Colour Normalisation in Digital Histopathology Images

Colour consistency in light microscopy based histology is an increasingly important problem with the advent of Gigapixel digital slide scanners and automatic image analysis. This paper presents an evaluation of two novel colour normalisation approaches against the previously utilised method of linear normalisation in lαβ colourspace. These approaches map the colour distribution of an over/under stained image to that of a well stained target image. The first novel approach presented is a multi-modal extension to linear normalisation in lαβ colourspace using an automatic image segmentation method and defining separate transforms for each class. The second approach normalises in a representation space obtained using stain specific colour deconvolution. Additionally, we present a method for estimation of the required colour deconvolution vectors directly from the image data. Our evaluation demonstrates the inherent variability in the original data, the known theoretical problems with linear normalisation in lαβ colourspace, and that a multi-modal colour deconvolution based approach overcomes these problems. The segmentation based approach, while producing good results on the majority of images, is less successful than the colour deconvolution method for a significant minority of images, is less successful than the colour deconvolution method for a significant minority of images as robust segmentation is required to avoid introducing artifacts.