def __load(self, index):
if self.patches_from_current_image > self.patches_from_single_image:
self.patches_from_current_image = 0
self.current_image_index = index
p, f = self.series[index]
self.image, self.label, affine = loader_helper.read_multimodal(p, f, self.annotation_path, True)
mask = self.image > 0
num_voxels = np.sum(mask,axis=(1,2,3))
#print(self.image[mask].shape)
mean = np.sum(self.image / num_voxels[:,None,None,None], axis=(1,2,3))
mean2 = np.sum(np.square(self.image) / num_voxels[:,None,None,None], axis=(1,2,3))
std = np.sqrt(mean2 - mean * mean)
#std1 = self.image.std(axis=(1,2,3))
self.image = (self.image - mean.reshape((self.image.shape[0],1,1,1))) / std.reshape((self.image.shape[0],1,1,1))#self.image.std(axis=(1,2,3), keepdims=True)#(self.image - self.image.mean(axis=(1,2,3), keepdims=True)) / self.image.std(axis=(1,2,3), keepdims=True)
#self.image[~mask] = -10
self.patches_from_current_image += 1
def __getitem__(self, index):
image, label, affine = loader_helper.read_multimodal(data_path=self.path, series=self.series[index], read_annotation=True)
#image = image - image.mean()
#image = image / image.var() ** 0.5
#image = image / 1000.
old_shape = image.shape
new_shape = tuple([loader_helper.closest_to_k(i,16) for i in old_shape[1:]])
new_image = np.full(shape=(old_shape[0],)+new_shape, fill_value=0., dtype=np.float32)
new_label = np.zeros(shape=new_shape, dtype=np.float32)
new_image[:,:old_shape[1],:old_shape[2],:old_shape[3]] = image
new_label[:old_shape[1],:old_shape[2],:old_shape[3]] = label
mask = new_image > 0
num_voxels = np.sum(mask, axis=(1, 2, 3))
mean = np.sum(new_image / num_voxels[:,None,None,None], axis=(1, 2, 3))
mean2 = np.sum(np.square(new_image)/ num_voxels[:,None,None,None], axis=(1, 2, 3))
std = np.sqrt(mean2 - mean * mean)
new_image = (new_image - mean.reshape((new_image.shape[0],1,1,1)))/ std.reshape((new_image.shape[0], 1, 1, 1))
#new_image[~mask] = -10
#new_image = new_image / new_image.std(axis=(1, 2, 3),keepdims=True) #(new_image - new_image.mean(axis=(1, 2, 3), keepdims=True)) / new_image.std(axis=(1, 2, 3),keepdims=True)
new_label_out = (np.eye(4)[new_label.astype(np.int32)]).transpose((3,0,1,2))
#new_label_out = (new_label > 0)[None]
wt = np.sum(new_label_out [1:], axis=0, keepdims=True)
tc = np.sum(new_label_out [[1,3]], axis=0, keepdims=True)
et = new_label_out[3,None]
new_label_out = np.concatenate([wt,tc,et],axis=0)
labels_torch = torch.from_numpy(new_label_out.copy()).float()
return [torch.from_numpy(new_image).float(),
], \
[ labels_torch,
]
def __cache(self):
# cache locations of the labels (bounding boxes) inside the images
print('Cache files')
for p, f in tqdm.tqdm(self.series):
image, label, affine = loader_helper.read_multimodal(p, f, self.annotation_path, True)
bbox = loader_helper.bbox3(label>0)
borders = np.array(label.shape)
borders_low = np.array(self.patch_size) / 2.0 + 1
borders_high = borders - np.array(self.patch_size) / 2.0 - 1
bbox[0] = np.maximum(bbox[0]-50, borders_low)
bbox[1] = np.minimum(bbox[1]+50, borders_high)
self.labels_location.append(bbox)
path_input = opt.data_path
path_output = opt.predictions_path
series = [f for f in os.listdir(path_input) if os.path.isdir(os.path.join(path_input, f))]
series.sort()
#series = series_val19
dice = metrics.Dice(input_index=0)
dicewt = metrics.DiceWT(input_index=0)
sum = 0
for f in series:
image, label, affine = loader_helper.read_multimodal(path_input, f, True)
predict = loader_helper.read_nii(os.path.join(path_output,f+'.nii.gz')).astype(np.uint8)
predict[predict==4] = 3
result = np.zeros(shape=(4))
for i in range(1, 4):
p = (predict== i).astype(np.float32)
g = (label == i).astype(np.float32)
numerator = (p * g).sum()
denominator = (p + g).sum()
r = 2 * numerator / denominator
models_root=opt.models_path,
rewrite=False,
connect_tb=False)
trainer.load_best()
trainer.state.cuda = True
series = [
f for f in os.listdir(path) if os.path.isdir(os.path.join(path, f))
]
series.sort()
print(series)
for f in series:
image, label, affine = loader_helper.read_multimodal(
data_path=path, series=f, read_annotation=False)
bbox = get_bbox(image)
image_crop = image[:, bbox[0, 0]:bbox[1, 0], bbox[0, 1]:bbox[1, 1],
bbox[0, 2]:bbox[1, 2]]
#data_crop = median_filter(data_crop, 3)
#=========================================
old_shape_crop = image_crop.shape[1:]
new_shape_crop = tuple(
[loader_helper.closest_to_k(i, 16) for i in old_shape_crop])
diff = np.array(new_shape_crop) - np.array(old_shape_crop)
pad_left = diff // 2
pad_right = diff - pad_left