Automated segmentation and morphological characterization of placental histology images based on a single labeled image

2025-05-02 0 0 4.17MB 26 页 10玖币

侵权投诉

Automated segmentation and morphological characterization of placental

histology images based on a single labeled image

Arash Rabbania,∗

, Masoud Babaeib, Masoumeh Gharibc

aThe University of Leeds, School of Computing, Leeds, UK

bThe University of Manchester, School of Chemical Engineering and Analytical Science, Manchester, UK

cMashhad University of Medical Sciences, Department of Pathology, Mashhad, Iran

Abstract

In this study, a novel method of data augmentation has been presented for the segmentation of placental

histological images when the labeled data are scarce. This method generates new realizations of the placenta

intervillous morphology while maintaining the general textures and orientations. As a result, a diversiﬁed

artiﬁcial dataset of images is generated that can be used for training deep learning segmentation models.

We have observed that on average the presented method of data augmentation led to a 42% decrease in

the binary cross-entropy loss of the validation dataset compared to the common approach in the literature.

Additionally, the morphology of the intervillous space is studied under the eﬀect of the proposed image

reconstruction technique, and the diversity of the artiﬁcially generated population is quantiﬁed. Due to the

high resemblance of the generated images to the real ones, the applications of the proposed method may not

be limited to placental histological images, and it is recommended that other types of tissues be investigated

in future studies. The image reconstruction program as well as image segmentation model are available

publicly.

Keywords: Placenta, Chorionic Villi, Morphology, Data augmentation, Semantic segmentation

1. Introduction

1.1. Human placenta

During pregnancy, the placenta is a temporary but vital organ in female humans that connects the baby

to the uterus [1]. The placenta forms soon after conception and adheres to the uterine wall on one side and

connects to the baby on the other side via the umbilical cord [1] as visualized in Fig. 1-c. The role of the

placenta includes providing a passage for the transport of oxygen, nutrients, hormones, and immune cells.

The umbilical cord bifurcates into numerous vessels and capillaries that reside in the villous space of the

placenta (Fig. 1 -b). Fetal capillaries pass close to the maternal blood pool in the intervillous space, and

controlled transport of oxygen and nutrients occurs at the cellular level through a set of selective membranes

known as the syncytiotrophoblast and cytotrophoblast which collectively are called the trophoblast[2, 3]

∗Corresponding author

Email address: a.rabbani@leeds.ac.uk | rabarash@gmail.com (Arash Rabbani )

Preprint submitted to arXiv.org October 10, 2022

arXiv:2210.03566v1 [eess.IV] 7 Oct 2022

(Fig. 1-a). In several studies, the microstructural morphology of the villous space has been found to have a

signiﬁcant correlation with the healthiness and development of the fetus [4, 5, 6]. Such observations justify

the need to take a closer quantitative look at the morphology of the placenta for diagnostic and prognostic

purposes.

Figure 1: Placental internal structure and its role in delivering oxygen and nutrients to the fetus. a) internal structure of

chorionic villus including fetal capillaries and trophoblast which acts as a selective membrane between maternal and fetal

blood, b) cross-section of placenta structure including umbilical cord, maternal arteriole, intervillous space, and chorion which

is the outermost membrane around the fetus, c) fetus.

1.2. Placental histology

Microscopical study of histological slides provides an invaluable visual representation of tissues and cells on

the micrometer scale. Histological images play a critical role in the study of body functions and the diagnosis

of certain diseases [7] from acute and chronic inﬂammation [8, 9] to autoimmunity [10] and cancer [11]. In

the same way, these images have been frequently used to shed light on the micro-structural morphology of

the placenta [12]. Several maternal conditions and diseases can aﬀect the structure and functionality of the

placenta, which can eventually lead to adverse eﬀects on the fetus development [13, 6]. As an example, using

histo-pathological images in combination with other imaging techniques, Facchetti et al. [13] observed that

SARS-CoV2 infection in mothers can cause neutrophil inﬁltration in the intervillous spaces and consequently

adversely aﬀect the performance of the placenta. As another example, type 1 diabetes has been found to

cause abnormal placentas that are enlarged, thick, and plethoric, with abnormalities of villous maturation

[14, 6].

