Timelapse-based 3D reconstruction of blastocysts reveals 3D morphologies of human blastocysts

Abstract

Assessing blastocysts from a three-dimensional (3D) perspective introduces a novel approach to clinical embryo evaluation. Currently, assisted reproduction laboratories lack clinically compatible methods to reconstruct 3D blastocyst structures without interfering with embryo culture. To address this, we developed an interference-free and non-invasive method that uses widely adopted time-lapse (TL) imaging systems to reconstruct 3D blastocyst structures directly from multi-focal images and quantitatively calculate various 3D morphological parameters, without requiring embryologist intervention. In this study, we reconstructed 3D structures for 2025 blastocysts using 22,275 TL images and analyzed the association between 20 3D morphological features and embryo transfer outcomes. Several parameters were significantly associated with clinical pregnancy and live birth (P < 0.05). This method is fully compatible with current clinical workflows, providing objective and comprehensive 3D data for blastocyst assessment. It holds strong potential for future implementation in standardized and automated embryo selection based on quantitative morphological criteria.

Introduction

Reliable, non-invasive assessment of embryo quality is critical for improving the success rate of in vitro fertilization-embryo transfer (IVF-ET). Among various embryo transfer strategies, blastocyst transfer has emerged as the predominant approach^1,2, with selection typically guided by the Gardner scoring system based on two-dimensional (2D) microscopic imaging³. However, this manual evaluation is inherently subjective and often fails to distinguish embryos with similar morphological grades^4–8. In a study involving ten embryologists from five clinics, a fair kappa score was observed for the overall agreement between the embryologists when grading inner cell mass (ICM) (Kappa = 0.349, 95% CI: 0.301–0.392) and trophectoderm (TE) (Kappa = 0.397, 95% CI: 0.356–0.423) for embryos with full blastulation⁵.

Conventional 2D imaging fails to fully capture or quantify the three-dimensional (3D) structure of blastocysts. The observed morphological characteristics of ICM and TE can vary depending on the imaging angle, further limiting the accuracy of morphological assessment⁹. To address these limitations, 3D reconstruction techniques have emerged as promising tools, offering a comprehensive and objective evaluation of embryo morphology without the need for invasive labeling.

Recent studies on 3D reconstruction of embryos at the cleavage and blastocyst stages underscore the potential of 3D morphological analysis in predicting clinical outcomes^9–12. For cleavage-stage embryos, reconstruction methods typically employ z-stack imaging at multiple focal depths to visualize individual blastomeres in three dimensions^10,11. Although blastomeres are generally larger and more uniform in size than ICM and TE cells, challenges, such as overlapping structures and cell fragmentation complicate accurate reconstruction. Current studies primarily focus on quantifying the number, morphology, and spatial distribution of blastomeres, while other important features of cleavage-stage embryos, such as multinucleation, degree of fragmentation, and temporal morphological dynamics, remain underexplored.

Compared to cleavage-stage embryos, 3D reconstruction of blastocysts using z-stacks at varying focal depths presents distinct challenges. Clinically available time-lapse (TL) incubators capture only a limited number of focal planes along the Z-axis, resulting in restricted vertical resolution. Additionally, the ICM displays complex morphological features that require higher reconstruction accuracy than the relatively uniform blastomeres observed at earlier developmental stages. Existing methods for clinical blastocyst assessment necessitate manual image acquisition and physical rotation of the blastocyst on Day 5 post-fertilization to obtain images from multiple angles for 3D model generation⁹. These procedures may disrupt the culture environment and are difficult to integrate into routine clinical workflows.

To overcome these limitations, the integration of artificial Intelligence (AI) driven 3D reconstruction algorithms with TL imaging systems offers a promising solution. Such an approach enables automated, standardized, and non-invasive blastocyst assessment, holds promise for improving the accuracy of embryo selection.

This study integrates TL imaging with AI-based 3D reconstruction for blastocysts, creating a novel 3D reconstruction model that quantitatively computes 20 3D morphological parameters using multi-focal-plane images captured by TL, without requiring additional work from embryologists. We reconstructed 3D models for 2025 blastocysts from 2025 frozen-thawed embryo transfer (FET) cycles using 22,275 TL images, reported baseline data for human blastocyst 3D morphological parameters, and evaluated associations between each parameter and blastocyst transfer outcomes. This study presents, to the best of our knowledge, the first comprehensive methodology for full-process 3D reconstruction of blastocysts and the large-scale quantification of 3D morphological parameters.

Results

Main descriptive features of the cycles

A schematic overview of the study workflow is shown in Fig. 1. In this study, we reconstructed 3D models of 2025 blastocysts from 2025 FET cycles, utilizing 22,275 TL images. Characteristics and clinical outcomes of these 2025 FET cycles are presented in Supplementary Table 1. Due to the very limited number of blastocysts with an ICM grade of C in FET cycles, our analysis focused exclusively on blastocysts with ICM grades of A or B. Gardner grading and demographic characteristics of the 2025 FET cycle blastocysts and 26 fluorescent reconstructed blastocysts are presented in Supplementary Table 2.

Fig. 1. Flowchart of the time-lapse-based blastocyst 3D reconstruction study.

The study workflow covers data acquisition, focal-plane image generation, ICM and TE segmentation, 3D reconstruction, validation of reconstructed models, and outcome analysis. FET, Frozen Embryo Transfer; 3D, Three-Dimensional; TE, Trophectoderm; ICM, Inner Cell Mass.

Three-dimensional parameters of blastocyst

Figure 2 provides definitions, units, and data distributions for 20 blastocyst 3D morphological parameters. Among these, the TE surface area represents the surface area of the TE on the side facing the blastocyst cavity, while TE density is defined as the TE cell number divided by the blastocyst surface area. The blastocyst space is divided into eight quadrants centered around the blastocyst’s center of gravity. The ICM quadrant is the quadrant containing the center of gravity of the ICM (Fig. 2c). Two ICM-TE relationship parameters are defined: ICM-TE relationship 1 indicates that the ICM quadrant aligns with the quadrant containing the maximum number of TE cells, and ICM-TE relationship 2 indicates that the ICM quadrant or its adjacent quadrants align with the quadrant containing the maximum number of TE cells. The definition of ICM shape factor, ICM minor-to-major axis ratio (Fig. 2a), and spatial distance between ICM and TE (Fig. 2b) are detailed in the methods.

Fig. 2. Definitions and distributions of the 20 blastocyst 3D morphological parameters.

