Purchase this article with an account.
Tobias Rudolph, Anna M. Broehan, Christoph A. Amstutz, Sebastian Wolf, Jens H. Kowal; Validation of a Real-time Retinal Image Registration Method for a Computer Assisted Laser Photocoagulation System. Invest. Ophthalmol. Vis. Sci. 2011;52(14):1042.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
An algorithm for the real-time registration of retinal video sequences from a scanning digital ophthalmoscope (SDO) to a fundus composite image is proposed and validated. Used in a computer assisted retinal laser photocoagulation system this method is will compensate for retinal motions and hence improves the accuracy, speed and patient safety of retinal laser treatments.
In order to align a video frame to the reference image, the algorithm combines intensity and feature-based registration techniques. Initially, the frame is tested for sufficient quality using an entropy-based grading algorithm. If the frame passes the test, the translational frame-to-frame motion between the preceding and the current frame is calculated by normalized cross-correlation. The initial transformation estimate is then constructed using this translation vector and the registration matrix of the previous frame. Next, the vessel points in the current frame are detected by looking for characteristic patterns in the intensity signal along horizontal and vertical grid lines. The vessel points are then iteratively aligned with the segmented vessel centerline of the composite image using a robust regression algorithm. This refinement procedure results in the final quadratic registration matrix. In order to measure the registration accuracy, a number of well-distributed feature points were selected in each composite image. After aligning the video frame, local, translation-only, cross-correlation based registrations were calculated at every feature point within the overlapping area. A region of 41x41 pixels around a feature point was chosen. The Euclidean distance serves as the error measure.
The algorithm was applied to ten different video sequences recorded from patients. The outlined validation method revealed an average accuracy of 2.47 pixels for 2764 evaluated video frames, which corresponds to approximately 23.3 micrometers in the reference image. On our test system the algorithm met the run-time constraint of 25 frames per second.
A real-time algorithm for multimodal video to reference image registration has been proposed, implemented and successfully validated using clinical data.
This PDF is available to Subscribers Only