Transfer Learning with Ensemble Feature Extraction and Low-Rank Matrix Factorization for Severity Stage Classification of Diabetic Retinopathy

The automatic classification of diabetic retinopathy (DR) is of vital importance, as it is the leading cause of irreversible vision loss in the working-age population all over the world today. Current clinical approaches require a well-trained clinician to manually evaluate digital colour fundus photographs of retina and locate lesions associated with vascular abnormalities due to diabetes, which is time-consuming. Recently, deep feature extraction using pretrained convolutional neural networks has been used to predict DR from fundus images with reasonable accuracy. However, techniques such as global average pooling (GAP), singular value decomposition (SVD) and ensemble learning have not been used in automatic prediction of DR. We propose to use a combination of deep features produced by an ensemble of pretrained-CNNs (DenseNet-201, ResNet-18 and VGG-16) as a single feature vector to predict five-class severity levels of diabetic retinopathy. Our results show a promising F1-measure of over 98% on the kaggle dataset and another dataset provided to us by an ophthalmic clinic. This is an improvement on the current state-of-the-art approaches in DR classification. We evaluated prominent CNN architectures (DenseNet, ResNet, Xception, InceptionV3, InceptionResNetV2 and VGG) that can be used for the task of transfer learning for DR. Moreover, we describe a technique of reducing memory consumption and processing time whereas preserving classification accuracy by using dimensional reduction based on GAP and SVD.