Towards Automated Fracture Detection: A Study on Feature-Based Transfer Learning with Xception for Hairline and Avulsion Fractures

Accurate diagnosis of bone fractures, particularly avulsion and hairline frac-tures, is essential for effective treatment and recovery. Traditional methods relying on radiologists’ expertise can be subjective and prone to errors. This study investigates the application of feature-based transfer learning with the Xception model to enhance fracture classification in X-ray images. A dataset of 80 hairline fractures, 100 avulsion fractures and 50 non-fracture images were used. Pre-trained Xception convolutional neural network (CNN) models extracted discriminative features, which were then classified using Support Vector Machine (SVM), Logistic Regression (LR), and k-Nearest Neighbors (kNN). The results demonstrated that both SVM and LR achieved high accu-racy, with SVM showing superior generalization due to its ability to handle complex, non-linear patterns. LR exhibited reliable performance but faced challenges with non-linear boundaries, while kNN was sensitive to noise and parameter selection. Despite these challenges, the study confirms that fea-ture-based transfer learning improves classification efficiency and accuracy compared to training CNNs from scratch. These findings highlight the po-tential of integrating deep learning and machine learning for developing au-tomated fracture detection systems to assist healthcare professionals. Future work should explore advanced architectures and refine model parameters to further enhance performance. This study lays a foundation for improving di-agnostic accuracy in medical imaging, contributing to better patient care and outcomes.