azamat1ch/rsna-knee-2nd-place

How azamat1ch/rsna-knee-2nd-place works

This repository implements a competition pipeline for detecting 12 knee MRI abnormalities. At inference, the prediction runner prepares and validates the deployment, runs four core model members and optional specialists, then produces a rank-weighted submission. The package also supports DICOM preprocessing, training, and report/image-derived labels; the external reader and PF specialist provide separate model paths. The README documents the prediction and blend workflow, while source excerpts show selected preprocessing and model-input details. Several training and inference components are unsampled, so their internal links are omitted where not evidenced.

Main entry points: infer.py Core inference · predict.py Prediction runner · tiles.py Crop inputs · windows.py Window inputs

Components 21

MRI inputs

Competition data

Core inference

Prediction runner predict.py
DICOM decoding knee/dicom.py
Series selection knee/series.py
Knee centering knee/centre.py
Crop inputs knee/tiles.py
Window inputs knee/windows.py
Core models knee/models.py
Core inference knee/infer.py

Specialist models

Reader members knee/external.py
Reader inference external/reader/infer_c.py
PF specialist external/v731/kaggle_infer.py
PF native inputs external/v731/prepare_native.py

Labels and training

Label generation knee/labels.py
Expert image panels knee/panels.py
Training stores knee/stores.py
Core training knee/train.py
PF training external/train/pf-specialist/train.py
Label tables

Prediction outputs

Rank blend knee/blend.py
Submission output

Connections 27

Competition data → Prediction runner: provides exams
Prediction runner → Core inference: runs core
Prediction runner → Reader members: runs specialists predict.py
Prediction runner → Rank blend: writes blend predict.py
Core inference → DICOM decoding: decodes exams knee/infer.py
Core inference → Series selection: selects series knee/infer.py
Core inference → Crop inputs: uses crop inputs knee/infer.py
Core inference → Window inputs: uses window inputs knee/infer.py
Core inference → Core models: runs models knee/infer.py
Crop inputs → Knee centering: uses centre knee/tiles.py
Crop inputs → DICOM decoding: reads pixels knee/tiles.py
Window inputs → Knee centering: uses centre knee/windows.py
Window inputs → DICOM decoding: reads pixels knee/windows.py
Reader members → Reader inference: runs reader knee/external.py
Reader members → PF specialist: runs PF model knee/external.py
PF specialist → PF native inputs: builds native inputs external/v731/kaggle_infer.py
Rank blend → Submission output: writes predictions
Prediction runner → Submission output: writes receipt
Label generation → Expert image panels: uses image panels knee/labels.py
Label generation → Label tables: produces targets
Competition data → Expert image panels: supplies MRI images knee/panels.py
Competition data → Training stores: supplies training data
Training stores → Core training: provides inputs
Label tables → Core training: provides targets
Core training → Core models: trains weights knee/train.py
Label tables → PF training: provides targets
PF training → PF specialist: trains specialist

Drawn by GitDiagram on November 12, 2026 from the repository's file tree, README and source files. Read as Markdown ·