An Automated Approach for Diagnosing Diabetic Retinopathy in Retinal Fundus Images

Diabetes mellitus and its associated complications, including diabetic retinopathy (DR) have been identified as a growing global public health problem. Early treatment and constant monitoring of DR can prevent 90 percent of visual loss. Manual screening of fundus images by ophthalmologists is tedious and time consuming. Hence, a computer-aided diagnosis system can significantly act as an efficient solution for the time consuming detection process. The propounded algorithm includes various stages such as pre-processing, feature extraction and classification. In literature, different approaches for preprocessing such as dual guassian, second derivative of Guassian and Gabor filter, mathematical morphology, K-mean clustering algorithms etc. are employed. In the proposed algorithm, preprocessing was done using Adaptive Histogram Equalization (AHE) and image de-noising using non-local means (NLM) filter. NLM takes the means of all the neighbouring pixels and weightage is given by checking the similarity between these pixels to the target pixel. Laplacian edge detection, by identifying the zero crossings of the second derivatives of the image intensity was used in both the methods. In the feature extraction stage, features like HOG, area, perimeter, mean and standard deviation are extracted. Lastly, Support Vector Machines (SVM) algorithm, k-nearest neighbours (KNN) and Artificial Neural Network (ANN) was used for the categorization of the images to healthy and diseased. The comparison on the functioning of AHE and NLM validates that NLM outperforms AHE. Parameters like accuracy, error rate, specificity, and precision were used to analyze the working of the classifiers. SVM with 0.875 accuracy, 0.12 error rate, 0.75 specificity and 0.80 precision outperformed well when analogized to the other methods.