Classification of glaucoma in retinal fundus images using integrated YOLO-V8 and deep CNN

Abstract

To propose a new system to identify glaucoma at an early stage with the help of deep learning-based AI method by utilizing Retinal Fundus Images (RFI). The method detects intrinsic key structures in the fundus images to predict retinal nerve layer thickness in order to improve the accuracy of glaucoma detection and classification. To learn complex and hierarchical image representations, the CNN model is used to identify the continuous value of retinal nerve layer thickness from RFI. The Binary Cross Entropy (BCE) loss function is used to perform multi-classification tasks to discover classes such as healthy eye, eye with glaucoma, and glaucoma suspect. In order to identify the local and global features in RFI, the YOLO-V8 object detection method is employed, which also helps to perform image localization, which includes image segmentation, deep optic disc analysis, and the extraction of ROIs. The main focus is given, especially for RNL thickness around OD regions and CDR measurement to perform glaucoma identification tasks. The PAPILA dataset is utilized with the ophthalmology records from 244 patients and includes 488 digital retinal fundus images, covering both left and right eyes for both male and female categories. The CNN model is trained on the PAPILA dataset with labeled RNL thickness values. The performance of CNN-BCE with YOLO-V8 is evaluated using MATLAB and compared against the prevailing approaches such as SVM, ADABOOST, and CNN-Softmax classifiers. The new model outperforms the existing methods with proven results of 98.88% accuracy rate, 0.9 dice-score, 97.74% and 98.03% sensitivity & specificity, 98.6% and 98.78% precision & recall, 98.06% f-score, and 0.92 true positive rates and 0.10 false positive rates under AUC-ROC. This clearly shows that the newly proposed CNN-BCE with YOLO-V8 detects and classifies glaucoma, which helps ophthalmologists perform potential screening and predict better treatments.