AI-Enabled Smart Monitoring of Bovine Embryo Development Using Time-Lapse Imaging and Transfer Learning

Abstract

Time-lapse incubation systems enable the continuous, non-invasive monitoring of embryonic development; however, identifying developmental stages still requires substantial manual assessment, making the process time-consuming, labor-intensive, and potentially subjective. This study evaluated the developmental kinetics of bovine embryos cultured in synthetic oviductal fluid (SOF) and Gx-TL™ media using MIRI time-lapse imaging, and investigated the effectiveness of transfer learning-based convolutional neural networks (CNNs) for automated embryo-stage classification. A total of 311 zygotes were individually cultured under standard in vitro fertilization conditions, including 152 embryos in SOF medium and 159 embryos in Gx-TL™ medium. In the SOF group, 81 embryos reached the two-cell stage, 10 developed to the morula stage, and 6 reached the blastocyst stage. In the Gx-TL™ group, 77 embryos reached the two-cell stage, 18 developed to the morula stage, and 10 reached the blastocyst stage. Time-lapse images were manually annotated according to key developmental stages, including the two-cell through eight-cell stages, morula, and blastocyst. The annotated image dataset was augmented to 5000 images and used to train three pretrained CNN architectures: ResNet18, DenseNet121, and EfficientNet-B0. All three models achieved 100% accuracy in the two-class classification task across both culture media. Classification accuracy ranged from 95% to 100% for the nine-class model and from 98% to 100% for the ten-class model. These findings provide preliminary evidence that transfer learning-based convolutional neural networks (CNNs) can support the automated classification of bovine embryonic developmental stages using time-lapse images.

1. Introduction

Commercial embryo transfer (ET) programs in cattle began in the early 1970s, and are the foundation of the techniques and advancements available to us today [1]. The early technology that triggered commercial adaptation is attributed to Tim Rowson and colleagues at the Agricultural Research council unit in Cambridge [1]. This group initiated surgical recovery and bovine ET, leading to the establishment of the first commercial ET centers in North America [1]. Another factor that accelerated the expansion of ET units in the early 1970s is the legal importation of European, “exotic” cattle to Canada [1]. By the late 1980s, the use of ET shifted toward other cattle breeds after the initial surge in interest in “exotic” breeds diminished [1]. Since then, methods for superovulation, oocyte pick up (OPU), fertilization, ET, and cryopreservation have developed significantly, consequently increasing in prevalence [1]. Yet limitations prevail in the pursuit of maximized efficiency and sustainability in the livestock industry.

The expanding population necessitates refinement and developments of current technologies to meet the growing demand for meat and dairy while maintaining biological and economic efficiency. A strong statement has been made by Pete Hansen, who declared that the true benefits of embryo technology in optimizing efficiency and sustainability cannot be reaped until ET replaces artificial insemination (AI) entirely [2]. Hansen supports this by emphasizing it’s utility in streamlining genetic selection and the facilitation of producing beef bulls from dairy cattle [2]. Moreover, Hansen also cites improved fertility with the use of ET in cases of heat stress and repeat-breeding cows, and the potential for improvement more broadly [2]. Yet, limitations are present when it comes to ET costs, which exceed the benefits of genetic or fertility improvements. This is exemplified by Kaniyamattam et al. [3] who used a stochastic, dynamic dairy model using multi-trait genetics to determine the optimal ratio of IVP-ET and artificial insemination (AI) in a system to maximize profitability. This study observed that in a 15-year model, the profit was maximized at only 40% ET and 60% artificial insemination (AI), necessitating more economic ET systems.

