Contextual anatomy-guided deep learning for accurate fovea segmentation in diabetic retinopathy fundus images

TL;DR

This paper introduces a new deep learning method for accurately identifying the fovea in diabetic retinopathy images by using anatomical context, improving segmentation and treatment decisions.

Contribution

The paper introduces a data-centric approach using anatomical landmarks and a new dataset and framework for robust fovea segmentation.

Findings

Incorporating anatomical landmarks like the optic disc and blood vessels improves fovea detection performance.

The proposed framework achieves high segmentation accuracy (fovea IoU = 0.812, F1 = 0.894).

The GEV-based augmentation technique outperforms baseline methods in detection rates.

Abstract

Accurate fovea segmentation in fundus images is a critical step in diabetic retinopathy screening; however, it remains a challenging task due to the indistinct boundaries of the fovea. Beyond simple localization, precise segmentation offers essential clinical value for Diabetic Macular Edema (DME) management, as treatment decisions–specifically the choice between intravitreal anti-VEGF injection for center-involved DME and laser therapy for extrafoveal edema–depend on the accurate delineation of the foveal region. While existing methods often rely on increasing model architecture complexity, the potential of anatomical context within the training process remains under-explored. This paper presents a data-centric approach that leverages contextual information to robustly identify the fovea. We demonstrate that progressively incorporating key anatomical landmarks–the optic disc, retina, and blood vessels–into training labels significantly enhances fovea detection. To facilitate this, we developed IDRiD-RETA-FV, a meticulously annotated dataset comprising 81 images (54 training, 27 testing) with complete anatomical structures (inter-observer F1=0.98), and introduce MNv4Fovea, a framework designed to explicitly exploit these anatomical inter-dependencies through a multi-class constraints mechanism. Evaluation on the held-out test set with verified ground truth demonstrates excellent segmentation performance (fovea IoU = 0.812, F1 = 0.894, AED = 4.06 pixels). To demonstrate the efficacy of our synthesis strategy, our GEV-based augmentation technique achieves a detection rate of 98.4% compared to 59.0% for baseline geometric augmentation (paired t-test: t = 8.536, p < 0.001, Cohen’s d = 1.093). Cross-dataset evaluation on REFUGE, MESSIDOR, and ARIA demonstrates competitive localization performance, achieving state-of-the-art Average Euclidean Distance on REFUGE (22.46 ± 18.73 pixels) and MESSIDOR (6.52 ± 5.89 pixels) with robust generalization across diverse imaging protocols. These results establish that explicit anatomical context, rather than mere model complexity, is key to accurate fovea segmentation, offering a robust paradigm for medical image analysis.

Figures