Abstract
Purpose: :
Multimodal and intermodal retinal image registration is important to clinical diagnosis and treatment. The purpose of this work is to develop an effective algorithm to register various types of fundus images, including OCT fundus images, fluorescein angiograms, fundus autofluorescence images, and color fundus photographs.
Methods: :
The two test images are first rescaled to the same scale, which can be estimated from common imaging techniques. Let I1 be the base image and I2 be the test image, and we will register I2 to I1. Our blood vessel ridge detection algorithm is used to obtain vessel ridge images, because compared to intensity vessels are stable features across modalities. The similarity between the two ridge images is defined as a variant of Hausdorff Distance. Brute force search on the translation parameter space is applied to look for an approximate registration location. Distance transform is used to reduce the computation time. After an approximate registration location is obtained, an Iterative Closest Point (ICP) algorithm is used to look for a more accurately registered affine transform. Human interaction is allowed to initialize registration, if necessary.
Results: :
We tested the algorithm on various types of fundus images, which included OCT fundus images, fluorescein angiograms, fundus autofluorescence images, and color fundus photographs. Experimental results show our algorithm perform well.
Conclusions: :
The developed algorithm is able to automatically or semi-automatically register various types of fundus images to provide a means for precisely correlating clinical pathology as imaged by different modalities.
Keywords: image processing • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound) • retina