Multiscale Recursive Graph Transformer for Retinal Vessel Segmentation
Javed, Syed Meijering, Erik Sowmya, Arcot Razzak, Imran
Javed, Syed
Meijering, Erik
Sowmya, Arcot
Razzak, Imran
Supervisor
Department
Computational Biology
Embargo End Date
Type
Conference proceeding
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Deep learning-based retinal vessel segmentation models often struggle to capture both local and global dependencies without explicitly modeling vessel connectivity. To address this challenge, we propose MRGT-Net, a novel Multi-Representation Guided Transformer network that unifies Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), and Transformers into a single framework. MRGT-Net leverages the strengths of multiscale CNNs for local feature extraction, GNNs for relational reasoning, and Transformers for capturing long-range dependencies and global context. The architecture adopts a U-Net-inspired encoder-decoder design. The encoder uses CNN blocks to extract hierarchical spatial features. To model contextual relationships between distant yet semantically similar regions, we introduce a Graph Convolution Block (GCB) that constructs a
-Nearest Neighbor (kNN) graph from high-level features. Furthermore, a Transformer Block (TB) enhances global context modeling in the feature space. We apply gamma correction and CLAHE (Contrast Limited Adaptive Histogram Equalization) during preprocessing to boost contrast and improve segmentation performance. Fused feature representations are decoded into precise segmentation maps. Extensive experiments on benchmark retinal vessel datasets demonstrate that MRGT-Net outperforms state-of-the-art methods, achieving F1 scores of 82.56% on CHASE-DB1 and 83.79% on DRIVE, surpassing prior methods by 0.93% and 0.58%, respectively, and achieving competitive F1 performance on STARE, with the highest accuracy across all three datasets.
Citation
S. Javed, E. Meijering, A. Sowmya, and I. Razzak, “Multiscale Recursive Graph Transformer for Retinal Vessel Segmentation,” Communications in Computer and Information Science, vol. 2757, pp. 176–187, 2026, doi: 10.1007/978-981-95-4100-3_13
Source
Neural Information Processing
Conference
32nd International Conference, ICONIP 2025
Keywords
Retinal vessel segmentation, Graph neural networks, Transformers, Contextual understanding, Medical image analysis
Subjects
Source
32nd International Conference, ICONIP 2025
Publisher
Springer Nature