Computerized Diabetic Retinopathy Diagnosis with Optimized Random Forest

Abstract

It is desirable to diagnose diabetic retinopathy at an early stage for developing a suitable treatment plan to prevent the condition from deteriorating. To provide an immediate diagnosis of the retina, various methods have been investigated to realize a time and cost effective classification of the fundus images. However, most diabetic retinopathy automated identification methods are structural based analysis. Moreover, Asian fundus images have larger optic disc and thicker retinal vessels compared with Caucasians. Hence, we explore a machine learning approach to the extraction of texture features for classification and the feasibility of this approach using texture parameters to complement current algorithms. Normal retina, non-proliferative diabetic retinopathy and proliferative diabetic retinopathy are identified in this paper. The first step is achieved with three groups of texture features such as gray level co-occurrence matric texture features, different statistical features and run length matrix texture features extracted. In the second step, these features are fed into an optimized random forest classifier for automatic classification. We test our system on two databases (D1 and D2) consisting of 432 and 579 fundus images from a diabetic retinopathy screening program consisting of Asians. The diabetic retinopathy is successfully diagnosed with sensitivity is 0.936 ± 0.019 for D1 and 0.941 ± 0.016 for D2, specificity is 0.917 ± 0.011 for D1 and 0.918 ± 0.011 for D2, positive predictive value is 0.924 ± 0.013 for D1 and 0.939 ± 0.012 for D2, when training on the same institutions, respectively.