Fast and stable color normalization of whole slide histopathology images using deep texture and color moment matching

Kyohei Sano, Daisuke Komura, Shumpei Ishikawa

Keywords: computational pathology, stain normalization, color transfer, deep texture representation

TL;DR: Our novel color normalization method for histopathology images is secured by estimation of a color transformation matrix based on reference and source patch pairs with similar deep texture representation.

Abstract:

Whole Slide Images (WSIs) are prone to color variations due to differences in fixation and staining conditions of tissue samples, as well as the scanning process. Such variations can adversely affect the image analysis, and in this paper, we propose a novel, fast and stable color normalization algorithm for WSIs called CONTEMM (COlor Normalization using deep TEx-ture and color Moment Matching). CONTEMM estimates color transfor-mation matrix based on pairs of reference and source patches with similar tissue components in the respective WSIs, which are selected using deep texture representations. The color transformation matrix is estimated quick-ly by fitting the second moment about white color.
Performance of CONTEMM algorithm was evaluated using histopathology images from different slide scanners and TCGA (The Cancer Genome Atlas) datasets. CONTEMM was shown to outperform the state-of-the-art methods in terms of variation (stability), accuracy, and computation time.