The left panel shows the distributions of each parameter (y-axis: number of blastocysts; x-axis: units of the parameter). The middle panel lists the parameter numbers, names, and definitions. a–c On the right illustrate the definitions of selected 3D morphological parameters. a ICM minor-to-major axis ratio, indicating how elongated the ICM shape is. b Spatial distance between the ICM and TE: R denotes blastocyst radius; d is the distance from the ICM centroid to the blastocyst centroid, reflecting ICM displacement. c Corresponding to parameters 19 and 20, the blastocyst is divided into eight equal quadrants (Q1–Q8) around its centroid. The ICM quadrant (Q_ICM) is defined as the quadrant containing the ICM centroid. These two parameters describe the spatial relationship between the ICM location and the spatial density distribution of TE cells. TE, Trophectoderm; ICM, Inner Cell Mass.

Verification of human blastocyst fluorescence staining

Figure 3 shows the Imaris software 3D reconstruction and verification results of blastocyst fluorescence staining.

Fig. 3. 3D reconstruction and validation of blastocysts using fluorescence staining.

a Comparison of 3D reconstructions obtained from fluorescence staining and from TL imaging, showing the morphology of the whole blastocyst, the ICM, and the TE. b Relative error between fluorescence-based and TL-based reconstructions. ICM, Inner Cell Mass; TE, Trophectoderm; TL, Time-Lapse.

To obtain the true values of 20 blastocyst 3D morphological parameters, the cell nucleus, trophoblast, cell membrane, and ICM were fluorescently stained and reconstructed (Fig. 3a), and the relative errors between fluorescence-staining 3D reconstructions and TL 3D reconstructions were compared (Fig. 3b). Compared with the fluorescence reconstruction results, our 3D morphological measurement method achieved a relative error of 2.13% ± 1.63% for the blastocyst surface area, 4.03% ± 2.24% for the blastocyst volume, 1.98% ± 1.32% for the blastocyst diameter, 4.83% ± 6.26% for the ICM surface area, 6.64% ± 12.83% for the ICM volume, and 10.00% ± 8.73% for the TE cell number.

The relationship between 3D morphological parameters and clinical outcomes

Table 1 illustrates the associations between 20 3D morphological parameters and clinical outcomes. Among the 2025 blastocyst samples, most 3D parameters followed a normal distribution. Parameters reflecting the overall morphology of the blastocyst were significantly associated with both pregnancy and live birth. Larger values for blastocyst surface area, volume, diameter, and blastocyst cavity volume, alongside a smaller blastocyst surface area/volume, were associated with higher probabilities of pregnancy and live birth (P < 0.001). Parameters related to TE quality also showed significant associations with pregnancy and live birth. Larger TE surface area, TE volume, TE cell number, and TE density were all linked to increased likelihoods of pregnancy and live birth (P < 0.001). Wilcoxon rank sum test revealed no significant relationships between ICM surface area, ICM volume, or ICM minor-to-major axis ratio and clinical pregnancy or live birth. The ICM shape factor, however, was significantly associated with both pregnancy and live birth (P < 0.05); a smaller ICM shape factor, indicating a shape closer to a sphere, correlated with higher probabilities of successful clinical pregnancy and live birth. Additionally, a larger spatial distance between ICM and TE (P < 0.05) and a smaller ICM volume/blastocyst volume (P < 0.05) were significantly associated with pregnancy outcomes. Embryos with a higher number of TE cells in the ICM quadrant were more likely to result in clinical pregnancy (P < 0.01). Conversely, a larger ICM surface area/volume and an increased number or proportion of TE cells within the ICM quadrant were associated with a higher risk of miscarriage (P < 0.05). The associations between ICM-TE relationship 1, ICM-TE relationship 2, and clinical outcomes were not statistically significant.

Table 1.

The associations between blastocyst 3D morphological parameters and clinical outcomes

3D parameters	Clinical pregnancy			Live birth			Miscarriage
	+ (1282)	- (743)	P-value	+ (1282)	- (743)	P-value	+ (215)	- (1067)	P-value
Blastocyst surface area (um2)	77,400.1 ± 14452.1	74,058.6 ± 14602.5	<0.001	77,615.1 ± 14483.7	74,569 ± 14554.4	<0.001	76,333 ± 14280.5	77,615.1 ± 14483.7	0.341
Blastocyst volume (um³)	1.99*106 ± 552,361.2	1.87*106 ± 546,603.4	<0.001	2.00*106 ± 554,441.5	1.89*106 ± 546,194.7	<0.001	1.95*106 ± 541,184.3	2.00*106 ± 554,441.5	0.313
Blastocyst diameter (um)	158.3 ± 14.2	154.6 ± 14.4	<0.001	158.5 ± 14.1	155.2 ± 14.4	<0.001	157.1 ± 14.1	158.5 ± 14.1	0.297
Blastocyst surface area / volume (1/um)	0.0399 ± 0.00392	0.0409 ± 0.00421	<0.001	0.0398 ± 0.00388	0.0407 ± 0.00419	<0.001	0.0402 ± 0.0041	0.0398 ± 0.00388	0.298
Blastocyst cavity volume (um³)	1.32*106 ± 419,788.3	1.22*106 ± 413,179.9	<0.001	1.33*106 ± 421,379.2	1.23*106 ± 413,365	<0.001	1.28*106 ± 410,472.7	1.33*106 ± 421,379.2	0.216
TE surface area (um2)	63,453.3 ± 12,518.3	60,261.1 ± 12,553.8	<0.001	63,674.8 ± 12,543.4	60,730.9 ± 12,535.4	<0.001	62,354.3 ± 12,363.5	63,674.8 ± 12,543.4	0.249
TE volume (um³)	589,148.6 ± 143,121	567,352.8 ± 145,104.2	0.001	589,828 ± 142,996.3	571,487.6 ± 144,991.8	0.006	585,776.8 ± 144,025.5	589,828 ± 142,996.3	0.750
TE cell number	98.9 ± 5.67	97.4 ± 5.92	<0.001	98.9 ± 5.6	97.8 ± 5.9	<0.001	99.1 ± 5.7	98.9 ± 5.6	0.293
TE density (/1000um2)	0.0013 ± 0.00024	0.0014 ± 0.00025	<0.001	0.0013 ± 0.00024	0.0014 ± 0.00025	<0.001	0.0013 ± 0.00024	0.0013 ± 0.00024	0.162
ICM surface area (um2)	10,509.2 ± 4357.7	10,788.8 ± 5476.5	0.974	10,550.6 ± 4158.6	10,679.9 ± 5425.8	0.303	10,304 ± 5241.7	10,550.6 ± 4158.6	0.087
ICM volume (um³)	97,549.4 ± 61,017	100,920 ± 76,695.6	0.869	97,909.1 ± 57,545.9	99,762.9 ± 76,546.3	0.221	95,764.3 ± 76,070.2	97,909.1 ± 57,545.9	0.065
ICM surface area / volume (1/um)	0.1225 ± 0.0345	0.1237 ± 0.0333	0.519	0.1216 ± 0.0339	0.1244 ± 0.0342	0.085	0.127 ± 0.0371	0.1216 ± 0.0339	0.043
ICM volume / blastocyst volume	0.0529 ± 0.0356	0.0583 ± 0.0443	0.027	0.0529 ± 0.0340	0.057 ± 0.0440	0.261	0.0524 ± 0.0425	0.0529 ± 0.0340	0.254
The number of TE cells in ICM quadrant	12.20 ± 1.80	11.98 ± 1.79	0.008	12.16 ± 1.80	12.08 ± 1.80	0.468	12.44 ± 1.80	12.15 ± 1.80	0.019
The proportion of TE cells in ICM quadrant	0.1253 ± 0.017	0.1248 ± 0.016	0.506	0.1249 ± 0.017	0.1254 ± 0.017	0.390	0.1275 ± 0.017	0.1249 ± 0.017	0.023
ICM shape factor	2.270 ± 0.052	2.278 ± 0.059	0.022	2.270 ± 0.051	2.278 ± 0.059	0.013	2.275 ± 0.056	2.269 ± 0.051	0.279
ICM minor-to-major axis ratio	0.7322 ± 0.139	0.7286 ± 0.147	0.792	0.7342 ± 0.137	0.7271 ± 0.147	0.461	0.7222 ± 0.147	0.7342 ± 0.137	0.370
Spatial distance between ICM and TE	0.2010 ± 0.0396	0.1967 ± 0.0425	0.024	0.2009 ± 0.0387	0.1977 ± 0.0428	0.106	0.2010 ± 0.0437	0.2009 ± 0.0387	0.734
ICM-TE Relationship 1	13.8%	12.5%	0.411	13.3%	13.4%	0.972	16.3%	13.3%	0.249
ICM-TE Relationship 2	52.4%	52.9%	0.836	53.0%	52.1%	0.666	49.3%	53.0%	0.316

