Question about preprocessing in the official implementation

[zzzyzh]

Hi, thank you for the great work!

I noticed the following preprocessing steps in the official code:

slice_arr = img_arr[:,:,slice_indx-2: slice_indx+3]
slice_arr = np.flip(np.rot90(slice_arr, k=1, axes=(0, 1)), axis=1)

mask_arr_2D = mask_arr[:,:,slice_indx-2: slice_indx+3]
mask_arr_2D = np.flip(np.rot90(mask_arr_2D, k=1, axes=(0, 1)), axis=1)

I was wondering if there is any specific reason or motivation behind this operation. Does this transformation have a particular significance for the training process?

Also, it seems that similar preprocessing is not applied during inference/testing. Would this inconsistency affect the final accuracy or introduce any bias?

Thanks in advance for your clarification!