An Explainable Attention-Enhanced Deep Learning Framework for Automated Bone Fracture Detection and Multi-Class Classification in Musculoskeletal Radiographs

Abstract

In medical diagnosis, detection of bone fractures in respective X-ray images is a primary core task. The order is high accuracy and consistency due to nuanced character of fracture patterns. A close reading of a radiologist is extensive and can lead to diagnostic ambiguity, especially in high-acuity cases. To overcome these obstacles, this paper recommends an increased focus and a deep learning framework for automatic fracture detection and classification. The proposed system uses convolutional neural networks based on ResNet architectures, extended with Convolutional Block Attention Modules (CBAM) to improve feature representation by highlighting significant local and channel information. Performed within the framework, a two-stage pipeline consists of binary classification for fracture identification followed by multi-class classification to separate the specific bone type. The Musculoskeletal Radiographs (MURA) dataset is used for training and evaluation, accounting for multi-regional anatomical data. To enrich image quality and improve feature readability, Contrast Limited Adaptive Histogram Equalization (CLAHE) is used during preprocessing. Besides that, sophisticated training paradigms including data augmentation, Mixup regularization, and adaptive learning rate scheduling are used to improve model generalization and robustness. Experimental results show that the proposed methodological framework achieves a fracture recognition accuracy of 77%, an improvement of 2.33% over the existing benchmark baseline accuracy of 74.67%. For multi-class classification, the model obtains an accuracy of 96.07%, exceeding the baseline performance of 95.7% by 0.36%. The parent system achieves an accuracy of 86.5%, indicating reliable performance across both detection and classification tasks. Besides that, Grad-CAM is used to generate visual explainability, amplifying clinical insight by highlighting the affected areas in the model's predictions. The proposed framework seamlessly combines preprocessing, attention mechanisms, and deep learning architectures to provide improved performance and interpretability, making it a viable framework for real-world medical imaging applications.