Hierarchical classifier for soft and hard exudates detection of retinal fundus images

Abstract

Nowadays, diagnosis of diabetic retinopathy has caught the eager eyes of enthusiastic experimenters, as it has emerged as the common cause of blindness in the working age group. Several works are available in the literature for the detection of normalities and abnormalities through retinal image processing. Recently, a variety of literatures are presented based on normal/abnormal detection using retinal images. In this paper, we have proposed an efficient technique to detect the hard/soft exudates from abnormal retinal images. At first, the preprocessing step is carried out using Gaussian filter for enhancing the input retinal image. Consequently, normal/abnormal detection is done using region segmentation, feature extraction and Levenberg-Marquardt-based neural network classifier. From the abnormal retinal image, we have detected the soft/hard exudates using fuzzy c-means clustering, feature extraction and Levenberg-Marquardt-based neural network classifier. Here, region segmentation is done by three ways; (i) blood vessel extraction (ii) optical disc extraction using curvelet transform and (iii) Damage area extraction. In soft/hard exudates detection, fuzzy c-means clustering is utilized for damage area extraction. Then, with the aid of segmented area, features such as mean, variance, area, perimeter, entropy, maximum intensity, minimum intensity, cross correlation, auto correlation and co-variance features are extracted. Once the features are computed, training of Levenberg-Marquardt-based neural network is done to classify the abnormal retinal images into soft or hard exudates. Here, the experimentation is done using Standard Diabetic Retinopathy Database and the performance is analyzed with the standard evaluation metrics of accuracy, specificity and sensitivity. The innovative technique is observed to achieve superb results and a comparison is also made with the existing method. The results have proved that the proposed technique has outperformed the existing method by having superior accuracy of 90.91% when compared with the existing methods.