Campus de Botucatu Instituto de Biociências PG-BGA UNIVERSIDADE ESTADUAL PAULISTA “Júlio de Mesquita Filho” INSTITUTO DE BIOCIÊNCIAS DE BOTUCATU “ANÁLISE DE TEXTURA: UMA POTENCIAL FERRAMENTA PARA DIFERENCIAR TUMORES CEREBRAIS PRIMÁRIOS E METÁSTASE CEREBRAL SOLITÁRIA” Sérgio Augusto Santana De Souza Discente Professora Associada Diana Rodrigues de Pina Departamento de Infectologia, Dermatologia, Diagnóstico por Imagem e Radioterapia da Faculdade de Medicina, UNESP (Campus de Botucatu) Orientadora Professor Associado Fabiano Reis Departamento de Anestesiologia, Oncologia e Radiologia da Faculdade de Ciências Médicas, UNICAMP Coorientador Tese apresentada ao Programa de Pós-Graduação em Biologia Geral e Aplicada do Instituto de Biociências, Campus de Botucatu, UNESP, como parte dos requisitos exigidos para a obtenção do título de Doutor em Biologia Geral e Aplicada, na área de concentração de Biologia Celular Estrutural e Funcional. BOTUCATU – SP 2022 Campus de Botucatu Instituto de Biociências PG-BGA SÉRGIO AUGUSTO SANTANA DE SOUZA “Análise De Textura: Uma Potencial Ferramenta Para Diferenciar Tumores Cerebrais Primários e Metástase Cerebral Solitária” Tese apresentada ao Programa de Pós-Graduação em Biologia Geral e Aplicada do Instituto de Biociências, Campus de Botucatu, UNESP, como parte dos requisitos exigidos para a obtenção do título de Doutor em Biologia Geral e Aplicada, na área de concentração de Biologia Celular Estrutural e Funcional. BOTUCATU – SP 2022 S729a Souza, Sérgio Augusto Santana de Análise de textura: uma potencial ferramenta para diferenciar tumores cerebrais primários e metástase cerebral solitária / Sérgio Augusto Santana de Souza. -- Botucatu, 2022 93 p. Tese (doutorado) - Universidade Estadual Paulista (Unesp), Instituto de Biociências, Botucatu Orientadora: Diana Rodrigues de Pina Coorientador: Fabiano Reis 1. tumores cerebrais primários. 2. metástase solitária cerebral. 3. parâmetros de textura. 4. aprendizado de máquina. I. Título. Sistema de geração automática de fichas catalográficas da Unesp. Biblioteca do Instituto de Biociências, Botucatu. Dados fornecidos pelo autor(a). Essa ficha não pode ser modificada. Campus de Botucatu Instituto de Biociências PG-BGA Agradecimentos Campus de Botucatu Instituto de Biociências PG-BGA Aos meus pais, Sérgio, Raimunda e meus familiares que mesmo frente a todas as dificuldades enfrentadas, sempre me apoiaram, acreditaram, incentivaram e investiram no meu crescimento pessoal e profissional. Obrigada por todo apoio, dedicação, incentivo e amor ao longo de toda vida. Sem eles não estaria onde estou hoje ou realizaria metade do que estou realizando. Amo vocês incondicionalmente! À minha companheira Isabel. Obrigado por todo apoio, amizade, companheirismo e irmandade durante todos estes anos. Obrigada pelas risadas, casa compartilhada, conselhos, conversas, e por todo amor. Sem você também não estaria onde estou hoje. À minha orientadora, Professora Associada Diana Rodrigues de Pina, por todo apoio, dedicação e oportunidades oferecidas e, além de tudo, por sempre acreditar em mim e não ter desistido. Foram muitos anos de aprendizado, crescimento profissional, intelectual e científico. Ao Professor José Ricardo de Arruda Miranda, ao professor Fabiano Reis, meus colegas de Laboratório LAFAR, Allan Felipe Fattori Alves, Raissa Alexia Guassu, Letícia Cotinguiba Silva, Matheus Alvarez, aos meus colegas do Laboratório de Biomagnetismo e todos que tem me apoiado, auxiliado e aconselhado a longo dessa caminhada. Obrigada pelo carinho, compreensão, cumplicidade, dedicação e, especialmente, por toda paciência ao me acompanhar durante esta jornada. O presente trabalho foi realizado com apoio da Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Código de Financiamento 001 Campus de Botucatu Instituto de Biociências PG-BGA Epígrafe “Quando a educação não é libertadora, o sonho do oprimido é ser o opressor”. Paulo Freire Campus de Botucatu Instituto de Biociências PG-BGA Abreviaturas Campus de Botucatu Instituto de Biociências PG-BGA AM - Aprendizado de Máquinas AT – Análise de Texturas RM – Ressonância Magnética FLAIR – Recuperação de Inversão Fluído Atenuada do inglês “Fluid-attenuated inversion recovery” ROI- Região de Interesse do inglês “Region of interest” SVM – Máquinas de Vetor de Suporte do inglês “Support Vector Machine” SGD – Gradiente Descendente Estocástico do inglês “Stochastic gradient descente” ROC – Característica de Operação do Receptor do inglês “Receiver operator characteristic curve” DWI – Imagem Ponderada em Difusão do inglês “Diffusion-weighted imaging” ML – Machine Learning TA – Texture analyses MRI – Magnetic Resonance Imaging LAFAR – Laboratório de Física Aplicada ao Radiodiagnóstico SNC – Sistema Nervoso Central GBM – Glioblastoma Multiforme MET – Metástase RF - Ondas de Rádio Frequência TR- Tempo de Repetição TE- Tempo de Eco T1 – Tempo de Relaxação spin-lattice T2- Tempo de Relaxação spin-spin DP – Densidade de Prótons GLCM - Gray-level Co-occurrence Matrix GLRL - Gray-level Run-length Campus de Botucatu Instituto de Biociências PG-BGA SRE - short run emphasis LRE - long runs emphasis GLN - gray level nonuniformity RP - run percentage RLN - run length nonuniformity LGRE - low gray level run emphasis HGRE - high gray level run emphasis DICOM - Digital Imaging and Communications in Medicine Campus de Botucatu Instituto de Biociências PG-BGA Resumo Campus de Botucatu Instituto de Biociências PG-BGA Tumores cerebrais são considerados um grave problema de saúde pública e sua diferenciação constitui um grande desafio diagnóstico. Geralmente, o diagnóstico definitivo só pode ser alcançado a partir de uma biópsia estereotáxica do cérebro, o que acarreta certo risco de mortalidade. Assim, é necessário o desenvolvimento de estratégias alternativas para a diminuição de procedimentos invasivos. Nesse aspecto, o aprendizado de máquinas (AM) aplicado a imagens biomédicas por meio de análise de texturas (AT) tem sido amplamente utilizado na medicina diagnóstica para classificação ou diferenciação de lesões. Essa pesquisa teve como objetivo a criação de um método de extração de texturas e aprendizado de máquina aplicado a neuroimagens para diferenciação de tumores cerebrais. O banco de dados foi constituído a partir de exames de Ressonância Magnética (RM) de encéfalo de pacientes do Hospital das Clínicas de Botucatu. O bando de dados foi dividido em dois grupos, tumores primários (n = 38) e metástase cerebral solitária (n = 58). 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. Radiologistas com mais de 15 anos de experiência realizaram na região tumoral, segmentações manuais nas regiões de interesse (ROI) de 10 x 10 pixels. Para cada ROI, 40 parâmetros de textura foram extraídos através do software Matlab® (v 2017) e aplicados a cinco métodos de AM diferentes: naive bayes, support vector machine (SVM), stochastic gradient descente (SGD), random forest, e tree. Os métodos de AM classificaram os grupos com boa diferenciação de até 97,5% da área sob as características do operador receptor (ROC) usando o SVM na sequência DWI e 94,9% de precisão usando o gradiente descendente estocástico na sequência ponderada em T1. O método proposto apresentou uma classificação confiável para a investigação de lesões tumorais e trouxe contribuições originais à comunidade cientifica com grande potencial de suporte ao diagnóstico clínico diferencial. Palavras-chave: tumores cerebrais primários; metástase solitária cerebral; parâmetros de textura, aprendizado de máquina. Campus de Botucatu Instituto de Biociências PG-BGA Abstract Campus de Botucatu Instituto de Biociências PG-BGA Brain tumors are considered a serious public health problem and their differentiation constitutes a major diagnostic challenge. Generally, a definitive diagnosis can only be reached from a stereotaxic brain biopsy, which carries a certain risk of mortality. Thus, it is necessary to develop alternative strategies to reduce invasive procedures. In this aspect, machine learning (ML) applied to biomedical images through texture analysis (TA) has been widely used in diagnostic medicine for classification or differentiation of lesions. In this way, we propose a texture extraction and machine learning method applied to neuroimaging to differentiate brain tumors. The database consisted of images from routine exams at “Hospital das Clínicas de Botucatu”. We used magnetic resonance imaging (MRI) scans of the brain. The dataset was divided into two groups, primary tumors (n=38) and solitary brain metastasis (n=58). The sequences used were: diffusion-weighted image (DWI), fluid-attenuated inversion recovery (FLAIR), T1-weighted, T1-weighted SE gadolinium-enhanced, and T2-weighted images. Radiologists with more than 15 years of experience performed, in tumoral regions, manual segmentation of regions of interest (ROI) of 10 x 10 tumor pixels. For each ROI, 40 texture parameters were extracted using Matlab® R2020b (license: 40962010) and applied to five different AM methods: naive bayes, support vector machine (SVM), stochastic gradient descent (SGD), random forest, and tree. The ML methods classified the groups with good differentiation of up to 97.5% of the area under the characteristics of the receiver operator (ROC). using the SVM on the DWI sequence and 94.9% accuracy using the stochastic gradient descent on the T1-weighted sequence. The proposed method presented a reliable classification for the investigation of tumoral lesions and brought original contributions to the scientific community with great potential to aid in the differential clinical diagnosis. Keywords: primary brain tumors; solitary brain metastasis; texture parameters; machine learning. Campus de Botucatu Instituto de Biociências PG-BGA Prefácio Campus de Botucatu Instituto de Biociências PG-BGA A presente tese de doutorado possui estrutura e formato divididos em capítulos. O primeiro capítulo destina-se a introdução, justificativa e relevância dos temas e os princípios que constituem a elaboração do artigo científico. O segundo capítulo destina-se aos fundamentos teóricos que fornecem todos os subsídios necessários para que o leitor possa compreender os objetivos, etapas, discussões e conclusões obtidas na presente tese. O terceiro capítulo destina-se aos objetivos do artigo científico. No quarto capítulo, apresentaremos