Identificação de patologias em pregas vocais por meio de classificador bag of features e redes neurais convolucionais
Carregando...
Data
2021-10-05
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Patologias de laringe, como cistos e nódulos, são caracterizadas como alterações benignas e têm seu diagnóstico dado por meio de análise visual e perceptivo-auditiva em telelaringoscopias, exame endoscópico ambulatorial para inspeções laríngeas auxiliadas por vídeo. Porém, uma inspeção visual incorreta, que ocorre por similaridade vocal, acústica e também em suas propriedades anatômicas, pode levar o diagnóstico de uma das patologias benignas a prejudicar o diagnóstico de outra. Visto que há possibilidade de inversão na determinação diagnóstica entre essas duas patologias, pela equipe médica responsável, propõe-se uma ferramenta para auxílio diagnóstico que visa identificar e classificar imagens de pregas vocais saudáveis e não saudáveis - nódulos e cistos - advindas de telelaringoscopias e demais bancos de imagens, utilizando classificador inteligente Bag of Features e Redes Neurais Convolucionais - AlexNet, ResNet e DenseNet. Foi realizada segmentação de 45 imagens e aplicação de técnicas de Data Augmentation, passando a ter 300 imagens, divididas em três grupos entre teste e validação; sendo 100 imagens no grupo “normais”, 100 imagens no grupo “cistos” e 100 imagens no grupo “nódulos”. Ambos os grupos foram submetidos a 5 testes com 19 épocas cada um em todas as ferramentas utilizadas e alcançou-se o melhor resultados na CNN ResNet com acurácia de 97,33% de robustez e erro válido de 2,67% nos testes e classificação de pregas vocais saudáveis - normais e não saudáveis cistos e nódulos.
Laryngeal pathologies, such as cysts and nodules, are characterized as benign changes and have their diagnosis given through visual and auditory-perceptual analysis in telelaryngoscopy. Outpatient endoscopic examination for video-assisted laryngeal inspections. However, an incorrect visual inspection, which occurs due to vocal and acoustic similarity and also in their anatomical properties, can lead to the diagnosis of one of the benign pathologies affecting the diagnosis of the other. Since there is a possibility of inversion in the diagnostic determination between these two pathologies, by the responsible medical team, a diagnostic aid tool is proposed, which aims to identify and classify images of healthy and unhealthy vocal folds - nodules and cysts - arising from telelaryngoscopy and other image banks, using Bag of Features smart classifier and Convolutional Neural Networks - AlexNet, ResNet and DenseNet. 45 images were segmented and Data Augmentation techniques were applied, which now has 300 images, divided into three groups between test and validation; 100 images in the “normal” group, 100 images in the “cysts” group and 100 images in the “nodules” group. Both groups were submitted to 5 tests with 19 times each in all the tools used and the best results were achieved in CNN ResNet with an accuracy of 97.33% of robustness and valid error of 2.67% in the tests and classification of healthy vocal folds - normal and unhealthy cysts and nodules
Laryngeal pathologies, such as cysts and nodules, are characterized as benign changes and have their diagnosis given through visual and auditory-perceptual analysis in telelaryngoscopy. Outpatient endoscopic examination for video-assisted laryngeal inspections. However, an incorrect visual inspection, which occurs due to vocal and acoustic similarity and also in their anatomical properties, can lead to the diagnosis of one of the benign pathologies affecting the diagnosis of the other. Since there is a possibility of inversion in the diagnostic determination between these two pathologies, by the responsible medical team, a diagnostic aid tool is proposed, which aims to identify and classify images of healthy and unhealthy vocal folds - nodules and cysts - arising from telelaryngoscopy and other image banks, using Bag of Features smart classifier and Convolutional Neural Networks - AlexNet, ResNet and DenseNet. 45 images were segmented and Data Augmentation techniques were applied, which now has 300 images, divided into three groups between test and validation; 100 images in the “normal” group, 100 images in the “cysts” group and 100 images in the “nodules” group. Both groups were submitted to 5 tests with 19 times each in all the tools used and the best results were achieved in CNN ResNet with an accuracy of 97.33% of robustness and valid error of 2.67% in the tests and classification of healthy vocal folds - normal and unhealthy cysts and nodules
Descrição
Palavras-chave
Classificação de Imagens, Bag of Features, Redes Neurais Convolucionais, Patologia de Laringe, Image classification, Bag of features, Convolutional neural networks, Laryngeal pathology