Improved classification of oral cancer through a personalized transfer learning CNN architecture

Introduction: Oral cancer presents a significant danger to worldwide health, resulting in high death rates and substantial suffering. Early detection is crucial to improving treatment outcomes. This study uses deep learning techniques, particularly Convolutional Neural Networks (CNNs) and transfer learning methodologies, to propose a dependable machine learning system for oral cancer detection. Materials and methods: The proposed model leverages CNNs and transfer learning, followed by two sequential fully connected operations: FC1 (Feature Embedding) to consolidate learned features and FC2 (Classification Head) for final classification. Despite a small dataset and imbalanced class distribution, the model was trained and evaluated using carefully selected performance metrics including accuracy, precision, recall, F1-score, and ROC-AUC. These metrics were specifically chosen to address the challenges of imbalanced datasets, where accuracy alone can be misleading. Class imbalance was addressed through the Synthetic Minority Oversampling Technique (SMOTE), data augmentation, and careful preprocessing strategies. The performance was validated through confusion matrices and AUC-ROC analyses to ensure reliability. However, external validation was not performed, representing a limitation of this study. Results: The model achieved an F1-score of 81.48 %, accuracy of 81.38 %, precision of 84.62 %, recall of 78.57 %, and a ROC-AUC score of 0.9082 on the test dataset. During training, it achieved higher metrics: accuracy of 96.94 %, precision of 97.92 %, recall of 96.17 %, F1-score of 97.04 %, and a ROC-AUC score of 0.9967. Conclusion: This research highlights how artificial intelligence can impact clinical workflows in detecting cancer early. The results offer a hopeful path for advancements in automated cancer diagnosis technologies.