Probabilistic graphical modeling of speckle statistics in laser speckle contrast imaging for noninvasive and label-free retinal angiography.

Abstract

This paper introduces a noninvasive and label-free approach for retinal angiography using Laser speckle contrast imaging (LSCI). Retinal vessel structure is segmented using a Hidden Markov Random Field (HMRF) based model. Prior to that, k-means clustering is used to obtain initial parameter set and labels for HMRF. Final parameter set for HMRF is estimated using expectation-maximization (EM) algorithm and final labeling is achieved using maximum aposteriori (MAP) algorithm. Clique energy for HMRF is computed from eigenvalue analysis of structure tensor for each pixel. This helps to get connectivity in the direction of strongest tangents in its neighborhood, facilitating the tracking of fine vessels in retinal vascular network. Quantitative evaluation shows an average vessel segmentation accuracy of 96.41% in normal condition with substantial improvement in tracking capability of fine vessels. Changes in blood flow can be tracked and observed at segmented output; particularly applicable for the study of different pathological conditions.