Plaque Tracking And Classification Based On Carotid Ultrasound Video

Cardiovascular and cerebrovascular diseases are among the diseases with the highest morbidity and mortality in the world.The rupture and shedding of unstable plaques are the main cause of acute cardiovascular and cerebrovascular diseases,which seriously endangers human health.Clinically,Doppler ultrasound imaging technology is widely used in the examination of carotid plaque due to its advantages of convenience and no radiation.However,due to the lack of a gold standard of diagnosis,the accuracy of diagnosis largely depends on the experience of clinicians and is of poor universality.Therefore,establishing a stable,efficient and intelligent method to accurately identify plaque types is of great significance for clinical auxiliary diagnosis of auxiliary plaque stability and effective prevention of cardiovascular and cerebrovascular diseases.Most traditional studies on plaque stability focus on static images.Since the carotid plaque area is small and has complex shape changes,it shows real-time changes in ultrasound video,and a single static sonographic view may not accurately present the status of plaque.For ultrasound video,conventional tracking network will increase the difference between different semantic patches.When the same plaque in different ultrasonic view present different semantics,such as morphology,properties,etc.,it will inevitably cause optimization contradiction in deep learning training,resulting in the model defining the changed composition as new plaques,or incorrectly linking similar components between different target plaques.Therefore,the study of carotid plaque tracking and echo classification based on ultrasound video remains a challenge.This paper focuses on the study of carotid artery ultrasound video,focusing on the identification of plaque through sapatio-temporal context information of the video.On the one hand,the plaques were located according to the temporal features of ultrasound video.On the other hand,the plaques and vascular environment of multi-scale context information between ultrasound views were combined to achieve accurate tracking and echo classification of the plaques.The main research contents and innovations are as follows:(1)Multi-scale Dilated Encoder(MDE)and Internal-Exterior Feature Decoupling(IEFD)module were proposed,and the carotid plaque ultrasound video tracking model MDTrack was established.On the one hand,considering the fact that plaque tracking was a small-target tracking,and that the extraction of highdimensional features was prone to missing detail information,this paper proposed MDE to expand the receptive field through dilated convolution,and to restore local finegrained features by combining temporal multi-scale features.On the other hand,considering the optimization contradiction caused by the semantic changes of plaques in the video,this paper used IEFD to decouple the semantics of plaques into the "interior-echo" semantics and the "exterior-identity" semantics.The interior semantic information was used to determine the plaque properties,and the exterior semantic information was used to distinguish the plaque identities.Corresponding sub-tasks were optimized by being assigned to different semantic information and the splitting of contradictions.In this paper,the optimal model architecture of MDTrack was determined through ablation experiments,and the performance was compared with the tracking method proposed in the latest relevant research literature.The experimental results showed that MDTrack had more satisfying tracking performance than the current mainstream tracking models.(2)A tracking-assisted 3D Attention Convolutional Neural Network(T3DACNN)was proposed for plaque echo classification in carotid ultrasound videos.On the one hand,to address the interference problems of ultrasonic artifacts and similar tissues in carotid ultrasound video,this paper proposed an MDTrack-based feature recombination strategy to locate and extract plaque features.On the other hand,in view of the importance of a complex vascular environment around plaque in clinical observation and evaluation,this paper proposed a dual-channel attention mechanism to extract plaque echo-related features in the surrounding environment while eliminating noise interference.Furthermore,a carotid plaque echo classification model of dualchannel sequential attention was proposed,in which the plaque features extracted by MDTrack-based feature recombination strategy were used in the local channel,and the environmental features extracted by original ultrasound video were used in the global channel.In particular,the model achieved an effective fusion of features between the two channels by combining a 3D attention mechanism and multi-layer spanning connection.Compared with other classification methods including single-channel and multi-channel,the experimental results showed that the proposed model was in higher accuracy and F1-score.In addition,this paper studied the influence of the trackingassisted mechanism on classical deep learning temporal classification model.The experimental results showed that the worst performance of model with tracking-assisted classification was more satisfying than the best performance of models without tracking-assisted classification,which indicated that tracking-assisted classification was an enhancing method with high universality and stability in the time series classification model.