Purchase this article with an account.
S. Lee, J. M. Reinhardt, M. Niemeijer, M. D. Abramoff; Comparing the Performance of Retinal Image Registration Algorithms Using the Centerline Error Measurement Metric. Invest. Ophthalmol. Vis. Sci. 2008;49(13):1833. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
to compare three frequently used coordinate transformation models for retinal image registration; affine, quadratic and radial distortion correction (RADIC) model. We compare the registration accuracy using a quantitative metric of vessel centerline error measurement (CEM), visually inspect the result, and finally suggest a transformation model that is optimal for retinal color images.
500 color fundus image pairs (1,000 images, 500 eyes) were randomly selected from a diabetic retinopathy screening program in the Netherlands. The rigid coordinate distortion in retinal image is described by the affine model. The quadratic model describes the spherical-to-planar projective distortion in retinal imaging. The RADIC model that we recently published is a combination of affine model following the radial distortion correction. The CEM is evaluated from the overlapping region, and the registered montage image is visually inspected to assess the geometric alignment in the non-overlapping region.
The test data are registered in a pair-wise manner to successfully generate 462 montage images and 38 pairs were not registered due to poor image quality or insufficient correspondence. Because of the nonlinear property, 145 pairs registered by the quadratic model show significant warping in the non-overlapping region. For the successfully registered pairs, the CEMs were 1.83±0.05 (affine), 1.61±0.07 (quadratic), and 1.72±0.04 (RADIC) pixels. Visual inspection indicated that even though the quadratic model has the lowest CEM, the RADIC model shows the least distortion in registering retinal images.
From these results, the RADIC model is found to have small error and the least distortion in retinal images registration. We are currently pursuing more objective techniques to compare the performance of retinal image registration algorithms.
This PDF is available to Subscribers Only