Semantic Segmentation of Atherosclerosis in Superficial Layer of IVOCT Images Using Deep Learning

Abstract

Labeling the lesion tissue pixel-by-pixel is a serious process for cardiovascular disease (CAD) doctors, which causes time-consuming and low effectiveness. We proposed an automatic method with deep learning technology to classify the vessel tissue on the intravascular optical coherence tomography (IVOCT) image with a pixel level. Considering that only the superficial layer contains valuable information about the tissue, we firstly segmented the region of interest (ROI) by using the level set method and cropped square patches from it as the input data of neural network for the purpose of utilizing the analyzable area and increasing the data volume to improve the generalization of the network model. We chose SegNet to implement the learning procedure and predicted the classification of each pixel of cropped patches. Finally, constructing a 3-D volume to place each prediction on each slice and finding out the maximum type number of every pixel as the final class of lesion tissue. The classification results show that Our method presents a considerable approach as a computer-assisted tool for doctors.