Review of Deep Learning Applications in Myopic Choroidal OCT Segmentation

Abstract

The incidence of myopia among Chinese children has been rising annually, with a notable trend towards younger onset. The choroid, a critical vascular structure for ocular homeostasis, exhibits a significant relationship between variations in its Thickness (ChT) and the pathological advancement of myopia. Optical Coherence Tomography (OCT), with its high resolution and non-invasive characteristics, has emerged as an essential instrument for assessing myopic choroidal lesions.

Automated choroidal segmentation underpins the quantitative analysis of OCT images, necessitating precise identification of Bruch's Membrane (BM) and the Choroid-Sclera Interface (CSI). Traditional manual annotation is labor-intensive and susceptible to considerable subjective bias, but Deep Learning (DL) offers a novel solution to this problem. This article evaluates the advancements in research within this domain: Among core models, Convolutional Neural Networks (CNNs) enhance boundary localization but experience spatial information loss; Fully Convolutional Networks (FCNs) facilitate end-toend segmentation yet lack instance-level segmentation capabilities; U-Net and its variants (e.g., Bio-Net, ADU-Net) are particularly adept at addressing the attributes of medical images. Mask R-CNN attains precise pixel-level predictions and demonstrates significant clinical applicability.

In practical applications, targeting typical pathological changes such as ChT thinning, DL models enhance robustness through multi-directional optimization. Specifically, ADU-Net and GCS-Net tackle the problem of blurred boundaries, while multi-task models simultaneously perform segmentation and calculation. Comparative studies have shown that Mask R-CNN demonstrates the best performance in OCT image segmentation.