3D, Three-Dimensional, ICM Inner Cell Mass, TE Trophectoderm.

After adjusting for transfer time (day 5 or day 6), female age (≤35 or >35), and endometrial thickness (≤7 or >7) using generalized estimating equation (GEE) analysis, the associations between the aforementioned 3D morphological parameters and clinical pregnancy and live birth outcomes remained significant (P < 0.05) (Supplementary Table 3).

As shown in Fig. 4a, the Cochran–Armitage trend test indicates a linear relationship between 3D morphological parameters related to blastocyst and TE quality and both pregnancy and live birth rates (P < 0.01). Additionally, a significant linear trend was observed between the ICM shape factor and both pregnancy and live birth rates, as well as between the ICM surface area-to-volume ratio and live birth rate (P < 0.05). In a subgroup analysis of grade 4 blastocysts (n = 1688), overall 3D parameters remained significantly associated with pregnancy and live birth rates (P < 0.01). For blastocysts with TE grade B (n = 1082), TE surface area, TE volume, and TE density continued to show significant relationships with pregnancy and live birth rates (P < 0.05), and TE cell count was significantly associated with pregnancy (P < 0.05). In blastocysts with ICM grade B (n = 1563), the ICM shape factor remained significantly associated with both pregnancy and live birth rates (P < 0.05), while the ICM surface area/volume was significantly associated with live birth rate (P < 0.05).

Fig. 4. The relationship between 3D morphological parameters and clinical outcomes.

a Bubble plots showing the associations between individual 3D morphological parameters and two key clinical outcomes—clinical pregnancy rate (solid circles) and live birth rate (open circles)—in the overall dataset and in specific subgroups. Bubble size indicates the number of blastocysts, and P values from Wilcoxon rank sum tests are shown within each plot. b Pie chart illustrating the distribution of blastocyst grades in the study cohort. c Bar plots showing the P values from Wilcoxon rank sum tests for the associations between each 3D morphological parameter and newborn gender. ICM, Inner Cell Mass; TE, Trophectoderm.

Current embryo quality grading systems often categorize many blastocysts similarly, especially 4BB blastocysts, which are commonly encountered in assessments at most reproductive centers. In our dataset of 2025 blastocysts, 4BB blastocysts accounted for 43.7% (n = 885), making further differentiation difficult using traditional grading methods (Fig. 4b). Within the 4BB blastocysts, several 3D parameters—including TE surface area and TE volume—remained significantly associated with pregnancy (P < 0.05). Additionally, blastocyst volume and the blastocyst surface area/volume were significantly associated with live birth rate (P < 0.05) (Fig. 4a).

The associations between 3D parameters and obstetric outcomes are outlined in Supplementary Table 4. Notably, certain 3D morphological parameters of the ICM exhibit significant differences based on newborn gender (P < 0.05) (Fig. 4c). Female embryos tend to have a larger ICM volume, a smaller ICM surface area/volume ratio, a larger ICM volume/blastocyst volume ratio, fewer TE cells in the ICM quadrant, a smaller ICM shape factor, and a larger ICM minor-to-major axis ratio. In summary, ICMs of female embryos are relatively larger and more spherical.

Discussion

Blastocyst transfer cycles are a crucial aspect of IVF procedures worldwide. Despite their importance, clinical practice still demands more refined, comprehensive, and automated evaluation criteria for the non-invasive assessment of blastocysts^13,14. The non-invasive 3D reconstruction of human blastocysts provides valuable new insights for clinical blastocyst selection and has emerged as a key focus in embryo selection research.

Currently, confocal laser scanning microscopy and multiphoton excitation laser scanning microscopy are widely employed to assess the 3D morphology of blastocysts via fluorescent labeling. However, these methods require pretreatment and dye labeling, leading to significant photodamage, rendering them unsuitable for clinical embryo selection. Beyond fluorescent labeling, most available blastocyst measurement techniques are tailored for laboratory applications and animal research, presenting substantial limitations for clinical use. For example, non-invasive Optical Coherence Tomography (OCT)¹⁵ and Full-Field Optical Coherence Tomography (FF-OCT)¹⁶ enable live imaging without fluorescent staining, which substantially reduces the light dose exposure to the sample. These methods are particularly valuable in studying early embryonic development. However, they cannot measure several 3D morphological parameters that may hold clinical value and still pose photodamage risks compared to the widely used TL imaging equipment in clinical practice. Additionally, certain techniques require supplementary apparatus, such as a rotation stage^17–19, which can complicate clinical setups. Pierré et al. introduced a compact 3D diffraction microscope designed for direct placement within a cell incubator, enabling long-term observation of developing organisms. This system reconstructs the sample’s optical properties (refractive index and absorption rate) in 3D using images recorded from an LED array under varying illumination angles²⁰. However, despite its innovative approach, the system remains complex and is currently unable to measure certain 3D morphological parameters of blastocysts.

