weird inference results in region-based training

[JunMa11]

Hey Nick and Fabian,

Hope you are doing well. Very happy to see you guys again in KiTS:)

I have a question about region-based training. I used nnUNetv2 to train a baseline on the kidney ROI with the following dataset.json

{
    "channel_names": {
        "0": "CT"
    },
    "labels": {
        "background": 0,
        "kidney": [
            1,
            2,
            3
        ],
        "masses": [
            2,
            3
        ],
        "tumor": 2
    },
    "numTraining": 964,
    "file_ending": ".nii.gz",
    "name": "Dataset120_KiTS23ROI",
    "description": "",
    "regions_class_order": [
        1,
        3,
        2
    ]
}

The data preprocessing and training go well but the inference results seem to have some interpolation errors.

Here is a segmentation example:

The data and trained model can be downloaded here https://drive.google.com/drive/folders/1VvPNUIWElkZYI6L7LapXVoIisC-1Khty?usp=sharing, https://drive.google.com/file/d/1ZtdkKft5OvYUdUc6e_8dQeAqa593SfGB/view?usp=sharing

Any comments are highly appreciated. Looking forward to seeing you guys in MICCAI 2023:)

[neheller]

Hi Jun! I hope you're doing well too :-)

Thanks for pointing this out! @FabianIsensee is certainly better qualified to answer this than I am. Any ideas, Fabian?

[neheller]

Just noticing that this issue is still open. @FabianIsensee did you ever have a chance to look at this?