This study proposes a framework for Explainable AI (XAI) designed to enhance early detection of oral cancer using optimized convolutional neural network (CNN) architectures. Four CNN architectures were evaluated, including three custom-designed architectures (Custom CNN 1, 2, and 3) and one pretrained model (EfficientNetB0) for oral cancer image classification. Performance evaluation was conducted using accuracy, precision, recall, and F1-score, with graphical representations. Custom CNN 1 performed best with an accuracy of 92.31%, with Custom CNN 2 and the pretrained model having generalization limitations; specifically, Custom CNN 3 could not classify non-cancer cases accurately. Model interpretability was enhanced through LIME (Local Interpretable Model-Agnostic Explanations) application, with significant regions contributing to prediction, particularly revealing model vulnerability to specific features. Despite a high training accuracy of 97.31%, this study reflects the necessity for model generalization improvement and a balanced precision-recall value. In future studies, model architecture will be optimized, and methodologies for balancing will be developed, with additional XAI techniques such as Integrated Gradients and SHAP incorporated for model interpretability improvement and establishment of model trust, allowing for effective real-life clinical practice in early oral cancer detection.