| Ischemic stroke is a common cerebrovascular disease with high incidence rate,high recurrence rate,high mortality,high disability rate and high economic burden,which seriously endangers people’s life and health.One of the most effective treatment methods for ischemic stroke is thrombolytic intervention in the acute phase.Timely treatment of ischemic stroke is of great significance to improve the prognosis of patients.With the development of computer technology,deep learning has been gradually applied to the field of medical image segmentation,providing a new idea for accurate segmentation of ischemic stroke related regions and tissues.This paper studies the key technologies of ischemic stroke segmentation based on UNet3+ network,and proposes several key technologies to improve the network in view of the shortcomings of UNet3+ network segmentation.The main work and innovation of this paper are as follows:1.Aiming at the insufficient ability of UNet3+ to fuse features of lesions with different scales,a stroke segmentation network based on multi-scale Attention feature fusion was designed.Multi scale attention(MSA)was introduced at the end of UNet3+network to obtain multi-scale feature maps from decoders.The importance of adaptive acquisition of features at different scales further enhances the representation ability of features in segmentation networks.Ablation and comparison experiments were performed on AIS and ISLES2022 data sets in the cooperative hospital.Compared with UNet3+,the Io U,DSC,SEN and PRE indexes increased by 1.48%,1.73%,3.33% and 2.79%,respectively.2.Although the full-scale jump connection in UNet3+ network can integrate multilevel features,the deep semantic features are still not fully utilized,which affects the accurate target location of the network.To solve this problem,a hybrid Attention stroke segmentation network based on UNet3+ is proposed.Firstly,Pyramid Squeeze Attention(PSA)is used to extract multi-scale features to obtain rich image context information and effectively mitigate the problem of information loss caused by downsampling.Then Coordinate Attention(CA)was introduced at the full-scale jump connection to emphasize useful features,improve the ability of network to acquire position information and channel information,and further improve the characterization ability.Finally,MSA was introduced at the end of the network to fuse the features of different scales and highlight the most significant feature maps of different scales to adapt to the adjustment of the current segmentation lesion size.Ablation and comparison experiments were also performed at AIS and open data set ISLES2022 at the partner hospital.Ablation experiments on AIS data set showed that the combination of PSA and CA modules increased by 1.89%,1.82%and 4.42% in Io U,DSC and SEN,respectively.With the introduction of MSA,PRE has been significantly improved,increasing by 2.99%.The ISLES2022 data set shows that Io U and DSC have increased by 2.85% and 2.48% respectively,while SEN and PRE have increased by 6.19% and 2.18% respectively.3.In order to further prove the effectiveness of the proposed network,the visualization of attention heat map and the introduction of mixed attention one by one can make the proposed network pay more attention to the focal areas and make correct segmentation decisions.Compared with UNet3+,the number of network parameters only increases by 4M,which is much smaller than that of UNet and Attention UNet,indicating that the network achieves the purpose of accurate segmentation at the expense of a little computational effort. |
