Stain separation in digital bright field histopathology

Digital pathology employs images that were acquired by imaging thin tissue samples through a microscope. The preparation of a sample from a biopt to the glass slide entering the imaging device is done manually introducing large variability in the samples to be imaged. For visible contrast it is necessary to stain the samples prior to imaging. Different stains attach to different compounds elucidating the different cellular structures. Towards automatic analysis and for visual comparability there is a need to standardize the images to obtain consistent appearances regardless of the potential differences in sample preparation. A standard approach is to unmix the the various stains computationally, normalize each separate stain image and to recombine these. This paper describes a modification to a standard blind method for stain normalization. The performance is quantified in terms of annotated expert data. Theoretical analysis is presented to rationalize the new approach.