Early detection and classification of oral lesions are essential for the prevention of oral cancers, and yet, manual diagnosis is still a challenge due to variations in the appearance of lesions, quality of images, and limited clinical datasets. This research explores the use of diffusion models, a recent class of generative models renowned for their stable training and high-fidelity reconstruction, to improve the automatic classification of oral lesion images. The proposed system includes dataset collection from open-source platform kaggle, preprocessing of dataset, a diffusion- based denoising and feature extraction pipeline, and finally, a classification stage to categorize the normal, precancerous, and cancerous lesions. By leveraging the forward and reverse process of diffusion, the model improves the clarity of the images and effectively extracts discriminative features, mitigating problems of noise, imbalance, and low-quality clinical images. In a deep learning approach combining CNN-based classification with the concept of enhancement provided by diffusion mechanisms, the generalization performance is boosted. The system will be evaluated based on accuracy, precision, recall, and F1-score, and the results provide promising improvements compared to the state-of-the-art traditional deep learning methods. This paper has found that diffusion models provide a robust, scalable, and clinically valuable pipeline for early oral lesion detection, with strong potential to be deployed in real-world diagnostic pipelines and future research on medical imaging and we obtained a very good accuracy i.e., 96% while training the model. This paper establishes the diffusion model as a promising approach for medical image analysis, particularly in the early detection and classification of oral lesions, paving the way for future research and clinical applications in healthcare.