Research On Segmentation And Recognition Of Intracranial Atherosclerotic Plaque In HRMR Image Based On Deep Learning

Stroke,also known as cerebrovascular accident,is caused by intracranial vascular obstruction or rupture.Stroke is one of the most serious cerebrovascular diseases in the world,and the rupture of intracranial atherosclerotic plaque（referred to as intracranial arterial plaque after this）is the leading cause of cerebrovascular accidents.High-Resolution Magnetic Resonance Imaging（HRMRI）can clearly show intracranial arterial plaque due to its high-precision imaging ability.It is one of the important methods of intracranial arterial image detection at present.The traditional diagnosis method of vulnerable plaque is that cardio-cerebrovascular experts first find the location of plaque from a large number of intracranial HRMR images,and then complete the diagnosis of pathology with their own experience.This process is time-consuming and laborious,and it is difficult to avoid the subjectivity of doctors.The accurate segmentation and recognition of intracranial arterial plaque through deep learning can help cardio-cerebrovascular experts make a more accurate judgment,so it has great research value.This paper takes the intracranial image of high-resolution magnetic resonance imaging as the research object.After extensively consulting the relevant literature at home and abroad and learning the pathological knowledge of intracranial artery plaque,the image acquisition,data set production,and image preprocessing of intracranial artery plaque are completed;On this basis,a segmentation method of UNet intracranial arterial plaque based on efficient channel attention and a recognition method of Retinanet vulnerable intracranial arterial plaque based on pyramid pooling and multi head self-attention are proposed.1)Making of intracranial arterial plaque data set;Data set is the core of deep learning.Without data sets,the neural network can not be trained.A data set with balanced positive and negative samples and correct label drawing can make the neural network achieve better results.Firstly,the collected 79210 intracranial images were screened one by one,1446 images containing plaque were selected,and the plaque contour was drawn by labelme.Then check the drawn contour information with the radiologist to ensure that the label of each image is accurate,which is used as the data set of image segmentation.Finally,according to the divided plaque,labelimg is used to label the vulnerable plaque as the data set of the recognition model.2)UNet intracranial arterial plaque segmentation based on efficient channel attention;Firstly,Eca Net is used to weight the feature channels containing plaque,so that the network can better learn the characteristics of plaque.Secondly,the feature fusion module is improved,the amount of network parameters is reduced,and the training and segmentation speed is improved,so that doctors can see the segmentation results faster.Finally,the loss function is improved to avoid abnormal loss in the training process.The experimental results show that compared with UNet segmentation method,the similarity of dice coefficients is improved by10.42%.3)The recognition method of Retinanet intracranial arterial vulnerable plaque based on pyramid pooling and multi head self-attention;For the problems of complex internal structure and difficult edge information extraction of vulnerable plaque,a feature extraction network based on Pyramid Pooling and Multi Head Self Attention Networks（PPSANet）is proposed.Firstly,the model of Atrous Spatial Pyramid Pooling（ASPP）is improved by adding dilated convolution with small expansion rate to reduce the loss of edge detail information;Secondly,the pyramid pooling model is combined into each residual block of the residual network to improve the feature extraction ability of the network for the plaque of different sizes.Finally,the bottom residual block of the residual network is replaced by Multi Head Self Attention（MHSA）,which enhances the ability of network feature learning and captures the correlation between data and features.The proposed feature extraction network is combined with Retinanet to identify vulnerable plaques.The experimental results show that compared with the residuals based Retinanet,it improves the m AP by 2.8percentage points.4)The design of the auxiliary diagnosis software of intracranial arterial plaque;The whole process of the software includes image denoising,image enhancement,lesion region segmentation and lesion type recognition.Firstly,through the pre-trained model in Chapter 3,the lesion area is segmented to help doctors locate the plaque quickly;Secondly,use the pre-trained model in Chapter 4 to recognize the segmented plaque,determine the type of lesion,assist the doctor to quickly grasp the patient’s condition and reasonably arrange the follow-up treatment.