Sophisticated snake algorithms help diagnose glaucoma
Glaucoma is a disease in which pressure inside the eye increases to the point that it impairs vision and, if left untreated, may cause blindness. To detect the disease in its early stages, optometrists use scanning laser opthalmoscopes (SLOs) to measure the optic-nerve head or blind spot of the eye. In patients with glaucoma, this normally circular structure changes shape due to increases in intraocular pressure. Measuring the shape of the bind spot is, therefore, a useful indicator of early sig
Sophisticated snake algorithms help diagnose glaucoma
Glaucoma is a disease in which pressure inside the eye increases to the point that it impairs vision and, if left untreated, may cause blindness. To detect the disease in its early stages, optometrists use scanning laser opthalmoscopes (SLOs) to measure the optic-nerve head or blind spot of the eye. In patients with glaucoma, this normally circular structure changes shape due to increases in intraocular pressure. Measuring the shape of the bind spot is, therefore, a useful indicator of early signs of the disease. Although SLOs can generate high-resolution volume images of reflectivity in the retina from which the optic-nerve head boundary can be accurately traced, SLOs are inherently expensive.
Because of this, Tim Morris and his colleagues at the department of computation and optometry of the University of Manchester Institute of Science and Technology (Manchester, England) have developed a less-expensive method based around off-the-shelf cameras, frame grabbers, PCs, and a novel algorithm that identifies the boundary of the optic-nerve head using dynamic contours or snakes. To digitize images, Morris used a Fundus camera from Carl Zeiss (Thornwood, NY) coupled to a DT2851 PC-based frame grabber from Data Translation (Marlboro, MA).
"The major problem with capturing image data was illuminating the retina adequately," says Morris. To do so, the patient`s pupil was first dilated and illuminated by a photographic flash and the video image captured. "Although this usually resulted in satisfactory images, image brightness was variable, and image-data preprocessing was required before dynamic contour processing could be performed."
After histogram equalization enhanced the difference between the bright nerve-head region and the darker surrounding retinal region, images were thresholded to remove pixels that were not part of the nerve head. To enhance edge details, a pyramid edge detector was applied to the equalized image. To determine this, pixels in a layer of the pyramid were computed as the average of groups of four pixels in the layer below. A pyramid of five layers was chosen to provide a compromise between speed and accuracy, with the fifth level containing 16 ¥ 16 pixels. After the pyramid was constructed, a Sobel filter operator was applied to the pyramid, starting at the highest level.
"If the edge detector gave a result greater than a specified threshold, the equivalent pixels in the layer below were examined until either the base layer was reached or the edge magnitudes fell below the threshold," says Morris. After this stage of preprocessing, Morris applied a dynamic contour or snake algorithm to detect the boundary of the optic nerve head.
"The algorithm`s major advantage," says Morris, "is that it is able to bridge discontinuities in the image feature being located." By varying the relative influences of contour length, stiffness, and image feature, 10 to 20 iterations of the snake algorithm allowed a trace of the boundary of the optic-nerve head to be found.
"Implementing the software on a PC using the Watcom C compiler from Powersoft (Concord, MA) provided direct support of the IBM 8514 graphics driver," says Morris. "This was used in preference to the standard VGA drivers supported by Borland--now Inprise (Scotts Valley, CA)--and Microsoft (Redmond, WA) as it allowed higher-resolution images to be displayed," he says.
