Prospective Assessment of Embryoid Body by Deep Learning on Label-Free Time-Lapse Images from the Microwell Array

Abstract

Background: Embryoid bodies (EBs) play a central role in organoid engineering, where their formation fidelity and size critically influence downstream differentiation outcomes. Current EB production workflows primarily rely on retrospective quality assessment, which limits reproducibility in high-throughput culture systems. Objective: This study aimed to develop a prospective, non-invasive framework that integrates early-phase bright-field time-lapse imaging with a three-dimensional convolutional neural network to predict EB formation outcomes and final EB diameter within the microwell platform. Methods: Time-lapse image sequences collected during the first hours after cell seeding on the microwell array were used to train 3D-CNN models for classification (formation vs. non-formation) and regression (final diameter). A balanced dataset was constructed through under-sampling, and five-fold cross-validation with data augmentation was applied to evaluate model performance. Results: The classification model achieved an accuracy of 96.5%, reliably distinguishing between successful and failed EB formation using short-duration image sequences. The regression model predicted the final EB diameter with a mean absolute error of ±7.1 µm, reflecting strong agreement with measured values and capturing seeding-density-dependent size variations. Conclusions: Early aggregation dynamics captured by bright-field time-lapse imaging contain sufficient spatiotemporal information to enable accurate, prospective EB quality prediction. The proposed framework provides a label-free and automation-compatible strategy for improving reproducibility in large-scale EB manufacturing and supports the future development of adaptive and closed-loop organoid culture systems for clinical applications.

1. Introduction

Three-dimensional (3D) cellular aggregates—including Embryoid bodies (EBs), spheroids, and organoids—have become indispensable platforms in stem-cell biology and regenerative medicine, as they enable the in vitro reconstruction of human developmental processes and disease phenotypes [1]. These multicellular structures recapitulate key architectural and microenvironmental features of native tissues, supporting cell–cell interactions and physicochemical gradients that are largely absent in conventional two-dimensional cultures. As a result, they provide a more physiologically relevant context for modeling tissue development and pathophysiology. Advances in induced pluripotent stem cell (iPSC) technology have further expanded the repertoire of organoid systems capable of reproducing organ-specific architectures and functions, including hepatic buds, cerebral tissues, and prostate epithelia [2,3,4]. In regenerative medicine, iPSC-derived cardiac spheroids have demonstrated therapeutic efficacy in rodent and porcine models of heart failure. More recently, cardiac spheroids have shown the ability to regenerate infarcted myocardium in nonhuman primates while maintaining sustained functional improvement with an acceptable arrhythmia risk [5,6]. The formation of EBs serves as a critical early intermediate in these systems, during which cell–cell interactions and microenvironmental cues guide lineage specification. Prior studies have shown that the initial EB size is a decisive determinant of differentiation trajectory and efficiency, with larger aggregates promoting hepatic or cardiomyogenic fates and smaller ones favoring vascular lineages [7,8].

Methods for producing uniform size-controlled EBs have progressed substantially over the past decade. Classical hanging-drop and static suspension cultures, although conceptually simple, exhibit poor reproducibility and limited scalability for translational or clinical manufacturing [9]. The introduction of microwell arrays and microfabricated containers—including concave and U-bottom plates—enabled more precise control of initial cell seeding by leveraging surface tension and gravitational settling [10,11]. Although these systems promote rapid aggregate formation, sustaining long-term culture remains difficult because medium exchange is restricted within confined microenvironments. Hydrodynamic bioreactors and rotary culture platforms addressed this limitation by providing continuous control over oxygen, nutrient, and morphogen gradients, thereby supporting large-scale EB production from embryonic stem cells and iPSCs for downstream differentiation and screening applications [9]. More recently, acoustic standing-wave platforms have enabled the generation of over 28,000 uniform EBs per batch, representing a significant advance toward clinically relevant manufacturing scales [12]. Despite these developments, most suspension-based approaches still lack the capacity to track individual EBs throughout culture, posing a persistent challenge for quality assurance and process standardization.

Among the platforms developed to enhance the reproducibility of EB formation, the Tapered Soft Stencil for Cluster Culture (TASCL) provides a uniquely effective approach for generating large numbers of size-controlled aggregates [13]. The tapered architecture confines cells within discrete microwells without flat interwell surfaces, thereby preventing unintended aggregation and enabling rapid, uniform EB formation driven primarily by gravitational settling [14,15]. The transparent porous membrane at the base of each microwell further facilitates efficient medium exchange within the microscale environment and permits continuous imaging of individual EBs. These design features support long-term culture stability, as demonstrated by Saito et al., who showed that TASCL enables controlled initiation of three-dimensional clusters, sustained propagation of intestinal organoids, and derivation of mature epithelial lineages [16]. Furthermore, TASCL ensures complete traceability of EBs.

