retinal vessel segmentation

Accurate segmentation of retinal arteries and veins is essential for retinal vascular analysis and systemic cardiovascular disease risk assessment (e.g., stroke, myocardial infarction). To address the challenge of fine-grained arteriovenous differentiation, we propose Attention-WNet—a W-Net variant enhanced with a channel-spatial joint attention mechanism that adaptively strengthens vessel structural feature representation. Evaluated on standard benchmarks (DRIVE, HRF), Attention-WNet achieves state-of-the-art performance, improving Dice score by +1.2% and AUC by +0.8% over prior methods, while demonstrating superior cross-dataset generalizability. The model delivers both high segmentation accuracy and intrinsic interpretability through attention visualization. This work provides a clinically viable, automated tool for arteriovenous segmentation in fundus images, enabling early detection of retinal and systemic microvascular pathologies.

Existing retinal artery-vein classification methods treat artery segmentation, vein segmentation, and overall vessel segmentation as three independent binary tasks, neglecting their anatomical coupling and leading to inconsistent predictions. To address this, we propose a collaborative learning framework: (1) a channel-wise vascular consistency loss that enforces feature-level alignment of the three segmentation outputs across channel dimensions; and (2) an intra-image pixel-level contrastive loss to enhance fine-grained discriminability. The method is built upon a standard CNN architecture, requiring no additional annotations or pretraining. Evaluated on three benchmark datasets—RITE, LES-AV, and HRF—it consistently outperforms state-of-the-art approaches, achieving significant improvements in classification accuracy and structural consistency. Our framework establishes an anatomically aware, interpretable, and robust paradigm for retinal image analysis.

Retinal vessel segmentation in optical coherence tomography angiography (OCTA) relies heavily on labor-intensive manual annotations, resulting in high costs and poor generalizability. Method: We propose BioVessel-Net, the first unsupervised generative framework for OCTA vessel segmentation. It explicitly encodes biostatistical vascular priors—including radius distribution and topological connectivity—via a vessel-radius-guided strategy and biostatistical consistency adversarial optimization, directly modeling anatomical structure rather than pixel-wise mapping. Additionally, we introduce RetinaMix, a high-resolution 2D/3D OCTA benchmark dataset. Contribution/Results: On RetinaMix and established benchmarks, BioVessel-Net achieves near-supervised segmentation accuracy (mean Dice ≥ 0.92), significantly outperforming state-of-the-art unsupervised, semi-supervised, and weakly supervised methods. Notably, it operates efficiently without GPU acceleration and offers strong interpretability through biologically grounded constraints.

This study addresses the limitations of manual, inefficient, and error-prone retinal vessel segmentation in multimodal fundus images by proposing the ELEMENT framework. This approach innovatively integrates vascular connectivity information into pixel-level machine learning classification, synergistically combining region-growing algorithms with multimodal image fusion to achieve highly accurate and efficient vessel segmentation. Evaluated on benchmark datasets including DRIVE, STARE, and CHASE-DB1, the method attains a peak accuracy of 98.35%, significantly outperforming 25 existing techniques—including six deep learning models—while simultaneously enhancing segmentation consistency and substantially accelerating the processing pipeline.

Retinal vessel segmentation and artery/vein (A/V) pixel-wise classification are critical for automated ocular disease screening, yet existing methods suffer from topological inconsistencies and A/V misclassification—especially at vessel crossings and fine branches. To address these challenges, we propose RR-Net, a recursive refinement network featuring a novel multi-stage recursive feature refinement mechanism and an A/V semantic consistency constraint. RR-Net employs a dual-branch decoupled architecture built upon an enhanced U-Net backbone, integrating multi-scale residual gated features. We further design a composite loss function combining Dice, Focal, and KL divergence terms to jointly optimize segmentation and classification. Evaluated on DRIVE, CHASE_DB1, and STARE, RR-Net achieves F1-scores of 98.2%, 97.6%, and 97.9%, respectively—surpassing state-of-the-art methods by over 2.1%—and improves A/V classification accuracy by 4.3%, significantly mitigating crossing misclassifications and branch omissions.

