Automated Musculoskeletal Fracture Analysis Using Hybrid CNN Classification and Attention-Enhanced Segmentation Models in X-ray Imaging

Objectives: This study aims to build an automated, reliable deep learning system to detect, locate and quantify musculoskeletal fractures from X-ray images. The framework is intended to enhance the diagnostic accuracy and facilitate effective clinical decision-making. Method: The proposed system is a complete deep learning system to identify fractures. A VGG19–AlexNet Adaptive Hybrid CNN is used for fracture classification, a transfer-learning model is utilized for fracture detection and an Attention-ResUNet architecture is used for fracture segmentation and localization. The integrated framework allows for an overall evaluation of fractures by classification, detection, segmentation and extraction of quantitative measurements. Findings: The performance was tested experimentally on various benchmark datasets and is found to be excellent. The proposed Hybrid CNN was tested on the MURA (WRIST) dataset with an accuracy of 99.63% and an AUC of 99.99%. In the FracAtlas dataset, fracture detection resulted in 91.03% accuracy and 93.46% AUC. The segmentation model attained Dice score of 0.384 and a pixel accuracy of 99.02%, and it allowed for meaningful extraction of the geometric measurements related to fracture. The framework also produces structured diagnostic outputs and enables deployment via app that relates to a FastAPI back end. Novelty: The proposed framework combines the classification, detection, localization and quantitative assessment of fractures in a single Hybrid CNN-based architecture, in contrast to conventional fracture analysis approaches that are based on only one task.