os materiais e a metodologia abordada no artigo científico, previamente submetido. No quinto capítulo, apresentaremos no anexo I, as diretrizes de submissão para o formato exigido pela revista “computer methods in biomechanics and biomedical engineering: imaging & visualization”, ISSN Online: 2168-1171, editora Taylor & Francis Online, para submissão do artigo científico, avaliação por pares e posterior publicação. O sexto capítulo destina-se a elaboração dos resultados científicos em formato de artigo científico, na qual o aluno participou durante o seu período de doutorado, junto ao grupo LAFAR (Laboratório de Física Aplicada ao Radiodiagnóstico). O sétimo capítulo destina-se as conclusões gerais e específicas do doutorado do aluno, dos resultados presentes no artigo científico elaborado e as referências. No oitavo capítulo, Anexo II, apresentaremos uma descrição expandida dos processos envolvidos no funcionamento e construção das técnicas em ressonância magnética. No nono capítulo, apresentaremos as referências da presente tese. No décimo capítulo, apresentaremos o parecer substanciado do Comitê de Ética em Pesquisa. Campus de Botucatu Instituto de Biociências PG-BGA No décimo primeiro e último capítulo, apresentaremos os anexos III, IV, V, VI, VII, VIII e IX, referentes aos artigos científicos publicados em periódicos, em que o aluno participou durante sua jornada no doutoramento. Campus de Botucatu Instituto de Biociências PG-BGA Sumário Campus de Botucatu Instituto de Biociências PG-BGA Sumário Capítulo I ................................................................................................................................................... 12 Introdução, Justificativa e Relevância do Tema ...................................................................................... 13 Capítulo II .................................................................................................................................................. 15 Fundamentos Teóricos ............................................................................................................................................ 16 Tumores Primários ............................................................................................................................................. 16 Metástase Solitária Cerebral ............................................................................................................................... 18 Ressonância Magnética ...................................................................................................................................... 19 Textura ............................................................................................................................................................... 20 Gray-level Co-occurence Matrix .................................................................................................................. 21 Gray-level Run-length ................................................................................................................................. 22 Transformadas Wavelets ............................................................................................................................. 22 Capítulo III ................................................................................................................................................ 26 Objetivos ................................................................................................................................................. 27 Capítulo IV ................................................................................................................................................. 28 Materiais e Métodos ................................................................................................................................ 29 Equipamento de Ressonância Magnética (RM) ................................................................................................. 29 Banco de Dados.................................................................................................................................................. 29 Infraestrutura ...................................................................................................................................................... 29 Capítulo V .................................................................................................................................................. 30 Anexo I: Diretrizes de Submissão ..................................................................................................................... 31 Capítulo VI ................................................................................................................................................. 37 Artigo Científico: Texture Analysis: A Potential Tool to Differentiate Primary Brain Tumors and Solitary 36 Title Page ........................................................................................................................................................... 38 Abstract .............................................................................................................................................................. 39 Introduction ........................................................................................................................................................ 40 Material and Methods ........................................................................................................................................ 41 Database ........................................................................................................................................................ 42 MRI Acquisition Parameters ......................................................................................................................... 43 Region of Interest Positioning and Feature Extraction ................................................................................. 44 Machine Learning ......................................................................................................................................... 46 Results ................................................................................................................................................................ 47 Discussion .......................................................................................................................................................... 50 Conclusions ........................................................................................................................................................ 52 Statements & Declarations ................................................................................................................................. 53 Article References .............................................................................................................................................. 54 Campus de Botucatu Instituto de Biociências PG-BGA Capítulo VII ............................................................................................................................................... 58 Conclusões ............................................................................................................................................................. 59 Capítulo VIII ............................................................................................................................................. 60 Anexo II: Ressonância Magnética Expandida ......................................................................................................... 61 Capítulo IX ................................................................................................................................................. 75 Referências .............................................................................................................................................................. 76 Capítulo X .................................................................................................................................................. 79 Parecer substância do Comitê de Ética em Pesquisa ............................................................................................... 80 Capítulo XI ................................................................................................................................................. 85 Anexo III ................................................................................................................................................................. 86 Anexo IV ................................................................................................................................................................. 87 Anexo V .................................................................................................................................................................. 88 Anexo VI ................................................................................................................................................................. 89 Anexo VII ............................................................................................................................................................... 90 Anexo VIII .............................................................................................................................................................. 91 Anexo IX ................................................................................................................................................................. 92 Fim .............................................................................................................................................................. 