A Bayesian‐Optimized Ensemble Deep Learning Framework for Automated Detection and Classification of Retinal Diseases in Ghana Using OCT Images

Retinal diseases pose a significant global health challenge due to their potential to cause severe visual impairment and blindness. This study aimed to develop a robust deep learning ensemble framework for the automated detection and classification of retinal diseases from optical coherence tomography (OCT) images. This study used OCT images from WATBORG Eye Services in Ghana, including glaucoma, macular edema, posterior vitreous detachment (PVD), and healthy eyes. The data preprocessing steps included augmentation, resizing, and one‐hot encoding. The dataset was divided into training (56%), validation (14%), and testing (30%) sets using stratified sampling. Six convolutional neural network (CNN) architectures, Visual Geometry Group (VGG)‐16, VGG19, DenseNet121, MobileNet, ResNet50, and InceptionV3, were evaluated as base models. To optimize hyperparameters for each model, Gaussian process–based Bayesian optimization was used. In addition to analyzing individual architectures, two ensemble learning strategies, soft ensemble and stacked ensemble, were implemented to combine the predictions from the models. Among the six base learners, VGG16, VGG19, InceptionV3, DenseNet121, MobileNet, and ResNet50‐VGG19 achieved the highest individual accuracy of 97.74%, followed by MobileNet (93.99%), InceptionV3 (93.61%), DenseNet121 (93.23%), VGG16 (91.73%), and ResNet50 (60.53%). The ensemble strategies outperformed most single models. The soft ensemble achieved an accuracy of 95.11% with high precision, recall, and specificity, indicating improved stability and reduced variance across disease categories. The stacked ensemble yielded the best overall results, achieving an accuracy of 97.37%, a precision of 97.78%, a recall of 97.93%, an F 1‐score of 97.85%, and a ROC–AUC of 0.9993. Bootstrap resampling was incorporated to assess the stability and reliability of the reported performance. This study demonstrates that integrating Bayesian‐optimized CNN architectures with ensemble learning significantly enhances retinal disease classification performance. By applying these methods to a Ghanaian dataset—an underrepresented population in retinal artificial intelligence (AI) research—this work highlights the potential of AI to enhance diagnostic accuracy, optimize resource allocation, and promote health equity in ophthalmic care. Future research should focus on scaling datasets, integrating multimodal data, and conducting clinical validations to support the real‐world deployment of these models.