Automatic Registration of Retinal Images Using Genetic Algorithms

Abstract

 

Purpose:

 

This research contributes to the development of a system that automatically registers multitemporal digital images of the human retina. This is the first step in the detection and analysis of structural changes, associated with retinopathy or systemic diseases.

 

Methods:

 

The proposed approach is based on the application of global optimization techniques to previously extracted binary maps of curvilinear structures (i.e., the retinal vessels) in images to be registered. The optimum transformation between the input and the reference image is estimated by using a genetic algorithm that represents a global optimization method that simulated evolutionary processes in natural systems. Specifically, an affine transformation model is chosen and its parameters are optimized by the genetic algorithm. Images were acquired from Icelandic patients attending a retinopathy screening service, by a Canon CR6-45NM fundus camera.

 

Results:

 

The algorithm was tested on eight pairs of grayscale images and four pairs of color images. For both cases, images were taken in the same visit or on different dates. The figure shows results for a pair of color images registered by a manual method (a) and by the proposed approach (b), by using a checkerboard representation, that reveals the registered image by uncovering squares placed like a checkerboard over the reference image. The registration accuracy can be checked by looking at the continuity of the image structures (i.e., the vessels) on the checker borders. Convergence to a solution was not possible only in one case, when dealing with images taken from very different view-points.  

 

Conclusions:

 

Comparison of results obtained with the genetic algorithm and with the manual method demonstrates that accurate registration is achieved with the proposed method. The genetic algorithm can be successfully used to register multitemporal images, taken of the same patient, before performing a change detection analysis for diagnosis of retinal diseases.

 

Keywords: imaging/image analysis: non-clinical • retina • image processing