How network structures affect the 2D-3D registration of cardiovascular images

Pre-operative and intra-operative vascular image registration plays an important role in the navigation of minimally invasive vascular interventional surgery (MIVI). Deep learning methods perform excellently in the balance of time consumption and registration accuracy. Optimal network structure is hard to construct due to the indirect correlation between convolution operation and vascular characteristics. In this paper, we compare the performance of various CNN models to find the most suitable network structure for vascular image registration. The effects of different convolution submodules are also discussed to optimize the network structures. Experiments on patients suffering from different cardiovascular diseases are conducted to evaluate the robustness of CNN models. Networks covered in this paper are all suitable for vascular image registration with the registration error within 1 mm or 1 degree, while Alex-reg outperforms the others and achieves the best registration result.

Automated summary

This paper explores the application of deep learning methods in vascular image registration, compares the performance of different CNN models, and discusses the optimization of network structures. The experiments demonstrate that these networks are suitable for vascular image registration, with Alex-reg achieving the best performance.