Retinal Vascular Segmentation Network Based on Multi-Scale Adaptive Feature Fusion and Dual-Path Upsampling

Abstract

Retinal diseases impair the normal function of the visual system, making accurate segmentation of retinal vessels crucial. This paper proposes an improved U-Net network, namely Mitigating Information Loss U-Net (MILU-Net), for retinal vessel segmentation. The network introduces the Multi-Scale Adaptive Detail Feature Fusion (MSADFF) module, ensuring effective fusion of features at different scales before skip connections to reduce information loss. Simultaneously, the Dual Path Upsampling (DPUS) module is employed to enhance image resolution and compensate for spatial and channel information. Experiments on the DRIVE/STARE datasets demonstrate that MILU-Net outperforms in accuracy, sensitivity, specificity, AUC, and F1-Score metrics. Compared to the original U-Net, MILU-Net shows improvements of 1.44% and 1.84% in AUC, as well as 7.15% and 6.35% in sensitivity. Compared to the advanced Attention U-Net, MILU-Net achieves increases of 1.20% and 0.45% in ACC, as well as 2.63% and 2.71% in F1-Score, respectively. These results indicate the significant advantages of MILU-Net in retinal vessel segmentation tasks.

Keywords

• Retinal blood vessels
• segmentation
• MILU-Net
• MSADFF
• DPUS