A significant factor reducing the success of ET is the accurate identification of the embryo’s ability to establish pregnancy. Thus, the methods of ET selection for embryos with the highest implantation and developmental potential must be robust in order to generate accurate predictions [4]. Traditionally, viability assessment is based largely on morphological appearance, with some consideration given to developmental rate [5]. This form of evaluation has been largely deemed as inadequate due to inherent subjectivity and the overlooking of critical developmental events [5]. Visual morphological embryo evaluation relies on subjective visual inspections, leading to inconsistencies due to variations in experience and interpretation [6,7]. Farin et al. [8] quantified this subjectivity by assessing the agreement between six experienced individuals while evaluating 40 bovine embryo timelapse videos. In the study, agreement among evaluators for the grading of in vitro embryos was 57.7%. The authors noted that agreement improved by extremes of quality grade (excellent or degenerated) and was reduced in conditions of abnormal morphology. Moreover, solely using morphological features provides a static and narrow understanding of embryo viability, as the embryo is only analyzed at one period in time [4]. Additionally, the observation of morphokinetic parameters with the use of conventional incubators is invasive, and hinders embryo development. This results from fluctuations in media pH, temperature, and general agitation during handling.

The inclusion of morphokinetic parameters would allow for the development of more robust predictions regarding embryo viability. Few research studies have used non-destructive techniques via spectroscopy and imaging in grading of bovine embryos, as summarized by Shivaani and Madan [9]. While spectroscopic techniques like Raman spectroscopy do not directly quantify the morphological features, they can be used as a pre-selection tool to evaluate the quality of bovine oocyte. The changes in proteins, carbohydrates, and lipids of developing oocytes were evident at 1004 cm⁻¹ and 1529 cm⁻¹, which helped in determining the indirect association with the structure of zona pellucida in bovine oocytes [10,11]. Similarly, Fourier transform infrared spectroscopy was used to evaluate the quality of bovine embryos by measuring the change in nucleic acids and amides. For amides the wavelengths between 1800 cm⁻¹ and 1400 cm⁻¹, and for nucleic acids (DNA) the wavelengths between 1240 cm⁻¹ and 950 cm⁻¹, helped to distinguish three types of blastocysts using principal component analysis and hierarchical cluster analysis [12]. In contrast, imaging techniques, specifically optical microscopic imaging when coupled with machine learning and deep learning algorithms, yielded classifications or gradings of bovine embryos with a success rate varying between 20 and 89% [13,14,15,16]. The wide spectrum in classification accuracies is mostly due to the differences in image pre-processing methods, the machine learning or deep learning model used in classification, and the size of the dataset. In another study, hyperspectral imaging was used to differentiate the on-time development of bovine embryos and fast-developing bovine embryos based on the metabolites produced during development. Regardless, there was no noticeable difference in the DNA damage seen in the two kinds of embryos [17]. Hyperspectral imaging in bovine embryo grading still remains in its infancy.

Time-lapse imaging is one of the available non-destructive imaging techniques used to obtain morphokinetic parameters by periodically taking images of the developing embryo. The time taken to reach developmental milestones can provide insights into embryo viability [4]. In addition, the safety and outcomes of time-lapse incubators have been researched in the context of human reproductive technologies, which encourages extension to bovine practice. A multicenter double-blind randomized controlled clinical trial conducted by Bhide et al. [18] explored the clinical outcomes of a disturbed standard culture and an undisturbed time-lapse culture (525 participants per group). In the standard group, embryos were removed from incubators for assessment. They determined no significant difference in live birth rate (LBR) between the groups, indicating the acceptable safety of time-lapse incubator use [18]. Similarly, Zhang et al. [19] performed a randomized controlled study, and noted no significant differences in cumulative implantation or LBR between time-lapse and conventional culture.

