Research on Retinal Vessel Segmentation Algorithm Based on Multiple Attention Mechanisms

Abstract

Retinal vessel segmentation a critical diagnostic procedure for the early screening of ophthalmic and systemic diseases is undergoing a profound transformation driven by the integration of advanced deep learning and adaptive attention mechanisms. Conventional segmentation approaches, constrained by localized receptive fields, uniform feature weighting, and susceptibility to pathological artifacts, are being superseded by attention-augmented architectures capable of dynamically recalibrating vascular responses, modeling long-range topological dependencies, suppressing lesion-induced noise, and preserving microcapillary continuity. This research examines the synergistic deployment of multiple attention modules across the segmentation pipeline from local channel-spatial refinement to global self-attention modeling evaluating both their representational advantages and computational trade-offs. Drawing on systematic experimental validation across the DRIVE and STARE datasets, the paper identifies optimal fusion strategies specifically a serial SE–Attention Gate pipeline and a bottleneck-deployed lightweight MHSA block and assesses their impact on segmentation fidelity and hardware efficiency. A comparative analysis against mainstream CNN and Transformer baselines is presented. The paper proposes a Structured Multi-Attention Fusion Framework and offers architectural guidelines for researchers and clinicians seeking to deploy efficient, interpretable models in resource-constrained environments. Findings confirm that the proposed architecture delivers substantial and measurable improvements in Dice coefficient, sensitivity, vascular connectivity, and inference efficiency, establishing a robust foundation for automated ophthalmic diagnostics.

Introduction

The text discusses the development of an advanced deep learning framework for retinal blood vessel segmentation using multi-attention mechanisms. Retinal vessel segmentation is essential for the early detection, diagnosis, and monitoring of eye and systemic diseases such as diabetic retinopathy, hypertensive retinopathy, and glaucoma. However, traditional manual segmentation is time-consuming, labour-intensive, and prone to human error and inconsistencies. Conventional computer vision methods also struggle with low image contrast, uneven illumination, variations in vessel size, and pathological abnormalities.

To overcome these challenges, the study proposes a Structured Multi-Attention U-Net Framework that integrates channel attention, spatial attention, and multi-head self-attention (MHSA). Channel and spatial attention improve the identification of important blood vessel features while suppressing irrelevant background information and pathological artifacts. Meanwhile, self-attention captures long-range relationships between blood vessels, helping preserve microcapillary continuity and overall vascular structure. The framework uses a ResNet-34 backbone and strategically places attention modules to balance segmentation accuracy and computational efficiency.

The literature review highlights the evolution of retinal vessel segmentation from traditional morphological filtering and machine learning to convolutional neural networks and attention-based architectures. Previous research demonstrates that U-Net models, attention mechanisms, and hybrid CNN-Transformer networks improve vessel detection and structural preservation. However, challenges such as high computational complexity, GPU memory limitations, feature misalignment, and increased inference time remain significant barriers to clinical deployment.

The study has seven main objectives, including investigating attention mechanisms, analysing their architectural contributions, comparing segmentation performance, conducting ablation studies, improving computational efficiency, enhancing model interpretability, and developing practical guidelines for clinical implementation.

The research adopts a quantitative experimental methodology using two publicly available retinal image datasets: DRIVE (40 images) and STARE (20 images). Image preprocessing includes green-channel extraction, Gaussian smoothing, and Contrast Limited Adaptive Histogram Equalization (CLAHE). Data augmentation techniques such as rotation, flipping, elastic deformation, and brightness and contrast adjustment are applied to improve model generalization. The network is trained using a combination of Binary Cross-Entropy and Dice loss, optimized through AdamW with cosine annealing learning-rate scheduling.

Ablation studies compare the baseline U-Net, serial channel-spatial attention, bottleneck MHSA, and the complete proposed framework. Performance is evaluated against established CNN and Transformer-based models using segmentation metrics, computational efficiency, inference latency, and memory requirements. Interpretability is assessed through attention maps, Grad-CAM visualizations, and segmentation overlays.

Finally, the proposed architecture aims to provide a lightweight, accurate, interpretable, and hardware-efficient solution for automated retinal vessel segmentation. By combining local feature refinement with global vascular context modelling, the research seeks to support reliable retinal screening, reduce clinicians' annotation workload, and enable deployment in resource-constrained healthcare environments.

