Localisation of luminal epithelium edge in digital histopathology images of IHC stained slides of endometrial biopsies
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.
Section snippets
Materials and methods
Endometrial biopsies were collected at University Hospitals Coventry and Warwickshire NHS Trust from patients suffering from recurrent pregnancy loss or recurrent IVF treatment failure. Written informed consent was obtained prior to tissue collection. The biopsies were taken in the mid-luteal phase and obtained using a Wallach Endocell™ sampler (Wallach, USA). The tissue was fixed in 10% formalin and embedded in paraffin wax. Sections (3 μm) were labelled with anti-CD56 monoclonal antibody and
Evaluation of the proposed detection methods
We evaluated our improved LIPSyM, the original LIPSyM, our proposed uNK cell nucleus detection method and a commercial software, VIS developed by Visiopharm [4], on 20 expert hand-marked ground truth images. VIS is a cloud application that can be used to segment various types of cell nuclei, glandular structures and epithelium edges in digital histopathology images. VIS requires training using nuclei and background regions from a sample image in order to perform detection of the different types
Conclusions and future work
In this paper, we proposed a complete solution to effectively detect stromal and uNK cell nuclei and localising luminal epithelium edge of endometrial biopsy slides in H&DAB stained digital histopathology images. A novel localisation method is used to localise the luminal epithelium edge using alpha shapes on a set of LIPSyM detections corresponding to epithelial cell nuclei and spline approximation of the edge. Our proposed algorithms perform with high accuracy (F1 score: 0.84) on a high
References (12)
- et al.
Uterine natural killer cells and angiogenesis in recurrent reproductive failure
Hum Reprod
(2009) - et al.
Elevated periimplantation uterine natural killer cell density in human endometrium is associated with impaired corticosteroid signaling in decidualizing stromal cells
J Clin Endocrinol Metab
(2013) - et al.
A feasibility trial of screening women with idiopathic recurrent miscarriage for high uterine natural killer cell density and randomizing to prednisolone or placebo when pregnant
Hum Reprod
(2013) Cloud Image Analysis - Image APPs: Endometrium
(2014)- et al.
Quantification of histochemical staining by color deconvolution
Anal Quant Cytol Histol
(2001) - et al.
Scale-space and edge detection using anisotropic diffusion
IEEE Trans Pattern Anal Mach Intell
(1990)
Cited by (5)
Multi-resolution cell orientation congruence descriptors for epithelium segmentation in endometrial histology images
2017, Medical Image AnalysisCitation Excerpt :Therefore, we propose to segment the glandular and luminal epithelial regions containing epithelial cells to discriminate between stromal and epithelial cells. Detection of UNK and stromal cells, and localisation of luminal epithelium from tissue boundaries was addressed in Li et al. (2014). However, in Li et al. (2014) epithelial regions were discarded manually and it does not perform automatic segmentation of glandular epithelium and luminal epithelium to discard the epithelial cells.
Biomedical Image Segmentation: A Survey
2021, SN Computer ScienceA survey on automated cancer diagnosis from histopathology images
2017, Artificial Intelligence ReviewRobust nucleus/cell detection and segmentation in digital pathology and microscopy images: A comprehensive review
2016, IEEE Reviews in Biomedical EngineeringA novel cell orientation congruence descriptor for superpixel based epithelium segmentation in endometrial histology images
2015, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)