Diagnosis of Diabetic Retinopathy using Convolutional Neural Network based model by EyePACS dataset

Abstract

Diabetic Retinopathy (DR) ranks as one of the most frequent causes of impaired vision in the world, and it has to be detected with accuracy and in time in order to be addressed properly. In the presented work we suggest a robust deep learning pipeline based on the Modified DenseNet-169 convolutional neural network and five-class classification of severity of DR based on retinal fundus images in the Kaggle EyePACS Diabetic Retinopathy dataset includes (filtered subset). Problems of blurred, truncated, and low-contrast images proved a problem during preprocessing and had an effect on the stability and accuracy of prediction. We solved these issues with the use of resizing, Gabor filtering, normalization, and sophisticated data augmentation methods in order to increase vessel visibility and diversity of data. The training method used augmented data with Adam optimizer and categorical cross-entropy loss, and the result is evaluated on accuracy, F1-score, precision, recall and support. In our model, we obtained a validation accuracy of approximately 95.4 %. We also observed performance differences between classes because of the imbalance in data as well as image quality. Class-wise performance could be clearly interpreted with visual analytics of the form of precision-recall-F1-sup-support graphs and confusion matrices. This workflow provides a systematic foundation to create enhanced DR detection mechanisms as well as to identify real-life issues in the medical image categorization systems.