Conclusion

This research has systematically investigated the integration of multiple attention mechanisms into retinal vessel segmentation, demonstrating that adaptive feature recalibration is not merely an incremental architectural modification but a fundamental paradigm shift in how vascular topology is extracted, preserved, and clinically validated. From the serial SE?Attention Gate pipeline that purifies skip-connection features and suppresses lesion-induced artifacts, to the bottleneck-deployed MHSA module that restores long-range capillary continuity without prohibitive computational overhead, the empirical evidence is unambiguous: strategically orchestrated attention mechanisms deliver substantial and measurable improvements in segmentation fidelity, microvascular connectivity, inference efficiency, and model interpretability. Validated across the DRIVE and STARE benchmarks, the proposed architecture consistently outperforms conventional CNN baselines and resource-intensive Transformer hybrids, achieving an optimal balance between diagnostic precision and hardware-aware deployability. The imperative for medical AI development is no longer to pursue accuracy at the expense of efficiency or transparency, but to engineer structurally coherent, clinically verifiable systems that seamlessly integrate into real-world diagnostic workflows. The Structured Multi-Attention Fusion Framework and the accompanying deployment roadmap proposed in this study provide a reproducible blueprint for researchers and engineers seeking to translate algorithmic innovation into scalable, edge-compatible ophthalmic screening tools. By enforcing hierarchical local refinement, restricting global self-attention to the semantic bottleneck, and embedding gradient-based interpretability, this work bridges the longstanding gap between representational capacity and clinical practicality. As the global burden of diabetic retinopathy and cardiovascular-related ocular pathologies continues to rise, automated vessel segmentation represents a critical enabler for high-throughput, equitable, and proactive eye care. While the current framework demonstrates robust performance on standardized fundus benchmarks, future work must address cross-domain generalization under severe pathological interference, extend the attention paradigm to multi-modal volumetric imaging, and explore self-supervised and federated learning strategies to reduce annotation dependency while preserving data privacy. Prospective multi-center clinical trials and biomarker-driven validation will further solidify the translational pathway from algorithmic prototype to regulated diagnostic aid.

References

[1] Z. Liu et al., \"Deep learning for retinal vessel segmentation: A systematic review of techniques and applications,\" Med. Biol. Eng. Comput., vol. 63, no. 8, pp. 2191–2208, 2025. [2] F. Neha et al., \"U-Net in medical image segmentation: A review of its applications across modalities,\" arXiv preprint arXiv:2412.02242, 2024. [3] J. Hu, L. Shen, and G. Sun, \"Squeeze-and-excitation networks,\" in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 7132–7141. [4] Z. Si, D. Fu, and J. Li, \"U-Net with attention mechanism for retinal vessel segmentation,\" in Proc. Int. Conf. Image Graph. (ICIG), 2019, pp. 321–332. [5] D. Wang et al., \"Hard attention net for automatic retinal vessel segmentation,\" IEEE J. Biomed. Health Informat., vol. 24, no. 12, pp. 3384–3396, 2020. [6] Z. Ma and X. Li, \"An improved supervised and attention mechanism-based U-Net algorithm for retinal vessel segmentation,\" Comput. Biol. Med., vol. 168, Art. no. 107770, 2024. [7] J. Li et al., \"DDAF-Net: A dual-direction attention fusion network for retinal vessel segmentation,\" Biomed. Signal Process. Control, vol. 88, Art. no. 105629, 2024. [8] M. Jiang, Y. Zhu, and X. Zhang, \"CoVi-Net: A hybrid convolutional and vision transformer neural network for retinal vessel segmentation,\" Comput. Biol. Med., vol. 170, Art. no. 108047, 2024. [9] L. Tong et al., \"LiViT-Net: A U-Net-like, lightweight Transformer network for retinal vessel segmentation,\" Comput. Struct. Biotechnol. J., vol. 24, pp. 213–224, 2024. [10] Y. Li et al., \"Global transformer and dual local attention network via deep-shallow hierarchical feature fusion for retinal vessel segmentation,\" IEEE Trans. Cybern., vol. 53, no. 9, pp. 5826–5839, 2022. [11] L. Jiang et al., \"Retinal vessel segmentation based on self-attention feature selection,\" Electronics, vol. 13, no. 17, Art. no. 3514, 2024. [12] X. Shen et al., \"Self-attentional microvessel segmentation via squeeze-excitation transformer Unet,\" Comput. Med. Imag. Graph., vol. 97, Art. no. 102055, 2022. [13] H. Huang et al., \"Channel prior convolutional attention for medical image segmentation,\" Comput. Biol. Med., vol. 178, Art. no. 108784, 2024. [14] S. Zhu et al., \"A multi-resolution hybrid CNN-Transformer network with scale-guided attention for medical image segmentation,\" IEEE J. Biomed. Health Informat., early access, 2025. [15] H. Tang et al., \"HTC-Net: A hybrid CNN-Transformer framework for medical image segmentation,\" Biomed. Signal Process. Control, vol. 88, Art. no. 105605, 2024.

Copyright

Copyright © 2026 Ngoma Mercilio, Zong Hui, A. Haq, Ngenda Eliot. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.