改进的大核卷积U-Net视网膜血管分割方法

Improved method of large kernel convolution U-Net for retinal vessel segmentation

眼底视网膜血管结构形态复杂、对比度低,且训练样本有限,易产生过拟合现象。针对以上问题,提出一种改进的大核卷积U-Net视网膜血管分割方法(large kernel residual U-Net, LKR-UNet)。减少U-Net下采样次数和每一层的通道数缓解模型过拟合和退化问题;使用大核残差卷积模块(large kernel residual convolution block, LKR-Block)充分提取视网膜血管的特征;通过级联空间通道注意力(cascaded spatial channel attention, CSCA)模块计算空间和通道注意力,提高分割的准确性。提出方法在DRIVE和CHASE_DB1数据集上进行消融实验,在两个数据集上的敏感度分别为84.04%和83.77%,AUC分别为97.82%和98.75%,F1-score分别为82.92%和84.67%。该方法较现有先进算法有一定提升,能有效进行视网膜血管分割。

The structural morphology of the retinal vessels in the fundus is complex,with low contrast and limited training samples,making it easy to produce overfitting.A retinal vessel segmentation method was proposed to address the above problems based on large kernel residual U-Net(LKR-UNet).The number of U-Net downsampling and the number of channels per layer were reduced to alleviate the problem of model overfitting.The large kernel residual convolution block(LKR-Block)was used to fully extract the features of retinal vessels.The spatial and channel attention was computed using the proposed cascaded spatial channel attention(CSCA)module to further improve segmentation accuracy.The ablation experiments were conducted on the DRIVE and CHASE_DB1 datasets,respectively.The sensitivity is 84.04%and 83.77%,the AUC is 97.82%and 98.75%,and the F1-score is 82.92%and 84.67%on the two datasets,respectively.Compared with the existing advanced algorithms,the proposed method has certain improvement and can segment retinal vessels effectively.