Abstract
Purpose: :
Severe retinopathy of prematurity (ROP) is usually accompanied by dilation of posterior pole blood vessels and/or increase in vessel tortuosity over time. To examine vascular changes seen in eyes with ROP, we have developed a computerized tool that automatically aligns fundus images captured at different times and facilitates identification of change in width or tortuosity of blood vessels. The tool does not require knowledge of camera parameters (e.g. focal length, position).
Methods: :
We have developed a technique that, given two fundus images I1 and I2, performs 2D alignment by modeling the retinal surface outside the optic disc as a quadratic surface, not as a planar surface as commonly employed. We first use SIFT (scale-invariant feature transform) and Harris (corner-like) features in the fundus images to find numerous and dense matching points between the two images to fit a model to the retinal surface. The parameters of the model define a spatially varying image warp that, when applied to image I1, aligns I1to I2. This alignment accounts for changes in camera internal parameters like focal length, as well as the position and orientation of the camera with respect to the eye. Therefore, retinal structures that have not changed over time will be in the same position in the aligned images. Flickering between these aligned images readily identifies change in vessel width or tortuosity between images.
Results: :
Applying this alignment process to 9 datasets, we found that flickering between aligned images enabled observers to detect interval changes easily.. These changes may be related to ROP progression, ROP response to treatment or even intra-session changes such as pulse-related variability in blood vessel caliber.
Conclusions: :
We have developed a novel fundus image alignment procedure that facilitates flicker-based identification of changes in width and tortuosity of retinal blood vessels. Since our alignment process implicitly accounts for unknown camera parameters, it has the potential to increase the sensitivity and precision of clinical estimates of disease progression or therapeutic response in ROP.
Keywords: retinopathy of prematurity • image processing • infant vision