Dec. 19 2005 4:00PM

Dec. 19 2005 4:20PM

According to recent research, auto-fluorescent images taken by means of Heildelberg Retina Angiograph (HRA) have promising value for the diagnosis of glaucoma, especially due to strong correlation between the size of auto-fluorescent zones and glaucoma. With respect to poor signal/noise ratio of this signal, noise is to be suppressed by computation of the average image from the acquired time series. Because of movements of patients during cquisition, it is crucial for good results to perform image registration prior to the image processing. Therefore this contribution deals with the HRA time series non-rigid image registration. The registration is performed by multi-resolutional optimization of the global similarity criterion based on mutual information and consists of two steps. First, both images are registered using affine model of image >> transformation in order to compensate global mis-registrations. In this step we use a controlled random search optimizer in low-resolution layers of the image pyramid and the Powell optimizer for high-resolution. Further, a B-spline transformation model compensating for local mis-registrations is done making use of limited memory Broyden, Fletcher, Goldfarb and Shannon optimizer (LBFGS). Precision of the algorithm has been evaluated on the HRA images with artificially introduced known misregistration and finally, the algorithm was tested on a set of 16 real time series each containing 9 images.

International Workshop on Data - Algorithms - Decision Making (Contact person: Martin Janžura)