Luiza Dri Bagesteiro, Msc, 2015

Classificação de Padrões Radiológicos Por Blocos em Imagens Não Segmentadas de Tomografia Computadorizada – Blockwise Classification of Radiological Patterns in Non-Segmented Images of Computed Tomography

Author: Luiza Dri Bagesteiro (Lattes)

E-mail: ldbagesteiro@inf.ufpr.br, lu.bagesteiro@gmail.com

Supervisor: Daniel Weingaertner

Co-supervisor: Lucas Ferrari de Oliveira

Abstract:

Diffuse lung diseases (DLDs) are comprised of more than 180 diseases that cause damage to the lungs. Patients who have these diseases present changes in lung tissue, which can be seen on Computed Tomography (CT) images. Thus, it becomes important to identify alterations in these images because, along with other information, they can define the type of disorder that the patient has. The aim of this study is to automatically recognize radiological patterns in non-segmented lung CT images, and also to classify the areas outside the lung region, in addition to five lung patterns: normal, emphysema, ground-glass opacity, fibrosis and micronodules. To achieve the objectives, a methodology that evaluates various descriptors is proposed. Searching for a complementarity of the results, combinations are performed among the classifiers. Using a public dataset containing 20.540 CT blocks, each descriptor was subjected to a classifier Support Vector Machine generating scores to represent the likelihood of the samples belonging to each of the six classes. After that, the classifiers were combined using fusion rules. The rotation invariant uniform CLBP classifier achieved the best individual result, and combined with top-hat transform descriptor, reached rates of 81.65% of sensitivity and 96.06% of specificity. Based on the relevance of the results obtained, the next step is to analyze the possibility of developing a support tool for diagnosis using the developed methodology.

Resumo:

As Doenças Pulmonares Difusas (DPDs) são compostas por mais de 180 patologias que causam danos nos pulmões. Os pacientes que possuem essas doenças apresentam alterações no tecido pulmonar, que podem ser visualizadas em imagens de Tomografia Computadorizada (TC). Dessa forma, torna-se importante identificar essas alterações nas imagens, pois juntamente com outras informações, elas podem definir o tipo de doença que o paciente possui. Com base nisso, o objetivo deste trabalho é reconhecer automaticamente padrões radiológicos em imagens não segmentadas de TC do pulmão, e também classificar áreas externas à região pulmonar, além dos cinco padrões pulmonares: normal, enfisema, opacidade em vidro-fosco, fibrose e micronódulos. Para atingir os objetivos, uma metodologia que avalia diversos descritores é proposta. Em busca de uma complementaridade entre os resultados, são realizadas combinações entre as respostas dos classificadores. Utilizando uma base de imagens pública, com 20.540 blocos de imagens de TC, cada descritor foi submetido a um classificador Support Vector Machine que gerou scores para representar a probabilidade das amostras pertencerem a cada uma das seis classes. Após isso, os classificadores foram combinados empregando algumas regras de fusão. O classificador CLBP invariante à rotação e uniforme obteve o melhor resultado individual, e combinado com a transformação top-hat, atingiram taxas de 81,65% de sensibilidade e 96,06% de especificidade. Com base na relevância dos resultados obtidos, o próximo passo é analisar a possibilidade de desenvolver uma ferramenta de auxílio ao diagnóstico utilizando a metodologia desenvolvida.

Related Topics:

  • Radiological patterns
  • Diffuse lung diseases
  • Computed Tomography
  • Pattern Recognition

 

Files:

Dissertation: “Classificação de Padrões Radiológicos por Blocos em Imagens Não Segmentadas de Tomografia Computadorizada”

  • Dissertation (PDF – in portuguese)
  • Presentation (PDF – in portuguese)
  • Citation to this dissertation:
  • BAGESTEIRO, Luiza Dri. Classificação de Padrões Radiológicos por Blocos em Imagens Não Segmentadas de Tomografia Computadorizada. 2015. 84 f. Dissertação (Mestrado) – Departamento de Informática, Universidade Federal do Paraná, Curitiba, 2015.

Article: “Blockwise Classification of Lung Patterns in Unsegmented CT Images”

  • Article (PDF – in english)
  • Presentation (PDF – in english)
  • Citation to this article:
  • Bagesteiro, L.D.; Oliveira, L.F.; Weingaertner, D., “Blockwise Classification of Lung Patterns in Unsegmented CT Images,” in Computer-Based Medical Systems (CBMS), 2015 IEEE 28th International Symposium on , pp.177-182, 22-25 June 2015.

 

Database:

We used the publicly available database provided by Depeursinge et al. The images can be acquired here.

 

Others:

Files of the extracted features (tar.gz)

Probabilities output files of the classifiers (tar.gz)