Frobenius deep feature fusion architecture to detect diabetic retinopathy

Diabetic retinopathy is a medical complication affecting the retina of patients prone to prolonged periods of Diabetes mellitus. Early detection and diagnosis are essential since the symptoms are subtle at the early stages. It can result in permanent vision impairment at advanced stages if not managed properly. Detecting various severity levels helps identify the level of treatment required for each patient, and directing them to the appropriate intensive care unit helps optimize the overall treatment efficacy. This work proposes a multi-model architecture by combining the features extracted from convolutional neural networks using novel Frobenius norm-based feature fusion with an ensemble of machine learning classifiers to perform the classification of binary and multi-class stages of Diabetic Retinopathy. The proposed approach delves into various phases- data collection and data pre-processing, feature extraction from VGG16 and Densenet201, feature selection using Random Forest, feature fusion using Frobenius norm, and classification using stacked ensembling of XGBoost classifier and ExtraTreeClassifier with SVC as meta-learner. The architecture is trained, and performance is assessed with the test set of the APTOS dataset, using accuracy, F1-score, precision, and recall metrics. The proposed architecture achieved the best accuracy of 98 percent for binary classification and 80.63 percent for multi-class classification. The performance of the recommended Frobenius-based deep feature fusion architecture achieved the highest accuracy for binary and multiclass classification compared to other related research works.