| Stroke is one of the most common cardiovascular and cerebrovascular diseases,and its morbidity,mortality,and recurrence rate remain high.The main reason is the prolonged accumulation of atheroma inside the carotid arteries,which will eventually block the blood vessels after it falls off.Early detection of the risk and timely intervention can effectively prevent it.This study uses deep learning technology to accurately segment the carotid artery and plaque tissue in the head and neck CTA image,assesses the degree of carotid artery stenosis,and classifies the plaque characteristics to assist doctors in diagnosis.The main research contents are as follows:(1)The network structure optimization of nn UNet based on residual,Squeeze-andExcitation Networks(SENet),and Atrous Spatial Pyramid Pooling(ASPP)module.The whole nn UNet framework is introduced from the perspectives of implementation mechanism,reasoning process and network structure.Aiming at the shortcomings of the insufficient receptive field and inability to extract multi-scale information in its underlying network,the residual module,SENet,and ASPP are introduced to optimize the network structure.Segmentation verification is performed on the brain tumor dataset.(2)Optimization of nn UNet segmentation algorithm based on image processing.According to the unique attributes of the CTA carotid artery plaque dataset,nn UNet is further optimized in the segmentation pipeline.In the preprocessing step,image denoising,irrelevant tissue removal,and more data augmentation are introduced to ensure the quality and diversity of the dataset.In the loss function,Top K loss,which is focusing on learning difficult samples is added.In the heuristic rule,the downsampling rule of nn UNet is modified to expand the size of the feature map in the bottom layer and retain more detailed information.In the post-processing,the isolated point removal is based on the eightconnected algorithm,and the reconnection of the broken blood vessel based on the topology refinement method is performed.The DICE coefficients on the carotid artery and plaque segmentation are 0.9400 and 0.7454,and the effectiveness of all optimizations is proved through ablation experiments.At the same time,the Bland-Altman consistency analysis is used to verify the high consistency between the segmentation results and the gold standard.Finally,the vascular stenosis is calculated.(3)Plaque classification algorithm.For carotid atherosclerotic plaques,radiomics and deep learning are used to classify the plaque characteristics.In radiomics,three methods,respectively based on positive and negative sample distance,recursive feature elimination,and random forest,are used to calculate the feature’s importance.Then,multiple classifiers are selected for classification.Moreover,the results of data augmentation are analyzed by the t-distributed stochastic neighbor embedding method.In deep learning,the Res Net series of networks are used to classify 3D plaque samples,and a shallow 3D classification neural network combined with the squeeze-and-excitation module is designed to classify the components of plaques,providing a reference basis for doctors to assess the risk of plaque.(4)Design and implementation of the medical image visualization tool.A medical imaging visualization tool is designed to visualize carotid artery plaque data in 2D and3 D views.Support real-time segmentation of vascular plaques,measurement of stenosis of the carotid artery,and component classification of plaques. |
