Document Type : Original Article

Authors

• Ivisson Alexandre Pereira da Silva ¹
• Carlos Alberto Correla Lessa Filho ²
• Anne Caroline dos Santos Barbosa ³
• José Marcos dos Santos Oliveira ⁴
• Matheus Henrique Alves de Lima ⁴
• Marcelo de Castro Meneghim Meneghim ¹
• Sonia Maria Soares Ferreira Ferreira ⁴

¹ Dept. of Public Health, Pediatric Dentistry and Orthodontics, Piracicaba Dental School, Universidade Estadual de Campinas, Piracicaba, Brazil.

² Dept. of Information Systems, Centro Universitário Cesmac, Maceió, Brazil.

³ Undergraduate Student, School of Dentistry, Centro Universitário Cesmac, Maceió, Brazil.

⁴ Postgraduate Program in Health Research, Centro Universitário Cesmac, Maceió, Brazil.

Abstract

Background: Oral cancer is one of the most common head and neck malignancies, often diagnosed at advanced stages, which compromises prognosis and increases treatment-related morbidity. Artificial intelligence (AI) has emerged as a promising tool to support early cancer diagnosis.
Purpose: To develop and validate an AI-based model incorporated into the Snap Oral Cancer mobile application to assist in the screening of oral cancer and potentially malignant lesions.
Materials and Method: This was an experimental technological study approved by the Research Ethics Committee (CAAE: 45329721.0.0000.0039). A total of 1,523 clinical images of SCC and potentially malignant disorders (leukoplakia, erythroplakia, and actinic cheilitis), collected between 2001 and 2022, were used. The model was built with Mobile-Net architecture, trained in Python with Keras/TensorFlow, and validated using performa-nce metrics, including sensitivity, specificity, accuracy, and area under the curve (AUC).
Results: The oral cancer model achieved excellent performance, with sensitivity of 98.4%, specificity of 87.7%, accuracy of 95.0%, and AUC of 0.93. Leukoplakia and erythroplakia models showed high sensitivity (100% and 96.3%, respectively), but low specificity (29.5% and 47.6%), resulting in higher false-positive rates. The actinic cheilitis model pre-sented intermediate performance, with sensitivity 77.2%, specificity 68.4% and AUC 0.73.
Conclusion: The AI model demonstrated high efficacy in detecting oral cancer, highlighting its clinical potential through the Snap Oral Cancer application. Despite limitations in precursor lesions, the findings reinforce the relevance of AI as an innovative and scalable tool for early diagnosis in dentistry. Future studies should expand the dataset and optimize algorithms to improve performance in precursor conditions and contribute to reducing inequalities in access to cancer diagnosis.

Keywords

• Oral Cancer
• Artificial Intelligence
• Early Diagnosis
• Deep Learning