import math
import os
import time
import cv2
import matplotlib.pyplot as plt
import numpy as np
import scipy
import scipy.misc
import scipy.ndimage
import scipy.signal
from imageio import imwrite
from skimage.measure import regionprops, label
from trainers import Metrics
from utils import image_utils, utils
def should(dictionary, key):
return key in dictionary and dictionary[key]
def get_eval_dictionary():
_eval = {
'x': [],
'reconstructions': [],
'diffs': [],
'epistemic_variance': [],
'labelmaps': [],
'reconstructionTimes': [],
'l1reconstructionErrors': [],
'l1reconstructionErrorMean': 0.0,
'l1reconstructionErrorSigma': 0.0,
'l2reconstructionErrors': [],
'l2reconstructionErrorMean': 0.0,
'l2reconstructionErrorSigma': 0.0,
'TP': 0,
'FP': 0,
'FN': 0,
'TN': 0,
'Dice': [],
'TPCC': 0,
'FPCC': 0,
'FNCC': 0
}
return _eval
def merge_eval_dictionaries(eval_dict, _eval_dict):
for k in eval_dict:
if isinstance(eval_dict[k], np.ndarray):
eval_dict[k] = np.concatenate((eval_dict[k], _eval_dict[k]), axis=0)
elif isinstance(eval_dict[k], list):
if isinstance(_eval_dict[k], list):
eval_dict[k] += _eval_dict[k]
else:
eval_dict[k] += [_eval_dict[k]]
return eval_dict
def is_float(s):
try:
float(s)
return True
except ValueError:
return False
def squash_intensities(img):
# logistic function intended to squash reconstruction errors from [0;0.2] to [0;1] (just an example)
k = 100
offset = 0.5
return 2.0 * ((1.0 / (1.0 + np.exp(-k * img))) - offset)
def add_colorbar(img):
for i in range(img.shape[0]):
img[i, -1] = float(i) / img.shape[0]
return img
def apply_brainmask(x, brainmask, erode=True):
strel = scipy.ndimage.generate_binary_structure(2, 1)
brainmask = np.expand_dims(brainmask, 2)
if erode:
brainmask = scipy.ndimage.morphology.binary_erosion(np.squeeze(brainmask), structure=strel, iterations=12)
return np.multiply(np.squeeze(brainmask), np.squeeze(x))
def postprocess_slice(x, x_rec, slice_skullmap=None):
if slice_skullmap is None:
brainmasks = np.ones(x.shape)
else:
strel = scipy.ndimage.generate_binary_structure(2, 1)
brainmasks = np.expand_dims(slice_skullmap, 3)
brainmasks = scipy.ndimage.morphology.binary_erosion(np.squeeze(brainmasks), structure=strel, iterations=12)
x_prior = np.squeeze(x < 0.6)
x_diff = np.multiply(np.squeeze(brainmasks), np.squeeze(x - x_rec))
x_diff[x_diff < 0] = 0
x_diff[x_prior] = 0
return x_diff
def apply_3d_median_filter(volume, kernelsize=5): # kernelsize 5 works quite well
volume = scipy.ndimage.filters.median_filter(volume, (kernelsize, kernelsize, kernelsize))
return volume
def filter_3d_connected_components(volume):
sz = None
if volume.ndim > 3:
sz = volume.shape
volume = np.reshape(volume, [sz[0] * sz[1], sz[2], sz[3]])
cc_volume = label(volume, connectivity=3)
props = regionprops(cc_volume)
for prop in props:
if prop['filled_area'] <= 7:
volume[cc_volume == prop['label']] = 0
if sz is not None:
volume = np.reshape(volume, [sz[0], sz[1], sz[2], sz[3]])
return volume
def compute_detection_rate(predicted_volume, groundtruth_volume):
tps = 0
fns = 0
fps = 0
num_slices = groundtruth_volume.shape[0]
