How ilyakolbafk/rsna-knee-abnormality-detection-4th-place works
This repository reproduces and retrains a competition-winning RSNA knee abnormality ensemble. The principal workflow runs Sixway inference, then deadline-bounded H6 inference, and combines their per-finding percentile ranks into a submission. Training uses historically preserved model code and prepared caches; H6 additionally depends on OAI and fastMRI data. Labels were created from image/report readings and public label tables. The graph follows documented workflow statements and the sampled inference orchestrator; detailed internals of unsampled model and cache modules are intentionally summarized.
Main entry points: predict_pipeline.py Inference orchestrator · train.py Training entry point · sources Historical model code · h6 H6 training code
Components 18
Inference
- Inference orchestrator
src/rsna_knee_final/predict_pipeline.py - Sixway ensemble
src/rsna_knee_final/inference/sixway_stage.py - H6 inference
src/rsna_knee_final/inference/h6_fast.py - Rank blend
src/rsna_knee_final/predict_pipeline.py - Submission outputs
src/rsna_knee_final/predict_pipeline.py
Model training
- Training entry point
train.py - Historical model code
training/sources - H6 training code
training/sources/h6
Data preparation
- Cache builders
data_prep/cache336 - H6 data preparation
data_prep/h6_v4
Label creation
- Core6 labels
labels/core6/build_master.py - Reader labels
labels/vm_reader
H6 auxiliary data
- OAI case pipeline
data_prep/h6_v4/oai/pipeline.py - fastMRI adaptation
data_prep/h6_v4/fastmri/fm384_v1_build/build_candidate.py
Other components
- Inference user
- Competition studies
- OAI archive
- fastMRI data
Connections 16
- Inference user → Inference orchestrator: runs
README.md - Competition studies → Inference orchestrator: provides studies
README.md - Inference orchestrator → Sixway ensemble: runs first
src/rsna_knee_final/predict_pipeline.py - Inference orchestrator → H6 inference: runs second
src/rsna_knee_final/predict_pipeline.py - Inference orchestrator → Rank blend: rank blends
README.md - Rank blend → Submission outputs: writes
README.md - Training entry point → Historical model code: stages model code
README.md - Training entry point → H6 training code: trains H6
README.md - Cache builders → Training entry point: prepares caches for
README.md - H6 data preparation → Training entry point: prepares H6 inputs
README.md - Core6 labels → Historical model code: supplies labels
README.md - Reader labels → Historical model code: supplies labels
README.md - OAI archive → OAI case pipeline: supplies cases
README.md - OAI case pipeline → H6 training code: provides training data
README.md - fastMRI data → fastMRI adaptation: supplies scans
README.md - fastMRI adaptation → H6 training code: provides adaptation data
README.md