ANALYSIS OF SVM PARAMETRIZATION IN THE CLASSIFICATION OF MAMMOGRAPHIC TEXTURE IMAGES

Álvaro Henrique de Araújo Rungue¹; Alexei Manso Corrêa Machado¹; Pedro Augusto Pinho Ferraz¹; Bernardo Augusto Godinho de Oliveira¹; Thiago Melo Machado-Coelho²; Willian Antônio dos Santos¹

¹ Pontifical Catholic University of Minas Gerais; ² Federal University of Minas Gerais

Baixar PDF doi:10.20906/CPS/CBA2018-1178

Resumo

Studies indicate that breast density is related to the risk of developing cancer since dense breast tissue can hide lesions, causing cancer to be detected at later stages. In this paper we classification method using support vector machines (SVM) associated to data reduction techniques to classify mammographic texture. An analysis of the parameters that influence the effectiveness of texture classification is also provided. Experiments were conducted on a set of 4,000 mammographic exams from which regions of interest representing the most significantly part of the texture of the breast tissue were extracted. Compared to other quantitative results found in the literature, the proposed multi-class SVM method using the radial basis function kernel and tuned parameters proved to be superior while classifying mammographic texture, reaching up to 99% of precision for 10% of recall.

Palavras-chave: Mammography; Image Classification; Texture descriptors; Support Vector Machines