Deep Learning-Based Classification of Retinal Diseases from OCT Images with LLM-Powered Patient Query Support

Retinal diseases such as Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), and Drusen often progress silently and may lead to permanent vision loss if not diagnosed early. Manual analysis of Optical Coherence Tomography (OCT) images is time-consuming and prone to human error, highlighting the need for automated diagnostic tools. This study proposes a deep learning-based platform using MobileNetV3, a lightweight CNN architecture, for the multi-class classification of retinal diseases. The model is trained on a publicly available Kaggle OCT dataset comprising 84,495 labeled images across four categories: CNV, DME, Drusen, and Normal. Extensive hyperparameter tuning was conducted by varying batch sizes (8, 16, 32) and optimizers (Adam, RMSProp, SGD). The optimal configuration—batch size 16 with Adam—achieved a test accuracy of 98.7%. Model performance was evaluated using accuracy, sensitivity, specificity, F1-score, and ROC-AUC. The results validate the model’s robustness and its potential for scalable, accurate, and real-time retinal disease detection in clinical practice. To enhance user interaction and accessibility, a Large Language Model (LLM) was integrated into the system to handle patient and practitioner queries related to disease understanding, symptoms, risk factors, and treatment options. Key words: Optical Coherence Tomography (OCT), MobileNetV3, Retinal Disease Classification, Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), Drusen, Deep Learning in Ophthalmology, Large Language Model (LLM).