RSNA 2026 Knee Abnormality Detection

Kaggle: rsna-knee-abnormality-detection · 12 binary findings per knee MRI study · macro ROC-AUC · code competition, <=9 h, internet off · final submission 2026-10-22

Start here

Doc	What it is
docs/STRATEGY.md	Problem, measured data facts, host intel, ranked ideas. Read first.
docs/FINDINGS.md	Phase 0 results — 819 k files scanned; the decisions they lock
docs/EXPERIMENTS.md	Run log — every result, runtime, and what it changed
docs/ROADMAP.md	The 75-day campaign, phase by phase, and the GPU-hour budget
docs/RESEARCH_AGENDA.md	What we still don't know, and which decision each answer changes
docs/PLATFORM.md	Where to run and why — 570 GB drives the decision
docs/DAY1.md	Research and decide. Plus the standing trap list

Decisions locked (8 Aug): full campaign to 22 Oct · target both the main and efficiency tracks · build our own pipeline rather than fork the public baseline · Kaggle-only compute.

The four things that matter

1. Only 58 of 4,407 training studies carry labels. The other 4,349 have a radiology report. This is a weak-supervision problem wearing a computer-vision costume.
2. Ground truth is image-derived, not report-derived — two MSK radiologists plus an adjudicator, using explicitly severity-thresholded criteria, with "on the fence" graded negative. Report-derived labels agree only ~82 %. That gap is systematic, not random.
3. Random K-fold inflates AUC by ~0.053 through scanner memorisation — and our first model showed a +0.136 grouped-vs-random gap, so the pixels leak site too. Group your folds.
4. Vision capacity is currently the bottleneck, not labels. On the 58 gold studies the text labeler scores 0.791 and the first vision model 0.674.

Pipeline

Everything runs Kaggle-to-Kaggle; nothing large passes through a local machine.

metadata scan (CPU)  ->  cache build (CPU)  ->  training (T4)  ->  submission
   819k headers          15.9 GB, 4407          mounts the           <=9 h,
   ~6.5 min              studies, ~1 h          build output       internet off

Each kernel mounts the previous kernel's output directly, so there is no dataset upload step between them.

Layout

data/          competition CSVs + sample DICOMs   (gitignored)
eda/           analysis scripts, run locally on CPU
src/           report_labeler.py — the multilingual severity extractor
kaggle/        one folder per Kaggle kernel, each with its kernel-metadata.json
docs/          strategy, findings, roadmap, experiment log
artifacts/     derived data (gitignored — embeds StudyInstanceUIDs)

Setup

python -m venv .venv && .venv/Scripts/pip install -r requirements.txt

Kaggle CLI must be authenticated (~/.kaggle/kaggle.json) and the competition rules accepted.

Two constraints that bite

Do not send report text to any hosted LLM API. Competition Rule 4.b (Data Security) plausibly forbids it and the host has not ruled. Use open-weights multilingual models locally or inside a Kaggle notebook. See docs/STRATEGY.md section 6.

Do not select the P100. Kaggle's current PyTorch ships no Pascal kernels, so the session dies at the first convolution. Set "machine_shape": "NvidiaTeslaT4".