Automated Diabetic Retinopathy Detection Using Horizontal and Vertical Patch Division-Based Pre-Trained DenseNET with Digital Fundus Images

Abstract

Diabetic retinopathy (DR) is a common complication of diabetes that can lead to progressive vision loss. Regular surveillance with fundal photography, early diagnosis, and prompt intervention are paramount to reducing the incidence of DR-induced vision loss. However, manual interpretation of fundal photographs is subject to human error. In this study, a new method based on horizontal and vertical patch division was proposed for the automated classification of DR images on fundal photographs. The novel sides of this study are given as follows. We proposed a new non-fixed-size patch division model to obtain high classification results and collected a new fundus image dataset. Moreover, two datasets are used to test the model: a newly collected three-class (normal, non-proliferative DR, and proliferative DR) dataset comprising 2355 DR images and the established open-access five-class Asia Pacific Tele-Ophthalmology Society (APTOS) 2019 dataset comprising 3662 images. Two analysis scenarios, Case 1 and Case 2, with three (normal, non-proliferative DR, and proliferative DR) and five classes (normal, mild DR, moderate DR, severe DR, and proliferative DR), respectively, were derived from the APTOS 2019 dataset. These datasets and these cases have been used to demonstrate the general classification performance of our proposal. By applying transfer learning, the last fully connected and global average pooling layers of the DenseNet201 architecture were used to extract deep features from input DR images and each of the eight subdivided horizontal and vertical patches. The most discriminative features are then selected using neighborhood component analysis. These were fed as input to a standard shallow cubic support vector machine for classification. Our new DR dataset obtained 94.06% and 91.55% accuracy values for three-class classification with 80:20 hold-out validation and 10-fold cross-validation, respectively. As can be seen from steps of the proposed model, a new patch-based deep-feature engineering model has been proposed. The proposed deep-feature engineering model is a cognitive model, since it uses efficient methods in each phase. Similar excellent results were seen for three-class classification with the Case 1 dataset. In addition, the model attained 87.43% and 84.90% five-class classification accuracy rates using 80:20 hold-out validation and 10-fold cross-validation, respectively, on the Case 2 dataset, which outperformed prior DR classification studies based on the five-class APTOS 2019 dataset. Our model attained about >2% classification results compared to others. These findings demonstrate the accuracy and robustness of the proposed model for classification of DR images.