A high-resolution network with adaptive spatial channel fusion for retinal vessel segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Accurate retinal vessel segmentation is vital for diagnosing ophthalmic diseases.
• Challenges include scale variations causing fractured topology and low-contrast boundaries leading to segmentation ambiguity.

Purpose of the Study:

• To propose an Adaptive Spatial Channel Fusion High-Resolution Network (ASCF-HRNet) for improved retinal vessel segmentation.
• To address challenges of vascular topology preservation and segmentation ambiguity.

Main Methods:

• Introduced a Spatial Semantic Enhancement (SSE) block with multi-scale kernels and spatial attention to maintain vascular topology.
• Designed a Channel Feature Enhancement (CFE) block for semantics-aware refinement before upsampling to reduce noise and ambiguity.
• Integrated SSE and CFE blocks into a high-resolution network architecture.

Main Results:

• ASCF-HRNet achieved leading AUC scores of 0.9880 (DRIVE), 0.9899 (CHASE_DB1), and 0.9828 (STARE).
• Achieved highly competitive F1-scores of 0.8263 (DRIVE), 0.8119 (CHASE_DB1), and 0.7781 (STARE).
• Demonstrated superior performance in preserving vascular topology and ensuring boundary fidelity.

Conclusions:

• The proposed ASCF-HRNet effectively addresses key challenges in retinal vessel segmentation.
• The network achieves state-of-the-art performance, particularly in maintaining topological integrity and boundary accuracy.
• ASCF-HRNet offers a promising solution for automated ophthalmic disease diagnosis through enhanced retinal image analysis.