# First, compute intersection of prediction and ground-truth to determine True Positives
intersected_volume = np.multiply(predicted_volume, groundtruth_volume)
for s in range(int(math.ceil(num_slices / 20))):
cc_intersected_volume = label(intersected_volume[s * 20:min((s + 1) * 20, num_slices), :, :])
props_intersected = regionprops(cc_intersected_volume)
cc_predicted_volume = label(predicted_volume[s * 20:min((s + 1) * 20, num_slices), :, :])
props_predicted = regionprops(cc_predicted_volume)
cc_groundtruth_volume = label(groundtruth_volume[s * 20:min((s + 1) * 20, num_slices), :, :])
# Filter cc_predicted_volume for any positives which have less than 8 voxels in size
for pidx, pprop in enumerate(props_predicted):
if pprop["area"] < 8:
cc_predicted_volume[cc_predicted_volume == pprop["label"]] = 0
# Then, remove all the TP connected components from cc_predicted_volume to later determine any False Positives (FPs)
# Do the same for TP connected components in cc_groundtruth_volume to later be able to determine any False Negatives (FNs)
for tpidx, tpprop in enumerate(props_intersected):
coords = tpprop["coords"][0]
label_in_cc_predicted_volume = cc_predicted_volume[int(coords[0]), int(coords[1]), int(coords[2])]
cc_predicted_volume[cc_predicted_volume == label_in_cc_predicted_volume] = 0
label_in_cc_groundtruth_volume = cc_groundtruth_volume[int(coords[0]), int(coords[1]), int(coords[2])]
cc_groundtruth_volume[cc_groundtruth_volume == label_in_cc_groundtruth_volume] = 0
# Recompute the regionprops on cc_predicted_volume and cc_groundtruth_volume to determine FPs and FNs
props_falsely_predicted = regionprops(cc_predicted_volume)
props_falsely_missed = regionprops(cc_groundtruth_volume)
# Done
tps += len(props_intersected)
fns += len(props_falsely_missed)
fps += len(props_falsely_predicted)
return tps, fps, fns
def postprocess_volume(volume):
volume = scipy.ndimage.filters.median_filter(volume, (5, 5, 5))
# subvolume = scipy.ndimage.filters.gaussian_filter(subvolume, 3, truncate=3.0)
# 3D Connected Component Analysis
return filter_3d_connected_components(volume)
def _evaluate(datasetObj, modelObj, sampleDir, options, split="TEST"):
os.makedirs(sampleDir, exist_ok=True)
# Determine the number of testing samples
num_testing_samples = datasetObj.num_batches(1, set=split) # batchsize is set to 1 here so we can evaluate per sample
print("Testing {} samples...".format(num_testing_samples))
# Setup eval Dictionary
eval_dict = get_eval_dictionary()
# Iterate over all patients, and therein, query the desired Nifti files and slices
patients = [datasetObj.patients[i] for i in datasetObj.get_patient_idx(split=split)]
for p, patient in enumerate(patients):
_eval_dict = get_eval_dictionary()
filtered_files = patient['filtered_files']
if type(filtered_files) is not list:
filtered_files = [filtered_files]
for n, nii_filename in enumerate(filtered_files):
if len(_eval_dict['diffs']) == 0:
nii, nii_seg, nii_skullmap = datasetObj.load_volume_and_groundtruth(nii_filename, patient)
prior_quantile = np.quantile(nii.data, 0.9)
# Sanity checks - if coregistration went wrong and shapes are bad, we skip this sample
if min(nii.shape()) < (datasetObj.options.sliceEnd - datasetObj.options.sliceStart):
