Bayesian K-SVD for H and E blind color deconvolution. Applications to stain normalization, data augmentation and cancer classification

Fernando Pérez-Bueno*, Juan G. Serra, Miguel Vega, Javier Mateos, Rafael Molina, Aggelos K. Katsaggelos

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

Abstract

Stain variation between images is a main issue in the analysis of histological images. These color variations, produced by different staining protocols and scanners in each laboratory, hamper the performance of computer-aided diagnosis (CAD) systems that are usually unable to generalize to unseen color distributions. Blind color deconvolution techniques separate multi-stained images into single stained bands that can then be used to reduce the generalization error of CAD systems through stain color normalization and/or stain color augmentation. In this work, we present a Bayesian modeling and inference blind color deconvolution framework based on the K-Singular Value Decomposition algorithm. Two possible inference procedures, variational and empirical Bayes are presented. Both provide the automatic estimation of the stain color matrix, stain concentrations and all model parameters. The proposed framework is tested on stain separation, image normalization, stain color augmentation, and classification problems.

Original languageEnglish (US)
Article number102048
JournalComputerized Medical Imaging and Graphics
Volume97
DOIs
StatePublished - Apr 2022

Keywords

  • Bayesian modeling
  • Blind Color Deconvolution
  • Histological images
  • Stain Normalization

ASJC Scopus subject areas

  • Radiological and Ultrasound Technology
  • Radiology Nuclear Medicine and imaging
  • Computer Vision and Pattern Recognition
  • Health Informatics
  • Computer Graphics and Computer-Aided Design

Fingerprint

Dive into the research topics of 'Bayesian K-SVD for H and E blind color deconvolution. Applications to stain normalization, data augmentation and cancer classification'. Together they form a unique fingerprint.

Cite this