Autonomous Diagnosis of Diabetic Retinopathy through Image Interpretation

Diabetic Retinopathy caused by diabetes mellitus, is among the most significant complications that lead to vision impairment and blindness without some form of timely identification. Traditional techniques for the diagnosis of DR are labor intensive and rely on specialized medical expertise, hence the difficulty in implementing widespread screening. Recent advances in state-of-the-art deep learning tools, particularly Convolutional Neural Networks, have been exploited with remarkable success for the automated identification of retinal fundus images with DR. [15] This paper presents CNN model for automatic detection and classification of severity of DR. The proposed model is trained and tested on the Kaggle Diabetic Retinopathy data set. Microaneurysms, exudates or haemorrhages can be found in retinal fundus images, all of which are signs of the DR potential phase, for a highly active diabetic population. Execution is Python programming Keras, OpenCV, concerned libraries and own NumPy one for mechanism functions to simplify image process and model functionality for them. Augmentations are applied to cater for the class imbalance and to contribute to model generalization. The analysis of using different CNN architectures such as ResNet50 and DenseNet201 have been carried out for capturing complex features in the retina.