PMFuse-Net: Enhancing Retinal Vessel Segmentation with Partial Multi-scale Convolutions and Attention Fusion

Retinal vessel segmentation is crucial for the early diagnosis of ocular and systemic diseases, yet it remains a challenging task due to the complex morphological features of vessels and varying imaging conditions. Existing deep learning methods, such as U-Net and its variants, face issues related to feature redundancy and inadequate multi-scale representation. To address these challenges, we propose PMFuse-Net, a novel network that incorporates two key modules: Partial Multi- scale Convolution (PMConv) and Multi-scale Attention Fusion (MSAF). PMConv reduces feature redundancy by selectively applying multi-scale convolutions to a subset of feature channels, while MSAF adaptively fuses multi-scale features using spatial attention mechanisms. Our extensive experiments on the DRIVE dataset demonstrate that PMFuse-Net outperforms state-of-the- art methods in terms of segmentation accuracy, sensitivity, and topological consistency. The results highlight the effectiveness of our approach in improving the segmentation of fine vascular structures and complex branching patterns.