A Multi-class Deep All-CNN for Detection of Diabetic Retinopathy Using Retinal Fundus Images

Diabetic retinopathy (DR) is a diabetes complication that affects retina due to diabetes mellitus. Funduscopy capture the phenomena, however, it needs to be augmented by effective algorithms that enable detection of severity levels of DR from fundus images. Researchers have mostly proposed conventional convolutional neural networks (CNNs) for binary classification with a very few attempts on multi-class detection of DR. In this paper, unlike other approaches, we propose a deep All-CNN architecture for the detection of DR and its five levels. We first correct the fundus images against the sensor parameters by using Gaussian filters and perform blending in order to highlight foreground features. This is followed by removal of retinal boundaries that further helps the detection process. The proposed pre-processing helps in visualization of intrinsic features of the images as well as builds trust in predictions using the model. Thus pre-processed images are fed into the proposed All-CNN architecture. This has 10 convolutional layers and a softmax layer for the final classification. It includes three convolutional layers with kernel size 3 x 3 at strides of 2 which are designed to work as pooling layers; while two convolutional layers with kernel size 1x1 at strides of 1 are constructed to act as fully connected layers. We apply the proposed methodology on the publicly available Kaggle dataset that has the five-class labeled information. Our model is trained on 30000 retinal fundus images while tested on 3000 images. The proposed architecture is able to achieve 86.64% accuracy, loss of 0.46, and average F1 score of 0.6318 for the five classes. Unlike other architectures, our approach: (1) is All-CNN architecture, (2) provides multi-class information of DR, and (3) has outperformed existing approaches in terms of accuracy, loss and F1 score.