A Hybrid CNN Model for Deep Feature Extraction for Diabetic Retinopathy Detection and Gradation

Abstract

Diabetic retinopathy (DR) is a complication of prolonged diabetic mellitus causing damage to the micro-vascular system in retina. Computer-aided detection of DR can significantly reduce time, effort, and cost in early diagnosis and thereby restrict further complications. In this paper, we present a deep learning based framework for automated DR stage classification from fundus retinal images.We design and develop a custom build Convolutional Neural Network (CNN) model named Model-IVS to extract meaningful and discriminative features from the retinal fundus images. The proposed hybrid-CNN combines the multi-scale pyramidal feature processing concept of inception model with the convolutional pipeline based feature extraction concept of VGG-Nets and adopts multi-scale feature concatenation based short-cut connections of DenseNet with learnable convolutions. For the binary classification tasks, we employ Model-IVS as a feature extractor and use Support Vector Machine (SVM) classifier on the reduced features obtained through principal component analysis. The proposed hybrid model is end-to-end trained incrementally on a class-balanced dataset curated from the train set of EyePACS dataset. The cross-dataset generalization of the proposed hybrid CNN based DR classification framework is evaluated on APTOS, MESSIDOR-1 and MESSIDOR-2 datasets. Experimental results demonstrate that our model has outperformed all the existing state-of-the-art methods for DR severity grading task on APTOS dataset. The proposed Model-IVS has achieved best accuracy, AUC-ROC, and kappa score of 84.33%, 0.978 and 0.899, respectively in DR severity grading task. In binary classification tasks like DR screening and referable DR predictions, our Model-IVS feature extraction network with SVM classifier on PCA features has also achieved comparable performance results w.r.t. the state-of-the-art methods. In DR screening task, we recorded an accuracy of 97.84%, 87.42%, and 89.67% on APTOS, MESSIDOR-1, and MESSIDOR-2 datasets, respectively. In referable DR prediction, we achieved AUC-ROC score of 0.987, 0.928, and 0.873 on APTOS, MESSIDOR-1, and MESSIDOR-2 datasets, respectively.