The transforms.SaveImaged saved nii segmentation mask have different origin and direction with original data

[auroua]

As shown in the following image, the transforms.SaveImaged saved nii segmentation mask have different origin and direction with original data.

I also used SimpleITK to read the origin and direction of the saved segmetntation mask, the values are different with the original nii data. How to keep the transformer.SaveImaged saved data have the same origin and direction with the original nii data.
Code for reading the origin and direction of nii data.

    original_itk_data = sitk.ReadImage(test_data_path)
    original_label_data = sitk.ReadImage(test_label_path)

    original_spacing = original_itk_data.GetSpacing()
    original_origin = original_itk_data.GetOrigin()
    original_direction = original_itk_data.GetDirection()
    print(original_spacing, original_origin, original_direction)

    pred_spacing = original_label_data.GetSpacing()
    pred_origin = original_label_data.GetOrigin()
    pred_direction = original_label_data.GetDirection()
    print(pred_spacing, pred_origin, pred_direction)

The output is :

(0.68359375, 0.68359375, 3.0) (0.0, 0.0, 0.0) (1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0)
(0.68359375, 0.68359375, 3.0) (-174.658203125, 174.658203125, -205.5) (1.0, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0, 1.0)

[Nic-Ma]

Hi @auroua ,

Thanks for your feedback.
Could you please help share your full program or at least transform chain for further analysis?
I think you can refer to below example to invert your transforms to align with original image:
https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/torch/unet_inference_dict.py#L70

@wyli I think maybe we can make a category / milestone to track all the metadata-related issues and discussions?
I feel we already got much feedback about metadata, then we can ask the users to help review when the metadata feature is ready?

Thanks in advance.

[auroua]

Hi @Nic-Ma, the following is the code used to generate the output masks.

    def visualize_3d(self, cfgs):
        datalist_json = os.path.join(cfgs.DATASETS.DATA_DIR, cfgs.DATASETS.JSON)

        test_org_transforms = transforms.Compose(
            [
                transforms.LoadImaged(keys="image"),
                transforms.EnsureChannelFirstd(keys="image"),
                transforms.Orientationd(keys=["image"],
                                        axcodes=cfgs.INPUT.ORIENTATION),
                transforms.Spacingd(keys=["image"],
                                    pixdim=(cfgs.INPUT.SPACING.X, cfgs.INPUT.SPACING.Y, cfgs.INPUT.SPACING.Z),
                                    mode="bilinear"),
                transforms.ScaleIntensityRanged(keys=["image"],
                                                a_min=cfgs.INPUT.SCALE_INTENSITY.ORIGINAL_MIN,
                                                a_max=cfgs.INPUT.SCALE_INTENSITY.ORIGINAL_MAX,
                                                b_min=cfgs.INPUT.SCALE_INTENSITY.TARGET_MIN,
                                                b_max=cfgs.INPUT.SCALE_INTENSITY.TARGET_MAX,
                                                clip=True),
                transforms.CropForegroundd(keys=["image"], source_key="image"),
                transforms.EnsureTyped(keys="image"),
            ]
        )
        test_files = load_decathlon_datalist(datalist_json,
                                             True,
                                             cfgs.DATASETS.TEST_TYPE,
                                             base_dir=cfgs.DATASETS.DATA_DIR)
        test_ds = data.Dataset(data=test_files, transform=test_org_transforms)
        test_loader = data.DataLoader(test_ds,
                                      batch_size=1,
                                      shuffle=False,
                                      num_workers=cfgs.DATALOADER.TEST_WORKERS,
                                      sampler=None,
                                      pin_memory=True,
                                      persistent_workers=True)

