Uso da inteligência artificial na caracterização nuclear e classificação entre leucoplasia bucal e leucoplasia verrucosa proliferativa

• Ano de defesa: 2020
• Autor(a) principal: Oliveira, Pedro Antônio de Ávila
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Biologia Celular e Estrutural Aplicadas
• Palavras-chave em Português: Leucoplasia bucal; Oral leukoplakia; Diagnóstico; Diagnosis; Inteligência artificial; Artificial intelligence; CNPQ::CIENCIAS BIOLOGICAS::MORFOLOGIA::CITOLOGIA E BIOLOGIA CELULAR; Citologia
• Link de acesso: https://repositorio.ufu.br/handle/123456789/30359; http://doi.org/10.14393/ufu.di.2020.710

Squamous cell carcinoma (SCC) of the oral cavity is one of the most common and deadliest head and neck neoplasms. Usually, SCC is preceded by lesions known as oral potentially malignant disorders (OPMDs). Among them, oral leukoplakia (OL) is one of the most prevalent and is characterized clinically by a white lesion and histologically by presenting hyperkeratosis and acanthosis. A variant of LB is a lesion known as proliferative verrucous leukoplakia (PVL), which has a higher malignant transformation rate than others OPMDs. However, the differential diagnosis between them is still a great challenge, in addition to the fact that both may present very similar histopathological aspects, especially in their early stages. Recently, artificial intelligence (AI) has proved to be very useful for the diagnosis and prognosis of malignant neoplasms and other diseases. Studies have shown that computational algorithms can detect tissue changes undetectable to a pathologist, hence helping them diagnose. However, for oral lesions, such as OL and PVL, there are no studies that use such a tool for diagnostic purposes. This study aimed to investigate cell nuclei from OL and PVL lesions through a computer system to elucidate whether this cell compartment is altered between them and a polynomial classifier capable of classifying the two lesions only with the extracted nuclear aspects. Sixty-one and three OL and PVL lesions, respectively, were gathered, and their H&E-stained slides were recovered and photographed for training and computational analysis. Clinicopathological and socio-demographic data were also raised from the requested pathological exam and then tabulated. The Mask R-CNN neural network was applied as a nuclear segmentation method and the polynomial classifier for OL and PVL classification based on the following nuclear information extracted by the network: area, perimeter, eccentricity, orientation, solidity, entropy and Moran Index. Clinicopathological and socio-demographic data from the OL-affected patients revealed that most of them were smokers and males, while the PVL-affected patients were female, and 1/3 underwent a malignant transformation. The neural network employed obtained an average accuracy of 92.95% in the identification of cell nuclei. Significant differences in 11 of the 13 nuclear characteristics studied were observed between OL and PVL, with the averages always higher in the LVP lesions, except for solidity. The polynomial classifier classified the two lesions with an average area under the curve of 97.06%. These data showed that the analysis of the nuclei through computational methods could be an essential tool to aid the diagnosis between OL and PVL lesions.