<p>Oral cancer is a major global health problem, with increasing incidence in recent years. Early detection and accurate diagnosis are critical to improving patient outcomes and reducing mortality associated with the disease. In recent years, deep learning particularly convolutional neural networks (CNNs), have shown significant potential in enhancing diagnostic accuracy in medical image analysis by enabling automated feature extraction, reducing observer variability, and improving classification performance across various imaging modalities. This study introduces a novel training approach for lightweight CNN models that integrates a weighted CNN ensemble with the knowledge distillation technique to improve the classification of oral cancer in histopathological images. Twelve pre-trained CNN architectures were fine-tuned using transfer learning to differentiate between normal and cancerous tissue. The three models with the best performance DenseNet-201, DenseNet-169 and VGG-16-BN were selected for the ensemble. Their ensemble weights were optimized via a particle swarm optimization algorithm, and the resulting weighted ensemble was employed as the teacher model in the knowledge distillation process. Then, three lightweight CNN architectures MobileNetV2, NasNetMobile, and MobileNetV3Small were trained using the proposed strategy that integrates the weighted CNN ensemble model with the knowledge distillation method. Experimental results revealed that these models achieved significantly improved classification accuracies of 95.24%, 92.86%, and 92.86%, respectively, outperforming their counterparts trained without the proposed approach, which yielded accuracies of 88.10%, 88.89%, and 85.71%. These findings confirm that combining weighted ensemble learning with knowledge distillation effectively improves lightweight CNN performance, offering a viable solution for oral cancer diagnosis in clinical and resource-constrained environments.</p>