ProjectStainSeparation

Chair for Computer Aided Medical Procedures & Augmented Reality
Lehrstuhl für Informatikanwendungen in der Medizin & Augmented Reality

Stain Separation and Structure-Preserving Color Normalization for Histological Images

Stain Separation and Structure-Preserving Color Normalization for Histological Images

Scientific Director: Nassir Navab

Contact Person(s): Tingying Peng

Keywords: Microscopic Image Analysis, IFL

Abstract

Staining and scanning of tissue samples for microscopic examination is fraught with unwanted variations that affect their color appearance. Sources of these variations include differences in raw material and manufacturing techniques of stain vendors, staining protocols of labs, and color responses of digital scanners. Color normalization of stained biopsies and tissue microarrays will help pathologists and computational pathology software while comparing different tissue samples. However, techniques that are used for natural images, such as histogram matching fail to utilize unique properties of stained tissue samples and produce undesirable artifacts. Tissue samples are stained with only a few reagents (frequently only two -- hemotoxylin and eosin or H\&E) and most tissue regions bind to only one stain or the other, thus producing sparse density maps composed of only a few components. This underlying structure of sparse stain density is biomedically important. We used these properties of stained tissue to propose a technique for stain separation and color normalization. Based on sparse non-negative matrix factorization (sparseNMF), we estimate prototype color and density map of each stain in an unsupervised manner to perform stain separation. To color normalize a given source image, we combine its stain density maps with the stain color prototypes of a target image whose appearance was preferred by pathologists. In this way, the normalized image preserve the biological structure encoded in the stain density of the source image. Both the proposed sparseNMF stain separation and color-normalization techniques yield higher correlation with ground truth than the state of the art. They are also rated qualitatively higher than other techniques by a group of pathologists. We further propose a computationally faster extension of this technique for large whole-slide images that selects an appropriately small sample of patches to compute the color prototypes of each stain instead of using the entire image. The fast scheme achieves a 20-folds acceleration, which does not only greatly enhance the analysis efficiency, but also allow its clinical applications to become practically feasible.

Team

Contact Person(s)

Tingying Peng
Dr. Tingying Peng

Working Group

Abhishek Vahadane
Abhishek Vahadane
Lichao Wang
Dr. Lichao Wang
Maximilian Baust
Dr. Maximilian Baust
Shadi Albarqouni
M.Sc. Shadi Albarqouni
Tingying Peng
Dr. Tingying Peng

Location



Technische Universität München
Institut für Informatik / I16
Boltzmannstr. 3
85748 Garching bei München

Tel.: +49 89 289-17058
Fax: +49 89 289-17059
Visit our lab at Garching.



Klinikum rechts der Isar
der Technischen Universitüt München
Ismaninger Str. 22
81675 München

IFL Lab - Room: 01.3a-c
Tel.: +49 89 4140-6457
Fax: +49 89 4140-6458
Visit our lab at Klinikum rechts der Isar.



internal project page

Please contact Tingying Peng for available student projects within this research project.

Edit | Attach | Refresh | Diffs | More | Revision r1.1 - 01 Sep 2015 - 10:22 - ShadiAlbarqouni

Lehrstuhl für Computer Aided Medical Procedures & Augmented Reality    rss.gif