        post_transforms = transforms.Compose([
            transforms.EnsureTyped(keys="pred"),
            transforms.Invertd(
                keys="pred",
                transform=test_org_transforms,
                orig_keys="image",
                meta_keys="pred_meta_dict",
                orig_meta_keys="image_meta_dict",
                meta_key_postfix="meta_dict",
                nearest_interp=False,
                to_tensor=True,
            ),
            transforms.AsDiscreted(keys="pred", argmax=True, to_onehot=5),
            transforms.SplitChanneld(keys="pred", channel_dim=0),
            transforms.SaveImaged(keys="pred_0",
                                  meta_keys="pred_meta_dict",
                                  output_dir=os.path.join(cfgs.OUTPUT_DIR, "inference_all_part1"),
                                  output_postfix="seg_0",
                                  resample=False),
            transforms.SaveImaged(keys="pred_1",
                                  meta_keys="pred_meta_dict",
                                  output_dir=os.path.join(cfgs.OUTPUT_DIR, "inference_all_part1"),
                                  output_postfix="seg_1",
                                  resample=False),
            transforms.SaveImaged(keys="pred_2",
                                  meta_keys="pred_meta_dict",
                                  output_dir=os.path.join(cfgs.OUTPUT_DIR, "inference_all_part1"),
                                  output_postfix="seg_2",
                                  resample=False),
            transforms.SaveImaged(keys="pred_3",
                                  meta_keys="pred_meta_dict",
                                  output_dir=os.path.join(cfgs.OUTPUT_DIR, "inference_all_part1"),
                                  output_postfix="seg_3",
                                  resample=False),
            transforms.SaveImaged(keys="pred_4",
                                  meta_keys="pred_meta_dict",
                                  output_dir=os.path.join(cfgs.OUTPUT_DIR, "inference_all_part1"),
                                  output_postfix="seg_4",
                                  resample=False)
        ])

        loader = LoadImage()

        with torch.no_grad():
            for test_data in test_loader:
                test_data["image"] = test_data["image"].cuda()
                img_name = test_data['image_meta_dict']['filename_or_obj'][0].split('/')[-1]
                self.logger.info("Inference on case {}".format(img_name))
                test_data["pred"] = sliding_window_inference(test_data["image"],
                                                             roi_size=(self.cfg.INPUT.RAND_CROP.ROI.X,
                                                                       self.cfg.INPUT.RAND_CROP.ROI.Y,
                                                                       self.cfg.INPUT.RAND_CROP.ROI.Z),
                                                             sw_batch_size=self.cfg.SOLVER.SW_BATCH_SIZE_TEST,
                                                             predictor=self.model,
                                                             overlap=self.cfg.SOLVER.INFER_OVERLAP)
                test_data = [post_transforms(i) for i in decollate_batch(test_data)]

[Nic-Ma]

Hi @wyli ,

As we just totally updated the image writing logic recently, maybe there is some compatible issue with inverting?

Thanks.

[wyli]

I didn't look into the details but it looks like the same issue as Project-MONAI/tutorials#481. I put a patch here #3882. I think we should merge #3882 for now and change it if we find a better default.

[adwaykanhere]

Hi, I'm still facing the same issue that @auroua mentioned. Please suggest how to fix it. I'm using the following script:

post_transforms_fl_liver = Compose([
    Invertd(
        keys="pred",
        transform=test_org_transforms,
        orig_keys="image",
        meta_keys="pred_meta_dict",
        orig_meta_keys="image_meta_dict",
        meta_key_postfix="meta_dict",
        nearest_interp=False,
        to_tensor=True,
        device="cpu",
    ),
    AsDiscreted(keys="pred", argmax = True),
    Orientationd(keys=["pred"], axcodes="RPI"),
    EnsureChannelFirstd(keys="image"),
    Spacingd(keys=["pred"], pixdim=(
            1.5, 1.5, 2.0), mode="bilinear"),    
    SaveImaged(keys="pred", meta_keys="image_meta_dict", output_dir="/content/s_all_outputs", output_postfix="f1_liver", resample=True),
])

model_liver.load_state_dict(torch.load(
    os.path.join(root_dir, "best_metric_model_liver_128.pth")))
model_liver.eval()

with torch.no_grad():
    for test_data in test_org_loader:
        test_inputs = test_data["image"].to(device)
        roi_size = (160, 160, 160)
        sw_batch_size = 4
        test_data["pred"] = sliding_window_inference(
            test_inputs, roi_size, sw_batch_size, model_liver)

        test_data = [post_transforms_fl_liver(i) for i in decollate_batch(test_data)]

[KumoLiu]

Hi @adwaykanhere, I noticed that you added some transforms after the invertd, if test_org_transforms have included these, it should be unnecessary to add, you can refer to this tutorial to modify your post_transforms_fl_liver . Also if your monai version is not the latest, maybe you can try to update to 1.1.0 and test again.
https://github.com/Project-MONAI/tutorials/blob/7403995f12f4198042eb1a52f3d2e1c88d53ae71/3d_segmentation/torch/unet_inference_dict.py#L65-L79

Hope it can help you, thanks!

[adwaykanhere]

Hi @KumoLiu I tried your suggestions and I'm still getting the same error.

[KumoLiu]

Hi @adwaykanhere, could you please share more details then I could reproduce the issue? Thanks!