Risk Stratification Of Carotid Atherosclerotic Plaque Using Ultrasound Images

Ischemic stroke is a disease,which seriously threatens global human health.Atherosclerosis is the main pathology leading to ischemic stroke.The carotid arteries provide an effective observation window of arterial disease and stratification of the risk of carotid plaque rupture leading to subsequent thrombosis resulting in ischemic stroke.Early diagnosis and risk stratification of carotid atherosclerotic plaques are clinically useful to formulate effective and reasonable treatment plans,prevent the occurrence of ischemic stroke,and reduce the morbidity and mortality of ischemic stroke.Ultrasound is a routine examination method for carotid atherosclerosis because of its safety,convenience,non-invasiveness,real-time,and high repeatability.Therefore,carotid ultrasound imaging is the main research subject for scholars to stratification the risk for stroke that carotid plaques pose.Deep learning has gradually become a preferred risk stratification method for researchers due to its outstanding feature learning and classification performance.However,deep learning still has some limitations in the risk stratification that carotid plaques pose.First,due to the cost of ultrasound image acquisition,processing,and labeling,the size of obtained carotid ultrasound datasets is often small,which is not optimal for training deep learning models.Second,because the prevalent convolutional neural networks(CNNs)have a uniform size requirement for the input image,uniform transformation(scaling and/or cropping)are performed without consideration of the different characteristics of the shape,size,and image intensity distribution of plaque images in different sections of the ultrasound images.Both scaling and cropping can destroy the original information in the image,resulting in a negative impact on the performance of risk stratification.Finally,carotid plaques tend to be multifocal,so only analyzing a single section(longitudinal or transverse section)of the carotid arteries will fail to integrate the features of multiple carotid plaques in both longitudinal and transverse sections for a comprehensive risk stratification,which are needed for accuracy stratification.In light of these issues,this study has been undertaken from the following aspects:(1)Automatic identification of carotid plaques from clinically collected ultrasound images and the establishment of a carotid plaque dataset are the crucial first steps in risk stratification of carotid plaques.Manual identification of carotid plaques by physicians is time-consuming,highly subjective,variable,and results in small dataset sizes.First,an automatic method using a residual network model is proposed for the identification of carotid plaques in ultrasound images,and then a secondary transfer learning method based on the optimal model is proposed to further improve the performance of the automatic identification of carotid plaques using a small dataset size.Then a generative adversarial network was built to generate carotid ultrasound images for data augmentation,which was compared with traditional data augmentation and the integration of data from multiple centers.The effects of different data augmentation methods on automatic identification performance of carotid plaques using a small dataset size were analyzed.Finally,a scheme to achieve the best automatic identification performance of carotid ultrasound plaques on a small dataset size was obtained.(2)To address the problem in risk stratification based on the echogenicity of ultrasound carotid plaques in longitudinal sections using existing convolutional neural network,a multi-level strip pooling module is proposed.CNNs equipped with this module can accept carotid ultrasound plaque images in longitudinal sections,extract features from the images and down-sampling the feature maps by multi-level strip pooling.This will result in a model that can enlarge the receptive field and collect the global and multi-scale long-range contextual information to improve the performance of risk stratification based on the identification of carotid plaque echogenicity in ultrasound images of longitudinal sections.(3)In light of the problem that only a single ultrasound image of a single plaque is used in the current approaches to carotid plaque risk stratification,a multi-path CNN model based on an object-specific pooling strategy is proposed,which is focused on patients.The model integrates features from plaque images of the bilateral carotid arteries in ultrasonic longitudinal and transverse sections of atherosclerotic patients.Then,the object-specific pooling strategies are employed for features down-sampling,and a fixed-dimensional feature vector is obtained and used for the risk stratification.The experimental results show that the model can effectively improve the performance of risk stratification of carotid plaques.This study focused on automatic identification of plaques from carotid ultrasound images using a small sample size,risk stratification based on echogenicity of carotid plaques in longitudinal sections,and risk stratification by fusing features from plaques of bilateral carotid arteries in patients with or without symptoms.The three aspects have provided a reference for assisting doctors in accurate diagnosis and treatment of carotid plaque and laid the foundation for the establishment of an end-to-end risk assessment system for ischemic stroke using carotid ultrasound images.