The rapid transition in ophthalmology from 35mm color film to digital media provides an opportunity to evaluate individual digital images for quality immediately after the photograph is taken. In other fields of medicine, such as radiology, image quality has been addressed in order to reduce unnecessary exposure to radiation. The goal of this project is to demonstrate a methodology that will evaluate a digital image from a fundus camera in real-time and give the operator feedback as to the quality of the image and the possible source of the problem in poor quality images.
The specific aims are to refine and apply a methodology that is both computationally efficient and highly effective in detecting poor or unacceptable quality images using a training set (N = 200) graded images, and to test it on a large (N = 800) set of retinal images. The approach is based on techniques that have been tested successfully on a preliminary data set and reported at the Association for Research in Vision and Ophthalmology (ARVO) 2007. The successful demonstration of the proposed methodology will lead to a significant reduction in poor quality retinal images that become part of a patient's record and/or a reduction in the negative impact on studies involving retinal images. By providing real-time feedback to the photographer, corrective actions can be taken and loss of data or inconvenience to the patient eliminated. Because the methodology uses parameters that are suggested by human perception qualities, the image quality model will produce results comparable to those of graders. The methodology will be based on image quality scores assigned by graders or ophthalmologists. Commercially, a real-time image quality assessment system is of interest to many manufacturers of fundus cameras and our methodology will be demonstrated to be scalable to any digital imager. Our methodology will also be of great value to screening centers where poor quality images can be reported immediately to the local or remote photographer. ? ? ?
