segfiles, patchsize, resultpaths,
uncertaintypaths)
import pickle
segmentations = segs.numpy()
picklepath = myconfig["pickleVar"]
with open(picklepath, 'wb') as f:
pickle.dump([segmentations, classifications, epistemicUncertainty, segfiles, subjects], f)
for i in range(0, len(subjects)):
seg = segs[i].numpy()
plotResultImage(dataset, resultpaths[i], uncertaintypaths[i], self.tmpdir, subjects[i], output_type=output_type)
for c in range(0,5):
truth = seg == c
positive = classifications[i] == c
(dice, TT, FP, FN, TN) = getPerformanceMeasures(positive, truth)
results = results.append({'sample':i_batch, 'class':c, 'subject':subjects[i], 'TP': TT, 'FP': FP, 'FN': FN, 'TN': TN, 'dice': dice}, ignore_index=True)
if c == 4:
all_dice_enhancing.append(dice)
class_whole = classifications[i] > 0
result_core = (classifications[i] == 1) | (classifications[i] == 4)
truth_whole = seg > 0
truth_core = (seg == 1) | (seg == 4)
(dice, TT, FP, FN, TN) = getPerformanceMeasures(class_whole, truth_whole)
(dice_core, TT_core, FP_core, FN_core, TN_core) = getPerformanceMeasures(result_core, truth_core)
all_dice.append(dice)
all_dice_core.append(dice_core)
self.logger.info('Nifti image segmented for ' + subjects[i] + '. Dice: ' + str(dice))
results = results.append({'sample':i_batch, 'class':'whole', 'subject':subjects[i], 'TP': TT, 'FP': FP, 'FN': FN,'TN': TN, 'dice': dice}, ignore_index=True)
results = results.append(
{'sample': i_batch, 'class': 'core', 'subject': subjects[i], 'TP': TT_core, 'FP': FP_core, 'FN': FN_core,
'TN': TN_core, 'dice': dice_core}, ignore_index=True)
if only_first: break
dice_mean = sum(all_dice)/len(all_dice)
dice_core = sum(all_dice_core)/len(all_dice_core)
dice_enhancing = sum(all_dice_enhancing) / len(all_dice_enhancing)
plt.boxplot(all_dice)
plt.savefig(os.path.join(self.tmpdir, 'boxplot_dice.png'))
plt.close()
results.to_csv(os.path.join(self.tmpdir, 'results_eval.csv'))
dataset.saveSplits(self.tmpdir)
self.logger.info('evaluation finished. Dice coefficient: whole: {}, core: {}, enhancing: {}'.format(dice_mean, dice_core, dice_enhancing))
self.tearDown()
