The Application of Artificial-based Models to Classify Oral Cavity Findings Based on Clinical Image Analysis

Document Type : Original Article

Authors

1 Oral Medicine and Periodontology Department, Faculty of dentistry, Cairo University.

2 Computer and Systems Engineering Department, Ain Shams University, Cairo, Egypt.

Abstract

Objectives: This study aims to collect a comprehensive and versatile dataset to provide a solid foundation to assess the performance of five promising models for early detection of oral cancer.
Materials and methods: Clinical photographs of the entire mouth were obtained from patients visiting the Oral Medicine clinic at Cairo University's Faculty of Dentistry. These images were labeled and prepared to evaluate five CNN models, examining various data processing methods. The study incorporated augmentation techniques for all models and tested each model both with and without oversampling.
Results: The dataset comprises 5, 616 intraoral images, which are subdivided according to the presence and type of oral lesion. These include 2,686 images classified as ‘Normal,’ 1,410 as ‘Benign and inflammatory,’ and 1,520 as ‘Potentially malignant and malignant.’ The findings indicate that oversampling (V1) significantly improved model performance, particularly for GoogleNet, which consistently ranked among the top models in precision, accuracy, recall, and F1-score. InceptionResNetvr2 performed better in all evaluation metrics without oversampling. EfficientNet b4 showed similar results with and without oversampling, while ViT was the least consistent.
Conclusion: These results highlight the dataset's variability and complexity, revealing challenges in processing large-scale clinical oral images. This advances versatile models for diverse diseases.

Keywords

Main Subjects

Files

Share

How to cite

Statistics