Placental histological samples are commonly obtained by biopsy in the chorionic villi section. Sampling

is usually performed in the 11th to 14th weeks of pregnancy [15] using a thick needle or a small tube through

the abdomen or cervix, respectively. The recovered samples are ﬁxed with paraﬃn or resin and some slices

are cut from the structure with a thickness of a few microns that helps light pass through [16]. Using light

microscopy, these thin slices can be visualized for qualitative and quantitative studies on placenta health

and function. To perform quantitative studies on placental histological images, manual or automated data

analysis methods are used. Considering the fact that the manual analysis of placental histological slides

by microscope is a costly and time-consuming activity, computer-aided analyzes oﬀer a fast and repeatable

result that eliminates inter- and intra-observer variability [17].

1.3. Image synthesis

Medical images are expensive to obtain and process both in terms of budget and time. A quick solution

to proliferate a variety of desired images at virtually no cost is image synthesis. Synthetic images have

been widely approached in litrature to help for automated diagnosis and classiﬁcation of several diseases

from liver lesion [18] and cardiac abnormalities [19] to brain tumors[20]. Such an approach relies on a set

of exemplar images to be used for training and then by capturing the main elements within the images a

diversiﬁed set of predicted images are generated that have a certain feature in common [21, 22, 23, 24].

Patch-based reconstruction of textural images as one of the classic image synthesis approaches have been

used in the present study to augment the availible image dataset [25, 26]. We have tailored the method to

ﬁt the purpose of this study to regenerate histological images of chorionic villi.

1.4. Application of deep learning

Image segmentation is the process of dividing an image into its structural elements that share common

characteristics [27]. Segmenting placental histological images into villous and intervillous spaces is a funda-

mental task in studying the morphology of the placenta [28]. Automated methods have been used for this

purpose to save time and eliminate operator bias from the diagnostic cycle [17, 29].

In addition to intervillous space segmentation, deep learning has been used for cellular phenotyping

based on placental histology images by Ferlaino et al. [30]. For this purpose, they have created a dataset

composed of thousands of high-conﬁdence, manually curated placenta cells from ﬁve classes. By combining a

nuclei-locator deep learning model and an image-based classiﬁer, they have been able to achieve an accuracy

of around 75% for the detection of diﬀerent types of placental cells. Deep learning can be also used for

improving the quality of the placental histology images as shown by Rabbani and Babaei [31]. They have

developed a convolutional neural network model that takes low-resolution as input and predicts the a residual

image that shows the diﬀerences between the low- and desired high-resolution images. They have shown that

such an approach does not only intensiﬁes main feautres of the image but alos, adds realistic-look details

that can helps for better understanding of the low quality images of placenta.

Another recent application of deep learning on placental histological images has been presented by

Mobadersany et al.[29]. Their developed deep learning model known as GestaltNet can estimate gestational

age by analyzing the scanned placental slides. They have been able to estimate gestational age by a mean

absolute error of 1.0847 weeks using a deep learning method known as attention-based feature aggregation

[29, 32]. To generate the required dataset for training, validation, and testing the GestaltNet, the authors

manually annotated 1918 regions on 154 histological slides [29], which is a considerably time-consuming task.

Common deep learning techniques rely on a relatively large dataset of manually segmented images for

training purposes. In the present study, we have minimized the requirement of manually segmented images by

relying on a single image for the training process. This task is done through an innovative data augmentation

technique that creates a diversiﬁed range of tissue textures that enables the model to perform at an acceptable

level of accuracy in the prediction of unforeseen image datasets with signiﬁcantly diﬀerent textures.

2. Methodology

This study presents a tailored method of histological data augmentation for AI–powered segmentation of

the placental intervillous space. We expect that the presented method is especially eﬀective in the scarcity

of labeled data. Labeling histological images is an expensive and time–consuming task, especially when the

presence of highly trained specialists is crucial. In this section, we initially describe the available data, and

then introduce the proposed image augmentation technique. In addition, the structure of four deep learning

models used for image segmentation is discussed. Finally, an image–based quantitative method is presented

that helps in characterizing the morphology of the segmented placental intervillous space.

