Abstract
Abstract: :
Purpose: Digital imaging of the neonatal retina provides new opportunities for studying retinovascular development and quantifying, sequentially, the features of retinopathy of prematurity. Automatic generation of a complete retinal montage would facilitate these studies. Methods: The system builds on work previously reported (Clay ST, et al. IOVS 2002;43:ARVO E–Abstract 4343). It detects blood vessels in images obtained with the RetCam 120 digital fundus camera (Massie Labs, Dublin, CA), using a scale space routine, and converts these into an abstract structure. A preliminary matching between two images is made using a clustering technique. The preliminary match is checked for accuracy by measuring how many landmarks in one image are near landmarks in the second (the Hausdorff fraction). Precise point correspondences are found using correlation. The images are then warped using thin–plate splines to provide accurate registration. Results: The vessel detection system has been tested on 847 images from 61 examinations of 22 eyes from 11 babies. It is very reliable if the image quality is good. The preliminary matching is successful in 50% of cases, but (N–1)(N–2)/2 of the matches between N images are redundant. The checking algorithm has a 91% true positive rate and a 16% false positive rate. Quantification of the accuracy of the final result is problematic, but the images appear well matched. Conclusions: Specific work is in progress to improve the reliability of the vessel detection system on poor quality images, and to provide feedback between the preliminary match and the match checker. Speed improvements are also under investigation. The results show an acceptable rate of success for image registration, and the montages produced show good agreement between the component images.
Keywords: imaging/image analysis: non–clinical