DEIT-Based Feature Extraction with Ensemble Machine Learning for Accurate Bone Fracture Detection in X-Ray Images

Abstract

Bone fracture detection using X-ray imaging has traditionally depended on expert radiological assessment, which is time-intensive and subject to inter-observer variability. With the increasing need for rapid and accurate diagnosis in clinical settings, automated computer-aided diagnostic systems have gained significant attention. Although advanced imaging modalities such as CT scan and MRI provide high diagnostic accuracy, they are costly and less accessible, making them unsuitable for large-scale screening. Consequently, X-ray-based intelligent systems have emerged as a scalable and cost-effective alternative. In this study, a hybrid framework is proposed by integrating transformer-based feature extraction with ensemble machine learning techniques. A pre-trained Data-efficient Image Transformer (DeiT) is employed as a fixed feature extractor to convert raw X-ray images into high-dimensional feature representations. These features are then used to train multiple classifiers, including Ridge Classifier, Passive Aggressive Classifier, and Nearest Centroid Classifier, along with a proposed Fast Interpretable Greedy-tree Sums (FIGS) model for enhanced decision-making. To improve transparency, an Explainable Artificial Intelligence (XAI) module is incorporated as a preliminary validation layer, ensuring that the input is a bone X-ray and providing contextual outputs such as fracture location, type, and severity. This enhances trust and interpretability in clinical decision-making. Performance evaluation using accuracy, precision, recall, and F1-score demonstrates the effectiveness of the proposed approach. The system is further deployed via a Telegram Bot interface, enabling real-time analysis and offering a practical, accessible solution for assisting medical professionals in fracture diagnosis.