Zomba/FSG-Net

Full-scale Representation Guided Network for Retinal Vessel Segmentation

This repository contains the Full-Scale Guided Network (FSG-Net), a novel approach for retinal vessel segmentation. FSG-Net introduces a feature representation module that effectively captures full-scale structural information using modernized convolution blocks. A guided convolution block then refines this information, leveraging an attention-guided filter similar to unsharp masking to enhance fine vascular structures. This architecture delivers competitive performance for retinal vessel segmentation across multiple public datasets.

The model was presented in the paper Full-scale Representation Guided Network for Retinal Vessel Segmentation (arXiv:2501.18921).

Code: https://github.com/ZombaSY/FSG-Net-pytorch

Experimental Results

The model achieves competitive performance across several public datasets:

Dataset	mIoU	F1 score	Acc	AUC	Sen	MCC
DRIVE	84.068	83.229	97.042	98.235	84.207	81.731
STARE	86.118	85.100	97.746	98.967	86.608	83.958
CHASE_DB1	82.680	81.019	97.515	99.378	85.995	79.889
HRF	83.088	81.567	97.106	98.744	83.616	80.121

Usage

For detailed instructions on environment setup, dataset preparation, training, and inference, please refer to the official GitHub repository. Pre-trained models for each dataset can also be found on the GitHub releases page.

Citation

If you find this work helpful or inspiring, please consider citing our paper:

@article{seo2025fullscalerepresentationguidednetwork,
  title   = {Full-scale Representation Guided Network for Retinal Vessel Segmentation},
  author  = {Sunyong Seo, Huisu Yoon, Semin Kim, Jongha Lee},
  journal = {arXiv preprint arXiv:2501.18921},
  year    = {2025}
}