Directional gradient based registration of retinal images

Abstract

Registration of retinal images provided by different modalities is required to facilitate diagnosis of the retina specifically in context with Diabetes Retinopathy (DR). Temporal registration is necessary in order to follow the various stages DR, whereas multimodal registration allows us to improve the identification of some lesions or to compare pieces of information gathered from different sources. This paper presents an algorithm for temporal and/or multimodal registration of retinal images based on directional gradient of salient vessel bifurcation points. Gradient along 0, +45 and 90 degrees around bifurcation point is invariant to rotation. We have validated this by using Fourier transform. This enables to estimate the correspondence of bifurcation points between the image pairs to be aligned. The retinal vessel tree is first segmented and vessel bifurcation points are located. A feature vector comprising of directional gradients along 0, +45 and 90 degrees around labeled bifurcation point in its 3 × 3 neighborhood is then computed in order to determine correspondence. The rotation is estimated from the coordinates of the matched vessel bifurcation points. The algorithm is tested on publically available DRIVE and STARE database. The overall registration error is observed to be 0.8% and 0.7% for STARE and DRIVE database respectively.