MAGEE ET AL.: CONTEXT AWARE COLOUR CLASSIFICATION 1 Context Aware Colour Classification in Digital Microscopy Derek Magee 1 D.R.Magee@leeds.ac.uk Darren Treanor 2 darrentreanor@nhs.net Phattthanaphong Chomphuwiset 1 scpc@comp.leeds.ac.uk Philip Quirke 2 P.Quirke@leeds.ac.uk 1 School of Computing, University of Leeds, UK 2 Pathology and Tumour Biology, Leeds Institute for Molecular Medicine, University of Leeds, Leeds, UK Abstract With the advent of digital histopathology imaging and automatic image analysis, colour constancy across multiple microscope slides has become an important issue. Colour variation due to chemical, user or protocol inconsistency is widespread. This paper presents an approach for computationally efficient context aware colour classification. A ‘context vector’ derived from the colour distribution of the complete image is com- bined with the per-pixel information to improve pixel classification performance. The context vector implicitly encodes global image information such as whether the slide is under/over stained, or cut thinly, or thickly. The method is evaluated for segmentation accuracy on two data sets with different stains, and as a pre-processing method for a cell nuclei detection algorithm. 1 Introduction Histopathology is the diagnosis of disease by examination of tissue. In order to visualise tissue sections (which are virtually transparent), tissue sections are prepared using coloured histochemical stains that bind selectively to cellular components. Colour constancy is a problem in histopathology based on light microscopy due to: variable chemical colour- ing/reactivity from different manufacturers/batches of stains, colouring being dependent on staining procedure (timing, concentrations etc.), and light transmission being a function of section thickness. Lyon et al. [5] outline the need for standardisation of reagents and proce- dures in histological practice. However, such rigorous standardisation is not practised in the majority of hospital laboratories and complete standardisation is not possible without purer (and less variable) reagents (requiring action from multiple chemical manufacturers, and an associated increase in cost). Current practises are limited to physical and procedural quality control methods, including subjective assessment of stain quality and inter-laboratory com- parisons of staining, in order to minimise the visible variability in staining and its impact on diagnostic quality. With the advent of digital imaging and automatic image analysis colour consistency in histopathology has become more of an issue. For example, many commercial automatic image analysis algorithms require parameters defining the expected colour of anatomy of interest and fail if these parameters are incorrect. This paper presents methods for taking c  2011. The copyright of this document resides with its authors. It may be distributed unchanged freely in print or electronic forms. 135