93 12 Campus de Botucatu Instituto de Biociências PG-BGA Capítulo I 13 Campus de Botucatu Instituto de Biociências PG-BGA 1. Introdução, Justificativa e Relevância do Tema O dia a dia da atividade médica é marcado por uma busca constante de um diagnóstico preciso e rápido para uma boa eficácia da terapêutica. Nesse contexto, doenças que merecem grande atenção são as que acometem o sistema nervoso central, consideradas um grave problema de saúde pública (Siegel et al., 2020). Dentre estas patologias, tumores cerebrais são algumas das mais letais para os pacientes, sendo que tumores cerebrais são a primeira causa de mortalidade por câncer para homens com idade até 40 anos e para mulheres com idade até 20 anos nos Estados Unidos da América (Siegel et al., 2020). A diferenciação de diversos tipos de tumores cerebrais é de suma importância, pois resultam em condutas médicas distintas, no entanto, constituem um grande desafio diagnóstico para especialistas da área de radiologia (Zacharaki et al., 2009) e são algumas das que mais necessitam de métodos auxiliares para o diagnóstico diferencial. Em muitos casos, o diagnóstico definitivo só pode ser alcançado a partir de uma biópsia estereotáxica do cérebro o que acarreta alto risco de mortalidade (Blanchet et al., 2011). Assim, o desenvolvimento de estratégias alternativas para o diagnóstico dessas doenças traz benefícios evidentes através da diminuição de procedimentos invasivos, sendo que o diagnóstico correto e precoce é de grande importância para a evolução favorável do paciente. Nesse aspecto, as modalidades de diagnóstico por imagem são essenciais para localização, determinação da etiologia e acompanhamento dessas doenças. A ressonância magnética (RM) é uma das mais promissoras modalidades devido a sua sensibilidade na diferenciação de tecidos. A RM possui dezenas de sequências e modos de aquisição de imagem, utilizadas no diagnóstico de patologias no sistema nervoso central. Entretanto, para alguns tipos de neoplasias, os achados neurológicos em RM possuem características semelhantes. Principalmente quando se procura a diferenciação entre tumores primários cerebrais e metástases solitárias cerebrais (Blanchet et al., 2011; Mouthuy et al., 2012). 14 Campus de Botucatu Instituto de Biociências PG-BGA Uma abordagem de método computacional, como o aprendizado de máquinas (AM), tem sido amplamente utilizada na medicina diagnóstica para classificação ou diferenciação de lesões (Bertsimas and Dunn, 2017; Senanayake et al., 2019; Turkki et al., 2019; Wildeboer et al., 2019). O AM é comumente aplicado a imagens biomédicas por meio de análises de recursos radiômicos quantitativos, como os extraídos por análise de texturas (AT), refletindo a heterogeneidade de uma Região de Interesse (ROI). A análise de textura estuda o comportamento de um atributo conhecido como textura que corresponde a um padrão visual de informação ou arranjo de estrutura, geralmente relacionado à distribuição de pixels em uma região de imagens digitais. Texturas geralmente contém informações significativas sobre o conteúdo da imagem, e sua análise já demonstrou potencial considerável na área de neuroimagens como uma estratégia objetiva para classificação e caracterização de lesões (Hwang et al., 2016; Kassner and Thornhill, 2010; Mohanty et al., 2012; Santos et al., 2015). Estas técnicas são úteis particularmente nos casos em que as regiões lesionadas não são distinguíveis das sadias com base nos protocolos padrão como os de ponderação em T1 e T2 (El-Zein R, 2002), atenuação de prótons ou de imagem ponderada em difusão (Kassner and Thornhill, 2010). Métodos de aprendizado de máquina (ML) podem ser aplicados para analisar características de texturas e têm sido usados em medicina diagnóstica para classificar ou diferenciar tumores cerebrais primários e metástases cerebrais solitárias (Zacharaki et al., 2009). Ainda há carência de artigos na literatura sobre a distinção entre tumores cerebrais primários e metástases cerebrais solitárias associadas a características de textura. Assim, este trabalho propõe uma abordagem inovadora para diferenciar tumores cerebrais primários e metástases cerebrais solitárias usando aprendizado de máquina aplicado a recursos de textura. 15 Campus de Botucatu Instituto de Biociências PG-BGA Capítulo II 16 Campus de Botucatu Instituto de Biociências PG-BGA 1. Fundamentos Teóricos 1.1 Tumores Primários Os tumores do sistema nervoso central (SNC) são responsáveis por 1,5% de todos os cânceres e 2,4% de todas as mortes relacionadas ao câncer anualmente (El-Zein R, 2002; MA, 2007). Os tumores cerebrais primários afetam entre 11 e 19 por 100.000 indivíduos (Health, 2017), enquanto as metástases cerebrais afetam 11 em cada 100.000 indivíduos da população em geral. Além disso, estudos em necropsia mostram que até 25% dos pacientes com neoplasias tem metástases no SNC (Cambruzzi et al., 2011). Desde meados da década de 1980, estudos epidemiológicos apontam para um aumento da incidência de tumores do SNC em diferentes países, sendo mais pronunciada em indivíduos idosos (Baldi et al., 2011). Algumas hipóteses para explicar essa observação incluem a disseminação de exames de tomografia computadorizada e doenças infecciosas (como AIDS), além de outras que ainda carecem de embasamento sólido. Tanto a incidência quanto a mortalidade, associadas a esses tumores têm aumentado na maioria dos países desenvolvidos, principalmente nos grupos com idade mais avançada. Na maioria desses países, a mortalidade associada a essas neoplasias ocupa a 12ª posição de todas as mortes (Dolecek et al., 2012). A localização supratentorial é o sítio SNC mais frequente dos tumores em adultos, sendo os gliomas (astrocitomas, ependimomas, oligodendrogliomas) responsáveis por mais de 30% dos tumores neste grupo, seguidos pelos meningiomas e schwannomas (MA, 2007). Embora a incidência de tumores do SNC seja menor que a de outros tumores, os tumores cerebrais assumem grande importância na oncologia e se destacam no contexto de novas pesquisas e tratamentos, devido aos grandes avanços nos campos genético e molecular de tratamento obtidos com novas descobertas sobre a etiologia desses tumores (Jung et al., 2020). Os gliomas são o tipo mais comum e potencialmente mais agressivo de tumores cerebrais primários em adultos. O glioblastoma multiforme (GBM) é a forma mais maligna de glioma. Seu padrão de crescimento específico é caracterizado por uma infiltração extensa e difusa de células 17 Campus de Botucatu Instituto de Biociências PG-BGA tumorais no neurópilo (ou seja, a rede atenuada de processos celulares neuronais e gliais entrelaçados). Esse aspecto é um fator importante no insucesso terapêutico cirúrgico (Mangiola et al., 2008). O crescimento de metástases cerebrais é um processo complexo de múltiplos estágios, mediado por mecanismos moleculares. As células cancerosas devem se transformar, crescer e ser transportadas para o sistema nervoso central, onde podem permanecer adormecidas antes de invadir e crescer ainda mais (Gavrilovic and Posner, 2005). Os sintomas clínicos associados ao GBM são semelhantes aos da metástase (MET) cerebral. No entanto, cada tipo de tumor tem uma natureza biológica diferente e, portanto, requer diferentes estratégias de tratamento (Claes et al., 2007; Giese et al., 2003). Na maioria dos casos, o diagnóstico definitivo é baseado na análise histopatológica de uma amostra de biópsia. No entanto, para alguns casos, a diferenciação de MET e GBM pode ser feita com RM convencional. Por exemplo, uma massa solitária abrangendo o corpo caloso e ambos os lobos frontais não pode ser um MET. Em outros casos, não é possível diferenciar esses dois tipos de tumores porque a visualização radiológica da frente invasiva dos glioblastomas difusos é muito difícil. Geralmente, o GBM é um tumor solitário e o MET não, mas há exceções a essa regra. Nesses casos, é difícil distinguir GBM e MET com base na RM. O uso de métodos não invasivos é preferível ou às vezes obrigatório quando a biópsia é impossível. Este é o caso quando a massa está localizada perto de uma área eloquente ou quando a idade avançada do paciente torna o procedimento muito perigoso. Uma melhor diferenciação dos dois tipos de tumor por métodos não invasivos seria, portanto, benéfica (Blanchet et al., 2011). 18 Campus de Botucatu Instituto de Biociências PG-BGA 1.2 Metástase Solitária cerebral As metástases cerebrais são as neoplasias intracranianas mais comuns no adulto, com mais de 100.000 novos casos diagnosticados a cada ano nos Estados Unidos (Nathoo et al., 2004). A incidência de metástases cerebrais está aumentando, não apenas como resultado do uso de terapias combinadas avançadas contra o tumor primário e da diminuição das taxas de mortalidade, mas também como resultado das novas técnicas de neurorradiologia (Barnholtz-Sloan et al., 2004; Gavrilovic and Posner, 2005). Qualquer tumor primário pode potencialmente metastatizar para o cérebro, e aproximadamente 10% dos pacientes com câncer desenvolvem metástases cerebrais (Kamar and Posner, 2010). Os tumores com maior tendência a metástases para a região intracraniana são, em ordem decrescente, câncer de pulmão, câncer de mama, melanoma, câncer colorretal e câncer urogenital(Le Chevalier et al., 1985). A via mais comum de disseminação para o cérebro é a hematogênica, particularmente através da artéria cerebral média (Adam et al., 2006). As vias menos comuns são pelo líquido cefalorraquidiano ou por contiguidade. Ao diagnóstico, 50% das metástases cerebrais são solitárias e surgem nos seguintes locais: hemisfério cerebral (60%), cerebelo (30%) e tronco cerebral (10%) (Jung et al., 2002). Quando solitárias e na ausência de uma primária conhecida, as metástases cerebrais entram no diagnóstico diferencial com uma ampla gama de condições, mais comumente com gliomas de alto grau, linfomas e abscessos cerebrais e, menos comumente, com lesões inflamatórias/desmielinizantes e lesões isquêmicas. Graças às suas capacidades multiparamétricas e multiplanares, resolução de alto contraste, invasividade limitada e uso de energia não ionizante, a ressonância magnética é a modalidade de escolha no estudo de lesões cerebrais expansivas em geral e de metástases em particular. No entanto, a RM nem sempre permite a caracterização completa da lesão, com consequências óbvias nas decisões de tratamento. 