Comparative Analysis of Pre-process Pipelines for Automatic Retinal Vessel Segmentation

Retinal vessel structure is an unique individual characteristic and important biology marker of many diseases, like Diabetic Retinopathy (DR), cardiovascular ailment, and so on. Automatic retinal vessel segmentation can be used to assist the early diagnosis of above diseases, but suffers from the poor quality and low contrast of fundus images. To eliminate the noise in the fundus images, many pre-process pipelines are designed to normalize and enhance the fundus images. However, the specific operations in pre-process pipelines of the fundus images haven't been distinguished from operations in normalization of natural images. This paper collects a dozen of pre-process pipelines from published retinal vessel segmentation researches, and proposes five general patterns of these pre-process pipelines. Furthermore, we test the flexibility of five classical pre-process pipelines on public retinal vessel datasets with a Dense-UNet model. Experiments demonstrate that the "Hiera" pre-process pipeline and the "DUNet" pre-process pipeline outperform the rest pipelines in assisting the Dense-UNet to segment the retinal vessels.