Identifying Carotid Plaque Composition in MRI with Convolutional Neural Networks

Abstract

Carotid plaques may cause strokes. The composition of the plaque helps assessing the risk. Magnetic resonance imaging (MRI) is a powerful technology for analyzing the composition. It is both tedious and error-prone for a human radiologist to review such images. Traditional computer-aided diagnosis tools use manually crafted features that lack both generality and accuracy. We propose a novel approach using Deep convolutional neural networks (CNN) to classify these plaque tissues. In order to accommodate the multi-contrast MRI images, we modify stateof-the-art CNN models to support different number of input channels, and also adapt the models to do pixel- wise predictions. On a dataset with 1,098 human subjects, we show that we achieve significantly better accuracy than previous models. Our result also indicates interesting relations between contrast weightings and tissue types