Currently, studies on 3D reconstruction of human blastocysts with a specific focus on clinical blastocyst selection are limited. Shan et al. developed a method in which embryos are fixed with two needles and rotated at fixed angles for image acquisition⁹. Through image segmentation, the TE and ICM are reconstructed by matching the positions of the pixel points, enabling feature extraction for both ICM and TE. This surface reconstruction approach allows for more precise TE cell counting and analysis of TE cell size. However, the ICM reconstruction remains relatively coarse. Additionally, this technique requires manual manipulation of blastocysts by embryologists and is incompatible with commonly used TL imaging systems. Factors, such as ambient light and the angle of rotation of each viewpoint affect the consistency of reconstruction results, limiting its clinical applicability.

This paper introduces a 3D reconstruction method for blastocysts that utilizes 11 (or more) focal-plane images directly from the TL system to reconstruct various 3D morphological parameters. This approach requires no additional equipment or operation procedures, enhancing its feasibility and practicality for clinical applications. Moreover, it efficiently utilizes the extensive TL imaging data already available for blastocysts. The method begins by interpolating additional images between the original 11 focal-plane images to improve reconstruction quality. Next, the ICM and TE are segmented from these multi-focal images and fitted into point cloud models. These point cloud models are then used to reconstruct surface models of the ICM and TE. Finally, texture features are mapped onto the surface reconstruction models. This spatial reconstruction-based method allows for the extraction of TE morphology, ICM morphology, and the spatial relationship between the ICM and TE, enabling analysis of their associations with clinical outcomes and providing deeper insights into the embryo development process. In addition, it lays the foundation for future quantification of texture information and the exploration of novel 3D features.

The primary challenge of the proposed 3D reconstruction method for blastocysts lies in segmenting the ICM and TE from out-of-focus images. The equatorial plane of the blastocyst serves as the optimal focal plane during imaging. However, capturing the upper surface often results in overexposure and overlapping artifacts, leading to blurring in images taken outside the optimal focal plane. This blurring complicates the accurate delineation of TE and ICM boundaries and the full region of the ICM. Additionally, when the ICM has a low degree of aggregation, its contour boundary is difficult to determine, especially in out-of-focus images. The segmentation process is further complicated by the presence of vacuoles and apoptotic cells, which obscure the ICM boundary. To mitigate these interferences and improve segmentation performance, we apply Richardson-Lucy with total Variation (RLTV) and Adaptive Histogram Equalization (AHE) techniques to enhance cell edge sharpness and overall image contrast.

During the reconstruction of TE cell contours, comparisons with fluorescent reconstruction results showed that surface reconstruction based on the TE mask from the segmentation network produced a blastocyst model with a volume and surface area larger than its actual size. Further analysis revealed that the reconstructed blastocyst surface contained numerous holes. This issue stemmed from the close alignment of the inner and outer TE contours in certain focal planes, indicating relatively thinner regions of the TE in these areas. To address this, we separately extracted the inner and outer boundary contours of the TE and performed surface fitting and reconstruction for each contour. This approach yielded a smoother blastocyst surface reconstruction based on the outer contour of the TE, which more accurately aligned with the actual blastocyst structure.

The substantial variation in ICM volume leads to considerable differences in the number of focal-plane images required, complicating the accurate fitting and reconstruction of its geometric structure. The primary challenges in segmenting the ICM across multiple focal planes arise from out-of-focus images (i.e., those distant from the optimal focal plane) and interpolated data. The ICM’s complex and variable shape is easily affected by vacuoles and cell debris, and low-aggregation ICMs often have indistinct boundaries across focal planes, hindering accurate segmentation. Furthermore, the unique imaging characteristics of Hoffman Modulation Contrast (HMC) make cells in the upper layer of the 0 focal plane susceptible to artifacts, complicating the identification of cell boundaries and regions in out-of-focus planes. This results in difficulties obtaining accurate segmentation masks for both manual annotation and neural network-based recognition. Occasionally, manual labeling errors also occur, as shown in Fig. 5h, where the model’s predictions more accurately capture the ICM boundary variations across focal planes. While simpler boundary-fitting functions may produce substantial shape discrepancies, more complex fitting functions require more extensive image data. To address these challenges, we developed an adaptive model that selects an appropriate surface fitting function based on the number of available focal planes in ICM and the trends that ICM changes across those planes. This approach enhances both the accuracy and generalizability of the 3D reconstruction model.

Fig. 5. Methods for 3D reconstruction of blastocysts.

a Focal-plane interpolation approach. b Schematic of spherical approximation for multi-focal plane images. c Spatial contour segmentation of TE. d Auxiliary circular markings on focal planes. e Process for fitting the geometric structure to the TE and ICM. f The focal depth estimation process. g Obtaining peak focus positions along the z-axis (quoted from Wang et al.)²⁹. h Comparison of ICM boundaries in different focal planes for manual labeling and model predictions. i Schematic of hemispherical texture projection (quoted from Jiang et al.)³⁰ j Textured map of the blastocyst’s external surface. k Texture-mapped 3D geometric model. l Various geometric shapes clarify the concept of the shape factor (quoted from Cao et al.)³³. m The improvement of cell edge sharpness and overall contrast by RLTV and AHE. n Performance of ICM and TE segmentation models in training and validation datasets. o Performance of the ICM segmentation model in the test dataset. p Performance of the TE segmentation model in the test dataset. AHE, Adaptive Histogram Equalization; IoU, Intersection over Union; RLTV, Richardson-Lucy with total Variation; TL, Time-Lapse.

This study presents a comprehensive series of quantitative 3D morphological parameters of blastocysts based on a large sample size, introducing new 3D morphological predictive indicators, such as the ICM shape factor and the spatial distance between the ICM and TE. These indicators offer a novel perspective for blastocyst assessment. By providing objective and quantitative measurements, 3D morphological parameters facilitate a more comprehensive evaluation of blastocysts from a 3D perspective and have the potential to aid in distinguishing blastocysts with subtle morphological differences that are difficult to detect in 2D microscopic images. Given the subjectivity often present in blastocyst evaluations, introducing 3D parameters as quantitative references for assessing blastocyst quality holds promise for enhancing consistency among embryologists and reducing subjective variability. The observed linear trends and quantitative associations between 3D parameters and clinical outcomes, along with the significant associations found in certain 3D parameters of 4BB blastocysts, suggest that this approach may hold potential to provide valuable numerical references for clinicians and embryologists in the future, for distinguishing blastocysts with identical ICM or TE grades under traditional scoring systems, such as the frequently encountered 4BB blastocysts in clinical settings.

