Deep Learning-Based Long Bone Fracture Classification Using Squeeze-and-Excitation ResNet

Abstract

This research explores the application of deep learning for the classification of long bone fractures, focusing on a novel architecture: Squeeze-and-Excitation ResNet (SE-ResNet). Traditional models, such as ResNet, Unet, and DenseNet, are evaluated to establish a benchmark for performance across key metrics, including Precision, Recall, F1-Score, and Accuracy. The comparative analysis reveals that while ResNet demonstrates robust classification capabilities, Unet significantly underperforms in this context. DenseNet offers moderate results but fails to surpass ResNet and the proposed SE-ResNet. The SE-ResNet model excels, achieving a Precision of 94.56, Recall of 96.78, F1-Score of 97.67, and Accuracy of 98.02. The integration of Squeeze-and-Excitation mechanisms enhances the model's ability to focus on pertinent features, significantly improving classification accuracy. This research underscores the potential of deep learning, particularly SE-ResNet, as an effective tool for accurate long bone fracture classification, which could lead to enhanced diagnostic practices in clinical settings.