def fit_normalization(self, num_sample=None, show_progress=False):
"""
Calculate the voxel-wise mean and std across the dataset for normalization.
Args:
num_sample (int or None): If None (default), calculate the values across the complete dataset,
otherwise sample a number of images.
show_progress (bool): Show a progress bar during the calculation."
"""
if num_sample is None:
num_sample = len(self)
image_shape = self.image_shape()
all_struct_arr = np.zeros(
(num_sample, image_shape[0], image_shape[1], image_shape[2]))
sampled_filenames = np.random.choice(self.filenames,
num_sample,
replace=False)
if show_progress:
sampled_filenames = tqdm_notebook(sampled_filenames)
for i, filename in enumerate(sampled_filenames):
struct_arr = utils.load_nifti(filename, mask=mask)
all_struct_arr[i] = struct_arr
self.mean = all_struct_arr.mean(0)
self.std = all_struct_arr.std(0)
def __getitem__(self, idx):
"""Return the image as a numpy array and the label."""
label = self.labels[idx]
struct_arr = utils.load_nifti(self.filenames[idx], mask=self.mask)
#struct_arr = utils.resize_image(struct_arr, (55,55,55), 1)
# TDOO: Try normalizing each image to mean 0 and std 1 here.
#struct_arr = (struct_arr - struct_arr.mean()) / (struct_arr.std() + 1e-10)
struct_arr = (struct_arr - self.mean) / (self.std + 1e-10) # prevent 0 division by adding small factor
struct_arr = struct_arr[None] # add (empty) channel dimension
struct_arr = torch.FloatTensor(struct_arr)
if self.transform is not None:
struct_arr = self.transform(struct_arr)
return struct_arr, label
def get_raw_image(self, idx):
"""Return the raw image at index idx (i.e. not normalized, no color channel, no transform."""
return utils.load_nifti(self.filenames[idx], mask=self.mask)
def image_shape(self):
"""The shape of the MRI images."""
return utils.load_nifti(self.filenames[0], mask=mask).shape