Plaque Recognition of Carotid Ultrasound Images Based on Deep Residual Network

In this paper, we explored a deep residual nets (ResNets) to automatically extract features of carotid ultrasound images and identify the plaques in the images. Our results showed that on a dataset with 1,828 carotid ultrasound images, a significantly better plaque recognition accuracy than the model based on support vector machine (SVM) was obtained. The accuracy of SVM and Resnets in the classification of carotid ultrasound images was 74.3% and 80.3%, respectively. In addition, we analyzed the impact of three different ROIs (regions-of-interest) on the recognition performance. The corresponding accuracies, which are classified by ResNets with three different ROIs are 80.3%, 85.3% and 92.5%, respectively. Furthermore, we have made an effort to improve the recognition performance through transfer learning in ResNets. There is a significant improvement in the performance of classification, for which the accuracies for the three different ROIs are 93.4%, 93.4% and 93.5%, respectively.