Shifted-Rectangle-Window Based Transformer for non-Displaced Femoral Neck Fracture Diagnosis

Abstract

Non-displaced femoral neck fracture (NFF) is a common type of hip fracture. Diagnosis and detection of NFF is a challenging task since fractures appear in stochastic directions and are accompanied by repetitive texture. Previous work paid little attention to the directional characteristics of fractures. In this study, we apply a shift rectangle window based transformer framework to automatically detect NFF. Specifically, both vertical rectangle windows and horizontal rectangle windows are constructed to capture directional features. Meanwhile, we introduce the deformable self-attention blocks and a pseudo-RGB preprocess into our framework. Furthermore, we build a mutilcenter dataset including 1606 radiographs to evaluate our framework. We performed one comparative experiment and two ablation studies. Experimental results demonstrate that our framework surpasses existing approaches in terms of accuracy, specificity, sensitivity, and AUC.