Boost diagnostic performance in retinal disease classification utilizing deep ensemble classifiers based on OCT

Damage to the retinal blood vessels is critical in diabetic retinopathy, a progressively emerging health concern that often advances quietly without explicit symptoms. Optical coherence tomography-OCT has emerged as a favored noninvasive imaging technique for diagnosing diabetic retinopathy promptly and accurately. However, timely and precise diagnoses from OCT images are essential in prevention of blindness. Moreover, accurate interpretation of OCT images is challenging. Single model learning debilitates in managing diverse data types and structures, constraining its adaptability to varied environments. Its limitations become apparent in tasks requiring expertise from multiple domains, delaying overall performance. Moreover, learning may exhibit susceptibility to overfitting with large and heterogeneous datasets, resulting in compromised generalization capabilities. In this study, we propose a hybrid learning model for the classification of four distinct classes of retinal diseases in OCT images with improved generalization capabilities. Our hybrid model is constructed upon the well-established architectural foundations of ResNet50 and EfficientNetB0. By pre-training the hybrid model on extensive datasets like ImageNet and then fine-tuning it on publicly available OCT image datasets, we capitalize on the strengths of both architectures. This empowers the hybrid model to excel in discerning intricate image patterns while efficiently extracting hierarchical prediction from various regions within the images. To enhance classification accuracy and mitigate overfitting, we eliminate the fully connected layer from the base model and introduce a concatenate layer to combine two objective learning prediction. A dataset comprising 84,452 OCT images, each expertly graded for illnesses. we conducted training and evaluation of our proposed model, which demonstrated superior performance compared to existing methods, achieving an impressive overall classification accuracy of 97.50%. Notably, we address the interpretability issue, providing clear and insightful results, making our model suitable for deployment in ophthalmology clinics, where it can significantly contribute to the diagnosis of retinal disorders with enhanced precision and efficiency.