Faster R-CNN for Detection of Carotid Plaque on Ultrasound Images

Carotid plaque is an important indicator in clinical diagnosis to whether it causes cardiovascular and cerebrovascular diseases. Due to the complexity of the principle of US imaging, doctors are cumbersome and mechanical in identifying plaque hardening inside blood vessels, it takes doctors some time to identify carotid plaque. In this study, Inception network is used as a pre-training network to extract the feature map of the image, and the recommendation information of plaque or sclerosis is extracted by the RPN network (Region Proposal Network) in Faster R-CNN to realize the automated target detection of carotid US image plaque or sclerosis. The carotid US dataset (687 images) was made and the marker categories were divided into: plaque, sclerosis, psudomorph and normal, enabling endto-end training in convolutional neural networks, with the two category of psudomorph and normal is to make the model more fault tolerant. Adjust the parameters, configuration, and network structure of the network to analyze the plaque target detection performance. Finally, the Average Precision(AP) with carotid plaque on the carotid US test set was 91.09%, and the mean Average Precision(mAP) was 58.62%. The stability and detection rate of the model met certain application requirements, and the better automatic target detection effect could be achieved.