Photomontage of fundus images by the vessel bifurcation matching and the retinal curvature correction

Abstract

 

Purpose

 

Fundus photographs are commonly used for diagnosing and monitoring retinal diseases. However, a single retinal image is not enough to reveal central and peripheral retinal changes simultaneously. Photomontage with multiple fundus photographs is widely used for this purpose, but results are sometimes not satisfactory if there are not enough distinctive features and overlapping regions. In this presentation, a new retinal image registration method will be introduced for the better performance of the photomontage procedure.

 

Methods

 

We adopted vessel bifurcations as distinctive features in matching adjacent fundus images. To enhance the performance, directional information of bifurcations is stored in sixteen directional bins, and this information of the secondary neighborhood is compared. If the directional information doesn’t match, then the comparison will be stopped, and only the matching points on the same direction are compared. Excessive deformation caused by quadratic polynomial transformation is compensated by an estimated retinal curvature. We evaluated the performances of the proposed method by comparing it with several different methods for extracting distinctive features and key-point matching algorithms for retinal image registration. Harris corner detector, multi-scale Laplacian of Gaussian and vessel bifurcation are compared to the proposed feature extraction method, and descriptor matching and template matching methods are compared to the proposed matching algorithm.

 

Results

 

By using the proposed algorithms, enough number of matched key-point pairs can be found effectively, and shows 5% or more enhanced performances in terms of the matching point detection and the calculation speed than the other methods evaluated. Peripheral retinal image warping is also properly corrected by the retinal curvature compensation. Because this compensation is performed on the virtual 3D relationship setting, result image can be viewed according to the visual axis rotation.

 

Conclusions

 

With the aid of the developed program, better retinal image photomontage can be provided for the clinical evaluation.

 

 

Keywords: 549 image processing • 550 imaging/image analysis: clinical • 688 retina