Unlike black-box AI models, this new approach reveals specific associations between blastocyst 3D morphology and clinical outcomes, facilitating further exploration of blastocyst morphological characteristics based on large-scale human embryo data, aiding in the exploration of embryo development. Compared to conventional evaluation methods, this approach offers improved objectivity and enhanced spatial resolution of 3D morphological structures. Looking ahead, automated and precisely quantified 3D measurements may contribute to the establishment of clearer, standardized criteria for blastocyst selection. Furthermore, the ability to generate transparent and quantifiable visualizations of embryo quality may enhance communication between embryologists and patients, improving understanding and trust.

One limitation of this study is the limited number of blastocysts included, particularly low-quality embryos. All blastocysts analyzed were from FET cycles, where the highest-quality embryos are prioritized for transfer, resulting in clinical pregnancy and live birth rates (63.3 and 52.7%, respectively) that are slightly higher than average clinical outcomes. Additionally, the 3D reconstruction method has certain requirements for imaging quality and the developmental conditions of blastocysts. In clinical TL embryo culture, interference with imaging quality is an inherent and unavoidable concern. For example, the ability to capture the entire blastocyst is constrained by both the design of the culture dish and the field of view of the TL imaging system. In some cases, once the embryo reaches the blastocyst stage, a complete image cannot be obtained, and reconstructions derived from such data may fail to accurately represent the true blastocyst morphology. In future clinical applications, these limitations could be mitigated through measures, such as manually adjusting the imaging field or embryo position, as well as optimizing the design of TL equipment and culture dishes. For low-quality blastocysts exhibiting distinct developmental abnormalities—such as collapse²¹, extensive fragmentation or vacuolation, or the absence of visible ICM-like cell clusters—the accuracy of reconstruction can be substantially compromised. In clinical practice, when no higher-quality embryos are available, such blastocysts may still be considered for transfer after thorough physician–patient discussion and informed consent. The reconstruction of these embryos constitutes one of the limitations of the present study. We propose that within the current TL culture system, increasing the number of focal planes may enable improved 3D reconstruction quality. Furthermore, due to the limited blastocyst sample size, GEE was not applied for confounder adjustment in the subgroup analyses. The associations between 3D morphological parameters and clinical outcomes warrant external validation in an independent, larger-scale dataset.

In conclusion, this study introduces a 3D reconstruction AI model for blastocysts, which is fully compatible with widely used clinical TL systems. Using multi-focal plane images captured automatically by TL systems, the model reconstructs a comprehensive set of 3D morphological parameters for blastocysts. These 3D parameters provide valuable predictive insights into blastocyst transfer outcomes by offering objective and quantitative data that may enhance clinical evaluation. By delivering additional 3D structural information, the model may improve blastocyst assessment in clinical settings and serve as a valuable complement to existing scoring systems.

Materials and Methods

Study design, patients, and data

This retrospective study included TL imaging data and patient information from 2686 FET cycles conducted at the Reproductive Medicine Center of Tongji Hospital between April 2018 and December 2021. There were no restrictions on patient inclusion criteria, and only blastocysts graded as Gardner stage 3 or 4 were included. All patients signed written informed consent and received routine clinical treatment at the Reproductive Medicine Center of Tongji Hospital. No additional interventions were conducted. This study adhered to the Declaration of Helsinki for Medical Research Involving Human Subjects and received approval from the ethics committee of the Reproductive Medicine Center of Tongji Hospital (Approval Number: TJ-IRB202412141).

All embryos were cultured in a TL incubator (EmbryoScope +), and 11 focal plane images from the final time point before transfer were selected as model inputs. Following preliminary image quality screening, 564 blastocysts were excluded due to incomplete imaging or abnormal images, and 97 blastocysts were excluded due to interference from granulosa cells, fragments, or vacuoles. Ultimately, 22,275 TL images of 2025 blastocysts from 2025 FET cycles were used for 3D reconstruction.

This study developed a 3D reconstruction pipeline for blastocysts capable of quantifying 20 distinct 3D morphological parameters. The pipeline consists of three main components: AI-assisted generation of multi-focal plane images, AI-assisted segmentation of the ICM and TE, and 3D reconstruction of the blastocyst structure. To validate reconstruction accuracy, 26 surplus blastocysts (Gardner stage 3 or 4), donated for research with written informed consent, were subjected to fluorescent staining. The resulting fluorescence-based 3D reconstructions were then compared to the TL 3D reconstruction models for validation. Subsequently, the associations between the 20 3D morphological parameters and clinical outcomes were systematically analyzed using a dataset comprising 2025 embryos from 2025 FET cycles. A schematic overview of the study workflow is shown in Fig. 1.

IVF, ICSI, embryo incubation, and timelapse monitoring

The methods for semen/cumulus-oocyte complex (COC) preparation, insemination, and embryo culture were conducted as previously described²². Briefly, semen was collected via masturbation and optimized through density gradient centrifugation. Sperm concentration, motility, and morphology were assessed following the fifth edition of the World Health Organization guidelines. For IVF cycles, sperm density was adjusted to 10,000 motile spermatozoa per COC, with insemination occurring 3–4 hours post-retrieval. In ICSI cycles, COCs were denuded using a brief exposure to HYASE (Vitrolife, Sweden) containing hyaluronidase, and sperm was injected into metaphase II (MII) oocytes. Zygotes were then transferred to G1 Plus (Vitrolife, Sweden) medium and cultured in the Embryoscope Plus TL microscope system (Vitrolife, Denmark), capable of capturing images at 11 focal planes. All fertilized oocytes were cultured at 37 °C in an environment of 6% CO₂, 5% O₂, and 89% N₂ until biopsy or cryopreservation on day 5 or 6. All embryo images were acquired on day 5 or 6 post-fertilization, prior to biopsy or cryopreservation. The grading of blastocysts was conducted according to the Gardner scoring system²³.

FET cycles

The FET procedure was conducted as previously described²⁴. Briefly, oral estradiol was administered in a stepped dosing regimen: 2 mg/day from Day 1 to Day 4, 4 mg/day from Day 5 to Day 8, and 6 mg/day from Day 9 to Day 12. On Day 13, 40 mg of progesterone was administered. The blastocyst was thawed and transferred after 5 days of progesterone administration. Each FET cycle involved single-blastocyst transfers (SBTs).

Outcome measures

All of these FET cycles were SBTs performed at the Reproductive Medicine Center of Tongji Hospital. Clinical pregnancy was defined as the presence of a gestational sac with fetal cardiac activity during ultrasound examination four weeks and six weeks after embryo transfer. Live birth outcomes, including gestational age, birth sex, and birth weight, were obtained through telephone interviews with parents.

