def __call__(self, voxel, seg):
shape = voxel.shape
voxel = voxel / 255. - 1
if self.train:
if self.move is not None:
center = random_center(shape, self.move)
else:
center = np.array(shape) // 2
voxel_ret = crop(voxel, center, self.size)
seg_ret = crop(seg, center, self.size)
angle = np.random.randint(4, size=3)
voxel_ret = rotation(voxel_ret, angle=angle)
seg_ret = rotation(seg_ret, angle=angle)
axis = np.random.randint(4) - 1
voxel_ret = reflection(voxel_ret, axis=axis)
seg_ret = reflection(seg_ret, axis=axis)
else:
center = np.array(shape) // 2
voxel_ret = crop(voxel, center, self.size)
seg_ret = crop(seg, center, self.size)
if self.copy_channels:
return np.stack([voxel_ret,voxel_ret,voxel_ret],0).astype(np.float32), \
np.expand_dims(seg_ret,0).astype(np.float32)
else:
return np.expand_dims(voxel_ret, 0).astype(np.float32), \
np.expand_dims(seg_ret,0).astype(np.float32)
def __call__(self, voxel, name):
shape = voxel.shape[1:] # 3 dims
y = np.load(
os.path.join(self.data_path, 'roi/roi/labels',
name + '.npz'))['arr']
loc = name.split('_')[-1]
if loc == 'L':
lb = 1 if 3 in y else 0
elif loc == 'R':
lb = 1 if 4 in y else 0
else:
print('error')
if self.move is not None:
center = random_center(shape, self.move)
else:
center = np.array(shape) // 2
voxel_ret = crop(voxel, center, self.size)
angle = np.random.randint(4, size=3)
voxel_ret = np.stack([
rotation(voxel_ret[0], angle=angle),
rotation(voxel_ret[1], angle=angle),
rotation(voxel_ret[2], angle=angle)
], 0)
axis = np.random.randint(4) - 1
voxel_ret = np.stack([
reflection(voxel_ret[0], axis=axis),
reflection(voxel_ret[1], axis=axis),
reflection(voxel_ret[2], axis=axis)
], 0)
return voxel_ret, lb
def __call__(self, ct):
voxel = ct['image']
voxel = np.expand_dims(voxel, 0).repeat(3, axis=0)
y = ct['coarse_mask']
shape = voxel.shape[1:] # 120*120*120
center = random_center_mask_mood(y)
voxel_ret, crop_pos = crop_at_zyx_with_dhw(voxel, center, self.size,
self.fill_with)
angle = np.random.randint(4, size=3)
voxel_ret = np.stack([
rotation(voxel_ret[0], angle=angle),
rotation(voxel_ret[1], angle=angle),
rotation(voxel_ret[2], angle=angle)
], 0)
axis = np.random.randint(4) - 1
voxel_ret = np.stack([
reflection(voxel_ret[0], axis=axis),
reflection(voxel_ret[1], axis=axis),
reflection(voxel_ret[2], axis=axis)
], 0)
return voxel_ret
def __call__(self, voxel, seg):
shape = voxel.shape
voxel = voxel / 255.0 - 1
if self.train:
if self.move is not None:
center = random_center(shape, self.move)
else:
center = np.array(shape) // 2
voxel_ret = crop(voxel, center, self.size)
seg_ret = crop(seg, center, self.size)
angle = np.random.randint(4, size=3)
voxel_ret = rotation(voxel_ret, angle=angle)
seg_ret = rotation(seg_ret, angle=angle)
axis = np.random.randint(4) - 1
voxel_ret = reflection(voxel_ret, axis=axis)
seg_ret = reflection(seg_ret, axis=axis)
else:
center = np.array(shape) // 2
voxel_ret = crop(voxel, center, self.size)
seg_ret = crop(seg, center, self.size)
# as resnet expects 3 channel input, then batch dimension is done in DataLoader
if self.copy_channels:
return np.stack([voxel_ret, voxel_ret, voxel_ret], 0).astype(
np.float32), np.expand_dims(seg_ret, 0).astype(np.float32)
else:
return np.expand_dims(voxel_ret,
0).astype(np.float32), np.expand_dims(
seg_ret, 0).astype(np.float32)