Phenotyping TPF via Self-Supervised Learning: A Label-Agnostic Framework with Expert Validation
Researchers have developed a novel label-agnostic framework for characterizing tibial plateau fractures using self-supervised learning. This approach bypasses the need for human-assigned labels, which are prone to inter-observer variability, by directly learning fracture representations from imaging data. The system, utilizing a fine-tuned ResNet-50 encoder with a SimCLR objective, identified four distinct fracture phenotypes. These phenotypes demonstrated robust stability and internal cohesion, with expert validation confirming their clinical interpretability and independence from traditional classification systems like Schatzker and AO/OTA. AI