TransFigX: An Explainable Transformer-Based System for Multi-Fracture Identification in Radiographic Images

Abstract

Fracture identification from X-ray images is challenging when multiple cracks, small hairline fractures, or low-contrast regions are present. Manual diagnosis often becomes slow in emergency rooms and rural healthcare centres where radiologists may not be immediately available. To address this real-time issue, our project proposes a Data Efficient Image Transformer (DEiT) model capable of detecting multiple fractures accurately from X-ray images. To strengthen feature extraction and decision-making, the system integrates Ridge regression for reducing overfitting, Passive Aggressive Classifier (PAC) learning principles for improving model generalization, and Nearest Centroid Classifier (NCC) to enhance similarity matching between fracture patterns. Additionally, Fast Interpretable Greedy Sums are used to generate lightweight, transparent scoring for model decisions, allowing doctors to understand why a fracture was detected. In real-time medical environments such as emergency rooms, trauma centres, and rural hospitals, patients with injuries often require immediate X-ray analysis to confirm bone fractures. Small or multiple fractures are especially difficult to detect quickly, increasing the risk of misdiagnosis and improper treatment.