This study addresses the scarcity of specialists and diagnostic subjectivity in screening for Plus disease associated with retinopathy of prematurity (ROP) in low-resource settings by systematically evaluating the complementary roles of image classification and retinal vessel segmentation in a Kenyan preterm infant cohort. We developed and compared multiple AI pipelines—including RGB-based classifiers, multiple instance learning, multitask joint models, and a segmentation-followed-by-classification pipeline—using patient-stratified nested cross-validation. The proposed probabilistic ensemble model achieved the best balanced performance at the eye level, with a sensitivity of 0.692, specificity of 0.914, and balanced accuracy of 0.803, significantly outperforming single-task classifiers. Additionally, vessel segmentation yielded a Dice coefficient of 0.533 and a high specificity of 0.979, effectively reducing unnecessary referrals.

This study addresses the challenge of discontinuous vessel responses in retinal vascular segmentation, where traditional filters such as Frangi’s often produce fragmented vessels, thereby compromising accurate extraction of vascular structures across multimodal medical images. To overcome this limitation, the authors propose LS-CF, an unsupervised post-processing method that, for the first time, integrates local sensitivity–aware connectivity constraints with a heuristic tolerance mechanism to evaluate and reconnect broken vessel segments at the pixel level—without requiring any training data, thus enabling generalization across diverse imaging modalities. Built upon Frangi responses, LS-CF combines local connectivity analysis with morphological operations to form a lightweight yet effective unsupervised filter. The method achieves state-of-the-art performance on multiple benchmark datasets—including OSIRIX, IOSTAR, DRIVE, STARE, and CHASE_DB—with particularly notable gains over existing unsupervised approaches on CHASE_DB.

This work addresses the challenge of deploying high-accuracy retinal vessel segmentation models on resource-constrained devices, where computational overhead remains a significant barrier. To this end, the authors propose LightVesselNet, a lightweight encoder-decoder network containing only 75K parameters. The architecture innovatively integrates channel and spatial attention mechanisms, multi-scale feature aggregation, sub-pixel upsampling, and edge-aware residual connections. Despite its sub-100K parameter count, LightVesselNet achieves Dice coefficients ranging from 0.8070 to 0.8649 and sensitivities between 0.8096 and 0.8640 across five benchmark datasets—DRIVE, STARE, CHASE_DB1, FIVES, and HRF—demonstrating performance comparable to much larger models. These results significantly enhance the feasibility of edge deployment while exhibiting strong cross-dataset generalization capabilities.

This work addresses the challenge of accurately segmenting low-contrast, small-scale peripheral retinal vessels, which existing methods struggle to detect due to their reliance on standard overlap-based losses. To overcome this limitation, we propose a hierarchical multi-scale attention network with four parallel branches that learns to fuse features across multiple resolutions through trainable weights. The model is jointly optimized using a composite loss function combining Dice, binary cross-entropy, and centerline Dice to balance region coverage and topological continuity. Furthermore, hard example mining is incorporated to enhance robustness. Evaluated on DRIVE, STARE, and CHASE_DB1 datasets, our method achieves an average Dice score of 88.72%, sensitivity of 90.78%, and AUC of 98.25%, with cross-dataset AUC consistently exceeding 95%. Notably, it demonstrates significantly improved recall for fine vessels compared to current state-of-the-art approaches.

Current methods lack scalable, standardized, and interpretable tools for the automatic extraction of retinal vascular biomarkers from color fundus photographs. To address this gap, this work proposes VascX, an open-source Python toolbox that constructs both directed and undirected vascular graphs from vessel segmentation masks, incorporates anatomical landmarks such as the macula and optic disc to enable spatial standardization, and introduces a region-aware mechanism to facilitate the automatic identification of non-computable biomarkers. The framework integrates vessel skeletonization, graph-based modeling, grid-based localization, and visualization techniques within a modular architecture to ensure interpretability and reproducibility. VascX efficiently supports the automated computation of diverse biomarkers—including vascular density, bifurcation angle, central retinal equivalent, tortuosity, and temporal angle—making it well-suited for large-scale clinical and epidemiological studies.