An obstacle preventing the utilization of the time-lapse incubator to monitor the moprohkinetics of bovine embryo development is that single-culture wells are required. In 1997, Donnay et al. discovered that single bovine embryos cultured in 20 µL of culture media did not progress to the blastocyst stage (20 µL/media) [20]. Embryos cultured in groups of three to six (6.7–3.3 µL/embryo) obtained a blastocyst rate of 6%. In groups of 20 embryos (1 µL/embryo), a blastocyst rate of 23% was reached [20]. Keefer et al. [21] came to a similar observation when comparing single vs. group-cultures of two-cell and eight-cell embryos. Single-culture two-cell embryos in 25 µL yielded lower rates of blastocyst formation (28%) than groups of five two-cell embryos (41%) [21]. Following this, single-culture eight-cell embryos in 25 µL had a blastocyst rate of 42% compared to 70% in groups of eight-cell embryos [21]. Further supporting these studies, Ferry et al. [22] compared cleavage development (day 2) and blastocyst development (day 8) in one, four, or forty zygotes cultured in 40 µL microdoplets. They concluded that the culture with forty zygotes had the highest blastocyst and cleavage rates, followed by the group of four then the individual culture [22]. Similarly, Khurana and Niemann [23] observed in 2000 that culturing groups of forty fertilized bovine zygotes in 500 µL of media resulted in higher rates of development compared to twenty zygotes in the same volume. Thus, the lack of developmental success in individual cultures is an obstacle blocking the observation of bovine morphokinetic parameters in the time-lapse incubator. Exploring different single-culture media can contribute to the development of more robust embryo viability predictions using morphokinetic parameters from the time-lapse incubator.

Objectives

• To develop and evaluate an AI-enabled smart monitoring system for assessing the developmental kinetics and stage-wise progression of bovine embryos cultured in SOF and Vitrolife Gx-TL™ media using MIRI time-lapse imaging, based on the time required and the proportion of embryos reaching each of the nine defined developmental stages;
• To evaluate the performance of transfer learning-based deep convolutional neural networks as an intelligent image analysis component of the smart monitoring system for the automated identification of bovine embryonic developmental stages from image frames extracted from MIRI time-lapse videos.

2. Materials and Methods

The protocol developed for “In Vitro Production of Bovine Embryos” by Elizabeth St John [24] from the Department of Biomedical Sciences at the University of Guelph (2024) was used in this study (Figure 1).

2.1. Ovary Collection

A thermos was preheated to 37 °C in a dry incubator overnight. On the day of ovary collection, the thermos was filled with 100 mL of a solution containing 10 mL Penicillin–Streptomycin (Pen/Strep) and 1 L of 0.9% saline. Once the ovaries were transported to the lab, they were washed with warm saline containing Pen/Strep before being distributed into beakers containing fresh saline and Pen/Strep solution. The beakers were then stored in dry incubators until use.

2.2. Oocyte Collection and In Vitro Maturation

2.2.1. Media Preparation

Four-well maturation plates were prepared using 500 μL of complete in vitro maturation (C-IVM) media per well. Additionally, two 35 mm search dishes and two 35 mm wash dishes were prepared, each containing 2.5 mL of C-IVM. All items were labeled with their contents and date, then placed in a 38.8 °C and 5.6% CO₂ incubator for equilibration for two hours prior to oocyte collection. The oocyte collection medium (F-10) was placed in a dry incubator at 37 °C for two hours before collection. Heaters in the aspiration room were turned on one hour prior to the procedure.

2.2.2. Aspiration Set Up

The aspiration bench was cleaned with ethanol and covered with bench coat paper. The vacuum pump, flask, and tubing were assembled. A 1 × 20 gauge multi-use blood needle was inserted into the rim of the vacutainer tube’s rubber lid and bent to approximately 35° using forceps, with the bevel facing the operator. A 1 × 18 gauge needle was inserted into the rim of the rubber lid, directly opposite the 20-gauge needle. The vacutainer tube was rinsed with 3 mL of F-10 collection media and then filled with 2 mL of F-10 collection media. A plate heater was set to 38 °C to maintain the temperature of the ovary beaker. Follicular fluid was aspirated and collected from 25 ovaries into a 50 mL conical tube. After aspiration was complete, the vacutainer tube was rinsed with F-10 collection media and the contents were transferred to a 50 mL conical tube. After 15 min, the oocytes settled to the bottom of the conical tube, and the supernatant was removed and discarded.