continue
# Iterate over all slices and collect them
subvolume = np.zeros(
[datasetObj.options.sliceEnd - datasetObj.options.sliceStart, options['train']['outputHeight'],
options['train']['outputWidth']])
subvolume_idx = 0
slice_start = 0
slice_end = nii.num_slices_along_axis(datasetObj.options.axis)
zoom_factor = 1.0
if datasetObj.options.sliceStart:
slice_start = datasetObj.options.sliceStart
if datasetObj.options.sliceEnd:
slice_end = min(datasetObj.options.sliceEnd, nii.num_slices_along_axis(datasetObj.options.axis))
for s in range(slice_start, slice_end):
slice_data = nii.get_slice(s, datasetObj.options.axis)
slice_seg = nii_seg.get_slice(s, datasetObj.options.axis).astype(int)
slice_skullmap = nii_skullmap.get_slice(s, datasetObj.options.axis).astype(int)
if datasetObj.options.sliceResolution is not None:
zoom_factor = tuple([i / j for (i, j) in zip(datasetObj.options.sliceResolution, slice_data.shape)])
slice_data = scipy.ndimage.zoom(slice_data, zoom_factor)
slice_seg = scipy.ndimage.zoom(slice_seg, zoom_factor, mode="nearest")
slice_skullmap = scipy.ndimage.zoom(slice_skullmap, zoom_factor, mode="nearest")
x = np.expand_dims(slice_data, 2)
labelmaps = np.expand_dims(slice_seg, 2)
_tmp = time.time()
# Monte Carlo Uncertainty Estimation
num_samples = 1
if should(options, "numMonteCarloSamples"):
num_samples = options["numMonteCarloSamples"]
x_recs = []
x_diffs = []
x_log_vars = []
results = None
for i in range(num_samples):
if num_samples > 1:
results = modelObj.reconstruct(x, dropout=True)
else:
results = modelObj.reconstruct(x)
x_rec_tmp = results['reconstruction']
if "log_var" in results:
x_log_vars += [results["log_var"]]
x_recs += [np.reshape(apply_brainmask(x_rec_tmp, slice_skullmap, erode=should(options, "erodeBrainmask")),
[1, *datasetObj.options.sliceResolution, 1])]
x_diffs += [
np.reshape(apply_brainmask(np.maximum(x - x_rec_tmp, 0), slice_skullmap, erode=should(options, "erodeBrainmask")),
[1, *datasetObj.options.sliceResolution, 1])]
x_recs = np.array(x_recs)
x_diffs = np.array(x_diffs)
x_log_vars = np.array(x_log_vars)
if x_log_vars.size == 0:
x_log_vars = np.zeros(x_diffs.shape)
x_recs_var = Metrics.combined_predictive_uncertainty(x_recs, x_log_vars, axis=0, log_var=False)
x_recs_var_epistemic = Metrics.combined_predictive_uncertainty(x_recs, np.zeros(x_recs.shape), axis=0, log_var=False)
x_recs_mean = np.mean(x_recs, axis=0)
x_recs_var = apply_brainmask(x_recs_var, slice_skullmap, erode=should(options, "erodeBrainmask"))
x_recs_var_epistemic * (2 * np.expand_dims(np.expand_dims(slice_skullmap, axis=0), axis=-1) - 1)
# values outside the brain are getting negative, while values on the brain stay the same
_eval_dict['reconstructionTimes'] += [time.time() - _tmp]
# Get a sample without dropout
x_rec = results['reconstruction']
l1err = results['l1err']
l2err = results['l2err']
if num_samples > 1:
x_rec = x_recs_mean
if should(options, "keepOnlyPositiveResiduals"):
x_diff = np.maximum(x - x_rec, 0)
else:
x_diff = np.abs(x - x_rec)
x_diff = np.reshape(apply_brainmask(x_diff, slice_skullmap, erode=should(options, "erodeBrainmask")),