19 Campus de Botucatu Instituto de Biociências PG-BGA 1.3 Ressonância Magnética A ressonância magnética (RM) é uma das mais promissoras modalidades devido a sua sensibilidade na diferenciação de tecidos. A RM possui dezenas de sequências e modos de aquisição de imagem, utilizadas no diagnóstico de patologias no sistema nervoso central. Entretanto, usando RM convencional (definida como a obtenção de imagens estruturais de corte transversal que fornecem predominantemente informações anatômicas) a diferenciação entre tumores primários cerebrais e metástases solitárias cerebrais é frequentemente difícil, e em muitos casos requerem biópsia ou imagem de acompanhamento(Blanchet et al., 2011; Mohanty et al., 2012) (Blanchet et al., 2011; Mouthuy et al., 2012). Na figura 1, podemos visualizar imagens de RM de diferentes tipos de tumores cerebrais, da esquerda para a direita: meningioma, glioma grau II, grau III, grau IV e metástase (Zacharaki et al., 2009). Podemos perceber que até para radiologistas experientes, essas lesões tumorais podem se tornar um desafio diagnóstico. Desse modo, análises de recursos extraídas da imagem podem auxiliar a ressonância magnética no diagnóstico clínico dessas lesões tumorais. Na 1ª linha da Figura 1, podemos observar: imagem T1 com a região tumoral de interesse. E na 2ª linha da Figura: imagem FLAIR (ampliada na região do tumor) sobreposta com um dos padrões texturais. Esses padrões texturais obtidos na imagem podem ser utilizados em métodos computacionais como o aprendizado de máquina e utilizados em técnicas de classificação para diferenciação de lesões tumorais (Blanchet et al., 2011; Mohanty et al., 2012; Zacharaki et al., 2009). No oitavo capítulo, apresentaremos uma descrição expandida dos processos envolvidos no funcionamento e construção das técnicas em ressonância magnética. 20 Campus de Botucatu Instituto de Biociências PG-BGA Figura 1. Imagens de RM de diferentes tipos de tumores cerebrais e exemplos de imagens de texturas extraídas de área edemaciada. Da esquerda para a direita: meningioma, glioma grau II, grau III, grau IV e metástase. 1ª linha: imagem T1 com a região tumoral de interesse. 2ª linha: imagem FLAIR (ampliada na região do tumor) sobreposta com um dos padrões texturais (λ = 8) utilizados no estudo de (Zacharaki et al., 2009). 1.4 Textura Podemos considerar textura as variações locais das intensidades de pixels que se repetem de maneira regular ou aleatória (Mohanty et al., 2012). A textura pode ser definida como um arranjo repetitivo de padrões sobre uma área que depende dos níveis de cinza nesta região (Haralick et al., 1973). Existem diversos métodos para extração de texturas de imagens: estrutural, espectral e estatístico. As texturas estatísticas consideram a distribuição de tons de cinza e sua inter-relação (Haralick et al., 1973). Os recursos de textura geralmente são extraídos de regiões de interesse (ROI). A seguir um exemplo dos possíveis processos envolvidos durante a análise de texturas de uma imagem. A partir da imagem original pode-se aplicar segmentação/realce manual ou automática de estruturas. Na região segmentada/realçada extrai-se diferentes texturas as quais são analisadas para depois serem utilizadas em um sistema de aprendizado de máquina. Os métodos de textura mais utilizados são: Gray-level Co-occurrence Matrix e Gray-level Run-length. 21 Campus de Botucatu Instituto de Biociências PG-BGA 1.4.1 Gray-level Co-occurence Matrix Gray-level Co-occurancy Matrix (GLCM) é um método estatístico de textura que considera a relação espacial dos pixels (Haralick et al., 1973) . O GLCM determina a probabilidade com que um pixel de certo nível de cinza ocorra adjacente a um pixel com outro valor de cinza em uma determinada região da imagem. Estas probabilidades dependem de dois parâmetros: distância inter-pixels (δ) e orientação (θ). A medida de probabilidade pode ser definida a partir das Equações 1 e 2. (Haralick et al., 1973) . Pr(𝑥) = {𝐶𝑖𝑗|(𝛿, 𝜃)} (Equação 1) Onde Cij é definida conforme descreve a equação 2. 𝐶𝑖𝑗 = 𝑃𝑖𝑗 ∑ 𝑃𝑖𝑗 𝐺 𝑖,𝑗=1 (Equação 2) Onde Pij representa o número de ocorrências de níveis de cinza i e j dentro de certo intervalo, para uma combinação de (δ, θ); e G é o número de níveis de cinza. A soma no denominador representa o número total de pares de níveis de cinza (i, j) dentro deste intervalo. A partir destas variáveis, é possível determinar muitas grandezas de texturas nas imagens conforme descrito por (Haralick et al., 1973). Neste trabalho serão determinadas grandezas de GLCM tais como uniformidade, entropia, dissimilaridade, contraste, correlação, energia, homogeneidade, momento de diferença, entre outros. Na Figura 2A, podemos observar a construção de um GLCM simples a partir de uma imagem estilizada de 3 × 3 pixels com três níveis de cinza (G = 3); distância, δ = 1 pixel; e direção, θ = 0° (ou seja, apenas esquerda ou direita). Começamos identificando o pixel de referência, com nível de cinza i, e contando o número de pixels vizinhos com nível de cinza j. A Figura 2B, por exemplo, olhando para a segunda entrada na primeira linha do GLCM e contando n (0,1), vemos 2 pixels com nível de cinza = 0 que têm vizinhos destros ou canhotos com nível de cinza = 1. Se olhamos para a entrada para (i,j) =(1,2) contamos 1 pixel com nível de cinza = 1, que tem um vizinho destro ou canhoto com nível de cinza = 2. 22 Campus de Botucatu Instituto de Biociências PG-BGA Figura 2. A, podemos observar a construção de um GLCM simples a partir de uma imagem estilizada de 3 × 3 pixels com três níveis de cinza (G = 3); distância, δ = 1 pixel; e direção, θ = 0° (ou seja, apenas esquerda ou direita). Começamos identificando o pixel de referência, com nível de cinza i, e contando o número de pixels vizinhos com nível de cinza j. B, por exemplo, olhando para a segunda entrada na primeira linha do GLCM e contando n (0,1), vemos 2 pixels com nível de cinza = 0 que têm vizinhos destros ou canhotos com nível de cinza = 1. Se olhamos para a entrada para (i,j) =(1,2) contamos 1 pixel com nível de cinza = 1, que tem um vizinho destro ou canhoto com nível de cinza = 2.2 (Kassner and Thornhill, 2010). 1.4.2 Gray-level Run-length Gray-level Run-length (GLRL) baseia-se na avaliação de tons de cinza idênticos em regiões consecutivas de uma imagem (Aggarwal and K. Agrawal, 2012; Kassner and Thornhill, 2010). Esta técnica tem sido amplamente utilizada para classificação de diferentes tipos de texturas em imagens digitais (Aggarwal and K. Agrawal, 2012; Kassner and Thornhill, 2010). Dentro de uma imagem, ou região de interesse, considera-se um gray level run como um conjunto consecutivo de pixels que possuem o mesmo tom de cinza, ou seja, o mesmo nível de intensidade. O gray level run pode ser computado em qualquer direção desejada dentro da imagem (horizontal – 0º, diagonal – 45º, vertical – 90º). No exemplo da Figura 3, um elemento de matriz (i,j) especifica o número de 23 Campus de Botucatu Instituto de Biociências PG-BGA vezes que um elemento de imagem contém uma sequência de pixels de tamanho j, em qualquer direção, com pixels que correspondem ao nível de intensidade i. Figura 3: Matriz representativa de uma imagem de 4 x 4 pixels com a determinação dos run- lengths nas direções de 0º e 45º (Kim et al., 2014). Analisando a imagem, na direção 0º, o valor de pixel igual a 0 aparece 4 vezes numa sequência de 1 pixel apenas. Nessa mesma direção, o valor de pixel 1 aparece 1 vez como uma sequência de 1 pixel e 1 outra vez como uma sequência de 3 pixels, e assim por diante. Dessa forma uma matriz de elementos é construída para todos os possíveis valores de pixels na imagem. A aquisição destas matrizes é uma tarefa simples do ponto de vista computacional. O número de elementos é proporcional ao número total de pontos na imagem. A partir destas matrizes, pode-se calcular sete grandezas as quais são descritas a seguir: short run emphasis (SRE), long runs emphasis (LRE), gray level nonuniformity (GLN), run percentage (RP), run length nonuniformity (RLN), low gray level run emphasis (LGRE) e high gray level run emphasis (HGRE). 24 Campus de Botucatu Instituto de Biociências PG-BGA 1.4.3 Transformadas Wavelets As Transformadas Wavelets são uma abordagem desenvolvida de grande destaque na área de processamento de sinais. No contexto de análise de imagens e processamento de sinais, as Transformadas Wavelets surgiram como uma abordagem alternativa para o processamento de imagens, cobrindo a deficiência da transformada de Fourier por permitir que tanto informações sobre a frequência quanto informações o tempo estejam presentes (Chaplot et al., 2006; El- Dahshan et al., 2010). As Transformadas Wavelets decompõem um sinal por meio de uma série de funções elementares, criadas a partir de escalas e translações de uma função base, chamada Wavelet mãe e por um sistema constituído de dois índices inteiros a que corresponde ao fator escala e b que corresponde ao fator translação, e funções de expansão (ou funções base) ψ a,b(t) que pode ser representada por (Chaplot et al., 2006; El-Dahshan et al., 2010): 𝜓𝑎,𝑏(𝑡) = 1 √𝑎 𝜓( 𝑡−𝑏 𝑎 ) a, b ϵ Ɍ (a ≠ 0) em que o fator 1 √𝑎 é um fator de normalização que mantém a mesma energia para todas as Wavelets independente da escala utilizada. Na figura 4, podemos ver a diferença entre uma (a) imagem de entrada e sua consequente (b) decomposição Wavelet. Figura 4. Diferença entre a (a) imagem de entrada e (b) decomposição Wavelet. 25 Campus de Botucatu Instituto de Biociências PG-BGA A decomposição Wavelet de um nível gera quatro sub-bandas, LL, HH, LH e HL. A primeira e a segunda denotam respectivamente as baixas e altas frequências; as duas últimas denotam as frequências intermediarias. Para cada sub-banda é calculada uma característica que descreve sua energia, dado por: 𝐹 𝑠𝑏=√ 1 𝑚 Ʃ𝑎(𝑥,𝑦)² onde sb denota uma das bandas LL, LH, HL ou HH, e a (x, y) representa o coeficiente da transformada na coordenada (x, y), obtido a partir de uma das sub-bandas, as quais são compostas de m coeficientes cada (Chaplot et al., 2006; El-Dahshan et al., 2010). Na Figura 5 é apresentada uma representação da decomposição Wavelet com um nível e com dois níveis de decomposição. Figura 5. Representação das sub-bandas obtidas pela decomposição Wavelet com (a) um nível de decomposição; (b) dois níveis de decomposição. 26 Campus de Botucatu Instituto de Biociências PG-BGA Capítulo III 27 Campus de Botucatu Instituto de Biociências PG-BGA 1. Objetivos O objetivo desta pesquisa foi a criação de métodos computacionais para classificação e diferenciação entre tumores primário cerebrais e metástases solitárias cerebrais. 