PROTÓTIPO COMPUTACIONAL PARA AVALIAÇÃO AUTOMÁTICA DA DOR NEONATAL POR PARTES DA FACE
Carregando...
Arquivos
Citações na Scopus
Tipo de produção
Trabalho de Conclusão de Curso
Data
2022-12-07
Autores
BRUNO AUGUSTO JOVENASSO
ERIK FERNANDO MENDES SILVA
RAFAEL TOCEGUI COMPRI
VANESSA GOMES FAVORITO
ERIK FERNANDO MENDES SILVA
RAFAEL TOCEGUI COMPRI
VANESSA GOMES FAVORITO
Orientador
Carlos Eduardo Thomaz
Periódico
Título da Revista
ISSN da Revista
Título de Volume
Citação
Texto completo (DOI)
Palavras-chave
Resumo
Atualmente a área de detecção de dor em neonatos a partir de imagens da face está em
constante desenvolvimento, possuindo principalmente como objetivo amenizar um dos problemas
que temos na área da neonatologia, os efeitos da dor a curto, médio e longo prazo em
recém-nascidos. Neonatos internados em unidades de terapia intensiva neonatal são expostos
constantemente a dor, tal fato pode causar danos no neurodesenvolvimento, afetando funções
cerebrais, sistema de estresse e lesões no processamento nociceptivo. Com o intuito de identificar
e classificar a presença de dor neonatal, são utilizadas escalas de avaliação de dor à beira
do leito por profissionais da área da saúde, e estão sendo criados métodos de inteligência artificial
e aprendizado de máquina para automatizarem este processo. Assim, este trabalho tem
por objetivo propor e implementar um protótipo computacional para a avaliação da dor neonatal
por partes da face, analisando cada parte da face separadamente e classificando a presença,
ou não, de dor. Para tal feito, o trabalho conta com algumas etapas sequenciais, começando
com um pré-processamento no banco de dados de imagens neonatais da Universidade Federal
de São Paulo, onde foi gerado uma face média e segmentada dos recém-nascidos. Na segunda
etapa, são realizados os treinamentos e criação do arcabouço, onde as imagens passarão pela
classificação de dor ou não dor, que será realizada por 12 redes neurais, uma para cada parte da
face, que serão analisadas e classificadas separadamente. Para a etapa de treinamento e criação
do arcabouço, foram utilizados 3 modelos de Rede Neural diferentes, sendo eles a VGG11, a
ResNet18 e a LeNet-5. A LeNet-5 apresentou um reconhecimento de dor por partes da face superior
aos outros modelos, sendo as partes da face com maior acurácia foram o nariz (79,66%),
olho esquerdo (79,44%) e boca (79,44%). Para a obtenção destes resultados, foi utilizado o
sistema de validação cruzada no treinamento das redes. Por fim, foi feita a união de todas as
etapas formando um sistema inédito e completo de detecção, segmentação e classificação de
dor em áreas de interesse faciais de neonato
Nowadays, the newborn pain detection by images of the face is being in constantly de- velopment, having principally as objective slow down one of the biggest problems that we have in the neonatology area that is the effect of the pain in short, medium and long term in new- borns. It is known that the newborns in intensive therapy units are exposed to constantly pain and this fact can cause several damages to the neural development, affecting cerebral functions, stress system and lesions in the nociceptive processes. With the intention to identify and clas- sify newborn pain presence, are used in pain assessment at bedside by healthy professionals and are being created in methods of artificial intelligence and machine learning to automate these processes. By this way, this conclusion paper has the objective of propose and implement a computational prototype to classification pain in parts of the newborn face, analyzing each part of the face separately and classify presence or not of pain. To make this, we have worked with some sequential tasks, starting with a preprocessing of the newborn image dataset provided by Federal University of São Paulo, where has been generated a segmented face of the newborn. In the second stage, training and creation of the framework are carried out, where these images will pass by the classification of pain or not pain, and 12 neural networks are the encharge of this classification, one neural network for each part of the newborn face to analyze and classify separately them. For the training and framework creating part were used 3 different models of neural network, VGG11, ResNet18 and LeNet-5. LeNet-5 show us a better classification of pain by parts from the face superior to the other 2 mentioned above, being the parts of the face with greater accuracy the Nose with 79.66% of accuracy, Left Eye with 79.44% of accu- racy and Mouth with 79.44% of accuracy. To obtain these results, we use a cross validation in training networks. Finally, was made the union of all phases, making an unprecedented and complete system of detection, segmentation and classification of pain in interest’s facial parts of newborns.
Nowadays, the newborn pain detection by images of the face is being in constantly de- velopment, having principally as objective slow down one of the biggest problems that we have in the neonatology area that is the effect of the pain in short, medium and long term in new- borns. It is known that the newborns in intensive therapy units are exposed to constantly pain and this fact can cause several damages to the neural development, affecting cerebral functions, stress system and lesions in the nociceptive processes. With the intention to identify and clas- sify newborn pain presence, are used in pain assessment at bedside by healthy professionals and are being created in methods of artificial intelligence and machine learning to automate these processes. By this way, this conclusion paper has the objective of propose and implement a computational prototype to classification pain in parts of the newborn face, analyzing each part of the face separately and classify presence or not of pain. To make this, we have worked with some sequential tasks, starting with a preprocessing of the newborn image dataset provided by Federal University of São Paulo, where has been generated a segmented face of the newborn. In the second stage, training and creation of the framework are carried out, where these images will pass by the classification of pain or not pain, and 12 neural networks are the encharge of this classification, one neural network for each part of the newborn face to analyze and classify separately them. For the training and framework creating part were used 3 different models of neural network, VGG11, ResNet18 and LeNet-5. LeNet-5 show us a better classification of pain by parts from the face superior to the other 2 mentioned above, being the parts of the face with greater accuracy the Nose with 79.66% of accuracy, Left Eye with 79.44% of accu- racy and Mouth with 79.44% of accuracy. To obtain these results, we use a cross validation in training networks. Finally, was made the union of all phases, making an unprecedented and complete system of detection, segmentation and classification of pain in interest’s facial parts of newborns.