[1, *datasetObj.options.sliceResolution, 1])
if should(options, "applyHyperIntensityPrior"):
x_diff[np.reshape(x, [1, *datasetObj.options.sliceResolution, 1]) < prior_quantile] = 0
subvolume[subvolume_idx, :, :] = np.squeeze(x_diff)
subvolume_idx += 1
# Fill eval array
_eval_dict['x'] += [x]
if num_samples > 1:
_eval_dict['epistemic_variance'] += [x_recs_var_epistemic]
_eval_dict['reconstructions'] += [x_rec]
_eval_dict['labelmaps'] += [np.squeeze(labelmaps)]
_eval_dict['l1reconstructionErrors'] += [l1err]
_eval_dict['l2reconstructionErrors'] += [l2err]
imwrite(os.path.join(sampleDir, '{}_{}.png'.format(p, s)), normalize_and_squeeze(x))
imwrite(os.path.join(sampleDir, '{}_{}_rec.png'.format(p, s)), normalize_and_squeeze(x_rec))
imwrite(os.path.join(sampleDir, '{}_{}_gt.png'.format(p, s)), normalize_and_squeeze(labelmaps)) # check if normalization is useful
imwrite(os.path.join(sampleDir, '{}_{}_diff.png'.format(p, s)), normalize_and_squeeze(x_diff))
imwrite(os.path.join(sampleDir, '{}_{}_rec_variance_combined.png'.format(p, s)),
np.squeeze(utils.apply_colormap(x_recs_var, plt.cm.jet)))
if x_log_vars.size > 0:
imwrite(os.path.join(sampleDir, '{}_{}_logvar.png'.format(p, s)), normalize_and_squeeze(np.mean(x_log_vars, axis=0)))
if should(options, "medianFiltering"):
subvolume = apply_3d_median_filter(subvolume)
_eval_dict['diffs'] += [subvolume]
for s in range(datasetObj.options.sliceStart, min(datasetObj.options.sliceEnd, nii.num_slices_along_axis(datasetObj.options.axis))):
imwrite(os.path.join(sampleDir, '{}_{}_diff_filtered.png'.format(p, s)),
normalize_and_squeeze(subvolume[s - datasetObj.options.sliceStart]))
squashed = squash_intensities(np.squeeze(subvolume[s - datasetObj.options.sliceStart]))
squashed = add_colorbar(squashed)
imwrite(os.path.join(sampleDir, '{}_{}_heatmap.png'.format(p, s)), np.squeeze(utils.apply_colormap(squashed, plt.cm.jet)))
if should(options, "exportVolumes"):
dezoom_factor = tuple([1]) + tuple(1 / np.asarray(zoom_factor))
subvolume_deprocessed = scipy.ndimage.interpolation.zoom(subvolume, dezoom_factor)
nii_seg.set_to_zero()
nii_seg.cast_to_float()
nii_seg.set_subvolume(datasetObj.options.sliceStart, datasetObj.options.sliceEnd, subvolume_deprocessed,
axis=datasetObj.options.axis)
nii_seg.save(os.path.join(sampleDir, '{}.nii.gz'.format(patient['name'])))
if options['threshold'] and is_float(options['threshold']):
nii_seg.data = np.asarray((nii_seg.data > options['threshold'])).astype(np.float32)
nii_seg.update_sitk()
nii_seg.save(os.path.join(sampleDir, '{}.binary.nii.gz'.format(patient['name'])))
# Update the total eval_dict
eval_dict['x'] += _eval_dict['x']
eval_dict['diffs'] += _eval_dict['diffs']
eval_dict['reconstructions'] += _eval_dict['reconstructions']
eval_dict['labelmaps'] += _eval_dict['labelmaps']
eval_dict['l1reconstructionErrors'] += _eval_dict['l1reconstructionErrors']
eval_dict['l2reconstructionErrors'] += _eval_dict['l2reconstructionErrors']
if "epistemic_variance" in _eval_dict and len(_eval_dict["epistemic_variance"]) > 0:
eval_dict['epistemic_variance'] += _eval_dict['epistemic_variance']
print("Done.")
