PA-Net: A hybrid architecture for retinal vessel segmentation

This paper proposes a hybrid architecture, PA-Net, which amalgamates the strengths of convolutional neural networks and the transformer model to enhance the precision of retinal vessel segmentation. We propose a novel component, the Lightweight Parallel Transformer (LPT), to augment the transformer's adaptability for the task of retinal vessel segmentation. This LPT addresses the shortcomings of standard transformer that are highly dependent on large datasets and computing resources, and can capture long-range dependencies to prevent slender vessels from breaking. Furthermore, we introduce an Adaptive Vascular Feature Fusion module to offset the vascular information loss induced by downsampling layers, thereby enhancing microvessel recognition. The effectiveness of PA-Net was assessed across four distinct datasets: DRIVE, CHASE_DB1, STARE, and HRF, with sensitivities of 0.8284, 0.8570, 0.8813, and 0.8497, respectively. The results suggest that the proposed method outperforms other state-of-the-art alternatives.