Return to the sanctuary

RSNA · MSK Reading

Multi-plane knee MRI abnormality detection across 4,396 teams.

Solo entry in the Kaggle RSNA Knee Abnormality Detection challenge (12 findings, offline T4 inference cap). Reached 0.944 public macro-AUC (rank 169 / 4,396 on 26 Sep 2026) through decorrelated LLM weak-label pipelines and 2.5D Attention-MIL.

Best public macro-AUC
0.944
as of · Kaggle public leaderboard
Public rank
169 / 4,396
as of · Kaggle public leaderboard

The problem

Detect 12 abnormality classes across multi-plane knee MRI studies under an offline 9-hour T4 inference cap.

Weak labels set the ceiling

Only 58 studies carried gold labels; decorrelating independent LLM report extractors yielded larger gains than scaling vision backbones.

Blind label audit

An agent-assisted blind review over the 58 gold-labeled studies isolated systematic label noise strictly for dataset auditing.

Experimental discipline

Single-variable experiments, pre-registered acceptance rules, gold strictly as veto, and ignoring public deltas below 0.01 (~1.4 SE).

Next iteration

Favor one well-calibrated 2.5D architecture over heavy ensembles, decorrelate weak labels from day one, and optimize AUC-per-minute.

PyTorchtimmCoAtNet2.5D Attention-MILWeak Supervision

Research use only. Not a clinical tool. All visuals shown here are synthetic; zero competition images or reports are displayed.