Research On Key Technologies Of Intelligence-assisted Diagnosis Of Carotid Ultrasound Imaging Plaque

Carotid atherosclerosis is one of the important causes of cardiovascular diseases such as coronary heart disease and stroke,early detection of vulnerable plaques due to atheromatous deposits in carotid arteries can change patient outcomes.Doctors analyze and diagnose lesions by observing ultrasound images,however,interpreting a large amount of image data brings great challenges to doctors.Inexperience and fatigue will lead to errors in disease judgment.Therefore,using deep learning model to segment plaque accurately in ultrasound videos and classify the stability of plaques is of great significance to assist doctors in clinical diagnosis.In actual diagnosis,doctors need to judge the condition by analyzing the state of carotid plaques along several cardiac cycles.And unlike natural images,ultrasound images usually have no obvious boundary features and are easily confused with other parts.Therefore,how to construct a high-precision segmentation model based on dynamic images has become a major challenge in carotid plaque research.At the same time,the acoustic characteristics of ultrasound imaging will cause speckle noise in the original ultrasound image,which will greatly affect the doctor’s judgment on the nature of the plaque.How to construct a carotid plaque classification model under the influence of noise is another major challenge in the research process.In response to the above challenges,this thesis studies the intelligent diagnosis technology of carotid plaque based on ultrasound images,and on this basis,develops an intelligent auxiliary diagnosis system for carotid ultrasound images,which is of great significance to the diagnosis process of doctors.There are three main research works in this thesis:(1)Aiming at the problem of accurate segmentation of objects in complex backgrounds,this thesis proposes a multi-layer feature complementary carotid plaque segmentation network.The cross-layer spatial positioning module is innovatively proposed to help the model achieve accurate patch positioning in moving images.In the coding structure of the model,adaptive refinement features are learned by calculating channel-spatial attention,and finally the patches are accurately segmented based on the gating network fusion of complementary information from different layers.The validity of the model and the necessity of each sub-module are demonstrated through a large number of comparative experiments and self-ablation experiments.(2)Aiming at the problem of classifying the carotid plaques accurately under noise interference,this thesis uses ResNet-18 as the basic backbone of classification network,the wavelet transform with integrated attention mechanism is used to replace part of the structure of the basic classification network to suppress the propagation of noise in the down-sampling process.An attention mechanism is introduced to the high-frequency components generated after wavelet transform,so that the model pays more attention to the target information in the high-frequency components,and the calculated attention value is then connected with the low-frequency components for classification tasks.The validity of the model and the necessity of each sub-module are demonstrated through a large number of comparative experiments and self-ablation experiments.(3)The ultimate goal of this thesis is to develop an intelligent assistance system that can help doctors in clinical diagnosis.Based on the above research content,this thesis develops a carotid ultrasound image intelligent auxiliary diagnosis system including system management module,patient management module and intelligent analysis module.The system can upload patient’s ultrasound image files,use the intelligent analysis module to analyze the uploaded image files,then send the generated medical records to remote experts,assisting experts to give professional diagnosis based on the analysis results of the system.