Research On Ultrasound Carotid Plaque Image Characterization Based On Deep Learning Fusion Network

Cardio-cerebrovascular diseases are one of the diseases with the highest mortality and disability rates in the world.It brings huge losses and economic burden to people’s health and society each year,and its prevention,monitoring and treatment have become a global research hotspot.Atherosclerosis is the main pathological basis of cardio-cerebrovascular diseases.The results of clinical studies have found that the ruptured carotid plaques can cause vascular embolism,which leads to ischemic stroke further.With non-radiation,non-invasiveness and convenience,ultrasound has become the most commonly used imaging method for clinical detection of carotid plaque.The purpose of this paper is to study a method for the characterization of ultrasound carotid plaque images.It is expected that various features,which can comprehensively represent plaques relatively,will be obtained from the ultrasound images,and provide an imaging evidence for the risk assessment of stroke.To this end,based on deep learning,a multi-level feature fusion network was proposed.This network takes ultrasound images of different sections and different sides as input,extracts and fuses the plaque features of different ultrasound images automatically,which greatly improves the identification accuracy between patients who experienced atherosclerotic events and those who are event-free.It will lay a foundation for establishing a risk prediction model of cerebrovascular events based on carotid ultrasound images.First,an acquisition standard for carotid ultrasound images was developed.Ultrasound imaging was performed on the left and right carotid arteries of each patient,and the largest plaque area images,obtained from the transverse and corresponding longitudinal views,were collected.The experiments covered a total of 1332 ultrasound images from 333 patients and prepared them into a carotid ultrasound image dataset for the study.Second,the multi-level feature fusion network proposed in this paper is composed of two feature generation sub-networks(G-T and G-L)and one feature fusion sub-network(F).The G-T and G-L extract plaque features from the transverse and longitudinal views of bilateral carotid ultrasound images separately.To provide an improved representation of the plaques,the F sub-network fuses the multi-level features generated from the G-T and G-L sub-networks hierarchically.In addition,a multi-output loss function was designed to further improve the classification performance of the network.By optimizing the weight parameters of the loss function automatically in the training procedure,a classification accuracy of about 91% and an area under the curve(AUC)of 0.95 were achieved.Compared with the texture features,Res Net and Mobile Net methods,using the same dataset,this result improves the accuracy by 6-7% and the AUC by 0.06.The results showed that our method was more suitable for the characterization of ultrasound carotid plaque images.The extracted fusion features may be used to identify plaque differences of different risks,which will provide strong basis and reference for the establishment of the early warning model of cerebrovascular events.