Research On Key Technologies Of Intelligent Recognition Of Cardiac Image Data Based On Deep Learning

At present,the mortality rate of coronary heart disease is in the primary position of the death of noncommunicable diseases,which has been a great threat to the life and health of all mankind.Coronary angiography is a common technique for the diagnosis of coronary heart disease.As a complement to coronary angiography,intravascular ultrasound can provide detailed information about atherosclerotic plaques on the vessel wall.In clinical practice,the combination of these two imaging techniques is often used in the diagnosis,treatment and prognosis evaluation of coronary heart disease.Therefore,the automatic identification of coronary angiography data and intravascular ultrasound data by intelligent recognition technology can assist doctors in the classification,detection and segmentation of lesions,which has important theoretical research significance and clinical practical value.Based on the deep learning theory,this paper systematically studies the detection of coronary artery stenosis and the segmentation of coronary arteries,as well as the segmentation of endovascular intima and medial membrane in intravascular ultrasound images and the classification of intravascular plaques.This work includes the following main contents and innovative achievements:(1)In the study of coronary artery stenosis detection,a coronary artery stenosis detection data set is constructed,and a coronary artery stenosis detection model based on large-scale feature fusion and attention mechanism is proposed.Based on the classical two-stage target detection model Faster-RCNN,the feature extraction part of this model is improved.A large-scale feature map output structure is constructed to alleviate the problem of missing detection of small and medium-sized narrow targets.In addition,a feature fusion structure from high level to low level and then from low level to high level is constructed,which increases the localization ability of multi-scale output features.After the output of high-level features,a spatial attention mechanism was constructed to alleviate the interference of bones and other body tissues in the background.The experimental results show that:the detection accuracy of the proposed model is greatly improved than that of the basic detection model;and the recall rate and average accuracy of the multi-scale narrow target mixed detection are 84.8%and 77.5%,respectively,which are also higher than that of other one-stage and two-stage target detection models.(2)In the study of coronary artery segmentation,the data set of coronary artery segmentation is constructed,and the coronary artery lightweight segmentation model based on the bottleneck residual module and attention mechanism is proposed.The bottleneck residual module is used to replace the traditional structure of U-Net model to reduce the number of parameters and the amount of calculation.The channel attention module is added to the bottleneck residual module to build the relationship between the feature channels.A block-focused attention module is proposed to model the relationship between each block in the feature map so as to enhance the feature representation of each block.In addition,considering that the vascular region in the coronary angiography image occupies a very small part of the entire coronary angiography image,in order to alleviate the problem of data imbalance,the weight cross entropy loss and Dice loss are adopted in the process of model training,simultaneously.The experimental results show that:these modules can greatly improve the segmentation performance of the model when in the scenarios of a few parameters are added or no parameters are added.The segmentation results of the model reached 87.7%,97.89%,97.29%and 99.10%in sensitivity,specificity,accuracy and AUC(Area under the Curve),respectively.In addition,the generalization ability of the model is verified by using clinical data sets.Finally,the experimental comparison with other segmentation models shows that,under the condition of similar segmentation accuracy,the proposed model has the lowest number of parameters,only 0.75M(million),thus saving computing resources.(3)In the study on the segmentation of intima and external membrane,a dataset is constructed,and an intravascular ultrasound segmentation model based on Transformer feature fusion and channel enhancement is proposed.The remote dependencies and local contexts of the features extracted by the Convolutional Neural Network(CNN)at different scales are modeled by using Transformer coding blocks.It makes up for the shortcomings of CNN’s inability to capture remote dependencies and Transformer’s lack of local context modeling ability,and further merges the fused features in the upsampling process.A channel enhancement module based on gate mechanism is proposed to enhance the feature channels with more valid information.The experimental results show that:the constructed intravascular ultrasound segmentation model has excellent performance in multiple evaluation indexes,whether for the intima of the vessel or the external membrane of the vessel.Among them,the indexes of Jaccard Measure(JM)and Hausdorff distance(HD)for intimal vascular segmentation are 91.9%and 0.201mm,respectively.The JM and HD indexes of vascular and medial membrane segmentation are 92.7%and 0.259 mm,respectively.In addition,the Pixel Accuracy(PA)is 98.49%.It is worth mentioning that the number of parameters of the constructed model is 10.09M(million),which is lower than other models in the comparison experiment.(4)In the study on the classification of intravascular plaques,the data set of multilabel classification of intravascular plaques is constructed,and a multi-label classification model of intravascular plaques based on semi-supervised learning is proposed.According to the characteristics of intravascular ultrasound images,a semisupervised learning framework based on different views of polar coordinates and Cartesian coordinates is constructed.The multi-scale feature fusion structure is used to better classify plaques.In order to make the output feature fusion of each stage more targeted,the stage attention module is proposed to adaptively select the effective information combination from the features of different scales,so that the final feature map can have the feature representation of both large and small size plaques,thus improving the classification accuracy of small size plaques.The experimental results showed that:compared with the semi-supervised framework under the same classification model,the proposed semi-supervised framework combined classification model ResNet50 outperforms the models constructed by other frameworks in the multilabel classification task of endovascular plaques.The Exact Match Ratio(MR),Overall Recall(OR)and Overall F1-measure(OF1)are 82.08%,93.26%and 96.02%,respectively.Compared with the classification models under the same semi supervised framework,combined with the semi supervised learning framework proposed in this paper,the classification performance of the classification model proposed in this paper is better than other models.The absolute matching rate of classification indexes,the overall recall rate and the overall F1 value reach 89.58%,97.04%and 97.84%respectively.