2.2.3. Oocyte Collection

The contents of the conical tube containing the oocytes were divided between two scored 100 mm dishes. The conical tube was then rinsed with 5–7 mL of F-10 collection media, which was added to the dishes to dilute the oocytes and debris. Good cumulus–oocyte complexes (COCs) were collected using a Wiretrol and placed into 35 mm C-IVM search dishes. Search dishes were replaced and returned to the incubator when pH changes were indicated by a color shift to hot pink. Once both scored dishes had been searched, the COCs were washed twice, once in each 35 mm C-IVM wash dish. Finally, the COCs were transferred to the four-well C-IVM maturation plate, with 50 COCs per well. The four-well maturation plate was then placed in the 38.8 °C and 5.6% CO₂ incubator for 22 to 24 h.

2.2.4. Coin Culture Degassing

The MIRI coin culture plate was removed from its packaging and left at room temperature to de-gas. A total of 12 mL of heavy embryo oil (5 mL per culture coin, with slight excess) was equilibrated in a T25 flask in a Tri-Gas incubator at 38.8 °C, 6% O₂, and 5.5–6.5% CO₂. SOF and Gx-TL™ were also equilibrated in the same incubator.

2.3. Fertilization

2.3.1. Media Preparation

Four-well plates were prepared using 480 μL of IVF TALP per well to accommodate 30 oocytes per well. Two 35 mm wash dishes were prepared with 2.5 mL of Hepes/Sperm TALP each, and one 35 mm wash dish was prepared with 2.5 mL of IVF TALP. Additionally, two 4 mL snap cap tubes were prepared, each containing 1.5 mL of Hepes/Sperm TALP. All dishes and tubes were equilibrated in a 38.8 °C, 5.6% CO₂ incubator for two hours prior to the swim-up. IVF TALP and Hepes/Sperm TALP were also placed in the same incubator with the lids loosened.

2.3.2. Swim-Up

After two hours, a semen straw was thawed in a 38.8 °C water bath and emptied into a snap cap tube. Eighty microliters of semen were transferred to the bottom of each Hepes/Sperm TALP snap cap tube, and both tubes were returned to the incubator for 45 min. During this period, the matured oocytes were washed twice in Hepes/Sperm TALP dishes and once in the IVF TALP dish. The matured oocytes were then transferred to the IVF TALP four-well plate, with 30 oocytes per well, and the plate was returned to the incubator. After the swim-up was complete, the upper layer containing motile sperm (above the sperm layer) was extracted and transferred to a 15 mL conical tube. Additional warm Hepes/Sperm TALP was added to bring the volume to 7 mL, followed by centrifugation for 5 min at 300× g. The pellet was retained and resuspended in 120–200 μL of IVF TALP, added drop by drop depending on the pellet size and the number of oocytes. Sperm concentration was assessed in an empty well, and 30–35 μL of the prepared sperm suspension was added to each well of the four-well plate containing oocytes. An additional 20 μL of penicillamine–hypotaurine–epinephrine (PHE) was added to each well. The fertilized oocytes were then returned to the 38.8 °C, 5.6% CO₂ incubator for 10 to 18 h.

2.3.3. SOF and Gx-TL™ Dish Preparation

The number of presumptive zygotes was counted, and 28 were subtracted for allocation to the MIRI coin culture plate. The number of four-well plates required to culture the remaining zygotes was calculated based on 30 zygotes per well, ensuring equal distribution between SOF and Gx-TL™. The required four-well plates were prepared with 400 µL of either SOF or Gx-TL™ in each well and covered with 1 mL of embryo oil. A 35 mm wash dish was prepared with 2.5 mL of SOF, and another with 2.5 mL of Gx-TL™. Additionally, a four-well wash plate was prepared with 400 µL of Hepes/Sperm TALP in each well. All prepared dishes, four-well plates, and Hepes/Sperm TALP were placed in a 38.8 °C, 5.6% CO₂ incubator with loose lids.

