Artificial Intelligence-Based Multi-Model for Risk Assessment of Oral Potentially Malignant Disorders Using Confocal Imaging of Exfoliative Cytology

The survival rates of oral cancer are heavily dependent on early diagnosis, particularly for oral potentially malignant lesions (OPMLs). However, the conventional diagnostic techniques, such as biopsies, are not only invasive but also resource-intensive and time-consuming [1]. Recent advancements in artificial intelligence (AI) based machine learning models offer promising alternatives for non-invasive screening. The efficacy of deep learning models such as convolutional neural networks (CNNs) in analyzing oral microscopic images with high accuracy has enhanced the decisions of oral pathologists, enabling risk assessment of oral lesions [2, 3]. However, gaps remain in integrating multimodal data into unified diagnostic platforms. The primary objective of this research was to develop a robust, AI-driven system capable of analyzing multiple types of data, including patients’ clinical records, confocal microscopic imaging of exfoliative cytology samples, and patterns of toluidine blue-stained macroscopic images to provide comprehensive risk evaluations of OPMLs.

Data Sources

Pilot data included 50 patients recruited from the outpatient clinic of the Oral Medicine, Oral Diagnosis, and Periodontology Department, Faculty of Dentistry, Ain Shams University. Patients were diagnosed clinically as having OPMLs. Data from each patient was categorized into 3 datasets: (1) Patients’ clinical records (CR dataset), including age, sex, smoking habit, alcohol consumption, and lesions site (2) Confocal microscopic imaging of exfoliative cytology samples from lesions (CC dataset), and (3) Toluidine bluestained clinical images of the lesions (TB dataset) (Figure 1).

AI Model Development

Three models were developed and integrated into a single web-based diagnostic tool:
1. Clinical Model: Clinical records were cleaned and normalized, with categorical variables (e.g., tobacco use, alcohol consumption) encoded numerically. A supervised machine learning classifier (AdaBoost) was trained on an anonymized CR dataset to categorize patients into mild, moderate, and severe risk cases.
2. Microscopic Image Model: A deep learning CNN (GoogleNet InceptionV3) was trained on the CC dataset to count the number of cells with mitotic figures, then classify images into mild, moderate, or severe dysplasia.
3. TB-Stain Pattern Model: A Roboflow-trained segmentation model (SVM) that classifies the TB dataset into non-stained, pale blue, or dark blue, indicative of mild, moderate, and severe dysplasia, respectively.
4. System Integration: The models were embedded into a React-based website, allowing users to upload data and receive real-time risk assessments (Figure 2). The backend, built with .NET Core, ensured secure data handling and scalability via Azure Cloud services.

The system was deployed and tested using a web interface by three independent oral medicine specialists.

Evaluation Metrics

Performance was assessed using accuracy, precision, recall, and F1-score. Output integration across the three models was used to generate a composite diagnostic recommendation.

• Clinical Model: AdaBoost demonstrated high performance (91% accuracy-90% precision) in predicting risk levels from clinical data.
• Microscopic Image Model: GoogleNet InceptionV3 showed superior performance (94% accuracy, 92% F1-score) in detecting cellular mitosis and classification.
• TB-Stain Pattern Model: SVM achieved 74.32% accuracy in staining pattern classification.
• Integrated Model: The oral medicine specialists reported ease of use and 98% agreement between the system output and their assessment.

The developed AI-driven system addressed critical limitations of traditional diagnostics by offering a non-invasive, scalable solution for OPMLs screening. The integration of multiple data modalities enhanced diagnostic precision, while the user-friendly interface ensures feasibility in clinical settings. Notably, the microscopic imaging model’s high accuracy (94%) highlighted the potential for deep learning in pathological data.

Challenges included the need for further validation in diverse populations and a larger dataset. Future work will focus on expanding datasets, incorporating real-time imaging technologies, and conducting multicenter clinical trials to validate the system’s efficacy.

This pilot implementation of an AI-powered screening and risk assessment system for oral potentially malignant lesions demonstrated high accuracy and usability in a clinical simulation.

The authors would like to thank engineers Amir Mamdouh Nassif, Mai Hossam Hassan Serageldin, Malak Mohamed Metwalli, Mirna Ihab Ramzy, and Nour Hany Abdallah for their help in creating and deploying the models. This work was based on facilities in the Central Lab of Stem Cells and Biomaterial Applied Research (CLSBAR) funded by the Science, Technology and Innovation Funding Authority (STDF) under grant [CB project ID (21747)].

The authors declare that they have no conflict of interest.

1. González Ruiz I, Ramos García P, Ruiz Ávila I, González Moles MÁ (2023) Early Diagnosis of Oral Cancer: A Complex Polyhedral Problem with a Difficult Solution. Cancers (Basel) 15: 3270.
2. Sukegawa S, Ono S, Tanaka F (2023) Effectiveness of deep learning classifiers in histopathological diagnosis of oral squamous cell carcinoma by pathologists. Scientific Reports 13: 1-9.
3. Li XL, Zhou G (2024) Deep Learning in the Diagnosis and Prognosis of Oral Potentially Malignant Disorders.