STARE (STructured Analysis of the REtina) is a system for computer-aided analysis of ocular fundus images and fluorescein angiograms. The combination of image analysis in conjunction with expert systems and neural networks is a sophisticated and innovative approach in ophthalmology. We propose to continue our development of STARE for a second three-year period. The process of transforming an image of the retina into a set of differential diagnoses is complex but manageable. The image is broken up into 250,000 picture elements (pixels) with 256 levels of gray or 16 million colors. Specialized image-processing algorithms form groups of pixels that have common attributes, such as texture or a range of brightness or color. Other algorithms assemble the lines and regions into different objects (e.g. optic nerve, blood vessels, lesions) that are not yet identified. Features, such as color, size, shape, and edge sharpness, that describe these objects are measured. Statistical classifiersm or expert systems may be used to identify each object. When all the relevant objects in the picture have been found and identified, each objects, its location, and its relationship to other objects are stored in a database which forms a coded description of the image. That description is used by a neural network or an expert system either to create a differential diagnosis list with the probability of each diagnosis or to detect change in a sequence of images. We have developed algorithms that perform a large number of image processing tasks, and we have applied many of them for the automatic location and identification of the optic nerve, blood vessels, the fovea, exudates, cotton-wool spots, drusen, and various intraretinal hemorrhages. We have also demonstrated that the type of information in a coded description of ocular images can successfully teach a neural network to diagnose diseases from new images. In the proposed continuation, we will expand the identification capability to more objects, thus achieving a fuller description of the image. We will develop neural networks and expert systems to determine which is superior for transforming objects into diagnoses. We will compare measurements of the same objects in a sequence of images to detect change. As our ability to identify objects becomes more complete, we will apply all the steps to diagnose one disease, diabetes, and then expand the number of diseases. Ultimately, we will have an automated ophthalmic image diagnosis system with a graphics user interface that will furnish decision support and teaching and extend the capability and productivity of the ophthalmologist and his/her assistants.
|Goldbaum, M H; Katz, N P; Nelson, M R et al. (1990) The discrimination of similarly colored objects in computer images of the ocular fundus. Invest Ophthalmol Vis Sci 31:617-23|
|Pavan-Langston, D (1990) Herpes simplex virus ocular infections: current concepts of acute, latent and reactivated disease. Trans Am Ophthalmol Soc 88:727-96|