融合小波特征的细节增强视网膜血管分割算法  

Detail enhanced retinal vessel segmentation algorithm based on wavelet feature and U-Net

作　　者：陆锡恒 宣士斌[1,2] LU Xiheng;XUAN Shibin(College of Artificial Intelligence,Guangxi Minzu University,Nanning 530006,China;Guangxi Key Laboratory of Hybrid Computation and IC Design and Analysis,Nanning 530006,China)

机构地区：[1]广西民族大学人工智能学院,广西南宁530006 [2]广西混杂计算与集成电路设计分析重点实验室,广西南宁530006

出　　处：《微电子学与计算机》2024年第12期29-39,共11页Microelectronics & Computer

基　　金：国家自然科学基金(61866003)。

摘　　要：针对视网膜血管分割任务中特征丢失、细血管提取困难等问题,在U-Net的基础上提出了一种融合小波特征的细节增强视网膜血管分割算法。首先,为了减少下采样过程中导致的血管特征丢失,设计了小波特征补偿模块,通过小波分析对眼底图像的高频、低频信息分开建模,以提供更丰富的多频特征。其次,考虑到眼底图像对比度低,细血管难以辨别的问题,提出了多维注意力模块增大血管与背景噪声的特征差异,结合不同深度的语义信息,分别在水平、垂直以及通道这3个维度对血管特征进行检测与探查,实现了空间与通道信息的有机融合,加强了算法的细节处理能力。最后,为了解决算法视野不足导致的血管表示不连续问题,设计了连接增强模块,将解码器各层输出相互结合,利用空洞卷积扩大算法感受野,捕获多尺度上下文信息,增强血管的连接性。在DRIVE与CHASE_DB1数据集上进行了算法的有效性测试,敏感度分别为0.8376和0.8453,F1分数分别为0.8347和0.8372,A_(UC)值分别为0.9886和0.9901。To address challenges such as feature loss and difficulty in extracting fine blood vessels in the task of retinal vessel segmentation,a detail enhanced algorithm for retinal vessel segmentation based on wavelet feature and U-Net is proposed.Firstly,to mitigate the loss of vascular features during the downsampling process,a wavelet feature compensation module is designed.This module employs wavelet analysis to independently model the high and low-frequency information in retinal images,thereby delivering a more enriched set of multi-frequency features.Secondly,considering the challenge of low contrast in retinal images and the difficulty in discerning fine blood vessels,a multi-dimension attention module is proposed to enhance the feature disparity between vessels and background noise.By integrating semantic information at different depths,this module conducts vessel feature detection and exploration along horizontal,vertical and channel dimensions.This approach achieves an organic fusion of spatial and channel information,thereby strengthening the algorithm's capability for detailed processing.Finally,to address discontinuities in vascular representation caused by insufficient algorithmic field of view,a connection enhancement module is designed.This module combines outputs from various layers of the decoder,utilizes dilated convolutions to expand the algorithm's receptive field,and captures multi-scale context information to enhance vascular connectivity.The algorithm's effectiveness is tested on the DRIVE and CHASE_DB1 datasets,yielding sensitivity values of 0.8376 and 0.8453,F1 scores of 0.8347 and 0.8372,and AUC values of 0.9886 and 0.9901,respectively.

关 键 词：U-Net 视网膜血管分割 小波特征 多维注意力 连接增强 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]