Dilation-Supervised Learning: A Novel Strategy for Scale Difference in Retinal Vessel Segmentation

Retinal fundus image segmentation based on deep learning is an important method for auxiliary diagnosis of ophthalmic diseases. Due to the scale difference of the blood vessels and the imbalance between foreground and background pixels in the fundus image, the deep learning network will inevitably ignore thin vessels when downsampling and feature learning. For the scale difference problem, this article aims to tackle its limitation from two perspectives: changing the supervised approach and adapting the feature learning. Correspondingly, a dilation-supervised learning method and an adaptive scale dimensional attention mechanism which are used to construct a two-stage segmentation model is proposed. Moreover, we introduce a quantitative approach to evaluate the scale difference of the blood vessels. With the help of the proposed weighted loss function, the segmentation results are refined, and the class imbalance problem between foreground and background pixels is resolved. Finally, the proposed adaptive threshold selection method is used in the postprocessing of segmentation results. The experiments on DRIVE, STARE, CHASE_DB1, and HRF datasets show that the proposed method achieves better segmentation performance compared with other state-of-the-art methods, and has good generalization ability and robustness.