Automated glaucoma diagnosis: Optimized hybrid classification model with improved U-net segmentation

Glaucoma is a leading cause of blindness, requiring early detection for effective management. Traditional diagnostic methods have challenges such as precise segmentation of small structures and accurate classification of disease stages remain. This research addresses these challenges by developing an optimized hybrid classification model for automated glaucoma diagnosis. At first, the preprocessing stage employs the histogram equalization technique known as Contrast Limited Adaptive Histogram Equalization (CLAHE) technique. Consequently, an improved U-Net segmentation process implemented with the proposed cross-entropy loss function is utilized. Then, features such as fractal features, cup-to-disc-based features, Inferior-Superior-Nasal-Temporal (ISNT) rule-based features and improved Pyramid Histogram of Orient Gradient (PHOG) based features are extracted. Further, a hybrid classification model, a combination of Improved Convolutional Neural Network (ICNN) and optimized Recurrent Neural Network (RNN) classifiers for diagnosing glaucoma disease. Also, to improve the performance of the diagnosis process, a new Opposition-based Learning-enabled Namib Beetle Optimization (OBL-NBO) approach is proposed to optimize the weights of the RNN classifier. Moreover, the ICNN classifier is employed for classifying the presence of glaucoma and non-glaucoma conditions. The proposed OBL-NBO scheme achieved an accuracy of 0.927 for dataset 1 and 0.945 for dataset 2 at an 80% training data.