2.3.4. Coin Culture Plate Preparation

Twenty-five microliters of SOF were added to each well of the coin culture plate, and the entire plate was gently covered with 5 mL of heavy embryo oil. The same steps were repeated using Gx-TL™ instead of SOF. The culture coins were then placed in the timelapse incubator with the lid on.

2.3.5. In Vitro Culture

A total of 1.5 mL of warm Hepes/Sperm TALP was added to a 15 mL conical tube. Zygotes were removed from the fertilization dish and transferred into the conical tube containing Hepes/Sperm TALP. The tube containing the presumptive zygotes was vortexed at high speed for two minutes. The contents were then emptied into an empty 35 mm dish, and the tube was rinsed with 1 mL of Hepes/Sperm TALP. Zygotes were transferred to the first well of a four-well Hepes wash plate. Zygotes lacking cumulus cells were washed through the subsequent Hepes wells. Using a pipette set to 100 µL, cumulus cells were denuded from those still attached to the zygotes. Once all zygotes were denuded, they were washed through the remaining Hepes wells. Prior to final transfer, half of the zygotes were washed in either the SOF or Gx-TL™ 35 mm dish, setting aside 14 for each coin culture plate. Finally, 40 zygotes per well were transferred equally into the four-well SOF and Gx-TL™ plates, and one zygote per well was placed in each coin culture plate.

2.4. Development of Deep Learning Classification Models

2.4.1. Developmental Stage Timing and Frame Extraction from Time-Lapse Video

The MIRI time-lapse imaging system automatically captured images of developing embryos every five minutes, and generated a video sequence for each embryo. The resulting video was in .avi format, and the duration of each embryo was approximately two minutes. The timing and progression of embryo development through the nine defined stages (2-cell (T2), 3-cell (T3), 4-cell (T4), 5-cell (T5), 6-cell (T6), 7-cell (T7), 8-cell (T8), morula, and blastocyst) were manually assessed for each embryo by carefully analyzing the video captured by the MIRI system.

Each time-lapse video was reviewed by the first author, and developmental stages were assigned based on the first frame in which the corresponding developmental morphology could be clearly identified. The T2–T8 stages were identified according to the number of discernible blastomeres. The morula stage was identified by the compaction and reduced visibility of individual blastomere boundaries, while the blastocyst stage was identified by the appearance of a blastocoel cavity. A non-developing embryo (NDE) was defined as an embryo that failed to cleave.

To support the automated classification, a Python script (version 3.11.10) was executed to extract representative frames from each video. These frames corresponded to various developmental stages in both developing and non-developing embryos and were used as input for further analysis and model training.

2.4.2. Image Augmentation

The number of original images for each developmental stage in each culture medium is provided in

Table 1. Developmental progression rate of bovine embryos.

Culture Medium	No. of Single Cultured Zygotes	Developmental Stages Reached and Clearly Identified in Images
		T2	T3	T4	T5	T6	T7	T8	Morula	Blastocyst
SOF	152	81	43	42	29	27	11	14	10	6
Gx-TL™	159	77	43	40	28	33	9	8	18	10

. For the SOF culture, the numbers were 81, 43, 42, 29, 27, 11, 14, 10, and 6 for the T2, T3, T4, T5, T6, T7, T8, morula, and blastocyst stages, respectively. Similarly, for the Gx-TL™ culture, the numbers were 77, 43, 40, 28, 33, 9, 8, 18, and 10, respectively. Therefore, the total number of original images was 529.

Due to the limited number of images obtained in this study, data augmentation was performed to increase the dataset to 500 images for each developmental stage. This resulted in a total of 5000 images (9 developmental stages × 500 images + 500 images of non-developing embryos). The augmentation techniques applied included brightness and contrast adjustment, the addition of 5% random noise, and minor rotation (±5 degrees) to enhance variability and improve model robustness.

