Anatomy-Aware Siamese Network: Exploiting Semantic Asymmetry for Accurate Pelvic Fracture Detection in X-ray Images

Haomin Chen, Yirui Wang, Kang Zheng, Weijian Li, Chi-Tung Chang, Adam P. Harrison, Jing Xiao, Gregory D. Hager, Le Lu, Chien-Hung Liao, Shun Miao ;

Trauma PXR are essential for instantaneous pelvic bone fracture detection. However, small, pathologically critical fractures can be missed, even by experienced clinicians, under the very limited diagnosis times allowed in urgent care. As a result, fracture CAD has very high demands to save time and assist physicians to detect (otherwise) missed fractures more accurately and reliably. In this work, we present a new approach to fracture detection that uses a Siamese network to take advantage of the anatomical symmetry of pelvic structures to improve fracture detection. We show that symmetric alignment at the network feature level makes Siamese learning more spatially accurate and also reduces the influence of imaging artifacts. Metric learning on Siamese deep features further improves CAD differentiation power with the anatomical symmetry cue. We evaluate our method on 2,359 PXR patients, reporting an area under the ROC curve value of 0.9771, the highest among state-of-the-art fracture detection methods. "