Detection of Diabetic Retinopathy using Transfer Learning

Diabetic retinopathy is the problem of diabetes that affects the eyes. It is caused when one of the parts of an eye called retina or say the blood vessels in the tissue at the back of the eye is damaged. The persons who had diabetes for more than 10 years are affected the most. It is observed that it is most common in the people of 70-79 years of age group. Early symptoms of diabetic retinopathy are found to be the blurriness, difficulties in perceiving colours and dark areas of vision. Chronic retinopathy can also cause blindness.As this disease shows very few symptoms, as a result the treatment is not given on time. Due to which, many people are being affected by this. The doctor who manually detects this disease is called ophthalmologists. The process of detecting diabetic retinopathy from the retinal images manually is a time-consuming process and also the diagnosis depends upon the expertise of the examiner.In this study, the authors collected two datasets for diabetic retinopathy, consisting of retinal images categorized into five groups. In the first dataset, the authors computed handcrafted HOG features and utilized the Support Vector Machine (SVM) algorithm for classification. They achieved an accuracy of 71.9% using this approach. In the second dataset, the authors employed a deep Convolutional Neural Network (CNN)-based framework to detect diabetic retinopathy and determine its severity level from retinal images. Early detection of the disease can lead to better disease management, and medical intervention can slow down disease progression. The proposed framework offers an easy method for detection.The paper utilized two datasets. In the first dataset, a machine learning algorithm was employed, achieving an accuracy of 71.9%. In the second dataset, three state-of-the-art transfer learning methods were employed. It was discovered that deep learning models outperformed the machine learning model. Among the deep learning models, the InceptionResNetV2 model exhibited the highest accuracy (89.5%), recall, and area under the curve (AUC) among the three models. However, the ResNet50 model demonstrated the highest precision.