Eficácia do ChatGPT no diagnóstico de lesões em odontologia através da análise descritiva de achados microscópicos

Abstract

Oliveira IA. Performance of ChatGPT in the diagnosis of oral lesions through descriptive analysis of microscopic findings. Campo Grande, 2025. [Trabalho de Conclusão de Curso – Universidade Federal de Mato Grosso do Sul].

Artificial intelligence (AI) has emerged as a promising tool in medicine and dentistry, particularly for tasks that require interpreting complex information such as clinical data, laboratory results, and histopathological descriptions. Among emerging technologies, large language models (LLMs) such as ChatGPT-5 stand out for their ability to process and generate natural language in a contextualized manner, simulating aspects of human clinical reasoning. This study investigated the diagnostic performance of ChatGPT-5, in its free version, in analyzing microscopic information from oral pathology reports at the Federal University of Mato Grosso do Sul (UFMS), collected between 2023 and 2025. A total of 135 lesions meeting the inclusion criteria were selected and divided into four groups: odontogenic lesions, potentially malignant disorders and squamous cell carcinoma (SCC), traumatic and reactive lesions, and lesions caused by biological agents. The model received detailed microscopic descriptions and generated primary and secondary diagnostic hypotheses. Clinical information was later added to assess its impact on diagnostic performance. In the first stage, considering only microscopic descriptions, ChatGPT-5 achieved a 74.8% accuracy rate for the primary diagnostic hypothesis. After including clinical data, the overall accuracy increased to 78.5%. Performance varied across groups: potentially malignant disorders and SCC achieved 97.7% accuracy, while odontogenic lesions reached 60.0% after the addition of clinical information. Traumatic and reactive lesions and those caused by biological agents showed accuracy rates of 76.0% and 60.0%, respectively. Some responses were conceptually correct but differed in nomenclature from the original diagnoses. These findings indicate that ChatGPT-5 can consistently interpret histopathological descriptions, especially when supported by clinical data, although it still struggles to distinguish between lesions with similar microscopic patterns or when clinical details are critical. Moreover, semantic errors and occasional hallucinations highlight the need for human supervision. In conclusion, ChatGPT-5 demonstrates strong potential as a complementary tool for analyzing histopathological reports in dentistry, supporting the diagnostic reasoning of pathologists, particularly for common and well-characterized lesions. Future integration with specialized databases and radiographic information may further enhance its accuracy and clinical applicability.