Abstract
Objective:To detect and classify carotid plaque ultrasound images based on different deep learning models of convolutional neural networks, and to compare the advantages and disadvantages of each model, with a view to providing a fast and accurate detection method for carotid atherosclerotic plaque ultrasound screening in stroke risk groups using artificial intelligence techniques.
Methods:A total of 5611 carotid ultrasound images of 3683 patients from the ultrasound departments of the Eighth People's Hospital of Shanghai, Fengxian District Central Hospital of Shanghai, the Second People's Hospital of Guangdong Province in Guangdong Province, and the People's Hospital of Huainan City in Anhui Province during the period of 17 September 2020 to 17 December 2022 were selected for the study.All carotid ultrasound image data redundant information was cropped, and two attending physicians with more than ten years of experience in cardiovascular ultrasound labelled and classified all the images for diagnosis. The total dataset was randomly split into a training set (3927 images) and a test set (1684 images) in a ratio of 7:3. Four deep learning models-YOLO V7 (ResNet 50) model, YOLO V7 (Inception V3) model, Faster RCNN (ResNet 50) model, and Faster RCNN (Inception V3) model-were used to detect and analyse the carotid artery plaque ultrasound images and to atherosclerotic plaques to identify and classify whether the carotid arteries are vulnerable plaques or stable plaques.The efficacy of the four deep learning models in classifying carotid atherosclerotic plaques was assessed using Accuracy (ACC), Sensitivity (SEN), Specificity (SPE), F1 scores, and Area under the curve of the working characteristics of the subjects (AUC), with P< 0.05 was taken as statistically significant difference.
Results:In this study, Faster RCNN model and YOLO V7 network base model were constructed using deep learning algorithms and two different feature extraction networks (ResNet 50 and Inception V3) were used to classify ultrasound images of carotid artery plaques.The Faster RCNN (ResNet 50) model in the test set had ACC, SEN, SPE, AUC were 0.88, 0.94, 0.71, and 0.91, respectively, which was the highest prediction efficacy for carotid atherosclerotic plaque classification among the four models. This study demonstrates the feasibility of deep learning for carotid plaque ultrasound image detection and classification, in which the Faster RCNN (ResNet 50) model has high accuracy and reliability.
Conclusion:In the diagnosis of carotid artery vulnerable plaque, the confidence level of the diagnosis using the deep learning Faster RCNN (ResNet 50) model is close to that of intermediate physicians, and the model can improve the diagnosis level of junior ultrasonographers, and also help clinics to formulate a more reasonable ischemic stroke prediction and early warning plan.