Despite recent progress in three-dimensional culture technologies, achieving a 100% yield of target EBs remains challenging due to intrinsic variability in procedural handling and cell-intrinsic properties. This variability poses significant barriers to the practical use of EB-derived aggregates in regenerative medicine and in applications intended to replace animal experimentation. For instance, generating cardiac spheroids for myocardial infarction therapy requires a multi-step differentiation protocol that involves repeated reagent administration over one to two weeks following EB formation. If the resulting spheroids fail to meet the required quality or quantity thresholds, the transplantation procedure may need to be delayed, underscoring the need for early and reliable indicators of EB quality.

To address this unmet need, the present study establishes an early-stage, image-based quality prediction framework for EBs generated using the TASCL platform. Recent advances in machine learning have transformed aggregate culture analytics with traditional endpoint assays increasingly replaced by deep learning models that extract prognostic morphological features from live imaging data. Convolutional neural networks (CNNs) trained on bright-field images of EBs can classify early mesodermal commitment with high accuracy, thereby obviating destructive immunostaining [17]. In organoid research, tools such as D-CryptO automate the assessment of colonic organoid maturation by quantifying crypt budding patterns, reducing analysis time from hours to seconds [18], while OrganoIDNet enables segmentation, growth tracking, and treatment-response prediction in pancreatic ductal adenocarcinoma organoids, supporting high-throughput drug screening [19]. Beyond classification, reinforcement learning-guided robotic systems such as SpheroidPicker autonomously select and transfer spheroids according to size and circularity, integrating computer vision with physical manipulation [20]. Deep learning models can also infer spheroid viability from single images, addressing the challenge of necrotic core formation in large aggregates and informing quality-control workflows in tissue engineering [21]. Building on evidence that TASCL supports highly uniform, scalable, and traceable EB formation, this study investigates whether time-lapse bright-field images acquired during the first few hours after cell seeding can predict two key quality metrics: (i) whether an EB will successfully form and (ii) the final EB diameter at 24 h (Figure 1). To this end, we employed a three-dimensional convolutional neural network (3D-CNN) to leverage the temporal and morphological information encoded in short-duration image sequences, enabling automated feature extraction and prediction without manual measurement [22,23,24]. By integrating TASCL-based high-throughput EB culture with deep-learning-driven predictive analytics, this work provides a prospective and quantitative framework for EB quality control and lays the groundwork for standardized large-scale production workflows.

Figure 1. Overview of EB formation ability prediction model or cell aggregate diameter prediction model using TASCL time-lapse images.

2. Materials and Methods

2.1. TASCL Device and EB Culture

TASCL (TASCL #1000, Cymss-bio, Nagoya, Japan) was used as the platform for high-throughput and size-controlled EB formation. The device consists of 1020 tapered microwells arranged on a culture insert (pore diameter 3 μm) for 6-well plate format, each comprising a 0.5 mm square top aperture and a 0.3 mm square bottom aperture connected by smoothly narrowing sidewalls.

Human induced pluripotent stem cells (#RPChiPS771-2, Reprocell, Kanagawa, Japan) were cultured on a 35 mm dish and dissociated into single cells with 300 μL 0.5× TrypLE Select (A12859-01, Life Technologies, Carlsbad, CA, USA) for 5 min. After adding 1 mL culture medium (Stem Fit, Ajinomoto, Japan) containing 10 μM Y27632 (#253-00511, Wako, Osaka, Japan), cells were collected into a 15 mL tube. Following 5 min of centrifugation, the supernatant was aspirated, the medium with Y27632 was added again to achieve the target concentrations. To introduce variation in the size of the resulting EBs, a total of 1.02 × 10⁶ or 3.06 × 10⁵ cells were seeded per TASCL plate, corresponding to approximately 1000 or 300 cells per microwell. The volume of culture medium was 1.5 mL per well. After cell seeding, cells were allowed to sediment into the microwells and cultured without agitation. After 24 h, the culture medium was replaced with a culture medium that did not contain Y27632, and then the medium was partially replaced in half every 24 h. To ensure precise measurement of viable areas, EBs were stained after 72 h using the Viability/Cytotoxicity Assay Kit (No. 30002, Biotium, Fremont, CA, USA), which contains Calcein-AM for identifying live cells (green fluorescence) and ethidium homodimer-1 for detecting dead cells (red fluorescence), in accordance with the manufacturer’s instructions.

