Localisation of luminal epithelium edge in digital histopathology images of IHC stained slides of endometrial biopsies

Abstract

Diagnosis of recurrent miscarriage due to abnormally high number of uterine natural killer (uNK) cells has recently been made possible by a protocol devised by Quenby et al. Hum Reprod 2009;24(1):45–54. The diagnosis involves detection and counting of stromal and uNK cell nuclei in endometrial biopsy slides immunohistochemically stained with haematoxylin for staining cell nuclei and CD56 as a marker for the uNK cells. However, manual diagnosis is a laborious process, fraught with subjective errors. In this paper, we present a novel method for detection of uterine natural killer (uNK) cells in the human female uterus lining and localisation of the luminal epithelium edge in endometrial biopsies. Specifically, we employ a local phase symmetry based method to detect stromal cell nuclei and propose an adaptive background removal method that significantly eases the segmentation of uNK cell nuclei regions. We also propose a novel method using alpha shapes for the identification of epithelial cell nuclei and B-Spline curve fitting on identified cell nuclei to localise the luminal epithelium edge. The objective of edge localisation is to avoid cell nuclei near the luminal epithelium edge being counted in the diagnosis process due to their non-relevance to the calculation of stromal to uNK cell ratio that determines the diagnosis of recurrent miscarriages in the end. The resulting algorithm offers a promising potential for computer-assisted diagnosis of recurrent miscarriage due to its high accuracy.

Introduction

Uterine natural killer (uNK) cells are immune cells found in the human female uterus lining. Normally, these cells make up no more than 5% of all cells in the womb lining. Recently, Quenby et al. [1] showed that there are abnormally high numbers of uNK cells in the uterus of women who suffer from recurrent miscarriages, a condition that affects one in every 100 women of reproductive age in the UK. High uNK cell density in the lining of the womb was associated with glucocorticoid deficiency [2] and a small randomised controlled trial suggested that women with high numbers of uNK cells are more likely to have a live birth if given glucocorticoids in lieu of placebo [3]. This means that uNK testing has clinical significance as it could direct clinicians towards effective treatment.

Computer-assisted diagnosis of recurrent miscarriage due to over-presence of uNK cells can be made by calculating the ratio of uNK cells to stromal cells in histopathology images of endometrial biopsy stained with haematoxylin and CD56, which stains uNK cells brown when used with DAB staining. This process is compounded by the fact that cell counting must be performed within image regions near the luminal epithelial edge, but the cells located 200 μm from the edge should not be counted as part of the diagnostic process. Currently, experts at our local hospital (University Hospitals Coventry and Warwickshire NHS Trust) count stromal and uNK cell nuclei manually. Manual counting of thousands of cells from a large image dataset on a regular basis is cost ineffective, potentially inaccurate due to subjective assessment, and involves the hassle of manually removing the epithelial edge from the image [3].

In this paper, we present a complete solution for detecting the stromal and uNK cell nuclei and for localising the luminal epithelium edge of endometrial biopsy samples. We improve a local phase symmetry based method [6] for detecting stromal cell nuclei and propose an adaptive background removal method for segmentation of uNK cell nuclei regions. We also propose a novel method for localising luminal epithelium edges, which fits a B-spline curve on epithelial cell nuclei identified using alpha shapes to mark a luminal epithelium edge. We evaluate our proposed methods and the state-of-the art commercial software VIS developed by Visiopharm [4] using expert hand-marked ground truth images. The results show that our proposed methods attain higher accuracy than VIS.