Image preprocessing

The data for non-invasive 3D reconstruction were obtained from 11 focal-plane images captured by a TL incubator (EmbryoScope +). The TL imaging system is Hoffman imaging with an image resolution of 800×800, and each focal plane is spaced 15 um apart on the z-axis. Although additional focal planes would provide more information for 3D reconstruction and bring the model closer to the actual structure, the TL imaging equipment has limitations that restrict data acquisition. To address this, the study employs a focal-plane interpolation approach to minimize 3D reconstruction errors caused by insufficient data (Fig. 5a). The focal-plane interpolation generation network was trained with 200 epochs based on Generative Adversarial Network (GAN) neural network, and the loss function design was based on Mean Square Error (MSE) and Structural Similarity (SSIM). Applying principles of a GAN, interpolated images are generated between adjacent focal planes, and a loss function calculates the error between interpolated and real data. This process iterates until the error reaches an optimal threshold, at which point the interpolated data are deemed a reliable representation of the real structure²⁵. After three iterations, the blastocyst’s focal planes are spaced at 2 μm intervals, corresponding to blastocyst volume. Experimental measurements indicate that most blastocysts contain 79–81 focal planes, so the final 3D reconstruction in this study utilizes an 81 focal-plane dataset.

ICM and TE segmentation

Accurate segmentation of the ICM and trophectoderm TE is essential for the 3D reconstruction of blastocysts, as segmentation performance directly affects the precision of the 3D reconstruction. The annotation of ICM and TE was performed in strict accordance with the Expert Consensus for the Construction and Quality Control of the AI Blastocyst Morphology Assessment Dataset Version 2024²⁶, with reference to the definitions of ICM and TE in existing studies²⁷. All annotations were reviewed and confirmed by two senior embryologists, Xinling Ren and Li Wu, each with over 15 years of experience in clinical embryo evaluation. For ICM segmentation, we used 2480 manually annotated images derived from 275 randomly selected embryos, split into training (95%) and validation (5%) sets. An independent test set comprised 868 manually annotated images from 94 additional embryos. For TE segmentation, 3,240 manually annotated images from 40 randomly selected embryos were utilized, divided into training (75%) and validation (25%) sets. The TE test set consisted of 220 manually annotated images from 20 held-out embryos, ensuring complete independence from the training/validation data.

Artifacts commonly appear between adjacent layers in HMC brightfield imaging, leading to significant segmentation errors for both ICM and TE. To mitigate this, and given the spherical morphology of blastocysts, we implemented an auxiliary circular marking strategy to aid in TE labeling. The multi-focal plane image data are approximated as horizontal cross-sections of a sphere (Fig. 5b), with auxiliary circular markings on the multi-focal planes (Fig. 5d). In each focal plane, the TE is assumed to lie within the boundaries of these auxiliary circular markings, with the segmented spatial contour of the TE shown in Fig. 5c.

UNet + + is a deep learning model optimized from the classical U-Net architecture, primarily used for medical image segmentation and other high-precision pixel-level classification tasks. While U-Net is widely used in segmentation for its ability to effectively utilize contextual information while preserving image details, UNet + + enhances segmentation performance by improving the feature extraction and fusion processes while maintaining computational efficiency. These advancements make UNet + + particularly effective for segmenting images with complex details and rich semantic information²⁸. In this study, UNet + + is employed to segment ICM and TE across multiple focal planes. The model’s performance is evaluated based on the average Intersection over Union (IoU) across different focal planes.

A total of 2480 images with marked blastocysts at 21 focal planes from 275 embryos were used to train the Unet + + network for ICM segmentation. The encoder backbone employs an SE-ResNeXt50_32×4 d architecture, initialized with ImageNet pre-trained weights, and uses a sigmoid activation function. This dataset comprised 11 images acquired from TL imaging and 10 generated through a focal-plane interpolation algorithm, and the image data were augmented by random image rotation, contrast enhancement, and brightness enhancement. And the entire data sample was expanded to 10,440 images. To improve the sharpness of cell edges and overall contrast, RLTV and AHE were applied (Fig. 5m). 95% of the data was allocated for training and 5% for validation. The model was trained over 50 epochs, and the version achieving the highest IoU on the validation set was selected as the final model. The optimal model yielded an IoU of 0.679 on the training set and 0.645 on the validation set, with corresponding Dice coefficients of 0.808 and 0.769, respectively (Fig. 5n). The test set comprised an additional 868 images from 94 labeled blastocysts (held-out data, independent of the original dataset).

TE segmentation was trained using 3240 annotated images of labeled blastocysts from 81 focal planes of 40 embryos (including 70 generated from a focal-plane interpolation algorithm). The model architecture was the same as that used for ICM segmentation. 75% of the data was allocated for training and 25% for validation. The model was trained over 100 epochs, and the version achieving the highest IoU on the validation set was selected as the final model. The optimal model achieved an IoU of 0.797 on the training set and 0.715 on the validation set, with corresponding Dice coefficients of 0.887 and 0.847, respectively (Fig. 5n). For testing, we used a separate dataset that included 220 images from an additional 20 embryos (11 images drawn at intervals from the 81 focal plane data), none of which overlapped with the training/validation data. After excluding blurred and unusable TE cell images, the final test set comprised 218 images.

All model training in this study was done on a DELL Tower 7920 server with Intel(R) Xeon(R) Silver 4216 @2.10 G CPU and NVIDIA GeForce RTX 3090 graphics card. The deep learning framework is based on PyTorch 2.0.1 and CUDA 11.7, and the segmentation network is built based on segmentation_model_pytorch 0.3.3, MMCV2.0.0rc4 and albumenttations1.0.3. In the training process, batch size is 4, optimizer is Adam, loss function type is Dice.

Performance test results of ICM and TE segmentation

The segmentation performance of ICM and TE is evaluated using IoU, Dice, and recall. For ICM, since the segmentation result of the best focal plane affects the 3D reconstruction result more, it is evaluated in a separate column. The model achieved a segmentation accuracy of 0.639 IoU, 0.765 Dice score, and 0.710 recall for all data, and for best focal, IoU, Dice, and recall are 0.722, 0.833, and 0.770, as shown in Fig. 5o. For the TE segmentation model, the average IoU, Dice, and recall for all data were 0.736, 0.879, and 0.856, respectively. For the 20 images with the best focal plane, the segmentation results show an IoU of 0.757, a Dice of 0.891, and a recall of 0.855. For the 40 images with one focal plane difference from the optimal focal plane, the IoU, Dice, and recall are 0.766, 0.896, and 0.857, respectively, while for the 40 images with two focal planes difference, the IoU, Dice, and recall drop to 0.736, 0.878, and 0.859, respectively. More details of the performance are shown in Fig. 5p.

