Abstract
Abstract: :
Purpose:Develop an automated system to identify anatomical features of the lens from scanned slit lamp photographs. The algorithm extracts 1) the location of the visual axis, (the antero-posterior line that bisects the nucleus horizontally, and 2) the location of anatomical features (cornea, cortices, sulcus, lentils) along the axis. Methods: The algorithm uses knowledge of the spatial relationship between the structures in the lens to facilitate image processing. An iterative approach is used: prominent features (cornea and anterior cortex) are used to determine an initial estimate of the visual axis, then feature detection locates the nuclear region and the line position is refined, repeating until the axis bisects the nuclear region. In detecting lens features, a multi-scale approach adapts processing based on spatial location. This is necessary to ensure the details of the nuclear region are extracted with high fidelity because that region is of particular interest for evaluating the degree of nuclear sclerosis (and the signal to noise ratio is low in the nuclear region). Problematic images are marked for human analysis. A user interface was developed to review, accept and/or modify the automated results. Results:Testing on an existing collection of images (n=1994) indicates the method is accurate. The axis and landmarks were successfully located (n=1962) or the image was correctly labeled as unacceptable, e.g. out of focus, (n=26). Failures were observed in 3 images. In 67 cases the image analysis was correct, but results were rejected because of poor image quality (focus, poor pupil dilation, other confounding pathology). Varying the sensitivity of the algorithm to uncertainty in the placement of the axis and detection of features can affect the performance: a higher tolerance to uncertainty will mark a higher percentage of images as requiring human intervention. Conclusion:A system has been developed to extract information from slit lamp images of the lens. A high degree of accuracy and repeatability of the measured location of anatomical features was obtained. We are currently evaluating the use of this algorithm in an epidemiological analysis of nuclear sclerosis.
Keywords: 429 image processing • 431 imaging/image analysis: non-clinical • 338 cataract