28 Campus de Botucatu Instituto de Biociências PG-BGA Capítulo IV 29 Campus de Botucatu Instituto de Biociências PG-BGA 1. Materiais e Métodos 1.1 Equipamento de Ressonância Magnética (RM) Os exames de RM foram adquiridos pelo ACHIEVA 3T scanner (Phillips Medical Systems) usando a bobina 5-channel- brain. As sequências utilizadas serão ponderadas em T1 com e sem contraste, T2, difusão e recuperação de inversão fluido-atenuada (FLAIR). 1.2 Banco de Dados O banco de dados foi constituído por 98 exames de ressonância magnética de encéfalo (em formato DICOM (Digital Imaging and Communications in Medicine), de pacientes do Hospital das Clínicas de Botucatu, no período compreendido entre 2020 e 2022 imagens. 1.3 Infraestrutura Essa pesquisa foi desenvolvida no Laboratório de Física Aplicada ao Radiodiagnóstico (LAFAR), utilizando computadores, com as seguintes configurações: processador Intel®Core™ i7- 9700KF CPU 3.60 GHz, memória RAM de 16GB e sistema operacional de 64bits, utilizando os softwares MATLAB® R2020b (licença: 40962010) e Orange Canvas®. 30 Campus de Botucatu Instituto de Biociências PG-BGA Capítulo V 31 Campus de Botucatu Instituto de Biociências PG-BGA 1. Anexo 1: Diretrizes de submissão Diretrizes de submissão exigido pela revista “computer methods in biomechanics and biomedical engineering: imaging & visualization”, ISSN Online: 2168-1171, editora Taylor & Francis Online, para submissão do artigo científico, avaliação por pares e posterior publicação. Structure Your paper should be compiled in the following order: title page; abstract; keywords; Main text introduction, materials and methods, results, discussion; acknowledgments; declaration of interest statement; references; appendices (as appropriate); table(s) with caption(s) (on individual pages); figures; figure captions (as a list). Word Limits Please include a word count for your paper. There are no word limits for papers in this journal. Style Guidelines Please refer to these when preparing your paper, rather than any published articles or a sample copy. Please use British (-ise) spelling style consistently throughout your manuscript. Any form of consistent quotation style is acceptable. Please note that long quotations should be indented without quotation marks. Formatting and Templates Papers may be submitted in Word or LaTeX formats. Figures should be saved separately from the text. To assist you in preparing your paper, we provide formatting template(s). are 32 Campus de Botucatu Instituto de Biociências PG-BGA available for this journal. Please save the template to your hard drive, ready for use. A is available for this journal. Please save the LaTeX template to your hard drive and open it, ready for use, by clicking on the icon in Windows Explorer. If you are not able to use the template via the links (or if you have any other template queries) please contact us here. Editing, which will ensure that your article is free of spelling and grammar errors, Translation, and Artwork Preparation. For more information, including pricing, visit this website. Checklist: What to Include 1. Author details. Please ensure all listed authors meet the Taylor & Francis authorship the cover page of the manuscript. Where available, please also include ORCiDs and social media handles (Facebook, Twitter or LinkedIn). One author will need to be identified as the corresponding author, with their email address normally displayed in the article PDF (depending on the journal) and the online article. Authors’ affiliations are the affiliations where the research was conducted. If any of the named co-authors moves affiliation during the peer- review process, the new affiliation can be given as a footnote. Please 2. Should contain an unstructured abstract of 100-150 words. Read tips on writing you r abstract. 3. You can opt to include a video abstract with your article. Find out how these can help your work reach a wider audience, and what to think about when filming. 4. No more than 3 keywords. Read making your article more discoverable, including information on choosing a title and search engine optimization. 5. Funding details. Please supply all details required by your funding and grant- awarding bodies as follows: https://authorservices.taylorandfrancis.com/contact/ https://www.tandfeditingservices.com/?utm_source=TCIV&utm_medium=referral&utm_campaign=ifa_standalone https://www.tandfeditingservices.com/?utm_source=TCIV&utm_medium=referral&utm_campaign=ifa_standalone https://authorservices.taylorandfrancis.com/editorial-policies/defining-authorship-research-paper/ https://authorservices.taylorandfrancis.com/editorial-policies/defining-authorship-research-paper/ https://authorservices.taylorandfrancis.com/publishing-your-research/writing-your-paper/using-keywords-to-write-title-and-abstract/ https://authorservices.taylorandfrancis.com/publishing-your-research/writing-your-paper/using-keywords-to-write-title-and-abstract/ https://authorservices.taylorandfrancis.com/research-impact/creating-a-video-abstract-for-your-research/ https://authorservices.taylorandfrancis.com/research-impact/creating-a-video-abstract-for-your-research/ https://authorservices.taylorandfrancis.com/research-impact/creating-a-video-abstract-for-your-research/ https://authorservices.taylorandfrancis.com/research-impact/#researchpapervisibility 33 Campus de Botucatu Instituto de Biociências PG-BGA For single agency grants This work was supported by the [Funding Agency] under Grant [number xxxx]. For multiple agency grants This work was supported by the [Funding Agency #1] under Grant [number xxxx]; [Funding Agency #2] under Grant [number xxxx]; and [Funding Agency #3] under Grant [number xxxx]. 8. Data availability statement. If there is a data set associated with the paper, please provide information about where the data supporting the results or analyses presented in the paper can be found. Where applicable, this should include the hyperlink, DOI or other persistent identifier associated with the data set(s). Templates are also available to support authors. 9. Data deposition. If you choose to share or make the data underlying the study open, please deposit your data in a recognized data repository prior to or at the time of submission. You will be asked to provide the DOI, pre-reserved DOI, or other persistent identifier for the data set. 10. Supplemental online material. Supplemental material can be a video, dataset, fileset, sound file or anything which supports (and is pertinent to) your paper. We publish supplemental material online via Figshare. Find out more about supplemental material and how to submit it with your article. 11. Figures. Figures should be high quality (1200 dpi for line art, 600 dpi for grayscale and 300 dpi for colour, at the correct size). Figures should be supplied in one of our preferred file formats: EPS, PS, JPEG, TIFF, or Microsoft Word (DOC or DOCX) files are acceptable for figures that have been drawn in Word. For information relating to other file types, please consult our Submission of electronic artwork document. https://authorservices.taylorandfrancis.com/data-sharing/share-your-data/data-availability-statements/ https://authorservices.taylorandfrancis.com/data-sharing/share-your-data/data-availability-statements/ https://authorservices.taylorandfrancis.com/data-sharing/share-your-data/repositories/ https://authorservices.taylorandfrancis.com/publishing-your-research/writing-your-paper/enhance-article-with-supplemental-material/ https://authorservices.taylorandfrancis.com/publishing-your-research/writing-your-paper/enhance-article-with-supplemental-material/ https://authorservices.taylorandfrancis.com/publishing-your-research/making-your-submission/submit-electronic-artwork/ 34 Campus de Botucatu Instituto de Biociências PG-BGA 12. Tables. Tables should present new information rather than duplicating what is in the text. Readers should be able to interpret the table without reference to the text. Please supply editable files. 13. Equations. If you are submitting your manuscript as a Word document, please ensure that equations are editable. More information about mathematical symbols and equations. 14. Units. Please use SI units (non-italicized). Using Third-Party Material You must obtain the necessary permission to reuse third-party material in your article. The use of short extracts of text and some other types of material is usually permitted. This journal uses Taylor & Francis' Submission Portal to manage the submission process. The Submission Portal allows you to see your submissions across Taylor & Francis' journal portfolio in one place. To submit your manuscript please click here. If you are submitting in LaTeX, please convert the files to PDF beforehand (you will also need to upload your LaTeX source files with the PDF). Please note that Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization uses Crossref™ to screen papers for unoriginal material. By submitting your paper to Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization you are agreeing to originality checks during the peer-review and production processes. On acceptance, we recommend that you keep a copy of your Accepted Manuscript. Find out more about sharing your work. https://authorservices.taylorandfrancis.com/publishing-your-research/writing-your-paper/mathematical-scripts/ https://authorservices.taylorandfrancis.com/publishing-your-research/writing-your-paper/mathematical-scripts/ https://authorservices.taylorandfrancis.com/publishing-your-research/writing-your-paper/mathematical-scripts/ https://www.bipm.org/en/si/ https://rp.tandfonline.com/submission/create?journalCode=%5BTCIV%5D https://rp.tandfonline.com/submission/create?journalCode=%5BTCIV%5D https://www.crossref.org/crosscheck/index.html https://authorservices.taylorandfrancis.com/research-impact/sharing-versions-of-journal-articles/ 35 Campus de Botucatu Instituto de Biociências PG-BGA Data Sharing Policy This journal applies the Taylor & Francis Basic Data Sharing Policy. Authors are encouraged to share or make open the data supporting the results or analyses presented in their paper where this does not violate the protection of human subjects or other valid privacy or security concerns. Authors are encouraged to deposit the dataset(s) in a recognized data repository that can mint a persistent digital identifier, preferably a digital object identifier (DOI) and recognizes a long-term preservation plan. If you are uncertain about where to deposit