2.4.3. Transfer Learning-Based CNN Models and Classification Approaches

Since the data obtained in this study are relatively few, transfer learning-based CNNs that were originally trained on large-scale image datasets (especially repositories such as ImageNet) were adapted and fine-tuned to our specific embryo classification task, so as to improve the robustness and reduce the training time.

Three well-established CNN architectures were used in this study: ResNet18, DenseNet121, and EfficientNetB0.

ResNet 18 uses residual connections to reduce the vanishing gradient problem to identify small differences between images, thus allowing stable training while remaining relatively lightweight. Likewise, DenseNet 121 utilizes densely connected layers that enable feature reuse and efficient gradient propagation, making it well suited for distinguishing subtle differences, especially in medical and biological images. On the other hand, EfficientNet-B0 was designed through a neural architecture search to maximize the balance between accuracy and computational efficiency (with a small number of parameters). Thus, EfficientNet-B0 is adaptable for biomedical imaging classification tasks, where small morphological differences are used for classification, resulting in higher classification accuracies with lower computational resource consumption.

These architectures balance interpretability, computational efficiency, and predictive performance, enabling a robust evaluation of CNNs for embryo stage classification.

Manually annotated images of embryo stages were used to develop classification models under the following approaches:

• Two-Class Model—Classifies embryos as either developing (positive) or non-developing (negative) at each stage;
• Nine-Class Model—Identifies developing embryos by their specific developmental stage, namely, T2, T3, T4, T5, T6, T7, T8, morula, or blastocyst;
• Ten-Class Model—Expands the nine-class model by including a non-developing embryo (NDE) class.

Each CNN was trained and evaluated across two individual media types (SOF and Gx-TL™) as well as in a combined dataset (irrespective of medium) to assess generalizability.

The training parameters were consistently set as follows:

• Batch size, 16;
• Learning rate, 0.0001;
• Number of epochs, 10;
• Data split ratio (train:validation:test), 70:15:15.

All deep learning CNNs were implemented in Visual Studio code using Python 3.11.0. The CNN models were built using PyTorch 2.5.1 and Torchvision 0.20.1. For data processing and evaluation, the libraries of NumPy (v2.3.0), Pandas (v2.3.0), scikit-learn (v1.3.0), and Matplotlib (v3.10.3) were used. The models were trained and executed on a laptop equipped with an NVIDIA GeForce RTX 3050 GPU, with CUDA (v12.1) and cuDNN (v9.1.0), providing GPU-accelerated computations.

3. Results

3.1. Effect of Culture Medium on Embryo Development

The developmental progression of bovine embryos cultured in SOF and Gx-TL™ is summarized in Table 1. Totals of 152 and 159 zygotes were individually cultured in SOF and Gx-TL™, respectively. In both media, a gradual decline in the number of embryos reaching successive developmental stages was observed. In SOF, 81 embryos reached the two-cell stage (T2), with only 6 progressing to the blastocyst stage. In comparison, Gx-TL™ supported the development of 77 embryos to the T2 stage, with 10 advancing to blastocysts. Although more Gx-TL™-cultured embryos reached the morula and blastocyst stages, these differences were not statistically significant.

Table 2. Average time to reach developmental stages.

Culture Medium	Mean developmental Time (h)
	T2	T3	T4	T5	T6	T7	T8	Morula	Blastocyst
SOF	44.50± 19.66	54.92± 21.23	56.02± 27.62	59.52± 23.43	57.25± 14.25	69.53± 25.00	86.77± 26.91	121.82± 11.01	190.71± 4.06
Gx-TL™	43.00± 12.40	52.05± 10.21	58.81± 21.17	62.72± 19.74	70.06± 24.70	84.15± 21.59	97.64± 25.70	135.69± 12.53	179.20± 15.71

presents the average time required for bovine embryos to reach each developmental stage. Embryos cultured in SOF reached the two-cell (T2) stage at an average of 44.51 h, while those in Gx-TL™ reached T2 slightly earlier, at 43.05 h. In general, across most stages, embryos in Gx-TL™ tended to progress slightly faster than those in SOF.