# Convert list of numpy arrays to numpy array
eval_dict['x'] = np.squeeze(np.array(eval_dict['x']))
eval_dict['reconstructions'] = np.squeeze(np.array(eval_dict['reconstructions']))
eval_dict['diffs'] = np.squeeze(np.array(eval_dict['diffs']))
if eval_dict['diffs'].ndim > 3:
eval_dict['diffs'] = np.reshape(eval_dict['diffs'], [eval_dict['diffs'].shape[0] * eval_dict['diffs'].shape[1],
eval_dict['diffs'].shape[2], eval_dict['diffs'].shape[3]])
eval_dict['labelmaps'] = np.squeeze(np.array(eval_dict['labelmaps']))
if "epistemic_variance" in eval_dict and len(eval_dict["epistemic_variance"]) > 0:
eval_dict['epistemic_variance'] = np.squeeze(np.array(eval_dict['epistemic_variance']))
# Computer average reconstruction error s etc
eval_dict['l1reconstructionErrorMean'] = np.mean(eval_dict['l1reconstructionErrors'])
eval_dict['l1reconstructionErrorVariance'] = np.var(eval_dict['l1reconstructionErrors'])
eval_dict['l2reconstructionErrorMean'] = np.mean(eval_dict['l2reconstructionErrors'])
eval_dict['l2reconstructionErrorVariance'] = np.var(eval_dict['l2reconstructionErrors'])
eval_dict['reconstructionTimes'] = np.mean(np.array(eval_dict['reconstructionTimes']))
return eval_dict, patients
def normalize_and_squeeze(x):
return np.squeeze(cv2.normalize(x, None, 0, 255, norm_type=cv2.NORM_MINMAX)).astype('uint8')
def evaluate(datasetPC, gan, options, epoch='last', description=None):
_time = {'evaluation': time.time()}
# Variables
histogram_range = (0.01, 0.075)
num_slices = options["sliceEnd"] - options["sliceStart"]
# Create eval folder
eval_dir = os.path.join(
options['train']['samplesDir'],
gan.network.__name__,
gan.model_dir,
'eval-' + str(epoch) + '-' + str(utils.timestamp()).replace(":", "-")
)
if description is not None:
eval_dir += "-" + str(description)
if not os.path.exists(eval_dir):
os.makedirs(eval_dir)
# EVALUATE LESION SAMPLES #
sample_dir = os.path.join(eval_dir, 'samples_test_PC')
if not os.path.exists(sample_dir):
os.makedirs(sample_dir)
eval_pc, patients_pc = _evaluate(datasetPC, gan, sample_dir, options, split="TEST")
print("Computing histogram for lesion testing difference images")
eval_pc['diffHistogram'], _ = np.histogram(eval_pc['diffs'], bins='auto', range=histogram_range)
utils.plot_histogram_with_labels(eval_pc['diffs'], eval_pc['labelmaps'], 'auto', histogram_range,
"Histogram of difference images in the lesion testing dataset",
exportPDF=os.path.join(eval_dir, 'testing_lesions_diffimages_histogram.pdf'))
print("Done.")
if "epistemic_variance" in eval_pc and len(eval_pc["epistemic_variance"]) > 0:
print("Computing uncertainty histogram for lesion testing difference images")
percentil_99 = np.percentile(eval_pc['epistemic_variance'][eval_pc['epistemic_variance'] >= 0], 99.8)
_range = (1e-5, percentil_99)
eval_pc['uncertaintyHistogram'], _ = np.histogram(eval_pc['epistemic_variance'], bins=50, range=_range)
utils.plot_histogram_with_labels(eval_pc['epistemic_variance'], eval_pc['labelmaps'], 50, _range,
"Histogram of Epistemic Variances images in the lesion testing dataset",
exportPDF=os.path.join(eval_dir, 'testing_lesions_epistemic_variances_histogram.pdf'))
print("Done.")