your data, please see this information regarding repositories. Authors are further encouraged to cite any data sets referenced in the article and provide a Data Availability Statement. At the point of submission, you will be asked if there is a data set associated with the paper. If you reply yes, you will be asked to provide the DOI, pre- registered DOI, where one or multiple data sets are associated with a manuscript. Publication charges There are no submission fees, publication fees or page charges for this journal. Colour figures will be reproduced in colour in your online article free of charge. If it is necessary for the figures to be reproduced in colour in the print version, a charge will apply. Charges for colour figures in print are £300 per figure ($400 US Dollars; $500 Australian Dollars; €350). For more than 4 colour figures, figures 5 and above will be charged at £50 per figure ($75 US Dollars; $100 Australian Dollars; €65). Depending on your location, these charges may be subject to local taxes. https://authorservices.taylorandfrancis.com/data-sharing-policies/basic https://authorservices.taylorandfrancis.com/data-sharing/share-your-data/repositories/ https://authorservices.taylorandfrancis.com/data-sharing-policies/citing-data-guide/ https://authorservices.taylorandfrancis.com/data-sharing/share-your-data/data-availability-statements/ 36 Campus de Botucatu Instituto de Biociências PG-BGA Copyright Options Copyright allows you to protect your original material, and stop others from using your work without your permission. Taylor & Francis offers a number of different license and reuse options, including Creative Commons licenses when publishing open access. Read more on publishing agreements. Complying with Funding Agencies We will deposit all National Institutes of Health or Welcome Trust-funded papers into PubMedCentral on behalf of authors, meeting the requirements of their respective open access policies. If this applies to you, please tell our production team when you receive your article proofs, so we can do this for you. Check funders’ open access policy mandates here. Find out more about sharing your work. My Authored Works If you have any queries, please visit our Author Services website or contact us here. Updated 28th June 2022 https://authorservices.taylorandfrancis.com/publishing-your-research/moving-through-production/copyright-for-journal-authors/ https://authorservices.taylorandfrancis.com/publishing-your-research/moving-through-production/copyright-for-journal-authors/ https://authorservices.taylorandfrancis.com/publishing-open-access/funder-open-access-policies/ https://authorservices.taylorandfrancis.com/research-impact/sharing-versions-of-journal-articles/ https://authorservices.taylorandfrancis.com/ https://authorservices.taylorandfrancis.com/contact/ 37 Campus de Botucatu Instituto de Biociências PG-BGA Capítulo VI 38 Campus de Botucatu Instituto de Biociências PG-BGA 1. Artigo Científico Title Page Texture Analysis: A Potential Tool to Differentiate Primary Brain Tumors and Solitary Brain Metastasis S. A. S. Souza a, R. A. C. Guassu a, A. F. F. Alves b, M. Alvarez b, L. C. C. Pitanga c, F. Reis c, A. Vacavant d, J. R. A. Miranda a and D. R. Pina e* aDepartment of Biophysics and Pharmacology, Sao Paulo State University Julio de Mesquita Filho, Botucatu, Brazil; bBotucatu Medical School, Clinics Hospital, Medical Physics and Radioprotection Nucleus, Sao Paulo State University Julio de Mesquita Filho, Botucatu, Brazil; c Department of Radiology, School of Medical Sciences, University of Campinas, Campinas, Brazil. d Institut Universitaire de Technologie Le Puy en Velay, Université Clermont Auvergne, France; eDepartment of Tropical Diseases and Imaging Diagnosis, Sao Paulo State University Julio de Mesquita Filho, Botucatu, Brazil; *Corresponding author e-mail: diana.pina@unesp.br, Telephone: +55 14 3880-1291. 39 Campus de Botucatu Instituto de Biociências PG-BGA Texture Analysis: A Potential Tool to Differentiate Primary Brain Tumors and Solitary Brain Metastasis Abstract We propose a machine learning (ML) approach applied to texture features to differentiate primary brain tumors and solitary brain metastasis. Magnetic resonance imaging (MRI) exams of 96 patients were divided into primary tumors (38) and solitary brain metastasis (58). MRI sequences used: diffusion-weighted image (DWI), fluid-attenuated inversion recovery, T1-weighted, T1- weighted SE gadolinium-enhanced, and T2-weighted images. Regions of interest (ROIs) of 10 x 10 pixels were positioned within the tumors. For each ROI, 40 texture features were extracted and applied to five different ML methods: naive bayes, support vector machine (SVM), stochastic gradient descent, random forest, and tree. The ML methods classified the groups with good differentiation of up to 97.5% of the area under the receiver operator characteristics (ROC). SVM was the best classifier, especially in the DWI sequence. The method has a reliable classification for the investigation of tumor lesions. Keywords: primary brain tumors; solitary brain metastasis; texture analysis; machine learning. 40 Campus de Botucatu Instituto de Biociências PG-BGA 1. Introduction Differentiating primary brain tumors and solitary brain metastasis is a major diagnostic challenge in clinical practice due to their similar imaging features (Lee et al., 2013). This differentiation is clinically relevant since medical staging, surgical planning, and medical decisions are quite distinct (Ahmed et al., 2014; Soffietti et al., 2017). Of all complications of systemic cancer, brain metastases are the most common, and half of them are solitary at diagnosis (Bauer et al., 2015). In many cases, the definitive diagnosis is based on a stereotactic brain biopsy, which carries a small but not insignificant risk (Blanchet et al., 2011). The development of alternative strategies for diagnosing these diseases brings clear benefits by reducing invasive procedures; the correct and early diagnosis is of great importance for the favorable evolution of the patient. The use of a non-invasive tool is recommended when the biopsy is difficult. This occurs when the mass is located in an eloquent area or if the procedure offers risk (Blanchet et al., 2011). Magnetic resonance imaging (MRI) is the preferred examination for the first evaluation of brain tumors (Binaghi et al., 2016). However, it still presents subjective difficulty, as the similarity among imaging features. The clinical manifestations associated with glioblastoma multiforme may be similar to brain metastasis. However, each tumor has a distinct biologic nature, needing different treatment strategies (Blanchet et al., 2011). Many decades ago, the diffuse infiltration of malignant gliomas was observed by neurosurgeons when super-radical resections using hemispherectomies failed to cure these tumors. This biologic characteristic can potentially be exploited as an anti-invasive treatment paradigm (Giese et al., 2003). Some works have been using several non-invasive strategies to help diagnose different diseases (Alves et al., 2020; Mei et al., 2015; Skogen et al., 2016; Tian et al., 2018). Automated computer analysis tools are more objective than human readers and can potentially lead to more 41 Campus de Botucatu Instituto de Biociências PG-BGA reliable and reproducible brain tumor diagnosis. Machine learning (ML) methods can be applied to analyze radiomics features and have been used in diagnostic medicine to classify or differentiate primary brain tumors and solitary brain metastasis (Sachdeva et al., 2013; Singh et al., 2021; Zacharaki et al., 2009). Radiomics is an emerging field and refers to the extraction of quantitative image characteristics from radiological images to create mineable databases. Radiomics process involves (a) image acquisition and reconstruction, (b) image pre-processing, (c) identification of regions of interest, (d) feature extraction and quantification, (e) feature selection, and (f) building predictive and prognostic models using machine learning (Rogers et al., 2020; Thawani et al., 2018). There is still a lack of articles in the literature regarding the distinction between primary brain tumors and solitary brain metastasis associated with texture features. Thus, this work proposes an innovative approach to differentiate primary brain tumors and solitary brain metastasis using machine learning applied to texture features. 2. Materials and Methods The patient image database was divided into two groups to differentiate primary brain tumors and solitary metastasis. Figure 1 shows the database according to group distribution and classification steps using machine learning and texture analysis. 42 Campus de Botucatu Instituto de Biociências PG-BGA Figure 1. Distribution of the image database in two different groups of patients according to group distribution and classification steps using machine learning and texture analysis. Classification results are approached by the area under the curve (AUC). 2.1 Database The study was conducted in accordance with the Declaration of Helsinki and developed with ethical approval from São Paulo State University Ethics Committee and CONEP - National Commission on Ethics in Research (Protocol Number: 49334821.5.0000.5411). The database consists of DICOM (Digital Imaging and Communications in Medicine) images. Brain MRI scans of 96 patients between 2020 and 2022 were used. Patients were divided into two groups: primary brain tumors group (38 patients, mean age 42.73 years ± 12.64, range 19-77 years). Solitary brain metastasis group (58 patients, mean age 56.36 years ± 10.32, range 29-81 years). The complete list of all pathologies selected for the present study, with the number of patients (n), is presented in Table 1. Table 1. Complete list of all pathologies selected for the present study, with the number of patients (n). Brain pathology Subtypes Number of patients (n = 96) Primary brain tumors (n = 38) Pilocytic astrocytoma 2 Low grade gliomas 6 High grade gliomas 10 Medulloblastomas 3 Glioblastoma 17 Solitary brain metastasis (n = 58) Esophagus 5 Colorectal 6 Liver and biliary tract 4 Breast 13 Melanoma 4 Kidney 4 Lung 14 Prostate 2 Others 6 43 Campus de Botucatu Instituto de Biociências PG-BGA Patients were selected according to the following inclusion criteria: patients with confirmed tumor diagnosis by histopathological analysis based on biopsy. Exclusion criteria: patients who had surgery or biopsy before image acquisition, lesions smaller than 10 mm, patients presenting contraindications to the use of venous contrast, and patients without all the sequences necessary for the analysis. 