Hierarchical Multi-Scale Mamba with Tubular Structure-Aware Convolution for Retinal Vessel Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Retinal vessel segmentation is vital for diagnosing retinal and cardiovascular diseases.
• Challenges include illumination variations and noise in fundus images, impacting vessel integrity.
• Accurate segmentation is foundational for computer-aided diagnostic systems.

Purpose of the Study:

• To introduce HM-Mamba, a novel hierarchical multi-scale Mamba-based architecture for improved retinal vessel segmentation.
• To enhance the extraction of local and global vascular features by incorporating tubular structure-aware convolution.
• To improve robustness and accuracy in segmenting complex vascular structures.

Main Methods:

• Developed a tubular structure-aware convolution to preserve vessel continuity and integrity.
• Designed a multi-scale fusion module to aggregate features across varying receptive fields.
• Integrated multi-branch Fourier transform with Mamba for long-range dependency and multi-frequency information capture.
• Proposed a hierarchical multi-scale interactive Mamba block for effective multi-scale semantic fusion.

Main Results:

• HM-Mamba demonstrated superior performance on five benchmark datasets (DRIVE, CHASE_DB1, STARE, IOSTAR, LES-AV).
• Achieved high Dice coefficients: 0.8327 (DRIVE), 0.8197 (CHASE_DB1), 0.8239 (STARE), 0.8307 (IOSTAR), and 0.8426 (LES-AV).
• The architecture effectively handles complex spatial patterns and reduces detail loss during downsampling.

Conclusions:

• HM-Mamba offers a robust and effective solution for retinal vessel segmentation.
• The proposed methods significantly improve the representation of vascular structures, including fine branches.
• This advancement holds promise for enhancing the accuracy and reliability of computer-aided diagnostic systems for eye and cardiovascular diseases.