U-Net based Multi-level Texture Suppression for Vessel Segmentation in Low Contrast Regions

Segmentation of retinal blood vessels is important for diagnosis of many retinal diseases. Precise segmentation of complete vessel-map is still a challenge in low contrast regions of fundus images. Vessel pixels belonging to these regions, such as, fine vessel-endings and boundaries of vessels, get merged in the neighboring vessel-like texture. This paper proposes a novel retinal vessel segmentation algorithm which handles the background vessel-like texture in a sophisticated manner without harming the vessel pixels. In this work, first we enhance all possible vessel-like features of fundus at different `levels' using 2-D Gabor wavelet and Gaussian matched filtering. At each `level', texture is suppressed using Local Laplacian filter while preserving the vessel edges. The resulting images are combined to produce a maximum response image with enhanced vessels of different thicknesses and suppressed texture. This handcrafted image is used to train the deep U-net model for further suppression of non-vessel pixels. Proposed segmentation method is tested on publicly available DRIVE and STARE databases. The algorithm has produced state-of-the-art results. It has performed outstandingly well in terms of sensitivity measure which is most affected with the correct segmentation of fine vessels and vessel-boundary pixels present in low-contrast regions.