DSP-Net: Deeply-Supervised Pseudo-Siamese Network for Dynamic Angiographic Image Matching

Xi-Yao Ma, Shi-Qi Liu, Xiao-Liang Xie, Xiao-Hu Zhou, Zeng-Guang Hou, Yan-Jie Zhou, Meng Song, Lin-Sen Zhang, Chao-Nan Wang

Abstract: During percutaneous coronary intervention (PCI), severe elastic deformation of coronary arteries caused by cardiac movement is a serious disturbance to physicians. It increases the difficulty of estimating the relative position between interventional instruments and vessels, leading to inaccurate operation and higher intraoperative mortality. Providing doctors with dynamic angiographic images can be helpful. However, it often faces the challenges of indistinguishable features between consecutive frames and multiple modalities caused by individual differences. In this paper a novel deeply-supervised pseudo-siamese network (DSP-Net) is developed to solve the problem. A pseudo siamese attention dense (PSAD) block is designed to extract salient features from X-ray images with noisy background, and the deep supervision architecture is integrated to accelerate convergence. Evaluations are conducted on the CVM X-ray Database built by us, which consists of 51 sequences, showing that the proposed network can not only achieve state-of-the-art matching performance of 3.48 Hausdorff distance and 84.09% guidewire recall rate, but also demonstrate the great generality to images with different heart structures or fluoroscopic angles. Exhaustive experiment results indicate that our DSP-Net has the potential to assist doctors to overcome the visual misjudgment caused by the elastic deformation of the arteries and achieve safer procedure.