Multi-scale retinal vessel segmentation using line tracking

In this paper an algorithm for vessel segmentation and network extraction in retinal images is proposed. A new multi-scale line-tracking procedure is starting from a small group of pixels, derived from a brightness selection rule, and terminates when a cross-sectional profile condition becomes invalid. The multi-scale image map is derived after combining the individual image maps along scales, containing the pixels confidence to belong in a vessel. The initial vessel network is derived after map quantization of the multi-scale confidence matrix. Median filtering is applied in the initial vessel network, restoring disconnected vessel lines and eliminating noisy lines. Finally, post-processing removes erroneous areas using directional attributes of vessels and morphological reconstruction. The experimental evaluation in the publicly available DRIVE database shows accurate extraction of vessels network. The average accuracy of 0.929 with 0.747 sensitivity and 0.955 specificity is very close to the manual segmentation rates obtained by the second observer. The proposed algorithm is compared also with widely used supervised and unsupervised methods and evaluated in noisy conditions, giving higher average sensitivity rate in the same range of specificity and accuracy, and showing robustness in the presence of additive Salt&Pepper or Gaussian white noise.