RVS Architectures: Performance and Efficiency Trade-offs in CNN, Attention, and Mamba Models

Abstract

Retinal Vessel Segmentation (RVS) is crucial for diagnosing eye diseases, requiring deep learning models to accurately segment complex vascular structures. While initial architectures like U-Net established the baseline, traditional CNNs struggle with capturing long-range dependencies necessary for full vessel connectivity. This Systematic Literature Review (SLR) critically analyzes the performance and efficiency trade-offs across conventional CNNs, Attention-based U-Net variants, and the novel Mamba/State Space Model (SSM) hybrids. Evaluation is based on segmentation quality (AUC, Sensitivity) and computational cost (Parameters, FLOPs). The findings confirm that advanced Attention-based models successfully boost Sensitivity (SE)—critical for fine vessels—over the baseline. Crucially, the emerging Mamba Hybrids (e.g., CRMA-UNet) offer superior efficiency, capturing global context with linear complexity rather than the quadratic cost of early Transformers. This shift, coupled with the necessity for lightweight designs (e.g., LDMRes-Net at 0.072M parameters), confirms that Mamba/SSM hybrids represent the optimal balance, making them the most promising future direction for efficient clinical deployment

Keywords

• Computer Science
• Data Science
• Retinal Vessel Segmentation
• Mamba
• U-Net
• Attention Mechanisms
• Computational Efficiency