Wavelet based Fine-to-Coarse Retinal Blood Vessel Extraction using U-net Model

Abstract

Segmentation of retinal blood vessels is a crucial preliminary step in the diagnosis of any retinal disease. While extracting vessel-map, the biggest challenge is to segment fine vessels which are in poor contrast with the non-vessel background. Key contribution of this work is a fine-to-coarse retinal vessel extraction model with high sensitivity. Proposed algorithm uses a directional wavelet to generate a novel multiscale, three-channel image. To generate this image, only the real coefficients of wavelet transform are used, which facilitate the extraction of fine vessel-ends. Multiple scales cover different thicknesses of vessels. The vessel-enhanced image catalyzes the learning of deep U-net model for pixel classification. This work uses STARE and DRIVE databases for experimentation. Algorithm has performed robustly well in cross-database testing, even in pathological environment. Proposed method has produced state-of-the-art results. The vessel segmentation is outstanding in terms of sensitivity measure which validates better extraction of fine vessels. In this paper, an elaborate comparison with the other existing methods is also presented.