2.2 MRI Acquisition Parameters MRIs were performed using a 3.0 T Siemens® MAGNETOM Verio Scanner with T1- weighted, T1-weighted spin echo (SE) gadolinium-enhanced images, T2-weighted, diffusion- weighted image (DWI), and fluid-attenuated inversion recovery (FLAIR). MRI acquisition parameters were: Sagittal T1-weighted (SE), 6 mm thick; 180° flip angle; repetition time (TR) of 430 milliseconds; echo time (TE) of 12 milliseconds; matrix 200 × 350, and field of view (FOV) of 25 × 25 cm. T2-weighted and proton density "fast spin echo" (FSE), 3 mm thick; 160° flip angle; TR of 4.800 milliseconds; TE of 108/18 milliseconds; matrix 256 × 256, and FOV of 22 × 22 cm. Axial T1-weighted SE, TR of 540 milliseconds and TE of 28 milliseconds. Axial T2- weighted fluid-attenuated inversion recovery (FLAIR) images, TR of 8.500 milliseconds and 2.000 or 100 milliseconds, and 2.200 milliseconds; TE of 72 or 90 milliseconds; matrix of 256 × 296, and FOV of 22 × 22 cm. T1-weighted SE gadolinium-enhanced images were obtained in three orthogonal planes and T1-weighted sagittal volumetric isotropic images. Diffusion-weighted images, TR of 22 milliseconds; TE of 9 milliseconds, and FOV of 230 × 250 mm, all acquired using a standard 8-channel head coil and with b value of 1000. 44 Campus de Botucatu Instituto de Biociências PG-BGA 2.3 Region of Interest Positioning and Feature Extraction Texture represents local variations of pixel intensities repeated regularly or randomly throughout the image (Mohanty et al., 2012). Textures are classified in: structural, spectral, and statistical approaches. Statistical textures consider the distribution of gray shades and their interrelationship (Haralick et al., 1973). Texture features are usually extracted from regions of interest (ROI). Statistical and Texture Analysis was performed using Matlab® R2021b (License number: 40962010) in a computer with Intel Core processor (TM) 2 Quad of 3.5GHz, 16 Gb of RAM, and 2 TB of hard-drive. One operator placed 10 x 10 pixels ROIs in each patient in regions previously segmented by an experienced neuroradiologist. In this procedure, ROIs were placed within the tumor, respecting a distance of at least 0.5 mm from the edges; Figure 2 shows the ROI previously segmented by an experienced neuroradiologist (A) and (C) the ROI placed within the tumor by the operator (B) and (D), for the primary tumor and solitary brain metastasis. Figure 2. T1-weighted SE gadolinium-enhanced image of primary tumor, with Radiologist (A) and operator (B) ROI positioning. T1-weighted SE gadolinium-enhanced image of solitary brain metastasis, with Radiologist (C) and operator (D). 45 Campus de Botucatu Instituto de Biociências PG-BGA The features extracted from the images were used to classify the lesions. For each ROI, a selection of features was extracted. There were 40 features, including first-order statistical ones such as mean, standard deviation, entropy, skewness, and kurtosis (Alobaidli et al., 2014). First- order statistical features provide important information about the distribution of pixel intensities. In addition, to capture the quantification of how pixels are located concerning one another, based on the average relationship between two pixels/voxels, we used second-order statistical features such as Gray Level Run-Length (GLRL) (Galloway, 1975) and Wavelet's Energy and Entropy (Materka and Strzelecki, 1998). Discrete wavelet energy and entropy were used to express the energy and entropy relative to each direction and frequency band. The 40 features extracted from each patient were divided into four categories and presented in Table 2. Table 2. Statistical Features, Gray Level Run-Length (GLRL), Wavelet’s Energy and Entropy features. 5 – 7 – 28 (40 features). Method Texture Feature Parameters Statistical Features mean, standard deviation, entropy, kurtosis, skewness GLRL short run emphasis (SRE), long runs emphasis (LRE), gray level non-uniformity (GLN), run percentage (RP), run length non-uniformity (RLN), low gray level run emphasis (LGRE) and high gray level run emphasis (HGRE) Wavelet’s Energy and Entropy wEntropy, energy ‘sym4’ (Ea_sym4, Eh_sym4_1, Eh_sym4_2, Ev_sym4_1, Ev_sym4_2, Ed_sym4_1, Ed_sym4_2, E_soma_sym4_1, E_soma_sym4_2), energy ‘haar’ (Ea_haar, Eh_haar_1, Eh_haar_2, Ev_haar_1, Ev_haar_2, Ed_haar_1, Ed_haar_2, E_soma_haar_1, E_soma_haar_2.), energy ‘bior’ (Ea_bior3.3, Eh_bior3.3_1, Eh_bior3.3_2, Ev_bior3.3_1, Ev_bior3.3_2, Ed_bior3.3_1, Ed_bior3.3_2, E_soma_bior3.3_1, E_soma_bior3.3_2) For all the Wavelet’s Energy and Entropy: [Ea] is the percentage of energy corresponding to the approximation, [Ed] is the vector containing the percentages of energy corresponding to the diagonal details, [Eh] and [Ev] which contains the percentages of energy corresponding to the horizontal and vertical details, respectively. In addition, we used the [E_soma], which is the sum of the vectors [Eh], [Ev] and [Ed]. For each method used, we used an index that indicates the level of decomposition (e.g., 1 and 2). Gray Level Run-Length (GLRL) 46 Campus de Botucatu Instituto de Biociências PG-BGA 2.4 Machine Learning Texture features were extracted within regions of interest and exported to spreadsheet files in the format “.xlsx” that were read and manipulated in the Orange Canvas® software to feed the learning algorithms. We used five different machine learning (ML) methods: random forest (Scornet, 2016), naive bayes (Zhang and Su, 2008), support vector machine (SVM) (Shawe-Taylor and Sun, 2011), tree (Bertsimas and Dunn, 2017), and stochastic gradient descent (SGD) (Sharma, 2018). The parameters used for each ML method were: for the random forest, we selected ten trees and did not split subsets smaller than 5; for the naive bayes numeric values were discretized to 4 bins with equal frequency; for the SVM we determined the radial basis function (RBF) Kernel with cost one and regression loss epsilon 0.10, the iteration limit was 100; for the tree, the minimum number of instances in leaves were 2, we did not split subsets smaller than 5, the maximal tree depth was limited to 100, and the classification stopped when majority reached 95%; for the SGD, to loss functions the classification was hinge and regression was squared loss, to regularization it was used Ridge (L2) with a strength of 0.00001, to optimization the learning rate was constant, the initial learning rate was 0.01, inverse scaling exponent was 0.25, the number of iterations was 1000 with tolerance of 0.0010 and data was shuffled after each iteration. These methods used the 40 textural features with a 10-fold cross-validation method applied to the dataset to avoid overfitting. Gain ratio and Gini index were used to rank all features according to their correlation with each class (Chen et al., 2008; Raileanu and Stoffel, 2004). Thus, we selected the five best-ranked features that achieved the highest classification scores for the five ML methods. To determine the efficiency of our classification models, we utilized parameters such as the area under the receiver operator characteristics (ROC) curve - AUC, accuracy (CA), F-score (F1), Precision, and Sensitivity. According to Carter et al., 2016, the ROC curve AUC is a sensitivity and specificity function prediction that can rank the test as excellent 47 Campus de Botucatu Instituto de Biociências PG-BGA (AUC higher than 90%), strong (AUC higher than 80%), reasonable (AUC higher than 70%) or non-useful (AUC below 70%) (Carter et al., 2016). Throughout this study, we used these definitions to assess the efficiency of the prediction models. 3. Results The differentiation results of the primary brain tumor versus solitary brain metastasis among the tested ML methods (support vector machine, naive bayes, random forest, tree and stochastic gradient descent) on FLAIR, T1-weighted, T1-weighted SE gadolinium-enhanced images (T1 C+), T2-weighted, and DWI image sequences are presented in Table 3. 48 Campus de Botucatu Instituto de Biociências PG-BGA Table 3. Classification results for the primary tumor versus solitary metastasis among the ML tested methods on FLAIR, T1-weighted, T1-weighted SE gadolinium-enhanced image (T1 C+), T2-weighted and DWI images with the area under the curve (AUC), accuracy, F-score (F1), precision, and sensitivity as percentage. Sequence Method AUC Accuracy F1 Precision Sensitivity FLAIR Random Forest 95.2 90.9 90.8 91.1 90.9 SVM 95.5 90.9 90.9 90.9 90.9 Tree 88.7 90.9 90.9 90.9 90.9 T1 SGD 94.9 94.9 94.9 94.9 94.9 SVM 96.8 90.3 90.3 90.3 90.3 NB 92.9 86.7 86.7 86.7 86.7 T1 C+ SGD 92.3 91.8 91.8 92.2 91.8 SVM 94.8 86.8 86.6 86.8 86.8 NB 78.6 70.6 70.9 73.6 70.6 T2 Random Forest 97.3 91.8 91.7 91.9 91.8 SGD 89.7 89.6 89.7 89.8 89.6 NB 78.6 70.2 70.5 73.0 70.2 DWI SGD 93.9 93.9 93.9 93.9 93.9 SVM 97.5 91.9 91.9 91.9 91.9 NB 78.5 70.3 70.3 70.7 70.3 SVM Support Vector Machine, SGD stochastic gradient descent, NB Naive Bayes. 49 Campus de Botucatu Instituto de Biociências PG-BGA For the FLAIR sequence, the SVM achieved the best AUC of 95.5%, both SVM and tree achieved the highest F1 value of 90.9%. All algorithms achieved at least 88.7% of AUC and 90.8% of F1. For the T1-weighted sequence, the SVM achieved the best AUC of 96.8% and the SGD achieved the best F1 of 94.9%. All algorithms achieved at least 92.9% of AUC and 86.7 of F1. For the T1 C+, the SVM achieved once again the best AUC of 94.8% and the SGD achieved the best F1 of 91.8%. All algorithms achieved at least 78.6% of AUC and 70.9 of F1. For the T2 sequence, the random forest achieved the remarkable AUC of 97.3% and the best F1 value of 91.7%. All algorithms achieved at least 78.6% of AUC and 70.5 of F1. For the DWI sequence, the SVM achieved an AUC value of 97.5% and the SGD achieved the best F1 value of 93.9%. All algorithms achieved at least 78.5% of AUC and 70.3 of F1. In Figure 3, the ROC curves of the 5 best-ranked features within each MRI sequence. Figure 3. ROC curves of the classifiers random forest, support vector machine (SVM), tree, stochastic gradient descent (SGD), and naive bayes. (A) ROC curve from a FLAIR image with the best-ranked features. (B) ROC curve from a T1-weighted image with the five best-ranked features. (C) ROC curve from T1-weighted SE gadolinium-enhanced image (T1 C+) image with 50 Campus de Botucatu Instituto de Biociências PG-BGA the five best-ranked features. (D) ROC curve from T2-weighted image with the five best-ranked features. (E) ROC curve from DWI with the five best-ranked features. 