Performance Evaluation of Oral Cancer Detection Using Deep Learning

Abstract

Oral cancer remains a major public health challenge, especially in developing regions where access to early screening facilities is limited. Conventional diagnosis depends heavily on clinical inspection and biopsy, which are time-consuming, invasive, and require expert evaluation. With recent advancements, deep learning has become a promising tool for automated medical image analysis, enabling faster and more reliable identification of cancer-related abnormalities. This study compares the performance of four deep learning models—CNN, ResNet152V2, EfficientNetB0, and VGG19—using an open-source oral cancer image dataset. After applying preprocessing, augmentation, and transfer learning, each model was evaluated using accuracy, precision, recall, and F1-score. Among the tested architectures, VGG19 achieved the highest performance with 95% accuracy and strong sensitivity for malignant lesion detection. The outcomes highlight the growing potential of deep learning as a supportive diagnostic tool for early oral cancer detection.

Introduction

Oral cancer is highly prevalent worldwide, particularly in India, due to risk factors like tobacco, betel-nut chewing, alcohol, and poor oral hygiene. Early detection is critical, but traditional biopsy-based diagnosis is invasive, resource-intensive, and sometimes inaccessible. Artificial Intelligence (AI), particularly deep learning with Convolutional Neural Networks (CNNs), offers a non-invasive approach to automatically detect cancerous lesions from oral images.

This study evaluates four deep learning models—Custom CNN, ResNet152V2, EfficientNetB0, and VGG19—on a Kaggle dataset of cancerous and non-cancerous images. Preprocessing included resizing, normalization, noise reduction, brightness correction, and data augmentation. Transfer learning with ImageNet weights was applied to all models except the baseline CNN.

Results: VGG19 achieved the highest performance with 95% accuracy, 94% precision, and 95% recall, outperforming ResNet152V2 (89%), EfficientNetB0 (90%), and the baseline CNN (84%). Deep learning models, especially VGG19, demonstrated strong potential for accurate, non-invasive oral cancer detection, addressing limitations of traditional screening methods.

Conclusion

The comparative study confirms that deep learning is a promising approach for oral cancer detection. Among the evaluated models, VGG19 achieved the best results with 95% accuracy, surpassing CNN, ResNet152V2, and EfficientNetB0. The system’s strong performance demonstrates its feasibility as a supportive diagnostic tool in clinical environments such as hospitals and dental screening centers.

References

[1] S. K. Baliarsingh, P. P. Dev, A. Bandyopadhyay, A. K. Dash and R. Pradhan et al.,” A Smartphone-based Deep Learning Framework for Early Detection of Oral Cancer Signs,” 2024 International Conference on Emerging Systems and Intelligent Computing (ESIC), pp. 181–186, 2024. doi: 10.1109/ESIC60604.2024.10481662. [2] S. Hemalatha, N. Chidambararaj and R. Motupalli et al.,” Performance Evaluation of Oral Cancer Detection and Classification using Deep Learning Approach,” 2022 International Conference on Advances in Computing, Communication and Applied Informatics (ACCAI), pp. 1–6, 2022. doi: 10.1109/ACCAI53970.2022.9752505. [3] R. Chavva, J. P. S and Mathu et al.,” Oral Cancer Detection Using Deep Learning,” 2024 International Conference on Science Technology Engineering and Management (ICSTEM), pp. 1–6, 2024. doi: 10.1109/ICSTEM61137.2024.10561172. [4] Xaviera A. L´opez-Cort´es, Felipe Matamala, Bernardo Venegas and C´esar Rivera et al.,” Machine-Learning Applications in Oral Cancer: A Systematic Review,” MDPI, vol. 18, p. 100472, 2022.

Copyright

Copyright © 2025 C V Keerthigha, Manasa K, Rachana M N, Amoghapriya R B, B. Sudha. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.