融合注意力机制的TransGLnet脉络膜自动分割  

Automatic segmentation of choroid by TransGLnet integrating attention mechanism

作　　者：黄文博[1] 屈超凡 燕杨[1] HUANG Wenbo;QU Chaofan;YAN Yang(School of Computer Science and Technology,Changchun Normal University,Changchun 130032,China)

机构地区：[1]长春师范大学计算机科学与技术学院,吉林长春130032

出　　处：《光学精密工程》2023年第23期3482-3489,共8页Optics and Precision Engineering

基　　金：吉林省科技厅科技发展计划与自然科学基金联合基金项目(No.YDZJ202101ZYTS147)。

摘　　要：针对脉络膜与巩膜间对比度低,脉络膜分割存在脉络膜下边界模糊,难以界定等问题,提出了融合注意力机制的TransGLnet脉络膜自动分割网络,在卷积层引入全局注意力模块(Global Attention Module,GAM),在特征之间应用矩阵乘法,在整体空间位置的多个特征之间建立非线性交互,在不使用大量参数的情况下提取全局特征;在卷积层和Transformer编码器之间引入局部注意力模块(Local Attention Module,LAM),以1/4特征图为基本单元探索局部特征,特征图元素位置移动规则为保持行位置的元素不变,将列位置的元素由大到小重新排列。两模块融合可令网络有效兼顾全局与局部特征。实验结果表明,TransGLnet网络的Dice值为0.91,准确率为0.98,平均交并比为0.89,F1值为0.90,豪斯多夫距离为6.56。与现有脉络膜自动分割方法相比,本文方法的各项性能指标均有提高。TransGLnet脉络自动分割网络具有较好的稳定性,可供临床借鉴。Addressing the challenge posed by the low contrast between the choroid and sclera in choroid segmentation,this research introduces the TransGLnet choroid automatic segmentation network,employing an attention mechanism.The incorporation of a Global Attention Module(GAM)within the convolutional layer involves matrix multiplication between features,establishing nonlinear interactions across multiple features in the global spatial context.This enables the extraction of global features without an excessive number of parameters.To explore local features,a Local Attention Module(LAM)is introduced between the convolution layer and Transformer encoder,focusing on a 1/4 feature graph.The movement rule for feature graph elements maintains row position consistency while rearranging elements in the column position from largest to smallest.The integration of these two modules ensures that the network effectively considers both global and local features.Experimental results showcase the efficacy of the proposed TransGLnet network with a Dice value of 0.91,accuracy at 0.98,equal crossover ratio of 0.89,F1 value reaching 0. 90, and a Hausdorff distance of 6. 56. Comparative analysis against existing automatic choroi-dal segmentation methods reveals notable improvements in performance metrics. The network presented in this study demonstrates robustness and stability, rendering it suitable for clinical reference.

关 键 词：医学图像处理 脉络膜自动分割 TransUnet 全局注意力 局部注意力 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]