Segmentação Interativa de Imagens de tomografia computadorizada pulmonar com COVID-19 via aprendizado profundo
Carregando...
Data
2021-06-02
Autores
Bruzadin, Aldimir José
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Neste trabalho, apresentamos um novo método de segmentação interativa de imagens baseado em aprendizado profundo, com foco em imagens médicas do pulmão de pacientes acometidos pela Covid-19. Inicialmente, aplicamos uma pré-segmentação nas imagens, utilizando-se para tal o modelo de Mistura Gaussiana visando predizer os pesos das bordas de cada imagem para posterior uso no algoritmo denominado Random Walker, que foi integrado como parte de uma arquitetura inteligente de Redes Neurais Convolucionais. Além disso, no processo de aprendizado, empregamos o método do Gradiente Descendente em conjunto com duas técnicas de simplificação, a saber: Gradiente Esparso e Poda do Gradiente, reduzindo assim a quantidade de cálculos a serem realizados durante a etapa de backpropagation no pipeline de aprendizado proposto. Já para a obtenção das segmentações após o aprendizado dos pesos das bordas, e tendo em mente que representações 3D, de Tomografia Computadorizada, podem chegar a reunir centenas de imagens 2D (slices), o método emprega apenas um mapa de pixels rotulados pelo usuário como entrada a fim de criar automaticamente novos mapas de rotulação para os demais slices 2D do órgão do paciente. Resultados experimentais e comparações com outras técnicas da literatura atestam a alta performance e acurácia da metodologia de segmentação de imagens proposta.
In this dissertation, we present a new interactive image segmentation method based on deep learning for pulmonary medical images, from patients affected by Covid-19. First, we apply a pre-segmentation step on the input images by using the so-called Gaussian Mixture Model in order to predict the edge weights of the images, which are used by the Random Walker algorithm coupled with an integrated Convolutional Neural Network-based architecture. During the learning process, we take the Descending Gradient method in conjunction with two simplification techniques: Sparse Gradient and Gradient Pruning, to reduce the amount of calculations in the back-propagation stage of the proposed learning pipeline. In order to obtain the segmentations after learning the edge weights, and keeping in mind that 3D representations from Computed Tomography can usually reach hundreds of slices, we employ only a user-labeled map of pixels as input data to automatically create new seed maps to all other slices of the patient’s organ. Experimental results and comparisons against existing techniques attest to the high performance and accuracy of the proposed image segmentation methodology.
In this dissertation, we present a new interactive image segmentation method based on deep learning for pulmonary medical images, from patients affected by Covid-19. First, we apply a pre-segmentation step on the input images by using the so-called Gaussian Mixture Model in order to predict the edge weights of the images, which are used by the Random Walker algorithm coupled with an integrated Convolutional Neural Network-based architecture. During the learning process, we take the Descending Gradient method in conjunction with two simplification techniques: Sparse Gradient and Gradient Pruning, to reduce the amount of calculations in the back-propagation stage of the proposed learning pipeline. In order to obtain the segmentations after learning the edge weights, and keeping in mind that 3D representations from Computed Tomography can usually reach hundreds of slices, we employ only a user-labeled map of pixels as input data to automatically create new seed maps to all other slices of the patient’s organ. Experimental results and comparisons against existing techniques attest to the high performance and accuracy of the proposed image segmentation methodology.
Descrição
Palavras-chave
Segmentação de imagens, Inteligência computacional, Aprendizado profundo, Covid-19, Imagens médicas, Image segmentation, Computational intelligence, Deep learning, Medical images