Segmentation performance was highest for images at the optimal focal plane and those one focal plane away, with accuracy declining as the distance from the optimal focal plane increased. This decline occurs because images further from the optimal focal plane exhibit blurrier contours of TE cells, making segmentation more challenging for both human observers and the model. Currently, the only solution to mitigate this issue is to acquire more focal planes. Therefore, depth estimation was introduced into the 3D reconstruction process to minimize this phenomenon.

The 3D reconstruction process of blastocyst

The 3D reconstruction workflow (Fig. 6) integrates the previously described interpolation and segmentation networks, forming a unified pipeline for 3D blastocyst model reconstruction. 3D reconstruction of blastocysts consists of two parts: geometric structure reconstruction and texture feature reconstruction.

Fig. 6. Workflow for 3D reconstruction of blastocysts.

The 3D reconstruction workflow consists of geometric structure reconstruction (shown in purple) and texture feature reconstruction (shown in blue). ICM, Inner Cell Mass; TE, Trophectoderm; TL, Time-Lapse.

For geometric structure reconstruction

Step 1: Image preprocessing. Since existing clinical TL equipment usually acquires images in only 11 focal planes, the information available for 3D reconstruction of blastocyst structure is limited. Therefore, focal plane interpolation was used to extend these 11 focal planes to 81 to generate additional data for 3D reconstruction. Then, the ICM and TE were segmented.

Step 2: Fitting 3D spatial structures.

For TE: The TE outline is extracted from segmented mask data to fit a point cloud model. Given the spherical structure of blastocysts, auxiliary labeling strategies are implemented before TE segmentation to assist in the localization of TE cells, which helps mitigate interference from artifacts between adjacent layers.

For ICM: The ICM boundary contours were similarly extracted and converted into a point cloud model.

Step 3: Surface reconstruction. The point cloud models of both the ICM and TE obtained from steps two are utilized for surface reconstruction to create a geometric structure model.

For texture feature reconstruction, depth estimation, and multifocal image fusion were performed on the original 11 focal plane images to generate surface depth maps and texture fusion maps of the embryos.

Finally, the geometric structure reconstruction model and the texture feature reconstruction model were aligned according to the point cloud position to generate a 3D model of the blastocyst with spatial structure and texture features. A visual demonstration of the reconstructed model is provided in Supplementary Movie 1.

Fitting and geometric structure reconstruction

Fitting the geometric structure to the TE and ICM requires obtaining their respective point cloud models, as illustrated in Fig. 5e. Step 1 is multi-focal plane ICM and TE segmentation. The 81 focal planes generated by focal-plane interpolation were segmented into TE and ICM, respectively, to produce corresponding mask images for each focal plane. Step 2 is to construct the point cloud models of TE and ICM separately. The multi-focal mask image obtained in the first step will be used for contour extraction. For TE, the contour information of TE in each focal plane was stacked to create the TE point cloud model. Due to the more irregular shape of the ICM, its contour points across focal planes were fitted in azimuthal order to form the ICM point cloud model. Step 3 is to transform the point cloud model into a mesh geometric structure model. The point cloud model obtained in step 2 is subjected to normal estimation and Poisson surface reconstruction to transform the point cloud model into mesh geometric structure model.

Depth estimation

To achieve a more accurate fitting of the blastocysts’ 3D model and to acquire the depth information of the blastocyst surface, it is essential to estimate the depth of both the upper and lower surfaces. This study utilizes a weighted evaluation algorithm based on a dynamic window method to estimate focal depth from the blastocyst’s multi-focal images. Figure 5f shows the focal depth estimation process, where the focal-plane images are divided into an upper layer (F_n-F₀) and a lower layer (F₀-F_-n), and depth calculations are conducted separately for each layer. Since depth coordinates in the image sequence are discrete, relying solely on these positions would result in suboptimal measurement resolution. To enhance measurement accuracy, a curve-fitting method is applied to identify the maximum focal point and calculate depth. During this process, it is essential to determine the depth values of the blastocyst surface points, focusing on obtaining peak focus positions along the z-axis. Taking pixel f(x, y) as an example (Fig. 5g, quoted from Wang et al.)²⁹ its focus values along the z-axis are calculated from the image sequence, Img 0, …, Img n. Following the sequence of image acquisition, the focus values are recorded as data points: (0, y0), (Δd, y1), (2Δd, y2), …, ((n-1)Δd, yn-1), (nΔd, yn). These data points generate the depth maps for the upper and lower layers.

Texture mapping

The upper and lower surface images of the blastocyst are fused to create textured images of the external surface, known as Down Textmap and Up Textmap. The process of multi-focal image fusion is illustrated in Fig. 5i-k. The texture information of the blastocyst’s external surface (Fig. 5j) is then projected onto the geometric structure model (Fig. 5k) using a hemispherical texture mapping method. In this approach, a circular texture plane acts as the projection plane, positioned parallel to the hemisphere’s equatorial plane and tangent at the positive pole, with the projection center located at the negative pole of the sphere (Fig. 5i, quoted from Jiang et al.)³⁰. This projection method establishes a relationship between the texture plane coordinates and the spherical coordinates. The concentric circles on the texture plane correspond to the latitude lines on the hemisphere, and the radial lines on the texture plane correspond to the longitude lines of the hemisphere. The mapping formula is as follows:

$$\mathrm{u}=\frac{2p}{1+\sqrt{1+p^2+q^2}}$$ 1

$$\mathrm{v}=\frac{2q}{1+\sqrt{1+p^2+q^2}}$$ 2

$$\mathrm{p}=\tan \phi \cos \theta$$ 3

$$\mathrm{q}=\tan \phi \sin \theta$$ 4

Delving into derived 3D morphological features

The 3D reconstruction of blastocysts enables evaluation from a 3D perspective, allowing the extracted features to more accurately represent the actual structure. Beyond basic 3D morphological parameters, we derived additional features from the 3D model of the blastocyst, offering a more comprehensive assessment. These features include the shape of the ICM, the number and density of TE cells, and the spatial relative position between the ICM and TE, as detailed below.

Feature 1: the shape of the ICM.

Compared to 2D space, a 3D space enables a more detailed assessment of the ICM shape, allowing for a more accurate representation of its 3D morphology. We developed two metrics to describe the ICM shape: the minor-to-major axis ratio and the sphericity.

A straightforward metric for describing the morphology of a 3D object is the minor-to-major axis ratio, which ranges from 0 to 1. Smaller values indicate a more elongated shape, while larger values suggest a more spherical form. Our study found that this parameter is independent of other parameters, such as ICM surface area, volume, and ICM surface area/volume (Supplementary Fig. 1).

