基于高灰度值注意力机制的脑白质高信号分割

White Matter Hyperintensities Segmentation Based on High Gray Value Attention Mechanism

脑白质高信号是脑小血管病的常见影像学表现,对脑小血管病患者临床诊断有重要参考价值。脑白质高信号分割是临床诊断的基础工作之一,往往需要极具经验的医师进行手动刻画,极其耗费时间且繁琐。脑白质高信号是脑核磁共振成像T2加权像或者液体衰减反转恢复序列图像(Fluid Attenuated Inversion Recovery,FLAIR)中的高信号影,其灰度值明显高于其它正常的脑部组织。为提高对脑白质高信号区域的关注,根据脑白质高信号的影像学特征,提出一种具有高灰度值注意力机制的网络模型。基于UNet网络,设计并引入高灰度值注意力模块,使网络模型更加关注于图像中灰度值较高的区域;为提高网络模型的特征提取能力,引入残差混合注意力模块。该方法明显地提升了脑白质高信号分割效果,DSC指标和Recall指标分别达到0.8330和0.8870,优于现有算法。消融实验也验证了高灰度值注意力模块和残差混合注意力模块的有效性。本文为基于FLAIR影像的脑白质高信号病灶分割提供了一种新方法,同时验证了传统图像分割方法与深度学习技术相结合的可行性。

White matter hyperintensities,commonly seen in the image of cerebral small vessel disease(CSVD),shed light on the clinical diagnoses of patients with cerebral small vessel disease.White matter hyperintensities segmentation,as a basic work in clinical diagnosis,often requires experienced doctors to carry it out manually,which is time-consuming and intricate.White matter hyperintensities,referring to the hyperintense shadows in T2 weighted magnetic resonance images of the brains or fluidattenuated inversion recovery sequence images,are of higher gray values than other brain tissues.To enhance the attention to areas of white matter hyperintensities,this paper proposes a network model of a high gray value attention mechanism in light of the imaging characteristics of white matter hyperintensities.The model,based on the UNet,introduces a module of high gray value attention so that it can pay more attention to the areas of relatively high gray values in the images.It also introduces a residual mixed attention module to enhance the ability for extracting features of the net model.As a result,it significantly enhances the segmentation effect of white matter hyperintensities,with its DSC and Recall indicators reaching 0.8330 and 0.8870,respectively,which is better than existing algorithms.Moreover,ablation experiments verified the effectiveness of the high gray value attention module and the residual hybrid attention module.This paper provides a new method for the FLAIR-based segmentation of white matter hyperintensities lesion,and verifies the feasibility of combining the traditional method for image segmentation with in-depth learning technology.