def preprocess_volume_3d(volume,
crop_size,
patches_per_side,
patch_overlap,
is_training=True):
result = []
w, _, _, _ = volume.shape
if is_training:
volume = crop_3d(volume, is_training,
(crop_size, crop_size, crop_size))
volume = pad_to_final_size_3d(volume, w)
for patch in crop_patches_3d(volume, is_training, patches_per_side,
-patch_overlap):
if is_training:
normal_patch_size = patch.shape[0]
patch_crop_size = int(normal_patch_size * (7.0 / 8.0))
do_flip = np.random.choice([False, True])
if do_flip:
patch = np.flip(patch, 0)
patch = crop_3d(
patch, is_training,
(patch_crop_size, patch_crop_size, patch_crop_size))
patch = pad_to_final_size_3d(patch, normal_patch_size)
else:
pass # lets give it the most information we can get
result.append(patch)
return np.asarray(result)
def preprocess_image(image, is_training, patches_per_side, patch_jitter,
permutations, mode3d):
label = random.randint(0, len(permutations) - 1)
if mode3d:
patches = crop_patches_3d(image, is_training, patches_per_side,
patch_jitter)
else:
patches = crop_patches(image, is_training, patches_per_side,
patch_jitter)
b = np.zeros((len(permutations), ))
b[label] = 1
return np.array(patches)[np.array(permutations[label])], np.array(b)
def preprocess_image_3d(image, patches_per_side, patch_jitter, is_training):
cropped_image = crop_patches_3d(image, is_training, patches_per_side,
patch_jitter)
return np.array(cropped_image)
def preprocess_image_crop_only(image, patches_per_side, is_training, mode3d):
if mode3d:
patches = crop_patches_3d(image, is_training, patches_per_side, 0)
else:
patches = crop_patches(image, is_training, patches_per_side, 0)
return np.stack(patches)