BarY7
commented
6 months ago A bit late but for anyone else struggling with the data preprocessing, there is a notebook in the commit history that gives some insight
As for your question the authors take the previous and next slice as input to the network, as shown in this snippet from the notebook:
for img_name in train_img_list:
label_name = 'label'+img_name[3:]
img = sitk.ReadImage(os.path.join(img_dir,img_name))
img = sitk.GetArrayFromImage(img)
label = sitk.ReadImage(os.path.join(label_dir,label_name))
label = sitk.GetArrayFromImage(label)
# mean,std = np.mean(img),np.std(img)
# img = z_score(img,-125,275)
# min_v,max_v = img.min(),img.max()
# img = (img - min_v) / (max_v-min_v)
# img = min_max_normalization(img,-1.0582,4.3029)
img = min_max_normalization(img,-125,275)
img = np.flip(img.transpose((1,2,0)),axis=0)
label = np.flip(label.transpose((1,2,0)),axis=0)
for index in range(img.shape[2]):
if index==0:
img_ = img[:,:,[0,0,1]]
elif index==img.shape[2]-1:
img_ = img[:,:,[index-1,index,index]]
else:
img_ = img[:,:,index-1:index+2]
seg = label[:,:,index]
assert img_.shape[2]==3
np.savez(os.path.join(site_dir,'train','{}_{}.npz'.format(img_name[3:7],index)),image = img_,label = seg)
When training the source domain model, I see that your config file says input_dim=3, and then the U-Net output is [batch_size，n_classes, H, W], does it mean that you input 3 slices to the network, and then only use 1 slice label to calculate the loss? Or I would like to ask what your data looks like, is the image of each .npz file 3 slices, and then the label is 1 slice? Thanks for your reply！