KDPCnet: A Keypoint-Based CNN for the Classification of Carotid Plaque

Abstract

AbstractClassification of carotid plaque echogenicity in ultrasound images is an important task for identifying plaques prone to rupture, thus for early risk estimation of cardiovascular and cerebrovascular events. However, it is difficult for normal classification methods to distinguish the plaque area and extract the feature of plaques, because the carotid artery plaque area accounts for a very small proportion of the entire ultrasound image, and the plaque boundary is fuzzy. In addition, the image usually needs to be resized before being fed to the neural network, resulting in information loss. In this work, a keypoint-based dual-branch carotid plaque echogenicity classification network (KDPCnet) is proposed to solve those problems. Our model consists of two parts. First, a lightweight sub-network is applied to identify the plaque’s center point. Then, a dual-branch classification sub-network is proposed to integrate global information of the entire ultrasound image and the local detail information of plaques without reducing the resolution of the plaque area and changing the aspect ratio of the plaque. On the dataset of 1898 carotid plaque ultrasound images from the cooperation hospital, the five-fold cross-validation results show that KDPCnet outperforms other advanced classification models and keypoint localization can effectively assist carotid artery plaque echogenicity classification.KeywordsCarotid plaqueUltrasoundKeypoint localizationClassification