Résumé
Hip fractures are a common cause of morbidity and mortality and are usually diagnosed from the X-ray images in clinical routine. Deep learning has achieved promising progress for automatic hip fracture detection. However, for fractures where displacement appears not obvious (i.e., non-displaced fracture), the single-view X-ray image can only provide limited diagnostic information and integrating features from cross-view X-ray images (i.e., Frontal/Lateral-view) is needed for an accurate diagnosis. Nevertheless, it remains a technically challenging task to find reliable and discriminative cross-view representations for automatic diagnosis. First, it is difficult to locate discriminative task-related features in each X-ray view due to the weak supervision of image-level classification labels. Second, it is hard to extract reliable complementary information between different X-ray views as there is a displacement between them. To address the above challenges, this paper presents a novel cross-view deformable transformer framework to model relations of critical representations between different views for non-displaced hip fracture identification. Specifically, we adopt a deformable self-attention module to localize discriminative task-related features for each X-ray view only with the image-level label. Moreover, the located discriminative features are further adopted to explore correlated representations across views by taking advantage of the query of the dominated view as guidance.Furthermore, we build a dataset including 768 hip cases, in which each case has paired hip X-ray images (Frontal/Lateral-view), to evaluate ourframework for the non-displaced fracture and normal hip classification task.