3D Fully Convolutional Network for Accurate Aortic Valve Segmentation in Cardiac CT Scans

抄録

Valve-sparing surgeries are becoming more and more popular because they allow a repair of the native aortic valve and a replacement of the diseased aortic root and the ascending aorta. While it has several advantages over valve-replacement surgeries, it is also technically more demanding. Accurate aortic valve segmentation in cardiac CT images is a robust strategy for delineation valve structures for sparing surgery planning. Recent developments of using deep learning for dense semantic segmentation of medical images, like 3D fully convolutional networks (FCNs) have achieved very impressing results. However, most of these network architectures have multiple down-sampling or cropping steps in order to enlarge the receptive fields to get enough context in the images for accurate segmentation due to the limitation of GPU memory. In order to solve this issue, we propose a combination of deep voxel-wise dilated residual network (DRN) and feature pyramid network (FPN), referred as DRN-FPN to precisely segment both internal and external structures of aortic valve.