Purchase this article with an account.
Y. Li, G. Gregori; Registration Error Analysis of the Ridge-Based Retinal Image Registration Algorithm for Oct Fundus Images and Color Fundus Photographs. Invest. Ophthalmol. Vis. Sci. 2010;51(13):3862.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To measure registration error of our previously developed ridge-based retinal image registration algorithm for registration between OCT fundus images (OFIs) and color fundus photographs (CFPs). The second purpose is to investigate that one possible reason to affect registration accuracy is hardly discernable distortion to the OFI caused by eye movements.
The used retinal image registration algorithm was previously developed by us, where detected blood vessel ridges serve as registration features. The image pair for registration consists of one CFP (50 degrees field of view) and one OFI (20x20 degrees field of view; taken by Carl Zeiss Meditec Cirrus HD-OCT Instrument). We tested the registration algorithm with one normal macular image dataset (8 image pairs) and one abnormal macular image dataset with a wide range of diseases (11 image pairs). Three transformation models (similarity, affine and quadratic models) are tested respectively. Registration errors are measured by the root mean square error (RMSE) of manually labeled control points on the registered image pair. The following method is to investigate that hardly discernable distortion to the OFI caused by eye movements may be another possible reason for registration error. Two groups of data are used: image pairs each consisting of one CFP and one OFI; image pairs each consisting of one CFP and one cropped CFP with a similar image area to the OFI. Similar blood vessel ridges are manually labeled for these images as registration features, and then image pairs are registered. Registration accuracies are compared between the two groups.
Experimental results showed that for the normal dataset, the quadratic model gave the best registration performance, with the average of root mean square error at 0.09 degree of field of view (about 27 microns). For the abnormal dataset with a wide range of diseases, the similarity model had lowest average of root mean square error at 0.24 degree of field of view (about 72 microns). For registration results based on manually labeled blood vessel ridges, we found that registration errors for the data between one CFP and one OFI are larger than that for the data between one CFP and one cropped CFP.
This work presents registration error measurement of our previously developed retinal image registration algorithm for a normal dataset and an abnormal dataset. We also claim that hardly discernable distortion to the OFI caused by eye movements is one possible reason to affect registration accuracy. These findings can help analysis of subsequent researches based on registration.
This PDF is available to Subscribers Only