Tumor localization in tissue microarrays using rotation invariant superpixel pyramids

Tumor localization is an important component of histopathology image analysis; it has yet to be reliably automated for breast cancer histopathology. This paper investigates the use of superpixel classification to localize tumor regions. A superpixel representation retains information about visual structures such as cellular compartments, connective tissue, lumen and fatty tissue without having to commit to semantic segmentation at this level. In order to localize tumor in large images, a rotation invariant spatial pyramid representation is proposed using bags-of-superpixels. The method is evaluated on expert-annotated oestrogen-receptor stained TMA spots and compared to other superpixel classification techniques. Results demonstrate that it performs favorably.