A Weakly Supervised Framework for 2D/3D Vascular Registration Oriented to Incomplete 2D Blood Vessels

Abstract

Registering preoperative 3D computed tomography (CT) to intraoperative 2D digital subtraction angiography (DSA) can complete the missing part of 2D vessels by 3D-2D projection, which is useful for the guidance of vascular interventional surgery. However, vessel excalation in DSAs remains hard for registration in clinical circumstances. To overcome this challenge, we proposed a weakly supervised 2D/3D vascular registration framework to improve the performance on DSAs with incomplete vessels. We combined a CNN regressor with a novel 3D-2D mask projection module to build the full network. Then a two-step training procedure is employed, including supervised pre-training on simulated images and weakly supervised finetuning on DSAs. In addition, a patch-based content loss is introduced in the finetuning step to give an appropriate similarity measure for images with incomplete vessels. We evaluated our method on both simulated and real images. Experiments prove that our proposed method is effective to improve the backbone regressor’s registration performance on incomplete vessels and has the ability to complete vessels in DSAs with high recall.