Article by A.L Nanayakkara, N. D Kodikara, A.S Karunananda, M.M Dissanayake
Diabetic retinopathy is a major health problem which is prevalent in a vast variety of diabetic patients. It will lead to eventual blindness due to changes of blood vessels in the retina. Retinal anomaly identification is a complex, time consuming task for ophthalmologists as they have to investigate a large portion of the area at once with the involvement of expert ophthalmologists or expensive equipment such as Fundus camera.
Diabetic retinopathy has 2 aspects such as NPDR & PDR. NPDR depends on the retinal lesion where there can be 2 kinds of retinal lesions namely exudates and microaneurysms. The other aspect is PDR which depends on the vascular network. Here easily breakable blood vessels begin to grow (Figure 1).
At present screening of diabetic retinopathy has been done by various machine learning techniques such as artificial neural network (ANN), Convolution neural network (CNN) etc.
1. Diagnosis of diabetic retinopathy using machine learning techniques. By Priya R et al (2013).
2. OpthoABM-An Intelligent Agent Based Model for Diagnosis of Ophthalmic Diseases. By Ranadive F (2014).
However these approaches failed to provide an acceptable reason for their anomaly classification. Very recently, Multi Agent technology has also been explored for biomedical applications.
Therefore it is of high importance to utilize automated techniques for the purpose of classification of stages of DR using fundus images with the acceptable reason behind the classification. In this study, we hope to extend the study to classify the stages of DR patients with the use of Multi Agent Technology.
For our initial experiments we have used data set of fundus images from National Eye hospital of Sri Lanka. This data set consist of 100 patients including each one from the four classes (Healthy, Mild, Moderate & severe). We hope to identify stages of diabetic retinopathy using image processing techniques and Multi Agent technology.
A sample set of 60 were chosen from this data set for the preliminary study. The sample contains series of fundus images of mild, moderate, severe DR patients with respect to the healthy patients. Those fundus images were pre-processed using MATLAB. Then extracted exudate or microaneurysm are used in detecting the anomaly stages of NPDR (Figure 2).
Microaneurysms are detected by analysing the dark lesions in localized blood vessels and exudates are detected only after localizing the optic disk (OD) to prevent OD capturing as a large exudate as contrast level of exudate regions in retinal images are quite similar to the OD. Finally these extracted features were converted into the region based statistical data using statistical model and output values were sent as input to the ANN. Here we used multi-layer perceptron (MLP) as our classifier model because of the statistical dataset.
We observed 85% of accuracy but it did not provide an acceptable reason for automated DR anomaly classification. Due to these drawbacks in ANN model we hope to utilize Multi agent solution for diabetic retinopathy classification as MAS is suitable for complex scenarios such as biomedical applications.
Once the proposed classification model for diabetic retinopathy is fully implemented it will significantly improve the classification of diabetic retinopathy and be beneficial for the entire world.