A Novel Hybrid Approach for Retinal Vessel Segmentation with Dynamic Long-Range Dependency and Multi-Scale Retinal Edge Fusion Enhancement

TL;DR

Retinal vessel segmentation remains challenging due to multi-scale vessel variability, curved morphologies, and low-contrast boundaries. The authors introduce HREFNet, a hybrid framework that preserves high-resolution representations and fuses multi-scale edge information via the MREF module, while employing the Dynamic Snake Visual State Space (DSVSS) block to capture curvature and long-range dependencies with eight-directional S-SS2D and dynamic weighting. The MREF module enhances boundary precision; DSVSS combines DSConv with Mamba-based state-space modeling to improve vascular continuity, especially for thin or fragmented vessels. Empirical results on DRIVE, STARE, and CHASE_DB1 show state-of-the-art or competitive performance across Dice, clDice, AUC, and HD95, with ablation studies confirming the individually beneficial contributions of DSVSS and MREF and the advantage of deeper stage designs. The approach offers a robust, clinically relevant solution for accurate retinal vessel analysis and lays groundwork for broader biomedical segmentation tasks.

Abstract

Accurate retinal vessel segmentation provides essential structural information for ophthalmic image analysis. However, existing methods struggle with challenges such as multi-scale vessel variability, complex curvatures, and ambiguous boundaries. While Convolutional Neural Networks (CNNs), Transformer-based models and Mamba-based architectures have advanced the field, they often suffer from vascular discontinuities or edge feature ambiguity. To address these limitations, we propose a novel hybrid framework that synergistically integrates CNNs and Mamba for high-precision retinal vessel segmentation. Our approach introduces three key innovations: 1) The proposed High-Resolution Edge Fuse Network is a high-resolution preserving hybrid segmentation framework that combines a multi-scale backbone with the Multi-scale Retina Edge Fusion (MREF) module to enhance edge features, ensuring accurate and robust vessel segmentation. 2) The Dynamic Snake Visual State Space block combines Dynamic Snake Convolution with Mamba to adaptively capture vessel curvature details and long-range dependencies. An improved eight-directional 2D Snake-Selective Scan mechanism and a dynamic weighting strategy enhance the perception of complex vascular topologies. 3) The MREF module enhances boundary precision through multi-scale edge feature aggregation, suppressing noise while emphasizing critical vessel structures across scales. Experiments on three public datasets demonstrate that our method achieves state-of-the-art performance, particularly in maintaining vascular continuity and effectively segmenting vessels in low-contrast regions. This work provides a robust method for clinical applications requiring accurate retinal vessel analysis. The code is available at https://github.com/frank-oy/HREFNet.

Figures (7)

• Figure 1: Challenges. (1) Multi-scale. Vessels vary in size from large to small, requiring accurate multi-scale representation. (2) Vessels structural variability. Vessels exhibit diverse structural characteristics, including variations in both morphology and orientation. (3) Edge Feature Ambiguity. Low contrast and unclear boundaries make accurate segmentation difficult
• Figure 2: Motivation. (a) The previous method relies solely on the base segmentation network, leading to edge feature ambiguity and vessel fracture (highlighted in red). (b) Our approach integrates scale fusion and edge awareness mechanisms, enhancing vessel continuity and reducing segmentation errors, particularly along thin and low-contrast vessels
• Figure 3: The overall architecture of the proposed HREFNet. (a) is overview of HREFNet, where $H$ and $W$ denote the height and width of the input image, respectively, and $C_i$ represents the number of channels. (b) takes Stage 2 and Stage 3 as an example to demonstrate multi-scale fusion details, while this fusion process actually occurs across Stage 2 to Stage 4. (c) represent the Multi-scale Retina Edge Fusion module, designed to enhance edge details and multi-scale feature aggregation
• Figure 4: Overview of the proposed DSVSS Block. (a) is the structure of the proposed Dynamic Snake Visual State Space Block, which consists of three branches: a dynamic branch with DSConv for adaptive receptive fields, an identity branch, and a DWConv branch with our 2D Snake-Selective Scan (S-SS2D) module for long-range modeling. Outputs from the three branches are fused and refined through a final convolution (Final Conv). (b) presents the Dynamic Snake Convolution, which adjusts the receptive field based on vessel curvature and direction, enhancing feature extraction. (c) depicts the S-SS2D module that performs dynamic weighted scanning across eight directions, enabling effective feature aggregation and improving segmentation of complex vessel structures
• Figure 5: Typical segmentation results of different methods on three classical datasets. DRIVE (top), STARE (middle), CHASE_DB1 (bottom). The red pixel indicates false positives, the yellow pixel indicates false negatives, and the green pixels represent the true positives