Normalized correlation speeds retinal diagnoses
Andrew Wilson
High-magnification and imaging-fundus-autofluorescence (IFA) techniques are being used by researchers to study a range of eye diseases. An IFA analysis takes advantage of the intrinsic autofluorescence of lipofuscin--the brown-colored material (commonly called "age-spots") made of damaged protein and fat that concentrates beneath the human skin. This material also accumulates in the retina of the eye and has been associated with hereditary diseases.
"To perform IFA analysis, newer techniques use confocal scanning laser ophthalmoscopes (cSLOs) that scan a small focused spot of the retina to generate an image, rather than illuminating a large area," says Linda M. Kelley, a clinical research coordinator with Retina Consultants of Southwest Florida (Ft. Myers, FL). This approach provides a high-contrast image from which an accurate medical diagnosis can be made.
Unfortunately, the signal-to-noise ratio of individual cSLO autofluorescence or high-magnification images is low. Consequently, sequences of images must be averaged to reduce noise and make the spatial structure more visible. According to Alex Wade of Stanford University (Palo Alto, CA), this method requires that each image in the sequence be aligned and registered with the others to compensate for eye movement. "Alignment is generally necessary for IFA imaging and is an absolute requirement for photoreceptor imaging where the effects of eye movement are more significant," he says. As a result, Wade and his colleagues have developed a cSLO-based computer system using off-the-shelf components specifically to deal with noisy images.
To digitize these images, a patient`s pupils are first dilated and then imaged using a prototype cSLO from Carl Zeiss Vision GmbH (Hallbergmoos, Germany). The images are then recorded on an S-VHS video system and digitized by a 133-MHz Pentium PC running Windows NT 4.0 from Microsoft Corp. (Redmond, WA) and a Pulsar PC-based frame grabber from Matrox Imaging (Dorval, Quebec, Canada).
To perform image alignment on the 768 x 576-pixel video images, Wade and his colleagues use a pattern-matching algorithm based on a normalized cross-correlation (NCC) program from the Matrox Imaging Library (MIL). "When presented with raw cSLO images, NCC algorithms fail because of the extremely high levels of electronic, thermal, and Poisson-distributed photon noise associated with cSLO imaging," explains Wade. To compensate for this, the image-alignment program initially copies the image frames and then filters one set of frames using a broad 12 x 12-kernal averaging function.
"Smaller convolution kernels can be used for images with strong landmark features," says Wade, "but we have found that the 12 x 12 kernel generally guarantees a successful pattern-matching operation." While the low-pass filtered images are being passed to the MIL NCC algorithm, image alignment is carried out on the original image frames, ensuring no loss of image data. The program uses another MIL library routine to search the filtered image data for a distinctive 200 x 200-pixel model. Once found, the model`s position is compared to its original position in the reference image, and the program calculates the amount of translation required.
"The translated, unfiltered image is then added to an accumulator buffer, and the process is repeated until all the images have been registered," says Wade. The completed system can generate averaged frames from video sequences 20 times faster than manual methods, and work is underway to increase the speed of the process.