2.2. Time-Lapse Imaging Setup

Time-lapse imaging was performed to capture the aggregation and growth dynamics of EBs formed within each TASCL microwell. The TASCL device in a 6-well plate was mounted on an inverted digital microscope (BZ-X 700, Keyence, Osaka Japan) equipped with an improvised on-stage environmental chamber (Blast, Tokyo, Japan) maintained at 37 °C, 5% CO₂, and saturated humidity. Bright-field images were acquired using a ×4 objective lens with a 2.83-megapixel monochrome CCD camera, providing a field of view that encompassed about 30 microwells simultaneously. To capture images of all 1020 microwells, the microscope stage was moved incrementally in the X and Y directions. Images taken at 54 positions were then merged into a single gray-scale image using the microscope’s built-in software (Ver.1.4.0), stored in uncompressed TIFF format to preserve spatial fidelity for downstream deep-learning processing. Images were collected every 30 min for a total duration of 72 h following cell seeding. Illumination settings were kept constant throughout acquisition to avoid brightness variability across frames and autofocus was disabled after initial adjustment to prevent focal shifts during long-term imaging.

2.3. Image Dataset Construction

Time-lapse image sequences obtained from TASCL cultures were processed to construct datasets for EB formation classification and final diameter prediction. Initially, the image corresponding to each microwell is extracted from the complete set of images at every time point, thereby generating a time-series image dataset for each microwell.

The dataset was composed of images obtained from seven TASCLs. Each individual microwell at the bottom aperture yields an image with dimensions of 130 by 130 pixels. For this image, the brightness values were normalized so that the mean was 64 and the standard deviation was 16.

For the classification dataset, each dataset was labeled as “successful” or “failed” based on the presence of a single, clearly formed spherical EB at 24 h. The success label was defined as a well in which a single EB consisting only of positive curvature in shape was formed in one well. Failure labels included wells with dispersed cell clusters, multiple EBs, or irregularly shaped EB. Both labels were defined by one annotator. For the regression dataset, the final EB diameter at 72 h was measured from fluorescence images of Calcein-AM using a calibrated pixel-to-micrometer conversion (0.49 pixel = 1 μm). The diameter was defined as the average of the major and minor axes of the EB boundary, measured using the built-in image analysis software of the microscope. It has been confirmed in the preliminary test that the success or failure of EB formation did not change over approximately 24 h, and that the timing at which diameter change ceased due to intercellular interactions between EBs was approximately 72 h.

2.4. Classification Model of EB Formation

3D-CNN is a method for recognizing video by considering spatiotemporal information through the convolution of three-dimensional image information: two-dimensional spatial information and one-dimensional temporal information, applied to the input time series images [22,23,24]. In this model, we inputted 2N images collected every 30 min up to N hours after cell seeding into a 3D-CNN to predict whether an EB would form within the microwell at 24 h post-seeding. The neural network was trained using information about whether EBs formed in each well as the ground truth. This task was essentially a binary classification problem. The structure of the 3D-CNN used for prediction is shown in Table 1A. Stochastic gradient descent was used as the optimization algorithm. The batch size was 32, the learning epoch was 150, and the L2 regularization parameter λ = 0.01. The activation functions for the convolutional layer and fully connected layer 1 were ReLU functions, and the sigmoid function was used as the activation function for fully connected layer 2. For comparison, we also performed predictions using only a single image taken at N hours post-seeding.

Table 1. Neural network structure in 3D-CNN. Example using images for up to 3 h post-seeding.

