Efficient Retinal Vessel Segmentation with 78K Parameters

Zhigao Zeng, Jiakai Liu, Xianming Huang, Kaixi Luo, Xinpan Yuan, Yanhui ZhuJ. Imaging 2025 · DOI

TL;DR

This paper introduces DSAE-Net, a novel, highly efficient deep learning model for segmenting retinal blood vessels. By employing a unique dual-stage architecture and innovative loss functions, DSAE-Net achieves state-of-the-art accuracy with a dramatically reduced parameter count, making it ideal for real-time diagnostics in resource-constrained clinical settings.

What Problem Does This Paper Address?

Accurate segmentation of retinal blood vessels is crucial for diagnosing conditions like diabetic retinopathy and glaucoma. However, existing deep learning models often face a trade-off: achieving high accuracy typically requires complex, computationally expensive architectures, which are impractical for deployment in many clinical settings. Conversely, lightweight models often sacrifice accuracy, especially for the fine details of capillaries. This paper aims to bridge this gap by developing a model that is both highly accurate and remarkably efficient.

Methods

The core of DSAE-Net is its Dual-Stage Attention-Enhanced Network (DSAE-Net) architecture. It's built upon a cascaded W-shaped topology that uses two modified U-Net modules (CAE-UNet) for progressive feature refinement.

Key innovations include:

  • Parameterized Cascaded W-shaped Architecture: This design allows for progressive feature refinement with significantly fewer parameters than a standard U-Net. The network depth and kernel sizes are tunable, enabling a balance between performance and computational cost.
  • Skeleton Distance Loss (SDL): To tackle the severe class imbalance in retinal images (where vessel pixels are much rarer than background pixels), a novel loss function is introduced. SDL leverages vessel skeletons to guide the learning process, overcoming limitations of traditional boundary losses that can generate detrimental negative weights.
  • Cross-modal Fusion Attention (CMFA) Module: This lightweight module enhances feature representation by combining grouped convolutions with dynamic weighting. It effectively expands the receptive field while preserving spatial precision.
  • Coordinate Attention Gates (CAGs): These plug-and-play modules are integrated into the decoder to refine skip connections. They use a coordinate attention mechanism to decouple directional pooling, capturing long-range dependencies and improving feature fusion.

Datasets

The proposed DSAE-Net was evaluated on four widely used public retinal vessel segmentation datasets:

  • DRIVE: A foundational dataset for general segmentation performance.
  • CHASE_DB1: Images from children, used to assess demographic robustness.
  • HRF (High-Resolution Fundus): High-resolution images from healthy and diseased subjects, testing performance on fine details.
  • STARE: Images with diverse pathologies, used to evaluate robustness.

Key Insights and Ideas

  • Efficiency is Not the Enemy of Accuracy: The paper demonstrates that it's possible to achieve state-of-the-art segmentation accuracy with a model that has a fraction of the parameters of existing methods. This is a significant step towards practical deployment.
  • Skeletonization for Loss Functions: Using vessel skeletons within the loss function is a clever way to address the class imbalance problem, providing a more robust signal for learning vessel boundaries.
  • Cascaded Refinement Works: The dual-stage architecture, where one stage refines the output of another, proves effective for progressively improving segmentation quality.
  • Attention Mechanisms for Lightweight Models: Integrating attention mechanisms like CMFA and CAG, even in a lightweight design, significantly boosts performance by improving feature representation and fusion.

Key Results

DSAE-Net achieved remarkable results:

  • Parameter Efficiency: It uses only 78K parameters, which is a 99% reduction compared to the average of the compared models and only 0.5% of a standard U-Net.
  • High Accuracy: DSAE-Net achieved state-of-the-art or highly competitive performance across all benchmark datasets. For example, on the DRIVE dataset, it achieved an Acc of 95.48%, AUC of 98.00%, and Dice of 82.44%, surpassing even much larger models.
  • Robustness: It demonstrated consistent performance across different datasets and age groups.
  • Speed: While not the absolute fastest, it offers a competitive inference time, especially considering its parameter count.

Limitations and Future Work

The authors acknowledge a couple of limitations:

  • Occasional Oversegmentation: SDL's spatial weighting strategy can sometimes lead to oversegmentation of complex vascular branches, potentially causing false positives in clinical screening.
  • Micro-vessel Underdetection in High-Resolution Images: Similar to other lightweight models, DSAE-Net showed some underdetection of subtle vessels in high-resolution HRF images, possibly due to resolution discrepancies between training and testing.

Future work will focus on:

  • Adaptive Kernel Selection and SDL Refinement: Optimizing SDL's terminal vessel weighting to improve micro-vessel delineation without sacrificing efficiency.
  • Multi-scale Augmentation: To address the resolution sensitivity.
  • Comparison with Vision Transformers: Further benchmarking against emerging lightweight vision transformer architectures.

Summary updated

BibTeX
@article{zeng2025efficient,
  author = {Zhigao Zeng and Jiakai Liu and Xianming Huang and Kaixi Luo and Xinpan Yuan and Yanhui Zhu},
  title = {{Efficient Retinal Vessel Segmentation with 78K Parameters}},
  journal = {J. Imaging},
  year = {2025},
  doi = {10.3390/jimaging11090306},
}
Figure