An Automatic Detection Scheme of Acute Stanford Type A Aortic Dissection Based on DCNNs in CTA Images

Abstract

As a severely life-threatening disease, acute Stanford type A aortic dissection brings patients with high potential of death. Computed Tomography Angiography (CTA) images become the effective diagnostic basis of aortic dissection. In this paper, we propose a scheme for automatic patient-level and site-level detection of aortic dissection based on Deep Convolutional Neural Networks (DCNNs) in order to relieve the burden on doctors and diagnose accurately and timely. Firstly, Mask R-CNN is adopted to precisely detect and segment the aorta from the original CTA images. Taking advantage of spatial continuity among CTA images, we put forward an area-wise screening method to remove a handful of unsatisfactory segmentation results. Then we process the segmented aorta images to extract the edges by Canny edge detector. Finally, two ResNets are employed in processed aorta images for aortic dissection detection. Specially, we can detect not only the ascending and descending aorta, but also the aortic arch, which may be the site of onset. We tested on over 3500 images, achieving an accuracy of 98.66% in slice-level detection, 96.51% in patient-level detection and 91.47% in site-level detection.