How to get segmentation result with Test.py about img0060 ~ img0080 on BTCV

[whduddhks]

In utils/data_utils.py

test_transform need label

research-contributions/SwinUNETR/BTCV/utils/data_utils.py

Line 119 in 4af80e1

test_transform = transforms.Compose(

But there is no label for test dataset.

And in test.py

research-contributions/SwinUNETR/BTCV/test.py

Line 91 in 4af80e1

val_inputs, val_labels = (batch["image"].cuda(), batch["label"].cuda())

need label same with data_utils.

So how to get segmentation result for test dataset?

Thank you,
whduddhks

[Mentholatum]

Same question, have you solved it now?

[whduddhks]

Maybe i solved it.

So, this is the utils.data_utils.py

test_transform = transforms.Compose(
    [
        transforms.LoadImaged(keys=["image"]),
        transforms.AddChanneld(keys=["image"]),
        # transforms.Orientationd(keys=["image"], axcodes="RAS"),
        transforms.Spacingd(keys="image", pixdim=(args.space_x, args.space_y, args.space_z), mode="bilinear"),
        transforms.ScaleIntensityRanged(
            keys=["image"], a_min=args.a_min, a_max=args.a_max, b_min=args.b_min, b_max=args.b_max, clip=True
        ),
        transforms.ToTensord(keys=["image"]),
    ]
)

And in line 133
test_files = load_decathlon_datalist(datalist_json, True, "test", base_dir=data_dir)

For Test.py

for img_path in raw:
      nii_img = nib.load(img_path)
      
      shape.append(nii_img.get_fdata().shape)
      affine.append(nii_img.affine)

with torch.no_grad():
      for i, batch in enumerate(val_loader):
            val_inputs = batch["image"].cuda()
            # original_affine = batch["label_meta_dict"]["affine"][0].numpy()
            
            h, w, d = shape[i]
            target_shape = (h, w, d)
            img_name = batch["image_meta_dict"]["filename_or_obj"][0].split("/")[-1]
            print("Inference on case {}".format(img_name))
            
            val_outputs = sliding_window_inference(
                val_inputs, (args.roi_x, args.roi_y, args.roi_z), 4, model, overlap=args.infer_overlap, mode="gaussian"
            )
            val_outputs = torch.softmax(val_outputs, 1).cpu().numpy()
            val_outputs = np.argmax(val_outputs, axis=1).astype(np.uint8)[0]
            val_outputs = resample_3d(val_outputs, target_shape)
            
            nib.save(
                nib.Nifti1Image(val_outputs.astype(np.uint8), affine[i]), os.path.join(output_directory, img_name)
            )

I got 61 ~ 80 segmentation result by this.

Thank you

[Dasha484]

Can I take a look at the complete content of the 'Test.py' file? I don't have any definitions related to 'raw,' so it will result in an error.

[weng-hb]

Just only can run.Do not sure must be correct
In utils.data_utils.py change:

    test_transform = transforms.Compose(
        [
            transforms.LoadImaged(keys=["image"]),
            transforms.AddChanneld(keys=["image"]),
            # transforms.Orientationd(keys=["image"], axcodes="RAS"),
            transforms.Spacingd(keys="image", pixdim=(args.space_x, args.space_y, args.space_z), mode="bilinear"),
            transforms.ScaleIntensityRanged(
                keys=["image"], a_min=args.a_min, a_max=args.a_max, b_min=args.b_min, b_max=args.b_max, clip=True
            ),
            transforms.ToTensord(keys=["image"]),
        ]
    )

in line 133 change: test_files = load_decathlon_datalist(datalist_json, True, "test", base_dir=data_dir)
In test.py:
add a new function:

import scipy.ndimage
def resample_3d(image, target_shape, mode="nearest"):
    zoom_factors = [t / s for t, s in zip(target_shape, image.shape)]
    return scipy.ndimage.zoom(image, zoom_factors, order=0 if mode == "nearest" else 1)

and then change the original with torch.no_grad():
change into:

    with torch.no_grad():
        for i, batch in enumerate(val_loader):
            test_inputs = batch["image"].cuda()
            original_affine = batch["image_meta_dict"]["affine"][0].numpy()
            img_name = batch["image_meta_dict"]["filename_or_obj"][0].split("/")[-1]
            target_shape = batch["image_meta_dict"]["spatial_shape"][0].tolist()
            print(f"Inference on case {img_name}")
            test_outputs = sliding_window_inference(
                test_inputs, (args.roi_x, args.roi_y, args.roi_z), 4, model, overlap=args.infer_overlap, mode="gaussian"
            )
            test_outputs = torch.softmax(test_outputs, 1).cpu().numpy()
            test_outputs = np.argmax(test_outputs, axis=1).astype(np.uint8)[0]
            test_outputs = resample_3d(test_outputs, target_shape, mode="nearest")
            nib.save(
                nib.Nifti1Image(test_outputs.astype(np.uint8), original_affine),
                os.path.join(output_directory, img_name)
            )
            print(f"Saved inference result for {img_name}")