Automated Classification for Breast Cancer Histopathology Images: Is Stain Normalization Important?

Breast cancer is one of the most commonly diagnosed cancer in women worldwide. A popular diagnostic method involves histopathological microscopy imaging, which can be augmented by automated image analysis. In histopathology image analysis, stain normalization is an important procedure of color transfer between a source (reference) and the test image, that helps in addressing an important concern of stain color variation. In this work, we hypothesize that if color-texture information is well captured with suitable features using data containing sufficient color variation, it may obviate the need for stain normalization. Considering that such an image analysis study is relatively less explored, some questions are yet unresolved such as (a) How can texture and color information be effectively extracted and used for classification so as to reduce the burden on the uniform staining or stain normalization. (b) Are there good feature-classifier combinations which work consistently across all magnifications? (c) Can there be an automated way to select reference image for stain normalization? In this work, we attempt to address such questions. In the process, we compare the independent texture and color channel information with that of some more sophisticated features which consider jointly color-texture information. We have extracted above features using images with and without stain normalization to validate the above hypothesis. Moreover, we also compare different types of contemporary classification in conjunction with the above features. Based on the results of our exhaustive experimentation we provide some useful indications.