Comparison of deep learning and machine learning architectures for early oral cancer diagnosis

Early detection plays a vital role in improving oral cancer outcomes, yet it remains a clinical challenge due to the inherent limitations of conventional screening methods, which often rely on subjective assessment and demonstrate poor consistency. Although histopathological analysis of a biopsy remains the gold standard for definitive diagnosis, Artificial intelligence (AI)-assisted tools can facilitate preliminary risk stratification and prompt earlier clinical suspicion, thereby supporting timely intervention and enhancing diagnostic efficiency. Recent advances in AI, particularly Deep Learning, offer promising solutions by enabling automated image-based diagnostics. This study proposes a comparative framework that evaluates the performance of Convolutional Neural Networks (CNN) architectures and hybrid models combining CNN-based feature extraction with traditional classifiers. Using two public datasets (Kaggle and Roboflow), the models were assessed across clinically relevant metrics: accuracy, sensitivity, specificity, loss, and diagnostic odds ratio. Inception-v3 achieved the most consistent diagnostic performance, with high accuracy, sensitivity, and diagnostic odds ratio, demonstrating strong suitability for clinical deployment. This makes it ideal for early screening scenarios, despite moderate specificity. Hybrid models improved specificity but underperformed in overall diagnostic balance, suggesting their complementary role in confirmatory diagnostics. Statistical analyses confirmed significant performance differences among models, reinforcing the reliability of deep learning approaches for oral cancer detection. These findings validate the potential of deep learning architectures for integration into preliminary diagnostic workflows and population-oriented telehealth platforms. They also highlight the need for further model optimisation, dataset expansion, and clinical validation to ensure generalisability and safe deployment in real-world healthcare environments.