Retinal vessel detection via second derivative of local radon transform

Abstract

For the automatic detection of retinal blood vessels a preprocessing of the noisy original images is necessary. Retinal blood vessels are assumed to be line-like structures and can therefore be enhanced via convolution with suitable, elongated kernels. Consequently we use the local Radon kernel as a prototype of an elongated kernel for this task. The Radon kernel is rotated at different angles and adapts via a maximisation procedure to the directions of the vessels. The proposed algorithm is easy to implement and combined with edge- and coherence-enhancing anisotropic diffusion as a preprocessing step it offers higher robustness than the Laplacian of Gaussian or Haralick operator. Furthermore, our algorithm detects vessels as connected structures with very few interruptions. The performance is evaluated in experiments on the publicly available databases DRIVE and STARE as well as on selected examples of our clinical database. Since our algorithm does not depend on a priori directional and branching models, in its generality it is capable to detect other vessel structures in the human eye such as the conjunctiva vessels.