Research On Recognition Of OCT Cardivascular Vulnerble Plaque Based On Deep Learning

Cardiovascular disease has become the leading cause of deaths among residents of our country,which has caused incalculable damage to national health of China and the country's economic development.Vulnerable plaques are one of the leading causes of cardiovascular disease,the leading cause of coronary heart disease.OCT imaging technology has a higher resolution than other existing imaging techniques,and can clearly observe the coronary arterial components(lipid,calcification,fibrous plaque),which play a major role in identifying vulnerable plaques in particular.Due to the time-consuming and laborious manual analysis of OCT cardiovascular images and subjective differences among different physicians,By using computer intelligent image recognition technology,it is of great significance to assist the doctor in the diagnosis and treatment of coronary heart disease by automatically identifying the locations of vulnerable plaques in the OCT cardiovascular image sequence of patients.Therefore,this course aims to study the automatic recognition algorithm of vulnerable plaques in OCT cardiovascular images.The main research content of this topic is to design OCT cardiovascular image vulnerable plaque recognition algorithm based on deep learning method.The main research content is divided into two parts: The first part deals with the deep learning segmentation network model based on the encoder-decoder structure.The backbone network in the encoder is replaced by Res Net and Res Next,which are widely used in various computer vision tasks,Dilated convolution are added to the network which increases the receptive field of the feature map.The decoder is inspired by the method of object detection.It adds the feature pyramid module and the spatial pyramid pooling module,which makes the network better integrate multi-scale feature information.The improved loss function was used during training process and auxiliary losses were added.In the experiment,by observing the performance of the algorithm in the validation data set,the model based on the encoder-decoder structure proposed in this paper can accurately segment the vulnerable patch area,and make inferring time short;the second part is mainly for U-Net which brings several innovative improvements that make it applicabl e to the vulnerable patch segmentation problem in this project: including replacing the U-net encoder part with a pre-trained classification network,adding the residual unit to the network that makes training of deeper network models easier and adding recurrent convolution units that makes use of the stronger feature maps brought about by feature pooling.By comparing the performance of pre-and post-improvement U-Net in plaque segmentation,it has been shown that the accuracy has been greatly improved.The results of this research show that the method proposed in this paper can achieve the highest score of 0.78 for MIOU index of vulnerable plaque segmentation on OCT cardiovascular dataset,according to the results of artificial segmentation by experts.The model presented in this paper basically meets the basic requirements for vulnerable plaque recognition in clinical analysis of OCT cardiovascular images,which has certain reference value for doctors to analyze OCT cardiovascular images and follow-up interventional treatment.