Calcification Detection of Coronary Artery Disease in Intravascular Ultrasound Image: Deep Feature Learning Approach

摘要

Coronary artery disease (CAD) is part of the non-communicable disease (NCD) in cardiovascular disease (CVD). The blood vessel area became narrow when the calcification with the plaque embedded in the coronary artery inner wall. The radiologists and medical practitioners used visual inspection to detect calcification on IVUS image. The presence of calcification will not be able to do the measurement to calculate the maximum diameter and the maximum area for the patient coronary artery either before treatment or after treatment. More than 100 frames per patient is needed to analyse the location of the calcification. In this study, our aim is to detect the presence and the absence of the calcification in the coronary artery using intravascular ultrasound (IVUS) images with catheter frequency of 20MHz. The IVUS images used were the original Cartesian coordinate image and the polar reconstructed coordinate image. In this study, three types of convolutional neural network (CNN) using Directed Acyclic Graph networks, were used together with five types of classifiers. The dataset used to demonstrate our framework is Dataset B from MICCAI Challenge 2011 that consists of 2175 coronary artery disease IVUS image where 530 are IVUS images with calcification and 1645 are IVUS images without calcification. The cross validation for testing and training, the k-fold value used was 2, 3, 5 and 10. The performance measures for the ResNet-50, the ResNet-101 and the Inception-V3 model shows an excellent result using support vector machine classifier and discriminant analysis for both types of images. A better improvement using polar reconstructed coordinate image when using decision tree classifier and Na?ve Bayes classifier whilst ResNet-101 architecture shows an excellent performance measure when applying images polar reconstructed images when using k-nearest neighbor classifier. However, Na?ve Bayes classifier has an excellent result when using Inception-V3 architecture.