Retinal Disease Detection Using Machine Learning Techniques

Author Name(s): Pooja M. Pawar, Dr. Avinash J. Agrawal
Author Email: *pooja27pawar@gmail.com

Abstract
One of the complications of diabetes is Diabetic Retinopathy. It occurs when high blood glucose damages the small retinal vessels which provide nutrients and oxygen to the retina. Initially, the diabetic retinopathy treatment results were recorded on the basis of Fundoscopic Exam. It is essential to have an automatic detection method for DR detection at an early stage to prevent severe vision loss. This paper provides a method for DR detection from retinal fundus image using the advanced machine learning techniques. Here we developed the screening system by performing classification using three different classifiers. The performance of proposed work has been tested by using MESSIDOR database. The input images are classified into three classes named Normal, NPDR and PDR. The original dataset has 1047 image, out of which 210 samples were used for testing. The proposed method achieves an accuracy of 74.28% using Decision tree classifier for multiclass classification and 82.85% on binary classification.

Introduction

Diabetic retinopathy (DR) is a rigorous and global eye disease which results in blindness and severe vision loss among the working age population. According to World Health Organization (WHO), approximately 10% population has diabetes and it will rise to 300 million around the world by 2025. In order to decrease this burden, it requires regular screening and early diagnosis.There are some distinct features in DR such as abnormality and leakiness of retinal blood vessels. When it moderately grows different lesions appear as microaneurysm, hard exudates, hemorrhages, cotton wool spots and red lesions etc. There are some tiny red dots around retinal blood vessels known as microaneurysms and the bigger blood clots are called as hemorrhages. Exudates are the important components which give information about diabetic retinopathy in primary phases. Exudates are again categorized into two classes named hard exudates and soft exudates (often called cotton wool spots). Ophthalmologists categorize DR into two major classes as Non-proliferative DR (NPDR) and proliferative DR (i.e. PDR). Microaneurysms and hemorrhages called red lesions belong to NPDR. Neovascularization occurs in PDR where the new abnormal blood vessels generated in the retina.

Conclusion

Diabetic retinopathy is a critical issue. This paper suggested different machine learning classifiers to identify the affected images. First, the input image is pre-processed and various features have been segmented by using image processing operations. Later, the feature vector is generated and classification is done. Three classifiers are used for classifying the input images such as decision tree, logistic regression, and AdaBoost. From the obtained results it is observed that the Decision Tree classifier outperformed the other three models. Decision tree gave 74.28% accuracy and ANN gave 66.79%. However, the efficiency of classification process can further be improved with some better feature values. The future work includes improving the performance of the proposed system by using the other machine learning classifiers to get more accurate results. We will attempt to increase the results by using some other deep learning techniques to get accurate feature values and classification outcome.

1,207 total views, no views today

Download PDF File

About the author: admin