The Problem
More than two million individuals in the United States have been diagnosed with glaucoma - the damaging of cells and fibers that form the optic nerve, largely through the build up of intra-ocular pressure - and it is estimated that another two million cases remain undiagnosed. Current diagnostic methods are primarily based on measuring the loss of retinal function. Since most of this reduction in eyesight is irreversible, there is a strong need for new methods of detecting glaucoma before its impact on vision can be measured.

Recent evidence suggests that visual comparison of images of the retina, taken months or years apart, can indicate subtle changes around the optic nerve head that are early signs of glaucoma. These images are taken through the pupil of the eye in a painless manner using a fundus camera. One image, taken before any potential onset of glaucoma, serves as a baseline, and with a second being taken during each office visit. Side-by-side viewing of the two images is one way to ascertain changes. A more effective method is to precisely overlay the two images on top of each other and then "flicker" the images back-and-forth, allowing subtle changes to "pop out" of the images.

The Challenge
The challenge then is to develop software that can automatically and precisely align two images of the retina, doing this so reliably that any two images of the same retina can always be aligned. This level of automation will allow such a flickering of images to be used for glaucoma diagnosis by every eye-case specialist who has access to a fundus camera and a computer.

The technical challenges to developing this fundus image alignment (registration) algorithm and software are substantial. The retina is a curved surface, but the images are flat. The position of the eye may be quite different in two different images. The image quality may be low, sometimes due to a lack of light during the imaging and sometimes due to clouding of the vitreous through which the retina is viewed. Finally, due to the effects of disease or aging, there may be substantial changes in the appearance of the retina over the intervening period between images. Despite these, a successful software solution must be virtually foolproof.

The Solution
DualAlign has developed such a solution based on its Automated Image Registration algorithm. Several underlying ideas make this possible. The first one is "starting small". The algorithm automatically detects and matches small regions of the two images that may correspond to the same region of the eye. Often, but not always, these are regions where the blood vessels of the retina branch or cross over each other. This initial matching produces several hypotheses about the relationship between the two images.

Next, each of the hypotheses is tested individually and in succession. For each hypothesis, the DualAlign algorithm generates a function that maps pixels from one fundus image onto the other. Initially this function will only be accurate in describing the relationship between the matched regions. Hence, the function must be refined to make it accurately describe the relationship between all visible parts of the retina. The algorithm achieves this gradually by expanding the small initial region, matching up parts of the retina that appear in the expanded region, and refining the mapping function based on these matches. In doing so, the algorithm automatically ignores what is inconsistent between the two images, generally due to changes in the retina, and it automatically accounts for the curvature of the retina. On the other hand, it also quickly determines when the initial hypothesis is wrong because there will not be sufficient analogous structure as the region is expanded. In this case it will immediately move on to the next hypothesis and test it.

The final stage of the algorithm occurs when a hypothesis has led to a refined mapping function that covers all of the images. In this case, a final test is applied to be sure that the two images, now aligned by the computed mapping function, are sufficiently similar to conclude that they are of the same eye and that the resulting alignment is sufficiently precise. When a mapping function passes this test, the two images are ready for the flickering process that will be used by the eye-case specialist to diagnose glaucoma.