In 2012, Žunić et al. introduced a compactness measure to assess how closely a 3D shape resembles a perfect sphere, with a compactness value of 1 indicating a perfect sphere and a minimum value of 0. Simply put, the compactness measurement by Žunić extends the concept of 2D circularity to 3D shapes^31,32. Cao et al. later defined a quantitative parameter that describes shape changes and is used to assess irregularities in cell shape³³, which is more applicable to ICM than Žunić's method. Based on Cao’s definition, we defined ICM shape factor as Eq. (5), a dimensionless parameter in 3D that is related to the intrinsic properties of the object, exhibiting rotational invariance. Notably, the shape factor of a sphere is minimal at 2.1991. However, as the number of small surface spikes increases on a 3D object, the shape factor rises, indicating a more complex shape. Figure 5l (quoted from Cao et al.)³³ illustrates various geometric shapes to clarify the concept of the shape factor.

$$\frac{\sqrt{I\mathit{CM\; surface\; area}}}{\sqrt[3]{I\mathit{CM\; volume}}}$$ 5

In our research, ICM shape factor values ranged from 2.21 to 2.65. Approximately 25% of the ICMs had a shape factor below 2.27, resembling an icosahedron or sphere. Conversely, about 25% exhibited a shape factor above 2.34, suggesting an octahedral or more edge-defined form. The majority of ICMs appeared similar to a dodecahedron. Importantly, our findings indicate that ICM shape is independent of parameters like ICM volume, surface area, ICM surface area/volume, and minor-to-major axis ratio (Supplementary Figure 1).

Previous works have used sphericity as a metric for ICM compactness in human blastocysts³⁴. Sphericity describes the degree of spheroidal-like shape of a polyhedral geometry, while shape factor further delineates the binding structure of a polyhedron. Although the two metrics are similar, the shape factor enables better differentiation of ICMs based on external surface morphology, allowing a more in-depth exploration of external surface morphological quality beyond mere sphericity.

Feature 2: TE cell number and TE density.

The implementation of TE cell counting in this paper is based on the Region Growing Algorithm (RGA) for point cloud. Originally introduced by Zucker in 1976 for 2D image processing³⁵, the region-growing method has since been adapted for 3D point cloud segmentation, enabling accurate boundary delineation and improved segmentation performance. The algorithm is a clustering algorithm based on domain relationships, which can divide the point cloud dataset into different clusters. The process begins with selecting the point of lowest curvature as the seed point. The algorithm then calculates the similarity between each domain point and the seed point to assess whether they belong to the same TE cell, ultimately achieving TE cell counting. However, this method is sensitive to the density distribution of the point cloud (i.e., if the distribution of the point cloud within a TE cell is discrete, it is not easy to form clusters), which can lead to the phenomenon of repeated counting. To avoid this phenomenon, we further filter out the discrete distribution of point cloud data by calculating the curvature of the point cloud to mark the concave region, so that the distribution of the point cloud located in the concave region is relatively aggregated, which makes it easier to form clusters. Using this approach, the TE cell count in this study ranged from 78 to 127.

Feature 3: Spatial distance between ICM and TE.

The spatial distance between the ICM and TE was defined as a dimensionless parameter, as described in Eq. (6), where R denotes the blastocyst radius and d is the distance from the centroid of the ICM to that of the blastocyst (Fig. 2b). This parameter ranges from 0 to 1, with larger values indicating a greater distance between the ICM and TE and smaller values indicating closer proximity. Notably, the spatial distance between ICM and TE is independent of factors, such as blastocyst volume, cavity volume, ICM volume, ICM morphology, and TE cell number (Supplementary Fig. 1).

$$\frac{R-d}{R}$$ 6

Verification of 3D parameters measurement results

To obtain accurate values for blastocyst 3D morphological parameters, we selected 26 human embryos at the blastocyst stage (2 graded as 3AB, 3 as 3BB, 10 as 4BB, and 11 as 4BC) and imaged them using a Hoffman imaging system, which enables non-invasive 3D reconstruction. The embryos were then fixed in a 4% paraformaldehyde PBS solution. We used DAPI (Sigma Aldrich) to stain the cell nuclei, SPY555-Actin Probe (Cytoskeleton) to label the cell membranes, and Rabbit anti-CDX2 (Invitrogen) along with FITC-anti-Rabbit (Jackson ImmunoResearch) to label the TE. Rabbit anti-CDX2 (Invitrogen) and Alexa Fluor 647-anti-Mouse (Invitrogen) were used to label the ICM. Subsequently, we captured 81 focal data sets using a confocal microscope. Imaris software was used for the fluorescent reconstruction of the embryos, enabling a comparison between the fluorescent reconstruction and the features of the non-invasive 3D reconstruction obtained through the TL system.

Statistical method

SPSS was mainly used for statistical analysis. Numerical data are expressed as the mean ± standard deviation (SD), and categorical data are expressed as the rate (%). Continuous parameters were compared via Wilcoxon rank sum test or one-way analysis of variance. During discretization, data points outside the mean ± 3 standard deviations are treated as extreme values. Following the removal of these extremes, the remaining data are divided into five groups using the equal interval method, resulting in four cut-off points. These four cut-off values classify all data for this variable from continuous to proportional variables. Proportional data are then analyzed with the Cochran–Armitage trend test, while the GEE is applied to account for the inherent associations among observations within the same cluster. Since embryos from the same patient are not independent, the GEE’s first level represents individual embryos, and the second level corresponds to the patients to whom they belong. P < 0.05 was considered statistically significant.

Author contributions

B.H., Z.S., and L.J. contributed to the design of the project. B.H. and K.S. were responsible for manuscript writing and data analysis. B.M., L.W., J.Y., and X.R. were responsible for assembling the TL data and data processing. J.W., W.Z., X.W., and Y.G. contributed to image processing and model construction. B.H., J.Y., L.J., and Z.S. were responsible for coordinating the study. All authors reviewed the manuscript.

Data availability

The datasets generated and analyzed during the current study are not publicly available due to patient privacy concerns. However, anonymized data may be made available to qualified researchers for academic purposes upon reasonable request. Requests should be directed to the corresponding authors at hb@tjh.tjmu.edu.cn or shizhenzhi@nbu.edu.cn.

Code availability

To preserve potential commercialization opportunities, the source code developed in this study is not publicly available. Researchers interested in academic and non-commercial use may request access by contacting the corresponding authors at hb@tjh.tjmu.edu.cn or shizhenzhi@nbu.edu.cn. Each request will be reviewed on a case-by-case basis, and access may be granted under the terms of a research use agreement.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-025-02028-9.