Feedback request: reproducible cardiac MRI SAX segmentation workflow with ACDC validation

[ShayLiu]

Hi MONAI community,

I’m building an open-source cardiac MRI SAX segmentation workflow and would appreciate feedback from people who work on medical imaging pipelines:

https://github.com/ShayLiu/cmr-ai-segmentation-workbench

Current status:

• Public ACDC fold-0 validation
• nnU-Net v2 2D CPU debug run with conservative anatomical postprocessing
• Dice and HD95 reporting for RV, myocardium, and LV
• QC overlays for expert labels and predictions
• Representative failure-case analysis
• No private patient data, DICOM/NIfTI files, or checkpoints included

Current ACDC CPU debug metrics:

• RV Dice: 0.8579; HD95: 7.44
• Myocardium Dice: 0.8547; HD95: 4.49
• LV Dice: 0.9093; HD95: 3.99
• Mean Dice: 0.8740; mean HD95: 5.31

This is not intended as a clinical model or a state-of-the-art claim. The main goal is to make the CMR segmentation workflow transparent and reproducible: data conversion, nnU-Net formatting, prediction, postprocessing, metrics, QC visualization, privacy boundaries, and failure analysis.

I’m especially interested in feedback on:

1. How to structure a clean MONAI baseline for the same ACDC workflow.
2. Whether the evaluation/reporting structure is useful for medical imaging researchers.
3. What failure-case or QC visualizations would be most helpful.
4. What would make this project more useful as a reproducible research template.

Any suggestions would be appreciated.