A Hybrid Transformer-Based Framework for Robust Bone Fracture Detection using Scalable EfficientNet-B3 Architecture

Abstract

This study provides a scalable deep learning architecture aimed at automatic bone fracture detection based on radio-graphic images. Specifically, the approach involves a Hybrid Transformer Model, which makes use of the EfficientNet-B3 network as its backbone feature extractor alongside the implementation of an attention module. In order to make the algorithm more resilient to the presence of clinical noise in images, a noise layer in the form of Gaussian noise was applied. Training and validation of the architecture were conducted using a data set comprising five thousand three hundred and ninety-six images belonging to the MURA musculoskeletal dataset. Findings: The outcomes of the experiments have demonstrated that the accuracy rate of the hybrid approach remains within the boundaries of seventy-seven and seventy-nine percent. The algorithm exhibits an excellent level of precision regardless of the used partition, which is crucial in order to reduce the number of false-positive diagnostics in the medical environment. As a result of examining the self-attention maps, it has been proven that the network is able to focus on clinically significant areas of interests. Originality: The novelty in this research is that this architecture integrates a scalable backbone network along with an attention mechanism-based weighting layer tailored to musculoskeletal radiology applications. This model strikes a balance between efficiency and interpretability. This proposed framework can be used as a secondary diagnostic tool by radiologists, helping to minimize any errors in diagnosis and facilitate workflow in stressful emergency departments.