An Adaptive Topology-enhanced Deep Learning Method Combined with Fast Label Extraction Scheme for Retinal Vessel Segmentation

Retinal vessel segmentation is vital for the eye disease diagnosis and treatment. Recently, the deep learning method has been commonly used to achieve vessel segmentation, but it usually requires manual labels, which is time-consuming and laborious. In this work, we propose a fast vessel label extraction scheme, which can detect labels accurately and automatically. We first construct vesselness maps using the vessel structure features detected by the optimally oriented flux (OOF) filter. Then, we extracted the main vessel structure by the threshold approach. Thirdly, we utilize the local maximum method to detect the vessel skeleton that contains the entire vessel structures, especially the thin vessels. Besides, the existing topological enhancement methods do not consider the differences between thick and thin vessels, which may lead to poor extraction of thin vessels. Hence, we propose an adaptive topology-enhanced loss function to increase the different weights of thick and thin vessels. The correct topology of vessels is guaranteed, especially for thin vessels. The segmentation performance of the method is analyzed through extensive experiments. The results show that the method extracts excellent vessel labels, and the model trained with the automatically extracted labels is comparable to the supervised learning method.