School of Computer Science and Technology, Hainan University, Haikou, 570228, China.
Affiliated maternal and child health hospital (Children's hospital) of Hainan medical university/Hainan Women and Children's Medical Center, Haikou, 570312, China.
J Med Syst. 2023 Sep 30;47(1):102. doi: 10.1007/s10916-023-01992-7.
Precise segmentation of retinal vessels is crucial for the prevention and diagnosis of ophthalmic diseases. In recent years, deep learning has shown outstanding performance in retinal vessel segmentation. Many scholars are dedicated to studying retinal vessel segmentation methods based on color fundus images, but the amount of research works on Scanning Laser Ophthalmoscopy (SLO) images is very scarce. In addition, existing SLO image segmentation methods still have difficulty in balancing accuracy and model parameters. This paper proposes a SLO image segmentation model based on lightweight U-Net architecture called MBRNet, which solves the problems in the current research through Multi-scale Bottleneck Residual (MBR) module and attention mechanism. Concretely speaking, the MBR module expands the receptive field of the model at a relatively low computational cost and retains more detailed information. Attention Gate (AG) module alleviates the disturbance of noise so that the network can concentrate on vascular characteristics. Experimental results on two public SLO datasets demonstrate that by comparison to existing methods, the MBRNet has better segmentation performance with relatively few parameters.
视网膜血管的精确分割对于眼科疾病的预防和诊断至关重要。近年来，深度学习在视网膜血管分割方面表现出了优异的性能。许多学者致力于研究基于彩色眼底图像的视网膜血管分割方法，但基于扫描激光检眼镜 (SLO) 图像的研究工作却非常匮乏。此外，现有的 SLO 图像分割方法在准确性和模型参数之间仍难以平衡。本文提出了一种基于轻量级 U-Net 架构的 SLO 图像分割模型，称为 MBRNet，该模型通过多尺度瓶颈残差 (MBR) 模块和注意力机制解决了当前研究中的问题。具体来说，MBR 模块以相对较低的计算成本扩展了模型的感受野，并保留了更多的详细信息。注意力门 (AG) 模块减轻了噪声的干扰，使网络能够专注于血管特征。在两个公共 SLO 数据集上的实验结果表明，与现有方法相比，MBRNet 在参数量较少的情况下具有更好的分割性能。
