Self-supervised learning reveals 4 new tibial fracture phenotypes without labels

Tibial plateau fracture classification (Schatzker, AO/OTA) suffers from inter-observer variability, causing supervised models to learn human disagreement rather than stable morphology. Researchers from the field designed a label-agnostic framework that eliminates this constraint by learning fracture representations directly from imaging data without observer-assigned labels. They fine-tuned a RadImageNet-pretrained ResNet-50 encoder on 154 cleaned knee radiographs using the SimCLR contrastive objective, followed by UMAP dimensionality reduction and k-means clustering to discover four imaging-derived phenotypes.

The four phenotypes demonstrated robust stability (bootstrap ARI=0.319±0.041), strong internal cohesion (silhouette=0.511), and blinded coherence ratings of 3-5/5 from two independent clinicians. One phenotype was unanimously identified as comminution—a complex fracture feature isolated without any supervisory signal. Inter-partition comparison against Schatzker labels yielded ARI=0.013, confirming orthogonality to conventional classification boundaries. Notably, reviewers anchored in established vocabularies perceived imaging-derived groups as heterogeneous precisely where Schatzker alignment was lowest, suggesting two orthogonal dimensions. These findings establish label-agnostic SSL phenotyping as a reproducible, clinically interpretable complement to conventional classification.