What it is
Three fold checkpoints and inference code for a 2.5D ConvNeXt-Tiny knee MRI reader, with public LB 0.929 standalone and 0.944 when blended. 3 ConvNeXt-Tiny cross-validation checkpoints plus end-to-end DICOM inference and training code for a 2.5D knee MRI reader; trained on report-derived soft labels, excluding 58 official gold-labelled studies.
Key results
- 3 cross-validation fold EMA checkpoints
- Includes preprocess, model, inference, and training code
- Uses compressed DICOM decoders for offline install
Useful for reproducible knee MRI classification baselines and raw-DICOM medical imaging inference pipelines.