Enhanced Lesion Localization and Classification in Ocular Tumor Detection Using Grad-CAM and Transfer Learning

Abstract

Ocular tumors pose significant diagnostic challenges due to their rarity and the subtle visual cues they present in fundus images. This paper introduces a novel deep learning framework, termed ELRC-GI (Enhanced Lesion Recognition and Classification with Grad-CAM Integration), and designed for accurate and interpretable ocular tumor detection. The proposed model integrates VGG19 and ResNet50 convolutional neural networks with Grad-CAM-based attention supervision, enabling both high classification accuracy and precise lesion localization. Unlike traditional CNN-based approaches, ELRC-GI incorporates a heatmap-guided loss function that aligns model predictions with interpretable visual explanations, thereby improving clinical trust and diagnostic reliability. We utilize transfer learning by initializing VGG19 and ResNet50 with pre-trained ImageNet weights, freezing the initial layers and fine-tuning the final layers using the RFMiD dataset to adapt the models to the ocular tumor detection task. The model maintains a true positive rate of 96% at a false positive rate of 0.3%, as evidenced by a robust ROC curve. Experimental results on the RFMiD dataset demonstrate the superiority of ELRC-GI, achieving 97% accuracy, 93% precision, 85% recall, and an AUC of 0.98, significantly outperforming baseline CNN models. Grad-CAM visualizations further validate the model’s capability to highlight tumor regions, even in the presence of overlapping ocular conditions. The ELRC-GI model thus offers a robust, explainable, and clinically viable solution for early ocular tumor detection, setting the stage for broader application in interpretable medical AI.