Abstract
Ocular tumors pose significant diagnostic challenges, requiring early and accurate detection to prevent severe visual impairment or loss. This paper introduces an innovative methodology for the first detection of ocular tumors using an enhanced lesion recognition and classification through gradient-weighted class activation mapping (Grad-CAM) integration (ELRC-GI) model. By leveraging deep convolutional neural networks (CNNs) combined with Grad-CAM, our approach focuses on precise lesion recognition and classification within ocular fundus images. The proposed model, based on VGG19 and ResNet50 architectures, demonstrates superior performance, achieving 97% accuracy, 93% precision, and 85% recall on the RFMiD dataset. Furthermore, the ELRC-GI model consistently outperforms traditional baseline models by maintaining a true positive rate (TPR) of 96% at a false positive rate (FPR) of 0.3, as evidenced by its robust ROC curve results. The integration of Grad-CAM not only enhances our model’s interpretability but also contributes to automated lesion detection, offering significant potential for real-world clinical applications in ocular tumor diagnosis and treatment outcomes, making it a valuable tool in clinical settings and paving the way for future research in medical image analysis.