RetCAD is a computer-aided detection system product whose main goal is to detect abnormalities related to Age-related Macular Degeneration (AMD) and Diabetic Retinopathy (DR) on color fundus (CF) images. RetCAD takes a CF image as input and produces several outputs: a quality assessment of the input image, heat maps indicating possible abnormal areas for AMD and DR, and scores for both these retinal diseases that are monotonically related to the likelihood that these diseases are present in the image. Users can take these outputs into account in their clinical work: they can decide if a new image should be acquired, in case the quality assessment indicates suboptimal image quality; they can decide to refer a patient for further testing for the presence of AMD, DR or other retinal abnormalities, in case the heat maps display suspicious regions that are verified by a human operator or when the scores are above certain thresholds.
RetCAD is software based on convolutional neural networks, a state-of-the-art technique in machine learning. In the process of analysing the input CF image, it compares regions in the image with regions extracted from normal and abnormal color fundus images. These latter images form the training data set of the software. A basic principle in machine learning is that the training data set is properly representative of the test data, otherwise the results may not be reliable. CF cameras from different manufacturers produce images of different quality because of hardware differences. In addition, image acquisition protocols can vary across acquisition sites, for example: the illumination, degree field of view and the resolution of the image can vary. Furthermore, the patients may originate from different populations in which the appearance of the retina, like color and pigmentation, may vary. Therefore, RetCAD has been trained with data from several populations and several CF camera’s.