AUTOMATIC MULTI-CLASS CLASSIFICATION OF COMMON INTRAORAL LESIONS USING CLINICAL IMAGES: A CONVOLUTIONAL NEURAL NETWORK APPROACH FOR DIAGNOSIS AND CLINICAL DECISION-MAKING ASSISTANCE

• Grant number: 25/05334-1
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate
• Start date: September 22, 2025
• End date: August 31, 2026
• Field of knowledge: Health Sciences - Dentistry
• Principal Investigator: Pablo Agustin Vargas
• Grantee: Cristina Saldivia Siracusa
• Supervisor: Marco Magalhaes
• Host Institution: Faculdade de Odontologia de Piracicaba (FOP). Universidade Estadual de Campinas (UNICAMP). Piracicaba , SP, Brazil
• Institution abroad: University of Toronto (U of T), Canada
• Associated to the scholarship: 22/13069-8 - ARTIFICIAL INTELLIGENCE FOR CLINICAL AND HISTOPATHOLOGICAL DIAGNOSIS OF INCIPIENT ORAL SQUAMOUS CELL CARCINOMA: A MULTICENTRIC INTERNATIONAL STUDY, BP.DR

Abstract

The aim of this study is to develop and assess the performance of Deep Learning models for multi-class automatic classification and decision-making assistance of common oral lesions using clinical images, including oral potentially malignant disorders and oral squamous cell carcinoma, and to explore Explainable Artificial Intelligence (XAI) methods to assess interpretability. Clinical images from cases of 15 highly prevalent intraoral lesions will be selected from the FOP-UNICAMP and from the Oral Pathology and Oral Medicine services from the University of Toronto. The development and application of the algorithms will be performed in collaboration with engineers and programmers of the University of São Paulo and University of Toronto. The FOP-UNICAMP dataset will be divided into training, model optimization, and internal testing subsets with an 8:1:1 proportion, while an external dataset will be kept for external testing. During training, data pre-processing based on augmentation techniques will be employed to enhance model robustness, and various architectures will be pre-trained using different transfer learning strategies. Performance will be assessed by mean accuracy, precision, recall, specificity, F1-score and AUROC values. Specialists and non-specialists will be tested using a REDCap survey to images from the internal testing set, to be compared with the CNN model obtained metrics. Qualitative analysis of XAI methods will be performed to assess explainability. We anticipate that these models could be impactful in future clinical practice for healthcare professionals' diagnostic assistance, particularly in regions around the world where access to infrastructure, resources and specialized oral medicine services are limited. (AU)