Sistema para Diagnóstico de Patologias do Sistema Respiratório Utilizando Redes Neurais
Carregando...
Data
2022-06-24
Autores
Watanabe, Fernanda Trevisan
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Atualmente o diagnóstico de um exame radiológico por imagem depende
inteiramente da interpretação visual do profissional de saúde, sendo este unicamente
responsável pelos resultados e tratamentos derivados desta interpretação. Ainda na
formação dos profissionais da saúde, a interpretação destas imagens pode ser de difícil
definição, uma vez que com o desenvolvimento de tecnologias que se agregam à
Radiologia, pode tornar o estudo desta mais complexo. Diante deste cenário, o presente
trabalho de graduação teve por objetivo criar um protótipo de uma ferramenta que auxilie
estudantes de medicina e enfermagem a examinar com maior precisão imagens de
radiografias torácicas para obter um melhor diagnóstico. O processo permite que uma
nova imagem radiográfica seja inserida em um aplicativo para dispositivos móveis, onde
será analisada e comparada a outras similares com o auxílio de uma rede neural
convolucional treinada, retornando ao usuário o diagnóstico mais preciso da patologia
presente nesta imagem. Esta interface foi desenvolvida utilizando os frameworks Jupyter
Notebook, Ionic e Angular, e implementada em um ambiente simulado em uma máquina
de desenvolvimento. Após todas as etapas de construção da ferramenta, o sistema foi
devidamente testado, inserindo imagens com patologias mapeadas pela rede neural,
patologias não mapeadas e imagens sem patologias detectáveis, a fim de se garantir o
resultado entregue ao usuário. Os resultados mostram que, para a atual modelagem, a rede
neural consegue prever uma imagem com aproximadamente 70% de acurácia geral,
considerando uma base de dados de radiografias torácicas sem patologias (normais), com
presença de pneumonia bacteriana e com pneumonia viral.
Currently, the diagnosis of a radiological imaging exam depends entirely on the visual interpretation of the health professional, who is solely responsible for the results and treatments derived from this interpretation. During the academic education of health professionals, the interpretation of these images can be difficult to define once the development of new technologies is being combined to the study of radiology, making its study more complex. Given this scenario, the present graduation paper aimed to create a prototype of a tool that helps medical and nursing students to examine more accurately chest radiographic images in order to obtain a better diagnosis. The process allows a new radiographic image to be inserted into an application for mobile devices, where it will be analyzed and compared to similar ones, with the help of a trained neural network, returning to the user the most accurate diagnosis of the pathology present in this image. This interface was developed using Jupyter Notebook, Ionic and Angular frameworks, and implemented in a simulated environment on a development machine. After all the steps to build the tool, the system was properly tested, inserting images with pathologies mapped by the neural network, unmapped pathologies and images without detectable pathologies, in order to ensure the result delivered to the user. The results show that for the current model, the neural network is able to predict an image with approximately 70% overall accuracy, considering a database of chest x-rays without pathologies (normal), with the presence of bacterial pneumonia, and with viral pneumonia.
Currently, the diagnosis of a radiological imaging exam depends entirely on the visual interpretation of the health professional, who is solely responsible for the results and treatments derived from this interpretation. During the academic education of health professionals, the interpretation of these images can be difficult to define once the development of new technologies is being combined to the study of radiology, making its study more complex. Given this scenario, the present graduation paper aimed to create a prototype of a tool that helps medical and nursing students to examine more accurately chest radiographic images in order to obtain a better diagnosis. The process allows a new radiographic image to be inserted into an application for mobile devices, where it will be analyzed and compared to similar ones, with the help of a trained neural network, returning to the user the most accurate diagnosis of the pathology present in this image. This interface was developed using Jupyter Notebook, Ionic and Angular frameworks, and implemented in a simulated environment on a development machine. After all the steps to build the tool, the system was properly tested, inserting images with pathologies mapped by the neural network, unmapped pathologies and images without detectable pathologies, in order to ensure the result delivered to the user. The results show that for the current model, the neural network is able to predict an image with approximately 70% overall accuracy, considering a database of chest x-rays without pathologies (normal), with the presence of bacterial pneumonia, and with viral pneumonia.
Descrição
Palavras-chave
Pulmões Doenças, Redes neurais (Computação), Processamento de imagens, Tórax Radiografia, Aplicativos móveis, Neural networks (Computer science), Image processing, Mobile apps