2.1. Material

In this research, H & E histological images of three placenta samples from healthy volunteers (HV)

have been used to demonstrate a possible improvement in automated segmentation techniques. The images

were obtained from the University of Michigan Histology and Virtual Microscopy archive [33]. In total, 34

zoomed images with a size of 256 ×256 pixels have been acquired from the original 3 histological slides

with a maximum magniﬁcation of 40X. All zoomed images were captured from locations with the presence

of chorionic villus and intervillous space. Fig. 2 illustrates three magniﬁcations of the histological image

obtained from HV #1. Fig. 2–a shows the complete slide with maternal to fetal tissues extended from right

to left. By zooming ﬁve times, Fig. 2–b is obtained, which illustrates the chorionic villus intertwined with

intervillous space near the decidual plates, which are compact maternal tissues protecting the fetus which

should not be included in the chorionic villus in the segmentation process. Finally, Fig. 2 –c shows the 40X

magniﬁed view of the chorionic villus that includes the fetal capillaries and syncytiotrophoblast, as well as

intervillous space with the presence of maternal blood cells that should be included as intervillous space in

the segmentation process.

Figure 2: Histological image of the placenta from a healthy volunteer in three length scales, a) overview of the sample which

shows maternal to fetal tissues extended from right to left, b) chorionic villus inter-twined with intervillous space in the vicinity

of decidual plates, c) 40X magniﬁed view of chorionic villus including fetal capillaries and syncytiotrophoblast, as well as

intervillous space with the presence of maternal blood cells.

2.2. Data augmentation

Considering the aim of this study to oﬀer automated segmentation of placenta slides with minimal

availability of labeled data, we have selected only one of the 256 ×256 histology images of HV #1 to

act as our training image (Fig. 3–a). The rest of the 33 available images will be used for validation

purposes. Considering the high demand of deep learning models for training data, artiﬁcial augmentation of

the available datasets is an inevitable step [34]. The current state-of-the-art in augmentation of histological

images includes color shift, noise addition, zooming, rotation, ﬂipping, translating, and elastic deformation

[35, 36] which is called base case in this paper. In this section, our aim is to develop an additional step in

relation to previous methods of image data augmentation to improve the diversity of the generated data,

which can lead to higher model performance. Fig. 3 describes the steps required in the proposed data

augmentation method. Initially, a dataset of sub–sample images is cropped from the original image to a

smaller size via a sliding-window approach. The sliding sampling window sweeps the area of the original

image with discrete 5 pixel steps in each direction (Fig. 3-a). Then, the obtained dataset of sub-samples (Fig.

3-b) is augmented by ﬂipping in horizontal (X), vertical (Y), and a combination of both directions (XY) (Fig.

3-c). Now, the reconstruction cycle begins by selecting a new random location in the new image to ﬁll (Fig.

3–d–1). Each of the locations is considered to have 5 pixels overlap with neighboring locations similar to the

sampling stage. If the edges of the selected location have not been occupied by neighbor blocks in advance,

we are free to select any of the sampled images in the dataset to ﬁll that speciﬁc location. In contrast, if the

selected location has predetermined edges from previous cycles, we need to look up in the created dataset

of sub–samples (Fig. 3-c) for entries with matching edges (Fig. 3–d–3). The search is done by temporarily

masking all parts of the image except the edges and then minimizing the mean squared error for the pixel

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

AutomatedsegmentationandmorphologicalcharacterizationofplacentalhistologyimagesbasedonasinglelabeledimageArashRabbania,,MasoudBabaeib,MasoumehGharibcaTheUniversityofLeeds,SchoolofComputing,Leeds,UKbTheUniversityofManchester,SchoolofChemicalEngineeringandAnalyticalScience,Manchester,UKcMashhadUniver...

展开>> 收起<<

Automated segmentation and morphological characterization of placental histology images based on a single labeled image.pdf

共26页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Automated segmentation and morphological characterization of placental histology images based on a single labeled image

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: