Retinal vessel segmentation using multiwavelet kernels and multiscale hierarchical decomposition

We propose a comprehensive method for segmenting the retinal vasculature in fundus camera images. Our method does not require preprocessing and training and can therefore be used directly on different images sets. We enhance the vessels using matched filtering with multiwavelet kernels (MFMK), separating vessels from clutter and bright, localized features. Noise removal and vessel localization are achieved by a multiscale hierarchical decomposition of the normalized enhanced image. We show a necessary condition to achieve the optimal decomposition and derive the associated value of the scale parameter controlling the amount of details captured. Finally, we obtain a binary map of the vasculature by locally adaptive thresholding, generating a threshold surface based on the vessel edge information extracted by the previous processes. We report experimental results on two public retinal data sets, DRIVE and STARE, demonstrating an excellent performance in comparison with retinal vessel segmentation methods reported recently.

► We identify multiwavelet kernels separating vessels from clutter edges (e.g., lesion).
► We perform an iterative segmentation to locate smaller and smaller vessels.
► We show a necessary condition to achieve the optimal number of the iterations.
► Our method does not require training, can thus be used on various images directly.