Purchase this article with an account.
carmen valverde, Maria I. Lopez, Maria Garcia, Roberto Hornero; Automatic Image Analysis based on Neural Networks to detect Hard Exudates and Red Lesions in Retinal Images. Invest. Ophthalmol. Vis. Sci. 2011;52(14):4040.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Diabetic retinopathy (DR) is an important cause of visual impairment. It is important to diagnose DR in its early stages, when treatment tends to be more effective. To ensure timely treatment, all diabetic patients should undergo periodic eye fundus examinations, which usually include digital retinal images. With the growing incidence of diabetes, the number of images that need to be examined by physicians can be prohibitively large. In addition, the high cost of examinations and the lack of specialists increase the time needed to obtain a diagnostic. For this reason, computer aided detection of DR lesions, like hard exudates (EXs) and red lesions (RLs), in retinal images could be an important aid in the screening and evaluation of DR.
We propose neural network (NN)-based methods for the detection EXs and RLs in retinal images. We examined four different types of NNs: multilayer perceptron (MLP), radial basis function (RBF) networks, support vector machine (SVM) and fuzzy ARTMAP. In addition, we developed two cooperative NN classifiers by combining individual NNs using majority voting and weighted majority voting schemes. Additional processing steps were also included to improve performance.
The majority voting scheme obtained the best trade-off between performance and complexity for both types of lesions. For this classifier, we obtained the following lesion-based and image-based statistics. Using a lesion-based criterion, we achieved a mean sensitivity (SEl) of 94.67% and a mean positive predictive value (PPVl) of 86.90% for EXs. For RLs we obtained SEl =82.74% and PPVl =53.29%. With an image-based criterion, a mean sensitivity (SEi) of 100%, a mean specificity (SPi) of 95.00% and a mean accuracy (ACi) of 97.50% were obtained for EXs. For RLs we achieved SEi =100%, SPi =56.00% and ACi =83.08%.
Automatic detection of clinical signs of DR could mean cost and time savings for health systems. Our results indicate that these automatic methods could be an important aid for ophthalmologists in the screening and evaluation of DR.
This PDF is available to Subscribers Only