01 — Research Overview
KNEE-SWARM AI is a research-stage framework investigating weakly supervised multilabel classification of knee MRI examinations. It combines multilingual radiology-report analysis, MRI representation learning, and proposed evidence-grounded assurance mechanisms.
The system targets twelve clinically important abnormalities. During development, radiology reports provide potential training supervision. The planned competition inference pipeline operates on MRI examinations without requiring test radiology reports.
02 — Dataset Audit
Dataset statistics derive from the project's Kaggle training-data audit. Automatic language identification found 11 language categories; these predictions have not yet been manually validated.
03 — Proposed System Architecture
MRI DICOM series + training radiology reports
Clinical Evidence Pipeline
Multilingual language identification, pathology extraction, negation analysis and weak-label confidence assessment.MRI Vision Pipeline
DICOM quality checks, anatomical-plane processing, pretrained visual encoders and sequence representation learning.Gold annotations + validated report-derived labels
2.5D convolutional encoders + multiple-instance aggregation
Image QC · Evidence consistency · Uncertainty analysis
12 study-level abnormality probabilities
04 — Prediction Targets
Ligaments
ACL, MCLMenisci
Medial Meniscus, Lateral MeniscusOsteoarthritis
Medial OA, Lateral OA, PF OAOther Findings
Effusion, Synovitis, Baker's Cyst, Contusion, Fracture05 — Research Methodology
| Component | Planned methodology |
|---|---|
| Study design | Weakly supervised multilabel classification |
| Supervision | Gold annotations and report-derived weak labels |
| Image representation | Pretrained CNN with multi-plane MRI aggregation |
| Validation | Five-fold gold-label cross-validation |
| Primary metric | Macro-averaged ROC-AUC |
| Assurance | Quality checks and uncertainty analysis |
| Inference | Offline MRI-only prediction |
06 — Development Status
Dataset structure audit: Completed
Gold-label distribution audit: Completed
Five-fold class coverage: Verified
Exact report-duplicate audit: Completed
Language identification: Completed
Clinical label extraction: In progress
Image model training: Pending
Kaggle submission: Pending
Patient-level independence has not been established. The proposed model and agent components have not yet been validated for predictive performance.
07 — Evaluation Protocol
The primary evaluation metric is macro-averaged ROC-AUC across twelve binary targets. Model development will use fixed study-level validation folds, with attention to weak-label contamination and duplicate examinations. Per-target performance and uncertainty will be reported where statistically estimable.