Automated segmentation of villi in histopathology images of placenta
Introduction
The human placenta is a vital organ of pregnancy, exchanging all gases, nutrients, and waste products between the mother and developing fetus. This complex organ is essential for fetal survival. Developmental abnormalities and/or in utero damage to this organ can have detrimental short- and long-term health effects for both the mother and her fetus. Placental damage or dysfunction is known to play a central role in the development of fetal intrauterine growth restriction, pre-term birth, stillbirth, preeclampsia, and has also been linked to long term health outcomes, such as premature cardiovascular disease in mothers and metabolic disturbances in offspring [[1], [2], [3], [4], [5], [6], [7], [8]]. The effective application of placental histopathology following a poor pregnancy outcome can inform ongoing clinical management of mothers and neonates, and offer insight into etiology of placental disease, the risk of recurrence in future pregnancies, and may provide indications regarding which mothers and offspring are at highest risk of cardio-metabolic disease later in life [[9], [10], [11], [12], [13], [14]]. However, unlike some other fields of clinical pathology, there are a limited number of perinatal pathologists who are highly specialized in the field of placental histopathology, and the quality and reproducibility of placental pathology examinations are currently very poor [[15], [16], [17]]. Collectively, this has limited the clinical utility of this diagnostic/prognostic tool. The ability to apply an unbiased quantitative approach to the field of placental histopathology would substantially reduce the high degree of inter-observer variability and quality control issues currently plaguing this field [[18], [19], [20]], allowing for highly reproducible and reliable findings from this clinical modality. It is anticipated that the contextually rich information that could be gained from automated forms of placental histopathology would allow for a highly integrated use of placenta pathology findings in the continuum of care – allowing for effective translation of findings from the microscope to the clinical management of mothers and babies.
In recent years, there has been increased interest in and research addressing automated detection and analysis of various features within the complex architecture of the human placenta [[21], [22], [23], [24], [25], [26]]. The chorionic villi are the basic functional units of the placenta, anatomically described as a branched “tree-like structure” covered in a multinucleated layer of syncytiotrophoblast cells, which encases the feto-placental vasculature, embedded within a non-cellular connective tissue core. These villi structures are bathed in maternal blood, found within the intervillous space, in vivo. Previous work has focused on automated detection of the feto-placental vascular space within the villi – as poor feto-placental vascular development has been linked to several obstetrical diseases. Almoussa et al. [22] used a segmentation approach based on artificial neural networks (ANNs) to automatically extract feto-placental blood vessel features from digital histological images of the placenta. The ANNs were successful in detecting these most prominent vascular spaces within the placental villi. Chang et al. [25] proposed an automatic filtering method, which locally detects pixels containing curvilinear structures and reduces non-vessel noise. Compared to ANN-based methods, Chang et al. [25] proposed a faster and more accurate approach for feto-placental vessel detection.
A fuzzy C-means clustering method was applied successfully to distinguish cellular vs. extracellular components of the chorionic villi and to identify areas normally filled with maternal blood (intervillous space) [26]. Kidron et al. [24] used ImageJ software (https://imagej.net/) to extract features, such as size and number of chorionic villi, from histological images of the placenta and tested the feasibility of automated diagnosis of delayed or accelerated villous maturation. However, the images analyzed appear to be selective (e.g., there was no indication that artifacts in the histological images were present). In addition, this method does not appear to properly segment villi that are in very close proximity, touching, or overlapping (e.g., Fig. 1 in Kidron et al. [24] shows touching villi identified as a single villous structure). Swiderska-Chadaj et al. [23] described a method of automatic segmentation of placental villi structures for assessment of edema within the placenta. These authors used texture analysis, mathematical morphology, and region growing operations to extract different structures from placental images. Although they presented a comprehensive pipeline for automated analysis of placental histology, a small sample size of 50 villi from placentas with a variety of pathologies was used, with selective/optimal image selection (e.g., no image artifacts).
The objective of this paper is to describe an image analysis pipeline for automated analysis and segmentation of chorionic villi structures in histopathology images of human placentas. The main contribution of this work is the development of a new automated, segmentation protocol, which operates with a high degree of accuracy over large sample size. We also address the issue of adjacent, touching, and overlapping chorionic villi in our work, a common clinical finding that has not been properly addressed in the literature. The proposed method is validated on a set of healthy control placentas, as well as those complicated with the placenta-mediated disease of preeclampsia, to ensure the developed algorithm performs effectively for analyzing placenta specimens from healthy and diseased subjects. We also compare our work with a previously published automated method to detect villi in histological images of the placenta [24]. We apply their provided source code to our dataset and compare the results to our method.
Section snippets
Placenta histopathology images
Our dataset comprises high-resolution digital scans of 12 placental histopathology specimens obtained from the Research Centre for Women's and Infants Health (RCWIH) Biobank (Mount Sinai Hospital, Toronto, ON). The ethics approval to perform sub-analyses on the Biobank samples was obtained from the Ottawa Health Science Network Research Ethics Board and the Children's Hospital of Eastern Ontario (CHEO) Research Ethics Board.