(A) Classification model of EB formation
	Input	Output	Filter Size	L2 Regularization
Convolution 1	100 × 100 × 6 × 1	96 × 96 × 5 × 32	5 × 5 × 2	-
Convolution 2	96 × 96 × 5 × 32	96 × 96 × 5 × 64	5 × 5 × 2	-
Max pooling 1	96 × 96 × 5 × 64	32 × 32 × 3 × 64	3 × 3 × 2	-
Convolution 3	32 × 32 × 3 × 64	32 × 32 × 3 × 64	5 × 5 × 2	Yes
Convolution 4	32 × 32 × 3 × 64	32 × 32 × 3 × 64	5 × 5 × 2
Max pooling 2	32 × 32 × 3 × 64	11 × 11 × 2 × 64	3 × 3 × 2	-
Fully connected layer 1	11 × 11 × 2 × 64	512	-	Yes
Fully connected layer 2	512	2	-	-
(B) Regression model of EB Diameter
	Input	Output	Filter Size	L2 Regularization
Convolution 1	100 × 100 × 6 × 1	96 × 96 × 5 × 32	5 × 5 × 2	Yes
Convolution 2	96 × 96 × 5 × 32	96 × 96 × 5 × 64	5 × 5 × 2	Yes
Max pooling 1	96 × 96 × 5 × 64	32 × 32 × 3 × 64	3 × 3 × 2	-
Convolution 3	32 × 32 × 3 × 64	32 × 32 × 3 × 64	5 × 5 × 2	Yes
Convolution 4	32 × 32 × 3 × 64	32 × 32 × 3 × 64	5 × 5 × 2	Yes
Max pooling 2	32 × 32 × 3 × 64	11 × 11 × 2 × 64	3 × 3 × 2	-
Fully connected layer 1	11 × 11 × 2 × 64	512	-	Yes
Fully connected layer 2	512	64	-	Yes
Fully connected layer 3	64	1	-	-

For model validation, a five-fold cross-validation scheme was applied. The dataset was partitioned into five subsets, each iteration using four subsets for training and one for testing, and the final performance was computed as the average across all folds. In this dataset, images of 585 wells that formed EBs and 585 wells that did not form EBs at 24 h post-seeding were used. In each cross-validation, 468 wells of successfully formed EBs and 468 wells of failed ones were used for training. To increase the effective size of the training dataset, data augmentation was applied exclusively to the training folds through horizontal flipping and rotations of 0°, 90°, 180°, and 270°. This augmentation increased the total number of training samples to 7488 wells, while 234 wells were retained as test samples across the five validation cycles. Afterall, each image was adjusted to be 100 pixels square. To evaluate the prediction accuracy of this model, we used precision and recall, metrics commonly employed for machine learning model evaluation. These are expressed by the following equations.

$$\mathrm{Precision}={\displaystyle \frac{\mathit{True} \mathit{Positive}}{\mathit{True} \mathit{Positive}+\mathit{False} \mathit{Positive}}}$$

(1)

$$\mathrm{Recall}={\displaystyle \frac{\mathit{True} \mathit{Positive}}{\mathit{True} \mathit{Positive}+\mathit{False} \mathit{Negative}}}$$

(2)

To verify which regions in the image contribute to the prediction, Gradient-weighted Class Activation Mapping (Grad-CAM) [25] was used.

2.5. Regression Model of EB Diameter

In this model, we input 2N images collected every 30 min up to N hours after cell seeding into a 3D-CNN to predict the diameter of EB at 72 h. The ground truth was the diameter calculated as described above, making this a regression prediction. The 3D-CNN architecture was constructed based on the network structure in Table 1B. The Adam optimization algorithm was used. The batch size was 32, the number of epochs was 120, and the L2 regularization parameter λ = 0.01. All activation functions were ReLU functions, and the loss function was the mean squared error function. Specifically, the network consists of two sets of two convolutional layers and one max-pooling layer each, followed by three fully connected layers. A dropout layer with a drop rate of 0.5 was introduced after each max-pooling layer to avoid overfitting. The output dimensions of each fully connected layer were set to 512, 64, and 1, respectively. Similarly to the classification model, a five-fold cross-validation scheme was applied. The obtained 2225 images of microwells with single EB at 72 h were increased by inversion and rotation, yielding 14,240 images for training and 445 images for testing. Afterall, each image was adjusted to be 100 pixels square. To verify which regions in the image contribute to the prediction, Grad-CAM was used.

3. Results

3.1. EB Formation Efficiency on the TASCL Device

EB formation behavior on the TASCL device exhibited high baseline efficiency. Under standard culture conditions, approximately 90% of microwells produced a single spherical EB after 24 h, whereas only about 10% failed to form an aggregate (Supplementary Figure S1). EBs seeded at 1000 cells/well had an average diameter of 169.8 μm with a standard deviation of 17.4, while those seeded at 300 cells/well averaged 138.4 μm with a standard deviation of 17.9.

3.2. Performance of the 3D-CNN for EB Formation Classification

Table 2 presents the precision and recall metrics for each trial conducted using the 5-fold cross-validation approach. Overlearning does not occur and learning converges appropriately (Supplementary Figure S2). These results reflect the model’s predictive performance on time-series images obtained within three hours following cell seeding when evaluated against the test dataset. The data indicate that both precision and recall consistently range from approximately 0.95 to 0.97. Furthermore, the relationship between elapsed time post-cell seeding and prediction precision—comparing both 3D-CNN and single-image methodologies—is analyzed (Figure 2). The precision values reported are averages calculated across all folds of cross-validation. Model performance depended on the temporal length of the input sequence. When only a small number of frames were provided, prediction accuracy decreased. In contrast, prediction precision improves over time after cell seeding, reaching 0.97 at the three-hour mark.

Table 2. List of precision and recall values for each trial in the 5-fold cross-validation method. These are the prediction results from the model trained by inputting time-series images up to 3 h after cell seeding into the 3D-CNN, applied to the test data.

Division 1			Precision	Recall
Successful EB Formation			0.974	0.949
Failed EB Formation			0.950	0.974
Average			0.962	0.962
Division 2			Precision	Recall
Successful EB Formation			0.949	0.949
Failed EB Formation			0.949	0.949
Average			0.949	0.949
Division 3			Precision	Recall
Successful EB Formation			0.966	0.974
Failed EB Formation			0.974	0.966
Average			0.970	0.970
Division 4			Precision	Recall
Successful EB Formation			0.991	0.949
Failed EB Formation			0.951	0.991
Average			0.971	0.970
Division 5			Precision	Recall
Successful EB Formation			0.966	0.983
Failed EB Formation			0.983	0.967
Average			0.974	0.974
Summary of the 5-fold cross-validation
	Precision		Recall
	Average	S.D.	Average	S.D.
Successful EB Formation	0.969	0.0136	0.961	0.0147
Failed EB Formation	0.961	0.0143	0.969	0.0136

Figure 2. The relationship between post-seeding time and prediction precision. The orange line shows the prediction precision using 2N images as input up to post-seeding time N (h), while the blue line shows the prediction precision using only the single image at time N (h) as input.

Figure 3A displays representative image sequences that were accurately classified by the 3D-CNN model during testing. The black, shadow-like areas observed in some wells correspond to air bubbles introduced during pipetting at cell seeding. Heatmap of Figure 3A shows a visualization of the regions in the time-series images that contribute to the predictions using Grad-CAM. The redder the areas, the higher their contribution to the prediction.

Figure 3. TASCL microscopy images input the 3D-CNN model (A) Representative image sequences that were accurately classified by the 3D-CNN classification model. (a,b) are image sequences of cases where EB formation succeeded, while (c,d) are image sequences of cases where EB formation failed. The heatmap shows the Grad-CAM results, where the red-colored regions were the image areas that contributed significantly to the classification. (B) Representative image sequences that were used by the 3D-CNN regression model. The heatmap of Grad-CAM shows that the bottom of the well is extracted as a feature, as was the case of classification model.

3.3. Regression Performance for Predicting Final EB Diameter

Training and prediction were performed using a dataset combining image sequences from wells seeded with cells at 1000 cells/well and 300 cells/well. The average diameter of EB was 158.3 μm, with a standard deviation of 23.2 at 72 h post-seeding in this dataset (Figure 4). Based on image sequences obtained up to several hours after cell seeding, the diameter of EBs at 72 h post-seeding is predicted by the regression 3D-CNN model (Figure 5). It was found that the prediction error decreases as time passes since seeding, reaches near a plateau after approximately 3 h post-seeding, and enables prediction with an error of ±7.1 μm after 6 h post-seeding. Heatmap of Figure 3B shows a visualization of the regions in the time-series images that contribute to the predictions using Grad-CAM. The redder the areas, the higher their contribution to the prediction.

Figure 4. Histogram of EB diameter at 72 h post-seeding used in this regression task Average diameter: 158.3 μm, Standard deviation: 23.2.

Figure 5. The relationship between post-seeding time and absolute error in predicted EB diameter. The minimum absolute error reached 7.08 μm at 6.0 h post-seeding.

4. Discussion

In the classification task predicting successful EB formation, examining values at 2 h post-seeding revealed that precision reached approximately 0.96 and plateaued with 3D-CNN model. In contrast, precision was around 0.94 when using a single image input, indicating that utilizing time-lapse images of the wells yielded higher accuracy. This suggests that early morphological evolution—including sedimentation rate, boundary continuity, and compaction kinetics—differed substantially between successful and failed wells. Such spatiotemporal differences supported the feasibility of using short-duration time-lapse data as predictive input for the subsequent 3D-CNN-based analyses. Representative well images taken several hours after cell seeding exhibit various patterns. Cases such as Figure 3A(a), where the EB forms normally, and Figure 3A(c), where the cell cluster exhibits poor cohesion, can be distinguished to some extent even by human experts. However, in cases like Figure 3A(b,d), where bubbles are present in the well and partially obscure the field of view, even the human experts find it difficult to determine the sucess of EB formation at 2 h after cell seeding. The developed 3D-CNN model, which can make accurate predictions even in such cases, can be considered highly useful. However, examining the values at 3 h after cell seeding reveals that the precision is approximately 0.96 for both the 3D-CNN model and the model trained using only a single image taken at 3 h post-seeding, indicating no significant difference. The reason for this lack of difference was considered to be that EB outlines had already formed within the wells. It is inferred that cell groups with a formed contour at this point are less likely to disperse over subsequent time periods. According to the visualization results using Grad-CAM, it was found that the edge of the wells contribute to the prediction. When compared with bright-field images, it was inferred that those regions were areas where cells are absent.

In the regression task predicting EB diameter, the margin of error for was ±7.3 µm by using 6 time-lapse images taken within the first 3 h after cell seeding. This indicates a strong relationship between early morphological changes and eventual size. Given the dataset’s average EB diameter of 158.3 μm, this level of error is sufficiently small for practical use. The prediction error remained nearly constant between 3 and 5 h post-seeding, but decreased even more from 5 to 6 h (Figure 5). Although this pattern is not straightforward to interpret, it may imply that subtle morphological shifts from 2D to 3D within cell distribution pattern—occurring after proper aggregation—contribute to greater predictive accuracy. Prior studies have shown that EB formation depends on rapid sedimentation, efficient cell–cell adhesion, and early boundary stabilization, processes that unfold within the first few hours after cell seeding and strongly influence downstream differentiation potential [7,8]. The present findings extend this understanding by demonstrating that these early subtle variations in intra-cellular dynamics are not only biologically important but also computationally recognizable through spatiotemporal patterns captured by 3D-CNN. The Grad-CAM results suggested that, similar to EB formation model, the appearance of the well bottom surface was important for diameter prediction as well.

The predictive nature of this approach also aligns with emerging trends in microfluidic [26] and feedback-controlled cell culture systems [27,28,29], in which early measurements guide real-time adjustments to improve yield and uniformity. Integrating predictive analytics with geometrically standardized microwells supporting long-term culture environments, a pathway toward fully automated, high-throughput organoid production pipelines can be realized. Such integration could reduce resource consumption, minimize operator-dependent variability, and accelerate the development of reproducible organoid-based disease models and therapeutic platforms.

5. Limitation and Future Work

While this research lays groundwork for predictive EB quality assessment, some limitations remain. One issue is the need to enhance prediction accuracy for wells that contain bubbles. Wells with misclassified categories contained a large amount of data with bubbles. Because wells with bubbles rarely occur when using TASCL for EB formation, future studies should focus on generating many images where bubbles are deliberately incorporated into sections of the well.

Second, this approach was characterized by combining a microwell device capable of culturing EBs at the same location over extended periods with the 3D-CNN model. It does not aim to generalize the model itself to other culture devices. When normalizing image data, the mean and variance values were hardcoded. There is a possibility that this can be optimized. Furthermore, the model was constructed using a single iPS cell line designated for our clinical purposes, and generalization to other cell lines was not validated. When making predictions with other cell lines, it is necessary to prepare a new dataset using those cell lines.

Third, the present analysis relied solely on bright-field imaging, and integrating additional modalities such as fluorescence markers or mechanical readouts may further enhance predictive accuracy.

Fourth, only samples manipulated by a specific technician were used as test data. This point may have an impact on generalization performance.

6. Conclusions

This study established a predictive framework that combines early-phase bright-field time-lapse imaging with 3D-CNN, facilitating prospective evaluation of EB quality within the microwell platform. The classifier demonstrated an accuracy of 96.5% in distinguishing success from unsuccessful EB formation, indicating that critical morphological features become apparent within the initial hours post-seeding. In parallel, the regression model predicted the final EB diameter with a mean absolute error of ±7.1 µm, capturing both intra-condition variability and seeding-density-dependent size differences. These results collectively show that early aggregation dynamics provide reliable and quantifiable indicators of EB quality during high-throughput production. This approach contributes to transitioning applications such as standardized EB manufacturing, organoid engineering, and automated culture systems from post-evaluation to proactive quality control.

While challenges remain to further improve accuracy and versatility, this work demonstrates the feasibility and value of data-driven prediction for improving the reproducibility and scalability of EB production workflows.