RRWNet: Recursive Refinement Network for effective retinal artery/vein segmentation and classification

TL;DR

RRWNet tackles the challenge of manifest classification errors in retinal artery/vein segmentation by introducing a two-subnetwork architecture: a Base subnetwork that produces initial A/V/BV maps and a Recursive Refinement subnetwork that iteratively corrects labeling to enforce topological consistency. The framework marries stacking and recursive refinement, and uses a loss that emphasizes accurate base predictions before refinement. Across RITE, LES-AV, and HRF, RRWNet achieves state-of-the-art A/V classification and BV segmentation, with large gains in topological metrics COR and INF, and demonstrates the RR module's value as a standalone post-processing tool for other methods. The method is efficient, generalizable, and publicly available, offering a robust solution for clinically meaningful vascular segmentation and potential extension to other topologically constrained medical imaging tasks.

Abstract

The caliber and configuration of retinal blood vessels serve as important biomarkers for various diseases and medical conditions. A thorough analysis of the retinal vasculature requires the segmentation of the blood vessels and their classification into arteries and veins, typically performed on color fundus images obtained by retinography. However, manually performing these tasks is labor-intensive and prone to human error. While several automated methods have been proposed to address this task, the current state of art faces challenges due to manifest classification errors affecting the topological consistency of segmentation maps. In this work, we introduce RRWNet, a novel end-to-end deep learning framework that addresses this limitation. The framework consists of a fully convolutional neural network that recursively refines semantic segmentation maps, correcting manifest classification errors and thus improving topological consistency. In particular, RRWNet is composed of two specialized subnetworks: a Base subnetwork that generates base segmentation maps from the input images, and a Recursive Refinement subnetwork that iteratively and recursively improves these maps. Evaluation on three different public datasets demonstrates the state-of-the-art performance of the proposed method, yielding more topologically consistent segmentation maps with fewer manifest classification errors than existing approaches. In addition, the Recursive Refinement module within RRWNet proves effective in post-processing segmentation maps from other methods, further demonstrating its potential. The model code, weights, and predictions will be publicly available at https://github.com/j-morano/rrwnet.

Figures (10)

• Figure 1: The proposed framework, RRWNet, here applied for the segmentation and classification of retinal arteries and veins, consists of a W-shaped fully convolutional neural network consisting of two subnetworks. The output of the first subnetwork (Base) is iteratively refined by the second (Recursive Refinement) through a recursive mechanism.
• Figure 2: Examples of manifest classification errors produced by a state-of-the-art FCNN-based method Morano_AIIM_2021. (1-3) While most of the vessel is classified as artery, the model misclassifies the last part as vein. (4-6) The model often confuses the classification of vessels in crossing areas. These errors are easily detected by a human observer because they are inconsistent with the overall structure of the vascular tree, hence the term "manifest".
• Figure 3: Proposed approach for the segmentation and classification of arteries and veins. The input image is fed to the Base subnetwork, which produces coarse segmentation maps of arteries (A), veins (V) and blood vessels (BV). Next, the A/V segmentation maps are fed to the Recursive Refinement subnetwork, which recursively refines them for a certain number of iterations $K$.
• Figure 4: U-Net architecture of the two subnetworks. $N$ represents the number of base channels. $C$ represents the number of output channels. In our case, $N=64$.
• Figure 5: Examples of retinography images from different datasets and their corresponding A/V segmentation maps. (a-d) RITE. (e,f) LES-AV. (f-h) HRF. The segmentation maps are visualized as RGB images composed of the segmentation maps of arteries (Red channel), veins (Green channel) and vessels (Blue channel). This composition makes arteries appear magenta, veins appear cyan, crossings appear white (because they are both arteries and veins at the same time), and uncertain vessels appear blue (because they are not assigned to either artery or vein class, but only to vessel). * Annotations used for training and testing. † Second expert annotations.