A machine learning approach to improve the diagnosis of diabetic retinopathy using fundus image

Diabetic retinopathy is the most common diabetic eye disease. There are more than 29 million people with diabetes mellitus (DM) as of 2012 in the U.S and approximately 40% of the patients with DM have at least mild diabetic retinopathy. Diabetic retinopathy is diagnosed through comprehensive eye exams where blood vessels are examined on fundus images. However, assessment of blood vessels on colored fundus images is a very time consuming and subjective process. In this research, we present an automated blood vessel segmentation algorithm to facilitate the evaluation of diabetic retinopathy through assessment of blood vessel abnormalities. The blood vessels are extracted using random forest based classification model combined with wavelet based features and local binary pattern (LBP) based texture information. Discriminant analysis is modified and adopted for selection of the significant features to train the proposed classification model. Results demonstrate that the proposed method achieves higher blood vessel segmentation accuracy compared to other supervised based methods. The main advantage of the proposed model is to provide robust and computationally efficient classification of the diabetic retinopathy.