DUF-Net: A Retinal Vessel Segmentation Method Integrating Global and Local Features with Skeleton Fitting Assistance

Assisted evaluation through retinal vessel segmentation facilitates the early prevention and diagnosis of retinal lesions. To address the scarcity of medical samples, current research commonly employs image patching techniques to augment the training dataset. However, the vascular features in fundus images exhibit complex distribution, patch-based methods frequently encounter the challenge of isolated patches lacking contextual information, consequently resulting in issues such as vessel discontinuity and loss. Additionally, there are a higher number of samples with strong contrast vessels compared to those with weak contrast vessels in retinal images. Moreover, within individual patches, there are more pixels of strong contrast vessels compared to weak contrast vessels, leading to lower segmentation accuracy for small vessels. Hence, this study introduces a patch-based deep neural network method for retinal vessel segmentation to address the issues. Firstly, a novel architecture, termed Double U-Net with a Feature Fusion Module (DUF-Net), is proposed. This network structure effectively supplements missing contextual information and improves the problem of vessel discontinuity. Furthermore, an algorithm is introduced to classify vascular patches based on their contrast levels. Subsequently, conventional data augmentation methods were employed to achieve a balance in the number of samples with strong and weak contrast vessels. Additionally, method with skeleton fitting assistance is introduced to improve the segmentation of vessels with weak contrast. Finally, the proposed method is evaluated across four publicly available datasets: DRIVE, CHASE_DB1, STARE, and HRF. The results demonstrate that the proposed method effectively ensures the continuity of segmented blood vessels while maintaining accuracy.