Multiscale Attention Gated Network (MAGNet) for Retinal Layer and Macular Cystoid Edema Segmentation

Abstract

Retinal optical coherence tomography (OCT) imaging is a mainstay in the clinical diagnosis of several sight-threatening diseases. Due to the wide variability in shape and orientation of retinal structures, analyzing and interpreting OCT images are complex tasks that require domain knowledge. Within the analysis process, delineating anatomical landmarks and pathological formations, i.e., segmenting OCT scans, is a labor-intensive task usually carried out by expert graders. Recently, several studies have proposed methods based on fully convolutional neural networks (FCN) to alleviate the burden of manual OCT segmentation. Despite the promising performance of FCN-based methods, the negative impact of the class imbalance problem on the segmentation of small foreground targets such as macular cystoid edemas remains a challenge. This article proposes a novel end-to-end automatic method for segmenting retinal layers and macular cystoid edema in OCT images. The proposed method introduces a novel FCN architecture that leverages spatial and channel-attention gates at multiple scales for fine-grained segmentation and a weighting loss approach to handle class imbalance. Results on a benchmark dataset that includes cases of severe retinal edema show the robustness of the proposed algorithm, which achieved state-of-the-art performance with a mean Dice score of 0.92 ± 0.03.

Keywords

• Deep learning
• semantic segmentation
• attention gates
• fully convolutional networks
• optical coherence tomography