融合注意力机制和U-Net的视网膜血管分割
Retinal Vessel Segmentation Based on Attention Mechanism and U-Net
-
摘要: 视网膜血管分割对于辅助医生进行临床筛查与诊断眼科疾病、糖尿病等大规模疾病具有重要意义.针对现有视网膜血管分割算法对细微血管分割不足、分割易受复杂背景区域影响的问题,提出一种融合注意力机制和U-Net的TCU-Net算法对视网膜血管进行分割.首先,通过单通道特征提取、图像裁剪等预处理方法,对原始眼底图像进行特征增强与数据扩充;然后,基于U-Net结构对预处理后的图像进行分割,编码器结合ResNet与Transformer对图像细节与全局特征信息进行捕获,可有效地增强血管特征提取能力,解码器引入改进通道注意力模块辅助上采样过程,细化分割结果.在DRIVE和CHASEDB1数据集上的实验结果表明,TCU-Net算法的准确率分别为0.968 4和0.974 8,灵敏度分别为0.789 9和0.825 6,受试者工作特性曲线下方的面积分别为0.982 0和0.987 6;综合分割性能较U-Net,CIEU-Net等算法均有较大提升.Abstract: Retinal blood vessel segmentation is of great significance to assist doctors in clinical screening and diagnosis of large-scale diseases such as ophthalmology and diabetes. Aiming at the problems of insufficient segmentation of fine blood vessels and the influence of complex background regions in the existing retinal blood vessel segmentation algorithms, a TCU-Net based on attention mechanism and U-Net is proposed to segment retinal vessels. Firstly, use single-channel feature extraction, image cropping and other preprocessing methods to perform feature enhancement and data expansion on the original fundus image. Secondly, the TCU-Net is proposed based on the U-Net to segment the preprocessed image. The encoder combines ResNet and Transformer to capture the image details and global feature information, which can effectively enhance the ability of vessel feature extraction, the decoder introduces an ICA block with improved channel attention to assist the upsampling process to refine the segmentation results. The experimental results on the DRIVE and CHASEDB1 datasets show that the accuracy of the proposed algorithm is 0.9684 and 0.974 8, the sensitivity is 0.7899 and 0.8256, and the Area Under Curve is 0.9820 and 0.987 6, the comprehensive segmentation performance is better than U-Net, CIEU-Net and other algorithms.