Trophectoderm segmentation in human embryo images via inceptioned U-Net

Highlights

• Four novel fully convolutional deep models are proposed based on the U-Net architecture for semantic segmentation of TE in human blastocyst images. Every of the proposed models introduces an effective approach to enhance the global receptive field without adding significant computational cost. While all of the proposed models outperform the base U-Net model, Inceptioned U-Net delivers superior performance compared to the state-of-the-art and establish a new standard for TE segmentation.

• A Multi-Scaled Ensembling strategy is proposed to ultimately enlarge the receptive field to cover the entire image in any existing architecture. This method trains five instances of the model in parallel using training images of five different resolutions. The final prediction map is generated using the aggregation of the five models by a weighted averaging scheme, according to the normalized Jaccard Index error on the training set. The proposed multi-scaled ensembling method offers a trade-off between the quantity and the quality of the spatial information. Moreover, it can be utilized for other segmentation applications regardless of the underlying architecture.

Abstract

Trophectoderm (TE) is one of the main components of a day-5 human embryo (blastocyst) that correlates with the embryo’s quality. Precise segmentation of TE is an important step toward achieving automatic human embryo quality assessment based on morphological image features. Automatic segmentation of TE, however, is a challenging task and previous work on this is quite limited. In this paper, four fully convolutional deep models are proposed for accurate segmentation of trophectoderm in microscopic images of the human blastocyst. In addition, a multi-scaled ensembling method is proposed that aggregates five models trained at various scales offering trade-offs between the quantity and quality of the spatial information. Furthermore, synthetic embryo images are generated for the first time to address the lack of data in training deep learning models. These synthetically generated images are proven to be effective to fill the generalization gap in deep learning when limited data is available for training. Experimental results confirm that the proposed models are capable of segmenting TE regions with an average Precision, Recall, Accuracy, Dice Coefficient and Jaccard Index of 83.8%, 90.1%, 96.9%, 86.61% and 76.71%, respectively. Particularly, the proposed Inceptioned U-Net model outperforms state-of-the-art by 10.3% in Accuracy, 9.3% in Dice Coefficient and 13.7% in Jaccard Index. Further experiments are conducted to highlight the effectiveness of the proposed models compared to some recent deep learning based segmentation methods.

Introduction

In the last four decades, the use of fertility services has increased dramatically. Couples are under pressure from delayed childbearing in combination with the natural decline in female reproductive capacity. By the age of 40, there is a significant reduction in women’s fertility and a high risk of miscarriage. The Canadian Community Health Survey (data for $2009-2010$) reports that about 15% (1.8 million) of Canadian women suffer from impaired fecundity. Despite the importance of infertility treatment, in particular for women, there has been little attention to the development of computer-based technologies to improve the treatment and/or outcomes.

In-Vitro Fertilization (IVF) procedure is one of the most common procedures for treating infertility problem. During IVF, a female’s ovaries are hyper-stimulated to produce multiple egg cells for external fertilization. The fertilized eggs (embryos) are incubated in controlled environmental conditions and imaged digitally using Hoffman Modulation Contrast (HMC) microscopes for about 5 to 6 days. These embryos are visually evaluated so that only embryos with the highest potentials are transferred to the uterus.

Clearly, the gold standard/ground truth for embryo quality is a healthy live birth. Quality assessment of an embryo is to identify its potentials leading to a healthy baby. As one of the most effective and commonly used quality assessment methods, morphological evaluation of an embryo is to assign grades for the quality of an embryo based on intrinsic morphological structures of its components such as the density, size and expansion level of various cells.

Another method for assessing an embryo’s quality is preimplantation genetic screening (PGS), which has excellent ability to predict non-implanting embryos (negative predictive value = 96%) (Scott et al., 2012) but modest positive predictive value ($41-57\%$ live birth rates) (Scott, Ferry, Su, Tao, Scott, Treff, 2012, Werner, Leondires, Schoolcraft, Miller, Copperman, Robins, et al., 2014). However, utilization of this technology remains low due to incremental cost with embryo biopsy and genetic testing (only 8.2% in Canada in 2015). Although the failure of implantation is often due to whole chromosome aneuploidy, the benefits of Preimplantation Genetic Testing for Aneuploidy (PGT-A) over conventional embryo morphology remains an ongoing debate. While earlier, smaller studies have demonstrated improved embryo implantation rates with PGT-A over morphology (Scott, Upham, Forman, Hong, Scott, Taylor, Tao, Treff, 2013, Yang, Liu, Collins, Salem, Liu, Lyle, Peck, Sills, Salem, 2012), increased costs with this technology and the invasiveness of the biopsy make it a less than an optimal method of embryo selection. More recent and larger studies have demonstrated a lack of benefit with PGT-A over embryo morphology, with no evidence of improved ongoing pregnancy or reduced miscarriage (Zegers-Hochschild et al., 2017b). Therefore, morphology assessment still remains the most common method of embryo/blastocyst selection (Gardner, Stevens, Sheehan, Schoolcraft, 2007, Ahlstrom, Westin, Reismer, Wikland, Hardarson, 2011, Luke, Brown, Stern, Jindal, Racowsky, Ball, 2014).

Here, we propose an automatic approach for segmentation of trophectoderm (TE) region in HMC microscopic day$-5$ human embryo (also known as a blastocyst) images using a novel deep learning approach. Automatic segmentation of TE regions in a blastocyst is the first step towards automatic grading of the quality of TE. The size and the quality of the TE region are strong indicators of an embryo’s quality (Gardner et al., 2000). The proposed methods outperform the best-reported results in this field to this date.

Before summarizing the previous work for segmentation of TE region, we review some of the specialist terminologies for an embryo’s progress during incubation. According to the 2017 International Glossary on Infertility and Fertility care definitions (Zegers-Hochschild et al., 2017a), an embryo is a biological organism resulting from the development of the zygote, until eight completed weeks after fertilization, equivalent to 10 weeks of gestational age. The blastocyst is the stage of pre-implantation embryo development that occurs around day $5-6$ after insemination. It contains a thick membrane that surrounds the developing embryo (zona pellucida, ZP), an outer layer of cells (trophectoderm, TE), a fluid-filled central cavity (blastocoele), and an inner group of cells (inner cell mass, ICM). ZP protects the embryo during the development and regulates communications of the inside TE (the component that develops into the placenta) and ICM (the component that gives rise to the structures of the fetus). The fluid-filled cavity of the blastocyst, viz., regions between the TE and ICM, is called the blastocoel or cavity. Fig. 1 depicts a sample blastocyst image with highlighted ZP (Fig. 1-a), TE (Fig. 1-b), and ICM (Fig. 1-c) regions highlighted in blue.