Hybrid CNN-ViT Model with Grad-CAM for Ocular Disease Detection

Timely diagnosis and effective treatment of ocular illnesses depend on accurate and automated detection. A hybrid deep learning model for identifying eye diseases from fundus images is presented in this study. It integrates Convolutional Neural Networks (CNN) with the Vision Transformer (ViT). The ViT improves the comprehension of global dependencies within images, whereas the CNN records local components. Enhanced medical image analysis and interpretability are achieved by the model's utilization of Gradient-weighted Class Activation Mapping (Grad-CAM), which highlights significant regions impacting classification decisions. Four datasets were used to assess the hybrid CNN-ViT model: the Retinal Disease Classification dataset, the ODIR-5K dataset, the ODIR dataset from Shang gong Medical Technology, and the Ocular Disease dataset by Amr Salem Helmy. On the ODIR-5K dataset, it obtained 87.25% training accuracy and 83.70% validation accuracy with a loss of 0.02. Using the ODIR dataset, the model showed 97% accuracy with validation and training losses of 0.252729 and 0.279614, respectively. The Ocular Disease dataset demonstrated 95.54% training and 92.56% validation accuracy, while the Retinal Disease Classification dataset produced 95.45% accuracy. These findings validate the model as a trustworthy instrument for identifying ocular diseases, assisting ophthalmologists in making clinical decisions.