Title: A Deep Learning-Based Multimodal Framework for Diagnosing Myopic Disorders
Authors: Okeke, Ogochukwu C. Udensi, Nkechi G. and Mgbeafulike, Ike J.
Volume: 9
Issue: 8
Pages: 8-17
Publication Date: 2025/08/28
Abstract:
Myopia is one of the most prevalent refractive disorders globally, with its increasing incidence posing a major public health concern. Early and accurate diagnosis is essential to prevent its progression into pathological complications. This study proposes a deep learning-based multimodal diagnostic framework that integrates retinal image data and structured clinical information to enhance myopia detection. The model comprises two 2D Convolutional Neural Networks (CNNs) for retinal image analysis and a 1D CNN for clinical data processing. Feature representations from each modality are fused to train a robust classification model capable of distinguishing myopic from non-myopic cases. To evaluate performance, the proposed model was trained over 40 epochs and compared against a unimodal baseline model by Yesugade et al. (2024), which utilized a single CNN and image-only input. Results revealed that the hybrid model achieved superior performance, with validation accuracy exceeding 91%, significantly lower training and validation losses, and a ROC-AUC score approaching 0.97-highlighting its strong generalization and discriminative capabilities. Accuracy and ROC-AUC graphs confirmed the model's enhanced convergence rate and learning stability, while also demonstrating minimal overfitting or underfitting. By incorporating both anatomical and contextual features, the hybrid model outperformed the unimodal approach in all key metrics. These findings underscore the importance of multimodal data integration in deep learning-based medical diagnostics and suggest that the proposed system can serve as an effective tool for community-based screening and tele-ophthalmology applications.