Global relationship memory network for retinal capillary segmentation on optical coherence tomography angiography images

Weili Jiang, Weijing Jiang, Lin An, Jia Qin, Lushi Chen, Chubin Ou

Abstract: Automatic retinal capillary segmentation is a necessary prerequisite for quantitatively analyzing retinal vessels. In recent years, active research has been using deep learning-based methods in this field. However, deep learning methods inevitably lose spatial information of vessels when downsampling, thereby limiting the segmentation performance for fine vessels. Additionally, existing methods must pay more attention to the dynamic correlations between feature mappings in deep learning frameworks, resulting in inefficient acquisition of multi-scale decoder features. To address these limitations, we propose a Global Relationship Memory Network (GRM-Net) that considers the relationship between frequency domain and decoder hierarchy. Specifically, we first design a frequency relation learning module to preserve fine details of vessels during downsampling. This module decouples encoder features into frequency domain features of different dimensions and employs globally learnable filters to better guide the network’s attention towards vessels of different sizes and shapes. Secondly, we investigate a hierarchical relation selection module that leverages gate mechanisms to dynamically adjust the collaboration between two adjacent decoder blocks, thereby adaptively aggregating multi-scale decoder features to address the issue of underutilized decoding information. Comparative experimental results on two retinal vessel datasets validate the effectiveness of the proposed GRM-Net segmentation method. Compared to other state-of-the-art methods (Unet, CS-Net, DeeplabV3, MiniSeg, and OCTA-Net), this method achieves more remarkable segmentation results, preserving more details in the tiniest retinal capillaries. Code is available at https://github.com/WeiliJiang/Global-Relationship-Memory-Network.