Ph.D. Theses

Feature-Based Retinal Image Registration: Algorithms and Validation

By Chia-Ling (Charlene) Tsai
Advisor: Charles V. Stewart
July 15, 2003

Many medical diagnostic tools rely on accurate registration algorithms. In ophthalmology, registering a pair of retinal images is often performed for analyzing, diagnosing and treating a number of diseases of the human retina. Areas of application include change detection from images taken at different times, montaging a sequence of images taken of a single retina, multimodal registration for information integration, and real-time image localization on a pre-computed reference map. There are several challenging issues involved in the development of a fully-automatic, retinal image registration algorithm. The focus of the thesis is on the design of feature-based registration algorithms to handle retinal images of diseased eyes, image pairs with low overlap, longitudinal images with physical changes, and real-time registration at the video frame rate.

Three major feature-based algorithms are introduced in this thesis: (1) a model-based landmark refinement algorithm for more accurate and repeatable registration features, (2) an off-line registration algorithm which can handle image pairs with as few as one common landmark, and (3) an incremental registration algorithm for real-time, frame-to-frame image tracking. The off-line registration algorithm is called Dual-Bootstrap Iterative Closest Point (DB-ICP). It initiates a local transformation from a single landmark or landmark pair correspondence, and gradually refines the transformation, extending from a low-order to a high-order transformation, and from a small image region to the entire region of overlap between the images based on the uncertainty of the transformation parameters. A prototype system for automatic vessel position change detection is developed as an application of DB-ICP. The incremental registration algorithm is called Iterative Multiple Closest Features (IMCF), which is a generalization of ICP involving multiple matches and a novel match-weighting scheme. Both registration algorithms are accurate to less than a pixel.

The model-based landmark refinement algorithm achieved repeatability of 1.05 pixels for the same landmark center location in different images. On tests involving approximately 6000 image pairs, DB-ICP successfully registered 99.5% of the pairs containing at least one common landmark, and 100% of the pairs containing at least one common landmark and at least 35% image overlap. In the disease-oriented, large-scale clinical validation, DB-ICP successfully aligned 99.5% of the image pairs having at least one common landmark, and 78.5% overall. Failures are mostly due to edema and scarring after laser eye surgery. IMCF was evaluated in terms of the domain of convergence. Comparing to ICP, the improvement is at least 1.85 times.

Return to main PhD Theses page