Detection And Classification Of Carotid Plaque Ultrasound Images Based On Deep Learning

Purpose:Different deep learning models based on convolutional neural network were used to detect and classify the ultrasound images of carotid plaques,and the advantages and disadvantages of each model were compared,in order to provide a fast and accurate detection method for ultrasound screening of carotid atherosclerotic plaques in high-risk population of stroke by using artificial intelligence technology.Methods:From September 17,2020 to December 17,2022,a total of 5611 carotid artery ultrasound images of 3683 patients from the Ultrasound Department of Shanghai Eighth People’s Hospital,Shanghai Fengxian District Central Hospital,the Second People’s Hospital of Yuebei in Guangdong Province,and Huainan People’s Hospital in Anhui Province were selected as the research objects.All the redundant information of carotid ultrasound image data was cut out,and all the images were labeled and classified by two attending physicians with more than 10 years of cardiovascular ultrasound experience.The total data set was randomly split into a training set(3927images)and a test set(1684 images)at a ratio of 7:3.Four deep learning models were used-YOLO V7(Res Net 50)model,YOLO V7(Inception V3)model,Faster RCNN(Res Net 50)model and Faster RCNN(Inception)model V3)model was used to detect and analyze the ultrasound images of carotid plaques,identify and classify the carotid atherosclerotic plaques,and determine whether the carotid plaques were vulnerable or stable.At the same time,junior physician A(with 1 year of work experience)and junior physician B(with 3 years of work experience)were randomly selected to diagnose carotid plaque images in validation set and test set.Accuracy(ACC),Sensitivity(SEN),Specificity(SPE),F1 score,Area under the receiver operating characteristic curve(AUC)were used to evaluate the accuracy,sensitivity(SEN),specificity(SPE),F1 score,and Area under the curve(AUC).Delong test was used to compare the diagnostic performance of different deep learning models and primary physicians.<0.05 was considered statistically significant.Results:(1)Comparison of the classification performance of the Faster RCNN model for carotid atherosclerotic plaque: ACC,SEN,SPE and AUC of Faster RCNN(Res Net50)model in the test set were 0.88,0.94,0.71 and 0.91,respectively.ACC,SEN,SPE and AUC of Faster RCNN(Inception V3)model in the test set were 0.83,0.91,0.59 and 0.85,respectively.Comprehensively,the detection indexes of Faster RCNN(Res Net 50)model are better than that of Faster RCNN(Inception V3)model,and its carotid plaque classification performance is better than that of Faster RCNN(Inception V3)model.(2)Comparison of the classification performance of the YOLO V7 model for carotid atherosclerotic plaque: ACC,SEN,SPE and AUC of YOLO V7(Res Net 50)model in the test set were 0.86,0.94,0.61 and 0.86,respectively.The values of ACC,SEN,SPE and AUC of YOLO V7(Inception V3)in the test set were 0.83,0.91,0.60 and 0.83,respectively.Comprehensively,the detection indexes of the YOLO V7(Res Net 50)model are superior to those of the YOLO V7(Inception V3)model,and its classification performance is better.(3)Comparison of diagnosis results of carotid atherosclerotic plaque by primary physicians: ACC,SEN,SPE and AUC diagnosed by two primary physicians in the test set were 0.71,0.70,0.62,0.65 and 0.73,0.74,0.65,0.70,respectively.The diagnostic efficacy of vulnerable plaques by primary physicians was lower than that of the four deep learning models.(4)Comparison between the optimal deep learning model and primary physicians in the diagnosis of carotid atherosclerotic plaque: The Faster RCNN(Res Net 50)model has the highest classification prediction efficiency among the four models.The diagnostic efficacy of the Faster RCNN(Res Net 50)model was higher than that of the two primary physicians((49)< 0.001),and the predictive diagnostic AUC of the model was higher than that of the two primary physicians for vulnerable carotid plaque.Conclusion:(1)In the diagnosis of vulnerable carotid plaque,the diagnostic reliability of artificial intelligence technology using deep learning is close to the diagnostic level of intermediate physicians,and the selected model can provide diagnostic assistance for junior physicians.(2)For carotid plaque detection and classification,artificial intelligence is more accurate.(3)Artificial intelligence technology based on deep learning has a promising application prospect in carotid plaque ultrasound images,which can effectively alleviate the workload of sonographers,improve the diagnostic level of primary sonographers,and contribute to the development of more reasonable prediction and early warning plans for ischemic stroke.Figure [19] table [7] reference [100]...