def rgb_to_rgb(self, mat):
''' Reference images are CT scans with highlighted OARs and PTVs. The
targets are 3D voxel maps with colourized dose intensities.
'''
dose_img = mat['dose_imgs']
ct_img = mat['ct_imgs']
d, w, h, nc = ct_img.shape
assert w == self.opt.loadSize, 'size mismatch in width'
assert h == self.opt.loadSize, 'size mismatch in height'
A = vox2tensor(ct_img).float()
B = vox2tensor(dose_img).float()
# ABs are 3-channel. Normalizing to 0.5 mean, 0.5 std
A = normalize3d(A, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
B = normalize3d(B, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
# flipping augments the dataset by flipping a bunch of the images.
if (not self.opt.no_flip) and random.random() < 0.5:
idx = [i for i in range(A.size(3) - 1, -1, -1)]
idx = torch.LongTensor(idx)
A = A.index_select(3, idx)
B = B.index_select(3, idx)
# flipped the width side. Q: is it worth it to flip on height as well?
if (not self.opt.no_flip) and random.random() < 0.5:
idx = [i for i in range(A.size(2) - 1, -1, -1)]
idx = torch.LongTensor(idx)
A = A.index_select(2, idx)
B = B.index_select(2, idx)
return A, B
def rgb_to_gray(self, mat):
''' Reference images are CT scans with highlighted OARs and PTVs. The
targets are dose intensity matrices.
'''
dose_val = mat['dose_imgs']
# dose_val = mat['dose_vals']
ct_img = mat['ct_imgs']
d, w, h, nc = ct_img.shape
assert (d, w, h) == dose_val.shape, 'size mismatch between dose and ct'
# I have to add some noise to ct_val because we are just training on
# a single patient and ct_val is the same all the time. I divide by
# 1000 to give stdev~=0.001, which is sufficiently small.
rnd_noise = randn(*ct_img.shape) / 500
ct_img = ct_img + rnd_noise
ct_img[ct_img <= 0] = 0.0
A = vox2tensor(ct_img).float()
A = normalize3d(A, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
B = torch.from_numpy(dose_val).float()
B = B.unsqueeze(0)
B.sub_(0.5).div_(0.5)
# flipping augments the dataset by flipping a bunch of the images.
if (not self.opt.no_flip) and random.random() < 0.5:
idx = [i for i in range(A.size(2) - 1, -1, -1)]
idx = torch.LongTensor(idx)
A = A.index_select(2, idx)
B = B.index_select(2, idx)
return A, B
def rgb_to_gray(self, mat):
''' Reference images are CT scans with highlighted OARs and PTVs. The
targets are dose intensity matrices.
'''
dose_val = mat['dose_vals']
ct_img = mat['ct_imgs']
d, w, h, nc = ct_img.shape
assert (d, w, h) == dose_val.shape, 'size mismatch between dose and ct'
# to handle aaron's weird uint format
if dose_val.dtype == np.uint8 or dose_val.dtype == np.uint16:
dose_val = dose_val / 256
if ct_img.dtype == np.uint8 or ct_img.dtype == np.uint16:
ct_img = ct_img / 256
A = vox2tensor(ct_img).float()
A = normalize3d(A, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
B = torch.from_numpy(dose_val).float()
B = B.unsqueeze(0)
B.sub_(0.5).div_(0.5)
# flipping augments the dataset by flipping a bunch of the images.
if (not self.opt.no_flip) and random.random() < 0.5:
idx = [i for i in range(A.size(2) - 1, -1, -1)]
idx = torch.LongTensor(idx)
A = A.index_select(2, idx)
B = B.index_select(2, idx)
return A, B