Advanced retinal imaging: Feature extraction, 2-D registration, and 3-D reconstruction

Abstract

In this dissertation, we have studied feature extraction and multiple view geometry in the context of retinal imaging. Specifically, this research involves three components, i.e., feature extraction, 2-D registration, and 3-D reconstruction. First, the problem of feature extraction is investigated. Features are significantly important in motion estimation techniques because they are the input to the algorithms. We have proposed a feature extraction algorithm for retinal images. Bifurcations/crossovers are used as features. A modified local entropy thresholding algorithm based on a new definition of co-occurrence matrix is proposed. Then, we consider 2-D retinal image registration which is the problem of the transformation of 2-D/2-D. Both linear and nonlinear models are incorporated to account for motions and distortions. A hybrid registration method has been introduced in order to take advantages of both feature-based and area-based methods have offered along with relevant decision-making criteria. Area-based binary mutual information is proposed or translation estimation. A feature-based hierarchical registration technique, which involves the affine and quadratic transformations, is developed. After that, a 3-D retinal surface reconstruction issue has been addressed. To generate a 3-D scene from 2-D images, a camera projection or transformations of 3-D/2-D techniques have been investigated. We choose an affine camera to characterize for 3-D retinal reconstruction. We introduce a constrained optimization procedure which incorporates a geometrically penalty function and lens distortion into the cost function. The procedure optimizes all of the parameters, camera's parameters, 3-D points, the physical shape of human retina, and lens distortion, simultaneously. Then, a point-based spherical fitting method is introduced. The proposed retinal imaging techniques will pave the path to a comprehensive visual 3-D retinal model for many medical applications.