Segmentation of HER2 protein overexpression in immunohistochemically stained breast cancer images using Support Vector Machines

17th International Workshop on Advanced Computing and Analysis Techniques in Physics Research

物理学;计算机科学

Pezoa, Raquel^1,2 ; Salinas, Luis^1,2 ; Torres, Claudio^1,2 ; Härtel, Steffen^3 ; Maureira-Fredes, Cristián^4 ; Arce, Paola^1

Department of Informatics, Universidad Técnica Federico Santa María, Valparaíso, Chile^1

Centro Científico Tecnológico de Valparaíso, Universidad Técnica Federico Santa María, Valparaíso, Chile^2

SCIAN-Lab, ICBM, BNI, University of Chile, Chile^3

Max Planck Institut für Gravitationsphysik (Albert-Einstein-Institut), Potsdam

D-14476, Germany^4

关键词: Classification performance;    Color and textures;    Deconvolution filters;    Haralick features;    Immunohistochemistry;    Over-expression;    Targeted treatment;    Tissue sections;   

Others  :  https://iopscience.iop.org/article/10.1088/1742-6596/762/1/012050/pdf
DOI  :  10.1088/1742-6596/762/1/012050

学科分类：计算机科学（综合）

来源: IOP

【 摘 要 】

Breast cancer is one of the most common cancers in women worldwide. Patient therapy is widely supported by analysis of immunohistochemically (IHC) stained tissue sections. In particular, the analysis of HER2 overexpression by immunohistochemistry helps to determine when patients are suitable to HER2-targeted treatment. Computational HER2 overexpression analysis is still an open problem and a challenging task principally because of the variability of immunohistochemistry tissue samples and the subjectivity of the specialists to assess the samples. In addition, the immunohistochemistry process can produce diverse artifacts that difficult the HER2 overexpression assessment. In this paper we study the segmentation of HER2 overexpression in IHC stained breast cancer tissue images using a support vector machine (SVM) classifier. We asses the SVM performance using diverse color and texture pixel-level features including the RGB, CMYK, HSV, CIE L∗a∗b∗ color spaces, color deconvolution filter and Haralick features. We measure classification performance for three datasets containing a total of 153 IHC images that were previously labeled by a pathologist.