print("Computing ROC curve for Lesion samples")
_time['ROC'] = time.time()
eval_pc['diff_AUC'], _fpr, _tpr, _threshs = Metrics.compute_roc(eval_pc['diffs'].flatten(), eval_pc['labelmaps'].astype(bool).flatten(),
plottitle="ROC Curve for Lesion Testing Samples",
filename=os.path.join(eval_dir, 'rocPC.png'))
_time['ROC'] = time.time() - _time['ROC']
print('Done in {} seconds'.format(_time['ROC']))
if should(options, "exportROC"):
_tmp = {"fpr": _fpr, "tpr": _tpr, "threshs": _threshs}
np.save(os.path.join(eval_dir, 'rocPC.npy'), _tmp, allow_pickle=True)
print("Computing Precision-Recall curve for Lesion samples")
_time['PRC'] = time.time()
eval_pc['diff_AUPRC'], _precisions, _recalls, _threshs = Metrics.compute_prc(
eval_pc['diffs'].flatten(),
eval_pc['labelmaps'].astype(bool).flatten(),
plottitle="Precision-Recall Curve for Lesion Testing Samples",
filename=os.path.join(eval_dir, 'prcPC.png')
)
_time['PRC'] = time.time() - _time['PRC']
print('Done in {} seconds'.format(_time['PRC']))
if should(options, "exportPRC"):
_tmp = {"precisions": _precisions, "recalls": _recalls, "threshs": _threshs}
np.save(os.path.join(eval_dir, 'prcPC.npy'), _tmp, allow_pickle=True)
# Quickly determine thresholds for different precisions to get the maximal possible recall
idx_precision70 = np.argmax(_precisions <= 0.7)
diffs_thresholded_at_precision70 = filter_3d_connected_components(np.squeeze(eval_pc['diffs'] > _threshs[idx_precision70]))
print("Computing DICE curve for Lesion samples")
_time['DiceCurve'] = time.time()
eval_pc['bestDiceScore'], eval_pc['bestThreshold'] = Metrics.compute_dice_curve_recursive(
eval_pc['diffs'].flatten(), eval_pc['labelmaps'].flatten(),
plottitle="DICE vs Thresholds Curve for Lesion Testing Samples",
filename=os.path.join(eval_dir, 'dicePC.png'),
granularity=10
)
_time['DiceCurve'] = time.time() - _time['DiceCurve']
print('Done in {} seconds'.format(_time['DiceCurve']))
if options["threshold"] == 'bestdice':
diffs_thresholded = eval_pc['diffs'] > eval_pc['bestThreshold']
else:
diffs_thresholded = eval_pc['diffs'] > options["threshold"]
diffs_thresholded_at_precision70 = diffs_thresholded
diffs_thresholded = filter_3d_connected_components(np.squeeze(diffs_thresholded))
eval_pc['thresholdType'] = options["threshold"]
eval_pc['DiceScore'] = Metrics.dice(diffs_thresholded, eval_pc['labelmaps'])
eval_pc['DiceScorePerPatient'] = []
eval_pc['PrecisionPerPatient'] = []
eval_pc['RecallPerPatient'] = []
for p, patient in enumerate(patients_pc):
subvolume_prediction = diffs_thresholded[p * num_slices:(p + 1) * num_slices, :, :]
subvolume_groundtruth = eval_pc['labelmaps'][p * num_slices:(p + 1) * num_slices, :, :]
eval_pc['DiceScorePerPatient'] += [Metrics.dice(subvolume_prediction, subvolume_groundtruth.astype(bool))]
eval_pc['PrecisionPerPatient'] += [Metrics.precision(subvolume_prediction, subvolume_groundtruth.astype(bool))]
eval_pc['RecallPerPatient'] += [Metrics.recall(subvolume_prediction, subvolume_groundtruth.astype(bool))]
# Choose a different operating point from the Precision Recall Curve!
# e.g. determine the threshold at 20% Precision and base don that, this lesion detection rate
_TPs, _FPs, _FNs = compute_detection_rate(np.squeeze(diffs_thresholded_at_precision70[p * num_slices:(p + 1) * num_slices, :, :]),
np.squeeze(subvolume_groundtruth.astype(bool)))
eval_pc['TPCC'] += _TPs
eval_pc['FPCC'] += _FPs
eval_pc['FNCC'] += _FNs
eval_pc['DiceScorePerPatientMean'] = np.mean(np.array(eval_pc['DiceScorePerPatient']))
