Measurement of retinal arteriolar diameters from auto scale phase congruency with fuzzy weighting and L1 regularization

Abstract

Manual measurements of small changes in retinal vascular diameter are slow and may be subject to considerable observer-related biases. Among the conventional automatic methods the sliding linear regression filter (SLRF) demonstrates the least scattered and most repeatable coefficients. For optimal performance it relies on the choice of the correct filter scale for different vessel sizes. A small scale extracts fine details at the expense noise sensitivity, while large scales have poor edge localization. Here we use auto scale phase congruency to select the filter scales with fuzzy weighting to reduce noise, and L1 regularization for edge smoothing. Our method uses a one dimensional analysis normal to the vessel and so is faster than the 2D phase congruency. In 65 vessels randomly selected from 20 images the proposed method showed better repeatability and over three times less scattering than conventional SLRF.