-
Respacing the source and target domain images (both should be gray) with the same XY plane resolutions, and crop/pad to the size of [512, 512, d] in terms of [width, height, depth].
-
Normlizae each 3D images to [0, 1], and extract 2D slices from 3D volumes along depth-axis.
-
Stack the list of 2D slices at zero dimension for the two domains respectively, resulting in 3D tensor with size of [N, 512, 512], and then save them as the follows:
.
└── DARUNET
└──datasets
├── A_imgs.npy
└── B_imgs.npy-
Respacing the source and target domain images and labels (both should be gray) with the same XY plane resolutions, and make sure the spacing ratio is about [1, 1, 4] and the size is [512, 512, d] in terms of [width, height, depth].
-
Normlizae each 3D images to [0, 1], rename the paired image and label to ensure the only name difference is 'img' and 'label', and save them as follows
.
├── TransUNet
└──datasets
├── source_training_npy
│ ├──2_img.npy
│ ├──2_label.npy
│ ├──3_img.npy
│ └──3_label.npy
└── target_training_npy
├──5_img.npy
├──5_label.npy
├──8_img.npy
└──8_label.npy- run stage_1.5_i2i_inference.py to generate the transferred source domain images