Automatic diabetic-retinopathy detection with fundus images using optimal hybrid image features

Backgrounds: Diabetic Retinopathy (DR) is a difficulty which impacting the eyes. Timely screening and treatment of severe diabetes are essential to mitigate its adverse effects, particularly its detrimental impact on the blood vessels within the retina, which can result in a spectrum of eye-related issues ranging from minor to severe. Methods: This investigation aims to create a framework for the examination of DR in images. Different phases of this approach include (i) image collection and preprocessing, (ii) deep- and handcrafted-feature extraction, (iii) feature optimization using Lévy-Hummingbird-Algorithm (LHA), and (iv) binary classification and verification using 5-fold cross-validation. This work considered the Pretrained Deep-Learning (PDL), Local Binary Pattern (LBP), and Pyramid Histogram of Oriented Gradients (PHOG) approaches to mine the features. The classification task is executed with; (i) individual, (ii) dual-deep, and (iii) Serially-Concatenated Features (SCF), and its performance is verified using the detection accuracy. Results: Timely detection is crucial for initiating treatment to effectively manage and potentially cure a disease. This research endeavor is focused on creating a DR detection tool leveraging deep learning techniques to analyze FRI. The proposed approach takes into consideration both DF and HF to enhance the accuracy of detection. Deep characteristics are obtained from the images using PDL models, while handcrafted features are obtained through LBP with several weightings and PHOG with different bins. To address the concern of overfitting, the research employs feature reduction based on LHA, which aids in identifying the optimal DF and HF. The DR detection process is executed separately using DF, HF, fused deep features (DDF), and concatenated features (DDF+OHF). The experimental results of this research confirm the accuracy of over 99% when the DDF+OHF-based classification is applied. Furthermore, this tool attains a perfect accuracy of 100% when utilizing the KNN classifier. These outcomes underscore the effectiveness of the developed scheme when applied to the chosen FRI database. In the future, it is advisable to assess the performance of the proposed tool on other benchmark datasets and clinically collected Fundus Retinal Images to further validate its efficacy. Conclusions: The outcomes of this study confirm that the K-Nearest Neighbor (KNN) helps to achieve better classification accuracy (100%) when SCF is considered.