SMIT CT Lung GTV segmentation model

[locastre]

Submitting the SMIT CT Lung GTV segmentation model for mHub

[LennyN95]

Thank you for your contribution 🚀

• Can we avoid using conda and instead install all requirements with uv? Our base image comes with uv installed and a virtual environment set-up. However, it is suggested that you create your own virtual environment with uv, e.g., uv venv -p 3.10 .venv310
• The contents of the meta.json are used to populate the model card on our website under mhub.ai/models. The more information, the better.
• To test-build the model and to move forward with our test routine, an mhub.toml file needs to be created.

[locastre]

@LennyN95 Thank you for the in-depth feedback! We're working to address the suggestions.

I have an additional question: when we upload our test data to Zenodo (for the mhub.toml), should we refer to the mHub DOI# 13785615 or create a new DOI from an independent/new entry?

[fedorov]

I would recommend renaming the PR to "SMIT CT Lung GTV segmentation model" (or something similar).

[LennyN95]

You're more than welcome. Thank you and your team for the great work.

I have an additional question: when we upload our test data to Zenodo (for the mhub.toml), should we refer to the mHub DOI# 13785615 or create a new DOI from an independent/new entry?

@locastre We mainly use Zenodo for it's reproducible storage mechanism, so you can create a new DOI for your sample & reference.

[LennyN95]

@locastre FYI; There are also some errors in the compliance check that need to be resolved:

Errors:

• Model folder models/msk_smit_lung_gtv does not contain a mhub.toml
  https://github.com/MHubAI/documentation/blob/main/documentation/mhub_models/model_folder_structure.md
  https://github.com/MHubAI/documentation/blob/main/documentation/mhub_contribution/testing_phase.md#test-procedure
• Model meta json is not compliant with the schema: 'non-contrast' is a required property
  https://github.com/MHubAI/documentation/blob/main/documentation/mhub_models/model_json.md
• WORKDIR must not be set to any other than /app as defined in our base image. WORKDIR ${WORK_DIR}/msk_smit_lung_gtv/src
  https://github.com/MHubAI/documentation/blob/main/documentation/mhub_models/the_mhub_dockerfile.md

[locastre]

@LennyN95 Hopefully we're ready for your review; we've addressed your feedback and tested the model within the mHub framework. Thanks for your help with this process.

[LennyN95]

Test Results

id: 7e53fc04-004d-4f0e-83cb-803fc3bd4b8f
name: MHub Test Report (default)
date: '2025-04-22 12:06:20'
missing_files:
- seg.dcm
summary:
  files_missing: 1
  files_extra: 0
  checks: {}
conclusion: false

[LennyN95]

Thank you Jue for the great work!

Unfortunately, the test failed due to what looks like a GPU incompatibility (see below). We use NVIDIA RTX A6000 GPU for testing and so far all models worked fine - can you have a look? Ideally we can find a solution that applies to a wide range of customer hardware.

/app/.venv39/lib/python3.9/site-packages/torch/functional.py:478: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at  ../aten/src/ATen/native/TensorShape.cpp:2895.)
return _VF.meshgrid(tensors, **kwargs)  # type: ignore[attr-defined]
/app/.venv39/lib/python3.9/site-packages/torch/cuda/__init__.py:146: UserWarning:
NVIDIA RTX A6000 with CUDA capability sm_86 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_70.
If you want to use the NVIDIA RTX A6000 GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

warnings.warn(incompatible_device_warn.format(device_name, capability, " ".join(arch_list), device_name))
val data size is  1
info: started to load weight:  trained_weights/model_Mhub.pt
info: model emb feature is :  96
info: Successfully loaded trained weights:  trained_weights/model_Mhub.pt
Traceback (most recent call last):
File "/app/models/msk_smit_lung_gtv/src/run_segmentation.py", line 252, in <module>
main()
File "/app/models/msk_smit_lung_gtv/src/run_segmentation.py", line 218, in main
val_data["pred"] = sliding_window_inference(val_inputs,
File "/app/models/msk_smit_lung_gtv/src/edit_inference_utils.py", line 120, in sliding_window_inference
importance_map = compute_importance_map(
File "/app/.venv39/lib/python3.9/site-packages/monai/data/utils.py", line 763, in compute_importance_map
importance_map = torch.ones(patch_size, device=device).float()
RuntimeError: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Done in 19.371 seconds.

[locastre]

@LennyN95 Hi Lenny, I added an explicit install command for torch==1.12.1+cu116 which should properly support sm_86. Let us know if this works. Thanks!

[LennyN95]

Test Results

id: aa683d35-ae95-4dd2-bdb9-4c7bc81cfe9c
name: MHub Test Report (default)
date: '2025-04-23 09:37:41'
missing_files:
- seg.dcm
extra_files:
- 1.3.6.1.4.1.14519.5.2.1.1.11635178980898764572976586249071182079/smit.seg.dcm
summary:
  files_missing: 1
  files_extra: 1
  checks: {}
conclusion: false

[LennyN95]

Thank you @locastre. The model now runs, we're almost there! The test is still failing, because the reference data doesn't match with the model output (see report). Please make sure, that the reference data looks like:

data/output_data/
`-- 1.3.6.1.4.1.14519.5.2.1.1.11635178980898764572976586249071182079
    `-- smit.seg.dcm

You can follow the steps here to prepare the output for the test procedure.

To increase the usability, could you change the filename from smit.seg.dcm to <model_name>.seg.dcm? We do not enforce this convention but it would make downstream automation a lot more convenient. Alternatively, rename the model and the folder consistently to smit.

[locastre]

@LennyN95 thanks for the catch. I've updated the data in the test zip and renamed the output SEG to msk_smit_lung_gtv.seg.dcm in default.yml, as you suggested.

[LennyN95]

Test Results

id: d5b328ee-037e-4909-9827-65e6e1c31a67
name: MHub Test Report (default)
date: '2025-04-23 19:08:01'
checked_files:
- file: msk_smit_lung_gtv.seg.dcm
  path: /app/data/output_data/1.3.6.1.4.1.14519.5.2.1.1.11635178980898764572976586249071182079/msk_smit_lung_gtv.seg.dcm
  checks:
  - checker: DicomsegContentCheck
    notes:
    - label: Segment Count
      description: The number of segments identified in the inspected dicomseg file.
      info: 1
summary:
  files_missing: 0
  files_extra: 0
  checks:
    DicomsegContentCheck:
      files: 1
conclusion: true

[LennyN95]

@locastre Amazing, we now passed all tests - yay! Thank you for the great work!
Do you have 1-5 example images you would like to have on the website? You can post them here :)

[locastre]

@LennyN95 Jue provided this screencap of representative segmentation. Will this do?

[LennyN95]

@locastre, perfect! I have one last request that I initially overlooked. Could you add the sample annotation to the default.yaml workflow? This addition will help our users understand how the input and output data are organized. It also gives us a chance to explain the role of each file and folder, which is quite self-explanatory in the case of DICOM. This is a fairly new feature I added, and we are gradually rolling it out for all legacy models as well. You can look at this model for more examples.

[locastre]

@LennyN95 sample annotation added to end of default.yml

[LennyN95]

@locastre The model is now online. Amazing work!! FYI; I shortened the title to SMIT Self-supervised Lung GTV Segmentation, please let me know if you prefer an alternative version.