Dynamic Statistical Attention-based lightweight model for Retinal Vessel Segmentation: DyStA-RetNet

Background and objective: Accurate extraction of retinal vascular components is vital in diagnosing and treating retinal diseases. Achieving precise segmentation of retinal blood vessels is challenging due to their complex structure and overlapping vessels with other anatomical features. Existing deep neural networks often suffer from false positives at vessel branches or missing fragile vessel patterns. Also, deployment of the existing models in resource-constrained environments is challenging due to their computational complexity. An attention-based and computationally efficient architecture is proposed in this work to bridge this gap while enabling improved segmentation of retinal vascular structures.

Methods: The proposed dynamic statistical attention-based lightweight model for retinal vessel segmentation (DyStA-RetNet) employs a shallow CNN-based encoder-decoder architecture. One branch of the decoder utilizes a partial decoder connecting encoder layers with decoder layers to allow the transfer of high-level semantic information, whereas the other branch helps to incorporate low-level information. The multi-scale dynamic attention block empowers the network to accurately identify different-sized tree-shaped vessel patterns during the reconstruction phase in the decoder. The statistical spatial attention block improves the feature learning capability. By effectively integrating low-level and high-level semantic information, DYStA-RetNet significantly improves the performance of vessel segmentation.

Results: Experiments performed on four benchmark datasets (DRIVE, STARE, CHASEDB, and HRF) exhibit the adaptability of DYStA-RetNet for clinical applications with a significantly smaller number of trainable parameters (37.19K) and GFLOPS (0.75), and superior segmentation performance.

Conclusion: The proposed lightweight CNN-based DYStA-RetNet efficiently extracts complex retinal vascular components from fundus images. It is computationally efficient and deployable in resource-constrained environments.