The placental specimens were fixed in paraffin wax, stained with
Results
As is typically observed in clinically relevant image datasets, our experimental image dataset contained artifacts and undesired objects, including imaging artifacts and maternal red blood cells in the intervillous space (12 out of 36 images were contaminated by imaging artifacts (Fig. 4)). In the villi classification step, the contours were classified based on their size and density of VC class and SC class. The proposed pipeline was able to recognize desired contours with 92.86% accuracy
Discussion
An initial step for automated analysis of placental images is segmenting the chorionic villi – the functional unit of the placenta – within the histological specimens. Villi segmentation allows extraction of key features (e.g., villi count, size distribution, shape feature) for an objective assessment of placental images. Due to the complexity of villous structures and the presence of imaging artifacts, segmentation of individual villi in a placental image is a difficult task. However, the
Conclusion
In this study, a fully automated method for segmentation of chorionic villi structures in histopathology images of the human placenta was presented. The proposed method has the ability to identify complex villous structures, including touching and overlapping villi, by performing color analysis on the detected concavities of the villi structures. Our proposed method yielded an F1 score of 80.76% and sensitivity of 82.18% for a dataset comprising of nearly 5000 sample villi, considerably higher
Conflicts of interest
None.
Acknowledgments
This work is supported in part by the Ontario Trillium Scholarship (OTS) and Natural Sciences and Engineering Research Council (NSERC) of Canada.
References (33)
Classification of placental lesions
Am. J. Obstet. Gynecol.
(2015)The clinical implications of placental diagnoses
Semin. Perinatol.
(2015)- (2008)
- et al.
Automated image analysis of placental villi and syncytial knots in histological sections
Placenta
(2017) - et al.
Image analysis of histological features in molar pregnancies
Expert Syst. Appl.
(2013) Placental pathology in pregnancy complications
Thromb. Res.
(2011)- et al.
Hypertension in pregnancy: taking cues from pathophysiology for clinical practice
Clin. Cardiol.
(2018) - et al.
The placental component in early-onset and late-onset preeclampsia in relation to fetal growth restriction
Prenat. Diagn.
(2012) - et al.
The placenta in preeclampsia
Pregnancy Hypertens
(2012) Role of the placenta in preterm birth: a review
Am. J. Perinatol.
(2016)
Causes of death among stillbirths
JAMA, J. Am. Med. Assoc.
Pre-eclampsia and risk of cardiovascular disease and cancer in later life: systematic review and meta-analysis
BMJ
The developmental origins of adult disease (Barker) hypothesis
Aust. N. Z. J. Obstet. Gynaecol.
Placental Pathologic Lesions with a Significant Recurrence Risk - what Not to Miss! APMIS
Neonatal outcomes following preterm birth classified according to placental features
Am. J. Obstet. Gynecol.
Association between placental pathology and neonatal outcome in preeclampsia: a large cohort study
Hypertens. Pregnancy
Cited by (12)
Proposing a care practice bundle for neonatal encephalopathy during therapeutic hypothermia
2021, Seminars in Fetal and Neonatal MedicineCitation Excerpt :Placental pathology should be performed in all cases, since placental pathologies are a frequent finding in the context of NE [25], and placenta pathology remains associated with severity of NE and adverse outcome despite TH [26,27]. Advent of new technologies for placental pathology analysis will make it more reliable and easily accessible worldwide [28,29]. Care at the time of birth and during the first few hours is critical.
Placental contribution to neonatal encephalopathy
2021, Seminars in Fetal and Neonatal MedicineCitation Excerpt :Fetal vascular malperfusion should increase suspicion forshould neonatal thrombosis, but this elevated risk may not be recognized unless placental histology is available [75]. The advent of comprehensive automated computer-based analysis of the placenta using artificial intelligence from placental photos [76] or from hematoxylin and eosin (H&E) stained placental images [77] may markedly shorten the time required for pathological placental assessment and in the near future may provide reliable results within days of birth, allowing refined, individualized treatment in addition to hypothermia. In combination with histological assessment of the placenta after delivery, the new available techniques using “-omics” for placental analysis may provide more individualized understanding of placental dysfunction and pregnancy disorders that contribute to NE.
Perinatal asphyxia from the obstetric standpoint
2021, Seminars in Fetal and Neonatal MedicineCitation Excerpt :Post-birth photographic computer recognition programs are being piloted and may provide an affordable way to identify which placentas should be selected for histologic examination and this may be relevant to neonates who go on to be diagnosed with disabilities later in childhood [76]. Utilizing computers to interpret histology may also prove a cost effective way to perform placental histologic examination going forward [77]. The Human Placenta Project was formed recently and is “a collaborative research effort to understand the role of the placenta in health and disease” [78].
GestAltNet: aggregation and attention to improve deep learning of gestational age from placental whole-slide images
2021, Laboratory Investigation