Purchase this article with an account.
M. Niemeijer, B. van Ginneken, M. Sonka, M.D. Abràmoff; Automated Classification of Exudates, Cottonwool Spots and Drusen From Retinal Color Images for Diabetic Retinopathy Screening . Invest. Ophthalmol. Vis. Sci. 2005;46(13):3468.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose:To evaluate an automated computer algorithm for classification of bright lesions in retinal color images into clinically relevant classes of exudates, cottonwool spots, or drusen. It is known that exudates and drusen often form characteristic groups or clusters. Methods:65 digital retinal color images were obtained (Topcon TRC–50 camera, 45° field of view, 768x576 pixels) from 65 patients diagnosed with diabetic retinopathy and/or drusen. All pixels in all images were classified into 4 classes as being either exudates, cottonwool spot, drusen (‘bright lesion’), or background, by a retinal specialist (MDA). Separately, microaneurysm and hemorrhage pixels were also classified. The automated algorithm proceeded as follows. All lesions in an image were extracted and clustered based on a distance metric. From each cluster, the following features were extracted: image intensity based measures; average lesion border strength, average image intensity within the lesions in the cluster and average image intensity around lesions, from the R, G and B color–planes separately; the average and standard deviation of the lesion area, the total number of lesions in the cluster, and the distance of the cluster to any red lesion in the image. A k–Nearest Neighbor classifier was trained with these features, using a leave–one–image–out protocol, and then used to classify unseen clusters into one of the three classes. Results:The 65 images contained 263 clusters (955 lesions) of exudates, 25 clusters (25 lesions) of cottonwool spots and 388 clusters (1782) lesions of drusen. The automated algorithm assigned each lesion cluster to one of three classes with classification correctness (true positives + true negatives / all clusters) of 84.2%. The classification correctness per lesion type was 90.1% for exudates, 52.0% for cottonwool spots and 82.2% for drusen. Conclusions:An automated algorithm was developed for classifying clusters of lesions as being either exudates, cottonwool spots or drusen, using lesion cluster based features. The method exhibits a high degree of classification correctness. This indicates that the characteristics of a cluster of lesions can be a useful addition to features extracted from the individual lesions. Employed in diabetic retinopathy screening programs, algorithms such as this may contribute to an early and more sensitive diagnosis of diabetic retinopathy.
This PDF is available to Subscribers Only