In brief
The 2026 RSNA Knee Abnormality Detection competition asks for multimodal knee-MRI classification across a defined set of clinically important abnormalities. Only 58 of the 4,407 training studies carry the 12 gold labels, but every training study has a free-text radiology report, so my pipeline runs report to labels to image model to image-only inference.
The model is a timm slice encoder feeding a BiGRU, attention pooling and series aggregation over 2.5D slice stacks, ending in 12 logits. Training uses DistributedDataParallel on two Quadro GV100 GPUs, patient-level GroupKFold, BCE, AMP, a cosine schedule and per-label AUC logging. On the data side, about 500 GB of raw DICOM was preprocessed in shards on Kaggle (windowed, rescaled uint8 arrays at 384 px) and synced to a GPU cluster.
Status
Status note
Ongoing as of 15 September 2026, with about a month to go. No medal, leaderboard position or result exists yet, and none is claimed. Open items: tiny gold-label folds, report-based pseudo-labelling in progress, no augmentation yet.
Placeholder
A validation curve or experiment table, once available.
The full case study, covering the problem, my contribution, the technical decisions, the evidence, the limitations and what I learned, follows.
Links
- Competition on Kaggle (external link)RSNA Knee Abnormality Detection, a research code competition.
- Kaggle profile (external link)Shows one competition entered.