Abstract
Abstract: :
Purpose: Diagnosis of retinal diseases requires routinely fundus image comparison for change detection, analysis and correlation. Currently, an accurate alignment of temporally different digital fundus images is not possible without operator’s input. An algorithm for non-assisted image registration was developed and tested. Initial validation results suggest that this automatic registration method is most promising. Methods: A new algorithm derived from the concept of oriented energies was developed and the accuracy and the efficacy were tested using a graphical user interface. Thirty pairs of fundus photographs obtained at different time intervals, were randomly chosen from the Scheie Eye Institute film archives. Conventional fundus photography (Zeiss FF4, 30 degree, 35 mm film), fluorescein angiography, and indocyanine green angiography studies had been obtained as indicated in the course of conventional patient management. Film-based color, red-free and fluorescein angiography studies were scanned at 1000 dots per inch, 8 bits per pixel for monochromatic images and 24 bits per pixel for color images to an image size of 1100X945. The initial test was performed on 30 pairs of images by two examiners in 21 standard image locations. Automatic registration was performed between color to color, fluorescein to fluorescein, color to fluorescein, indocyanine green to fluorescein fundus photographs. Results: The evaluation proved high accuracy and robustness of the registration algorithms. From the initial 30 pairs of images, 29 were perfectly registered. In only 1 case the misregistration of 1-5 pixels occurred inferiorly, due to poor image quality. A second registration effort on the same set of images yielded the same degree of accuracy. The average time for registration on Pentium III computer was 40 seconds using non-optimized code. A surprising capability of the software was the ability to register late frames of fluorescein angiograms where the usual landmarks were barely discernable. Conclusion: A new highly accurate robust algorithm was developed for automatic computer alignment of fundus images. The software will allow for rapid evaluation, fundus feature change detection and comparison.
Keywords: 429 image processing • 430 imaging/image analysis: clinical • 432 imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)