eval_pc['DiceScorePerPatientStd'] = np.std(np.array(eval_pc['DiceScorePerPatient']))
eval_pc['PrecisionPerPatientMean'] = np.mean(np.array(eval_pc['PrecisionPerPatient']))
eval_pc['PrecisionPerPatientStd'] = np.std(np.array(eval_pc['PrecisionPerPatient']))
eval_pc['RecallPerPatientMean'] = np.mean(np.array(eval_pc['RecallPerPatient']))
eval_pc['RecallPerPatientStd'] = np.std(np.array(eval_pc['RecallPerPatient']))
# Threshold diffs and compute Confusion matrix, TPR, FPR and VolumeDifference
eval_pc['TP'], eval_pc['FP'], eval_pc['TN'], eval_pc['FN'] = Metrics.confusion_matrix(
diffs_thresholded, eval_pc['labelmaps'].astype(bool))
eval_pc['TPR'] = Metrics.tpr(diffs_thresholded, eval_pc['labelmaps'].astype(bool))
eval_pc['FPR'] = Metrics.tpr(diffs_thresholded, eval_pc['labelmaps'].astype(bool))
eval_pc['VD'] = Metrics.vd(diffs_thresholded, eval_pc['labelmaps'].astype(bool))
if eval_pc['TPCC'] + eval_pc['FNCC'] > 0:
eval_pc['TPRCC'] = eval_pc['TPCC'] / (eval_pc['TPCC'] + eval_pc['FNCC'])
else:
eval_pc['TPRCC'] = 0.0
if eval_pc['TPCC'] + eval_pc['FPCC'] > 0:
eval_pc['PrecisionCC'] = eval_pc['TPCC'] / (eval_pc['TPCC'] + eval_pc['FPCC'])
else:
eval_pc['PrecisionCC'] = 0.0
for idx in range(0, eval_pc['x'].shape[0]):
tmp = image_utils.augment_prediction_and_groundtruth_to_image(eval_pc['x'][idx],
diffs_thresholded[idx],
eval_pc['labelmaps'][idx])
p = math.floor(float(idx) / num_slices)
s = datasetPC.options.sliceStart + (idx % (datasetPC.options.sliceEnd - datasetPC.options.sliceStart))
imwrite(os.path.join(sample_dir, '{}_{}_vis.png'.format(p, s)), np.squeeze(cv2.normalize(tmp, None, 0, 255)).astype('uint8'))
# Store evalPC to disk
eval_pc.pop('x')
eval_pc.pop('diffs')
eval_pc.pop('labelmaps')
eval_pc.pop('l1reconstructionErrors')
eval_pc.pop('l2reconstructionErrors')
eval_pc.pop('reconstructions')
eval_pc.pop('diffHistogram')
np.save(os.path.join(eval_dir, 'evalPC.npy'), eval_pc)
_time['evaluation'] = time.time() - _time['evaluation']
# Export to TXT
f = open(os.path.join(eval_dir, 'evalPC.txt'), "w")
f.write(str(eval_pc))
f.close()
def determine_threshold_on_labeled_patients(dataset_pc, model, options, epoch='last', description=None):
# Create eval folder
eval_dir = os.path.join(
options['train']['samplesDir'],
model.network.__name__,
model.model_dir,
'eval-' + str(epoch) + '-' + str(utils.timestamp()).replace(":", "-")
)
if description is not None:
eval_dir += "-" + str(description)
if not os.path.exists(eval_dir):
os.makedirs(eval_dir)
sample_dir = os.path.join(eval_dir, 'samples_val_PC')
if not os.path.exists(sample_dir):
os.makedirs(sample_dir)
if not isinstance(dataset_pc, list):
dataset_pc = [dataset_pc]
eval_pc_val = None
patients_pc_val = None
for i, ds in enumerate(dataset_pc):
if i == 0:
eval_pc_val, patients_pc_val = _evaluate(ds, model, sample_dir, options, split="VAL")
else:
_eval_pc_val, _patients_pc_val = _evaluate(ds, model, sample_dir, options, split="VAL")
eval_pc_val = merge_eval_dictionaries(eval_pc_val, _eval_pc_val)
patients_pc_val += [_patients_pc_val]
print("Computing DICE curve for Lesion Validation samples")
eval_pc_val['bestDiceScore'], eval_pc_val['bestThreshold'] = Metrics.compute_dice_curve_recursive(
eval_pc_val['diffs'].flatten(),
eval_pc_val['labelmaps'].flatten(),
plottitle="DICE vs Thresholds Curve for Lesion Testing Validation Samples",
filename=os.path.join(eval_dir, 'dicePC_VAL.png'),
granularity=10
)
return eval_pc_val['bestDiceScore'], eval_pc_val['bestThreshold']