É possível diferenciar tumores primários e metástases cerebrais utilizando análises de textura?
Carregando...
Data
2022-11-22
Autores
Guassu, Raissa Alexia Camargo
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A diferenciação entre tumores cerebrais primários e metástases solitária cerebral são de suma importância, pois requerem condutas de tratamentos diferentes. O diagnóstico definitivo em procedimentos invasivos, como a biópsia cerebral estereotáxica, apresenta certo risco aos pacientes. Dessa forma, alternativas para diagnóstico não invasivo estão sendo intensivamente estudadas para auxiliar na diferenciação dessas lesões, de modo a otimizar na relação risco benefício do paciente e custo para a instituição. Essa pesquisa, teve como objetivo a diferenciação entre tumores cerebrais primários e metástases solitária cerebral. Para atingir esse objetivo, foram utilizados métodos computacionais, para classificação e diferenciação dessas lesões cerebrais, com aprendizado de máquina (AM) aplicado em análises de texturas nas regiões de interesse (ROI). O banco de dados foi constituído a partir de exames retrospectivos de Ressonância Magnética (RM) de encéfalo de pacientes do Hospital das Clínicas de Botucatu (HCFMB)-UNESP. As sequências utilizadas foram: imagem ponderada em difusão (DWI), sequência de recuperação de inversão atenuada por fluido (FLAIR), imagem ponderada em T1 com e sem contraste endovenoso, imagem ponderada em T2. ROIs de 10x10 pixels e 32x32 pixels foram posicionadas nas imagens. Para cada ROI, 40 parâmetros de textura foram extraídos através do software Matlab® e aplicados aos seguintes métodos de AM: naive bayes, support vector machine (SVM), stochastic gradient descente (SGD), random forest (RF), tree e neural network (NN). Com a ROI de 10x10 pixels, foram obtidos diferenciação entre os casos estudados de até 97,5% de AUC usando o SVM na sequência DWI e 94,9% de F1 usando a SGD na sequência ponderada em T1. Para o estudo com ROI de 32x32 pixels, o melhor resultado alcançado foi na sequência T2 com uma AUC de 99,8%, obtendo uma excelente diferenciação e 98,5% de F1, ambos com o classificador NN. A ROI de 32x32 pixels apresentou resultados consistentes e ligeiramente superiores ao da ROI 10x10 pixels e menor variação entre os classificadores analisados. Essa pesquisa traz contribuições originais para a comunidade científica, além de apresentar um grande potencial técnico, como uma ferramenta para auxiliar no diagnóstico diferencial clínico na análise de imagens de ressonância magnética do sistema nervoso central.
Differentiating between primary brain tumors and solitary brain metastases is paramount, as they require different treatments. Definitive diagnosis in invasive procedures, such as stereotactic brain biopsy, poses a specific risk to patients. Thus, alternatives for non-invasive diagnosis are being studied to assist in the differentiation of these lesions, to optimize the risk-benefit ratio for the patient and cost for the institution. This research aimed to differentiate between primary brain tumors and solitary brain metastases. Computational methods can be used to classify and differentiate these brain lesions, with applied machine learning (AM) and texture analysis in regions of interest (ROI). The database was constituted from retrospective magnetic resonance imaging (MRI) scans of the brain of patients at Hospital das Clínicas de Botucatu (HCFMB)-UNESP. The sequences used were: diffusion-weighted image (DWI), fluid-attenuated inversion recovery sequence (FLAIR), T1-weighted image with and without intravenous contrast, and T2-weighted image. ROIs of 10x10 pixels and 32x32 pixels were placed on the images. For each ROI, 40 texture parameters were extracted using Matlab® software and applied to the following AM methods: naive bayes, support vector machine (SVM), stochastic gradient descent (SGD), random forest (RF), tree, and neural network (NN). With the ROI of 10x10 pixels, differentiation between the studied cases of up to 97.5% AUC using the SVM in the DWI sequence and 94.9% accuracy using the SGD in the T1-weighted sequence were obtained. For the study with ROI of 32x32 pixels, the best result was achieved in the T2 sequence with an AUC of 99.8%, obtaining excellent differentiation and 98.5% accuracy, both with the NN classifier. The ROI of 32x32 pixels presented consistent results, was slightly superior to the ROI of 10x10 pixels, and showed minor variation between the analyzed classifiers. This research brings original contributions to the scientific community and presents great technical potential as a tool to assist in the clinical differential diagnosis in the analysis of magnetic resonance images of the central nervous system.
Differentiating between primary brain tumors and solitary brain metastases is paramount, as they require different treatments. Definitive diagnosis in invasive procedures, such as stereotactic brain biopsy, poses a specific risk to patients. Thus, alternatives for non-invasive diagnosis are being studied to assist in the differentiation of these lesions, to optimize the risk-benefit ratio for the patient and cost for the institution. This research aimed to differentiate between primary brain tumors and solitary brain metastases. Computational methods can be used to classify and differentiate these brain lesions, with applied machine learning (AM) and texture analysis in regions of interest (ROI). The database was constituted from retrospective magnetic resonance imaging (MRI) scans of the brain of patients at Hospital das Clínicas de Botucatu (HCFMB)-UNESP. The sequences used were: diffusion-weighted image (DWI), fluid-attenuated inversion recovery sequence (FLAIR), T1-weighted image with and without intravenous contrast, and T2-weighted image. ROIs of 10x10 pixels and 32x32 pixels were placed on the images. For each ROI, 40 texture parameters were extracted using Matlab® software and applied to the following AM methods: naive bayes, support vector machine (SVM), stochastic gradient descent (SGD), random forest (RF), tree, and neural network (NN). With the ROI of 10x10 pixels, differentiation between the studied cases of up to 97.5% AUC using the SVM in the DWI sequence and 94.9% accuracy using the SGD in the T1-weighted sequence were obtained. For the study with ROI of 32x32 pixels, the best result was achieved in the T2 sequence with an AUC of 99.8%, obtaining excellent differentiation and 98.5% accuracy, both with the NN classifier. The ROI of 32x32 pixels presented consistent results, was slightly superior to the ROI of 10x10 pixels, and showed minor variation between the analyzed classifiers. This research brings original contributions to the scientific community and presents great technical potential as a tool to assist in the clinical differential diagnosis in the analysis of magnetic resonance images of the central nervous system.
Descrição
Palavras-chave
Tumores cerebrais primários, Metástase solitária cerebral, Parâmetros de textura, Aprendizado de máquina, , ,