4. Discussion Similar imaging features lead to difficulty in differentiating between primary brain tumors and solitary brain metastases. The distinction is clinically relevant due to its difference in clinical outcomes and specific treatment strategies (Ahmed et al., 2014; Soffietti et al., 2017). We tested a new approach to differentiate primary brain tumors and solitary brain metastasis using texture features. Zacharaki et al. (2009), in a similar analysis, investigated the use of pattern classification for distinguishing some types of brain tumors, such as primary gliomas from metastases, and for grading gliomas. Using an SVM-based classification system, the authors obtained accuracy, sensitivity, and specificity values of 85%, 87%, and 79%, respectively. Such results are comparable to our findings, although we used 4 sequences, for SVM method, FLAIR, T1- weighted, T1-weighted SE gadolinium-enhanced image, and DWI accuracy/precision/sensitivity values are 90.9%/90.9%/90.9%; 90.3%/90.3%/90.3%; 86.8%/86.8%/86.8% and 91.9%, 91.9%, 91.9% respectively. In our model the results were slightly higher. These analyses can be classified as excellent, according to Carter et al. 2016. Sachdeva et al. (2013) performed a multiclass brain tumor classification using 428 T1- weighted SE gadolinium-enhanced MRI images from 55 patients. These images included primary brain tumors such as diffuse astrocytoma, glioblastoma, medulloblastoma, meningioma, metastasis, and normal regions. Two hundred eighteen intensity and texture features were extracted from these ROIs. The groups were classified using an artificial neural network (ANN), obtaining an overall accuracy of 85.23%, a strong classification value that demonstrates that ANN is a powerful classification tool, and should be studied further. Our accuracy for T1-weighted SE 51 Campus de Botucatu Instituto de Biociências PG-BGA gadolinium-enhanced MRI images using SGD, SVM, and naive bayes were 91.8%, 86.8% and 70.6%, respectively. In our model the results were slightly higher. However, Sachdeva et al. (2013) used a greater number of MRI images in their model, which may indicate greater generalization of the data. X. Li et al. (2020) performed discrimination between glioblastoma and solitary brain metastasis using a comparison of inflow-based vascular-space-occupancy (which does not need an exogenous contrast agent) and dynamic susceptibility contrast MRI. Inflow-based vascular- space-occupancy proved the potential to discriminate glioblastoma from solitary brain metastasis (AUC 83%; intratumoral AUC 90%). Despite these significant results, that study used only one type of primary brain tumor. Here, we used several primary brain tumors, including glioblastoma. We achieved values of AUC higher than 90% in the sequences without the use of exogenous contrast agent (T2-weighted, FLAIR, DWI and T1-weighted) and AUC values of up to 94.8% in the T1-weighted SE gadolinium-enhanced image. These similar results for sequences with and without contrast may indicate that alternative non-invasive techniques (even without the use of a contrast agent) have the potential to contribute to clinical routine in an equivalent way to invasive methods. The ML methods applied here had an excellent performance for this task, and have proven to be very promising for brain tumor classification tasks. Emphasizing that our database is quite heterogeneous, comprising eleven different types of primary tumors and nine types of solitary brain metastases, still we achieved high differentiation values for all five classifiers using five different MRI sequences. Remarkable, in this model, all sequences obtained at least one classification result above 90% of AUC. In this work, a “hold-out” method was not performed after the cross-validation step at the machine learning analysis. The lack of a “hold-out” method may positively alter the results and 52 Campus de Botucatu Instituto de Biociências PG-BGA lead to overfitting. Future works should consider using a “hold-out” method in machine learning analysis. This approach is not intended to replace the radiologist's role or prevent biopsy analysis from being performed. However, machine learning classification tools act as diagnostic decision support, for the clinical physician and the neuroradiologist, in moments that precede invasive intervention and classification scoring criteria. 5. Conclusions This study used machine learning applied to textural features to differentiate primary brain tumors and solitary brain metastasis. The tool developed here achieved an excellent differentiation of up to 97.5% AUC using the SVM in the DWI sequence and 94.9% accuracy using the stochastic gradient descent in the T1-weighted sequence. The method has a reliable classification for the investigation of brain tumor lesions, and its applications are valuable and meaningful in clinical practice, assisting in clinical decisions. Further research involving multi-centers and larger datasets may improve the results obtained in this work. 53 Campus de Botucatu Instituto de Biociências PG-BGA Statements & Declarations Acknowledgements: We would like to thank all the technicians of Magnetic Resonance Imaging service at Botucatu Medical School, including neuroradiologists experts that contributed to this study. The Coordination for the Improvement of Higher Education Personnel (CAPES), the National Council for Scientific and Technological Development (PQ/CNPq 303509/2019-8) and The São Paulo Research Foundation, (FAPESP 2020/05539-9) for the financial support throughout this study. Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results. Funding: This work was supported by the [São Paulo Research Foundation #1] under Grant [number 2020/05539-9]; [Brazilian National Council for Scientific and Technological Development #2] under Grant [number 303509/2019-8]; and [Coordination of Superior Level Staff Improvement #3] under Grant [number 001]. Data availability statement: No patient data (DICOM images) analyzed during this study are available on request. Our ethics committee only approved its use in this particular study. Consent to participate: Informed consent was waived by the ethics committee in this research, as patient data were obtained retrospectively, and completely anonymized. Consent for publication: Informed consent was waived by the ethics committee in this research, as patient data were obtained retrospectively, and completely anonymized. 54 Campus de Botucatu Instituto de Biociências PG-BGA Article References Ahmed R, Oborski MJ, Hwang M, Lieberman FS, Mountz JM. 2014. Malignant gliomas: current perspectives in diagnosis, treatment, and early response assessment using advanced quantitative imaging methods. Cancer Manag Res. 6:149-170. Alobaidli S, McQuaid S, South C, Prakash V, Evans P, Nisbet A. 2014. The role of texture analysis in imaging as an outcome predictor and potential tool in radiotherapy treatment planning [Review]. The British journal of radiology. 87(1042):20140369. eng. Alves AFF, Miranda JRA, Reis F, de Souza SAS, Alves LLR, Feitoza LM, de Castro JTS, de Pina DR. 2020. Inflammatory lesions and brain tumors: is it possible to differentiate them based on texture features in magnetic resonance imaging? The journal of venomous animals and toxins including tropical diseases. 26:e20200011. Bauer AH, Erly W, Moser FG, Maya M, Nael K. 2015. Differentiation of solitary brain metastasis from glioblastoma multiforme: a predictive multiparametric approach using combined MR diffusion and perfusion. Neuroradiology. 57(7):697-703. Bertsimas D, Dunn J. 2017. Optimal classification trees. Machine Learning. 106(7):1039-1082. Binaghi E, Pedoia V, Balbi S. 2016. Meningioma and peritumoral edema segmentation of preoperative MRI brain scans. Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization. 6(4):362-370. Blanchet L, Krooshof PW, Postma GJ, Idema AJ, Goraj B, Heerschap A, Buydens LM. 2011. Discrimination between metastasis and glioblastoma multiforme based on morphometric analysis of MR images. AJNR American journal of neuroradiology. 32(1):67-73. 55 Campus de Botucatu Instituto de Biociências PG-BGA Carter JV, Pan J, Rai SN, Galandiuk S. 2016. ROC-ing along: Evaluation and interpretation of receiver operating characteristic curves [Review Research Support, Non-U.S. Gov't]. Surgery. 159(6):1638-1645. eng. Chen J, Huang H, Tian F, Tian S. 2008. A selective Bayes Classifier for classifying incomplete data based on gain ratio. Knowledge-Based Systems. 21(7):530-534. Galloway MM. 1975. Texture analysis using gray level run lengths. Computer Graphics and Image Processing. 4(2):172-179. Giese A, Bjerkvig R, Berens ME, Westphal M. 2003. Cost of migration: invasion of malignant gliomas and implications for treatment. J Clin Oncol. 21(8):1624-1636. Haralick RM, Shanmugam K, Dinstein IH. 1973. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics. SMC-3(6):610-621. Lee EJ, Ahn KJ, Lee EK, Lee YS, Kim DB. 2013. Potential role of advanced MRI techniques for the peritumoural region in differentiating glioblastoma multiforme and solitary metastatic lesions. Clin Radiol. 68(12):e689-697. Materka A, Strzelecki MJTuol, institute of electronics, COST B11 report, Brussels. 1998. Texture analysis methods–a review. 10(1.97):4968. Mei PA, de Carvalho Carneiro C, Fraser SJ, Min LL, Reis F. 2015. Analysis of neoplastic lesions in magnetic resonance imaging using self-organizing maps. J Neurol Sci. 359(1-2):78-83. Mohanty AK, Senapati MR, Beberta S, Lenka SK. 2012. Texture-based features for classification of mammograms using decision tree. Neural Computing and Applications. 23(3-4):1011-1017. 56 Campus de Botucatu Instituto de Biociências PG-BGA Raileanu LE, Stoffel K. 2004. Theoretical Comparison between the Gini Index and Information Gain Criteria. Annals of Mathematics and Artificia