Automatic Aortic Wall Segmentation and Plaque Detection using Deep Convolutional Neural Networks

Abstract

Abnormal aortic wall thickness and the presence of aortic plaque have been linked to various types of cardiovascular disease. QuantiĄcation of both indicators currently depends on manual or semi-automatic methods which suffer from limited quality and long acquisition times. This work presents various fully automatic state-of-the art solutions to two medical image processing problems: aortic wall segmentation and plaque slice detection. A u-net derived residual convolutional neural network (CNN), a cascaded pipeline of two CNNs and a 3D CNN architecture are used for aortic wall segmentation. Plaque detection is performed by a standard multilayer residual CNN classification architecture, a u-net derived CNN classifier and a capsule CNN. The experiments show that the u-net inspired residual CNN performs best at the aortic wall segmentation task with a Dice score of around 0.8 while the capsule CNN achieves the best results in slice-wise plaque detection with a precision of 0.74 and an accuracy of 0.68.