Retinal vessel diameter measurement using multi-step regression method

Study of retinal blood vessel features, especially the caliber, has been widely used for risk level assessment of certain disease conditions. In this work a vessel diameter measurement technique based on Gaussian modeling has been proposed. This method adaptively combines a series of second and higher order Gaussians to model the vessel profile and uses the sigma parameter of the best fitted Generalized Gaussians to the boundaries, to measure the vessel caliber. One advantage of this technique is that, unlike other methods, it does not assume the vessel profile as a symmetrical Gaussian-like shape with similar blurriness levels at the edges. The technique has been tested on 580 cross-sections from normal and pathological public datasets with mixed quality fundus images. Comparison was made against two gold standard techniques and a ground truth obtained manually by three experts to sub-pixel accuracy.