Contrastive Learning-Driven Representation and Feature Selection for Spinal Fracture Detection on CT Images

Abstract

In this study, the effect of contrastive learning-based representations and feature selection on the detection of spinal fractures from lumbar spine CT images was investigated. A two-class dataset consisting of 1,504 labeled CT images, including both fractured and non-fractured cases, was collected from Elazig Fethi Sekin City Hospital. Deep representations (embeddings) were extracted from unlabeled images using three different contrastive learning methods: Barlow Twins, MoCo, and SimCLR. Feature selection was then performed using the MRMR and NCA algorithms. The selected features were classified using classical machine learning algorithms, including SVM, Logistic Regression, KNN, XGBoost, and GBM. According to the experimental results, the Barlow Twins + NCA + Logistic Regression combination achieved the highest success, with 98.34% accuracy and a 99.82% ROC-AUC value. The findings reveal that fracture detection with high accuracy is possible even with limited labeled data when contrastive learning-based representations are combined with appropriate feature selection.