Retinal vascular segmentation network based on dual-scale morphological enhancement

Abstract

Retinal vessel segmentation is crucial for diagnosing ocular diseases, but current methods struggle with fine vessels and complex structures. This paper proposes a full-resolution network based on dual-scale morphological enhancement (TED-SCNet) to improve the accuracy and robustness of retinal blood vessel segmentation. Initially, a dual-scale shape enhancement module was designed and integrated into the segmentation network to achieve end-to-end contrast enhancement, highlighting fine vascular structures. Subsequently, a feature aggregation module based on dynamic snake-shaped convolutions is introduced, which adapts convolution paths to achieve the fusion of features at different levels. Following this, a deep semantic enhancement module is devised to maximize the enhancement of semantic information. Finally, an over-segmentation-guided classification network is constructed to reduce the misclassification of avascular pixels. This study employs a combination of binary cross-entropy and structural similarity loss functions to accurately generate binary segmentation results. In the DRIVE, CHASE_DB1, and STARE datasets, TED-SCNet achieves AUCs of 98.90%, 99.02%, and 99.16%, respectively, outperforming traditional methods by 0.66%, 0.21%, and 0.32%. The F1 scores for the respective categories are 84.3%, 80.49%, and 84.45%, revealing a considerable progression in the fine vessel segmentation process. The combined loss function enhances the Dice similarity coefficient of thin tubular structures by a margin of 1.23%. In essence, TED-SCNet demonstrates high precision and robustness in retinal vessel segmentation, particularly excelling in fine vessel and complex structure segmentation while maintaining thin tubular structure connectivity.