BONE FRACTURE CLASSIFICATION IN X-RAY IMAGES USING SWIN TRANSFORMER

Abstract

Recent and precise diagnosis of the fractures of the bones is essential in the realization of an ideal therapeutic performance; however, the manual examination of radiographic materials remains labor-intensive and prone to mistakes. This paper introduces an automated binary classification of bone fractures using deep-learning-based methods based on the Swin Transformer Base (Swin-B) architecture using radiographic modalities. Leveraging the Kaggle Bone Fracture Dataset comprising 10,580 labeled X-ray images (Fracture vs. Non-Fracture classes), we implement a hierarchical vision transformer with shifted window-based self-attention mechanisms combined with transfer learning from ImageNet. We utilize a comprehensive graph of image-processing and data-augmentation approaches, such as rotation and horizontal flipping, contrast regulation and optimize a binary cross-entropy loss objective operational through the Adam optimizer. On the held-out test split, the model achieves state-of-art results with an accuracy of 98.45%, a precision of 98.32, a recall of 98.58, an F1-score of 98.45, and an area-under-curve of 0.9912 of the receiver-operating-characteristic. Comparative training with CNN-based baselines, i.e. ResNet-50 and DenseNet-121, shows that Swin-B architecture achieves better results in the local and global features capturing. Our results underscore the effectiveness of vision transformers for medical image classification and provide a robust, clinically-applicable framework for fracture detection in resource-constrained settings.