| Cardiovascular disease(CVD)is a common disease with a high mortality rate,which has been causing great harm to people’s health for a long time.Automatic segmentation of coronary plaque in Computed Tomography Angiography(CTA)images is of great significance for the treatment and prevention of cardiovascular diseases.In recent years,remarkable progress has been made in automatic coronary plaque segmentation through deep learning networks.However,these methods generally have problems such as excessive resource consumption during training and low segmentation accuracy in scenes with small calcification areas and complex distribution.To this end,this study took cardiovascular CTA images as the research object,and proposed two calcified plaque segmentation algorithms based on deep learning.The main research contents of this study are as follows:(1)The existing deep learning methods have the problems that the training process consumes more resources and the segmentation effect in the small calcified plaque area is unpleasant.This study improved on the basis of the dual attention mechanism network DANet,and proposed a double attention network CGNet that integrates the global attention mechanism.CGNet used Res Net as the backbone network,and reduced the computational parameters and computational costs by introducing 3?3convolution layers to replace the 7 ?7 convolution layer.By introducing the channel attention module CAM and the global attention module GAM in parallel after the backbone network,the intrinsic dependence between different pixels in the image was modeled in the channel dimension and the spatial dimension,resulting in ability of network capture imformation of calcified plaque region was improved.The experimental results show that when Res Net50 is used as the backbone network,compared with DANet,CGNet reduces 866 MB in memory cost and 2.46 ms in inference time respectively,and improves the segmentation accuracy on Dice and MIo U by 0.0075 and 0.0064 respectively.The under-segmentation was reduced.(2)In order to further improve the segmentation accuracy of calcified plaque in cardiovascular CTA images,the accuracy of contour fitting of the network in complex calcified areas and the segmentation of calcified plaque shape and size was improved.This study improved on the basis of the dual attention mechanism network DANet,and proposed a network MDNet that integrates dynamic multi-scale fusion module DAM and spatial attention mechanism PAM.MDNet introduced a parallel dynamic multiscale fusion module DAM and a spatial attention module PAM after the backbone network Res Net.DAM can capture semantic information at different scales to improve the segmentation effect,and PAM can model the spatial dependence of images to improve performance.The experimental results show that when Res Net50 is used as the backbone network,compared with DANet,MDNet reduces 113 MB in memory cost,and improves the segmentation accuracy on Dice and MIo U by 0.0325 and 0.0381 respectively.MDNet also improves the accuracy of shape segmentation and contour in complex areas.Finally,by comparing the segmentation results of different Transformer networks on cardiovascular CTA images,the effectiveness of MDNet in the segmentation of calcified plaques in cardiovascular CTA images was further proved. |