The average times required to reach each developmental stage for bovine zygotes that successfully developed to the blastocyst stage are given in

Table 3. Average time to reach developmental stages for zygotes that reached the blastocyst stage (n = 6 for SOF and n = 10 for Gx-TL™).

Culture Medium	Mean Developmental Time (h)
	T2	T3	T4	T5	T6	T7	T8	Morula	Blastocyst
SOF	33.76± 6.45	45.43± 3.74	43.94± 8.01	56.63± 2.85	56.10± 6.08	61.72± 3.65	99.98± 7.22	120.39± 8.41	190.71± 4.06
Gx-TL™	36.48± 5.63	46.29± 7.98	50.59± 7.00	57.09± 10.45	64.85± 18.30	83.93± 21.34	104.88± 15.02	132.37± 12.57	179.20± 15.71

.

3.2. CNN-Based Classification

3.2.1. Two-Class Model

The classification accuracy and processing time for two-class analysis using the SOF model, Gx-TL™ model, and the combined (SOF + Gx-TL™) model are presented in

Table 4. Classification accuracy and processing time for two-class classification of bovine embryos using SOF model.

Developmental Stage	Average Inference Time/Image (Sec)			Total Training + Evaluation Time (Sec)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	1.133	1.460	1.141	170.0	219.1	171.2	100	100	100
T3	1.170	1.378	1.150	175.6	206.8	172.6	100	100	100
T4	1.132	1.375	1.153	169.9	206.3	173.0	100	100	100
T5	1.121	1.354	1.123	168.1	203.2	168.4	100	100	100
T6	1.147	1.373	1.129	172.1	205.9	169.4	100	100	100
T7	1.115	1.352	1.181	167.3	202.8	177.2	100	100	100
T8	1.169	1.361	1.167	175.4	204.1	175.1	100	100	100
Morula	1.183	1.335	1.122	177.5	200.3	168.3	100	100	100
Blastocyst	1.248	1.369	1.163	187.2	205.4	174.5	100	100	100

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0.

,

Table 5. Classification accuracy and processing time for the two-class classification of bovine embryos using the Gx-TL™ model.

Developmental Stage	Average Inference Time/Image (S)			Total Training + Evaluation Time (S)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	1.155	1.394	1.140	173.2	209.1	171.1	100	100	100
T3	1.228	1.387	1.141	184.2	208.1	171.2	100	100	100
T4	1.091	1.362	1.220	163.7	204.3	183.0	100	100	100
T5	1.197	1.375	1.137	179.6	206.4	170.6	100	100	100
T6	1.148	1.353	1.182	172.2	203.0	177.4	100	100	100
T7	1.171	1.359	1.119	175.7	203.9	167.8	100	100	100
T8	1.238	1.377	1.153	185.5	206.5	172.9	100	100	100
Morula	1.161	1.376	1.117	174.1	206.4	167.6	100	100	100
Blastocyst	1.150	1.377	1.193	172.5	206.6	179.0	100	100	100

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0.

and

Table 6. Classification accuracy and processing for the time two-class classification of bovine embryos using the combined model (SOF + Gx-TL™).

Developmental Stage	Average Inference Time/Image (S)			Total Training + Evaluation Time (S)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	1.013	1.303	1.166	303.9	391.1	349.9	100	100	100
T3	0.997	1.359	1.117	299.1	407.9	335.3	100	100	100
T4	0.990	1.390	1.124	297.1	417.1	337.2	100	100	100
T5	1.162	1.402	1.127	348.8	420.8	338.1	100	100	100
T6	1.315	1.425	1.155	394.7	427.5	346.6	100	100	100
T7	0.985	1.438	1.103	295.7	431.4	330.9	100	100	100
T8	0.990	1.369	1.120	297.0	410.9	336.2	100	100	100
Morula	0.990	1.484	1.107	297.2	445.3	332.1	100	100	100
Blastocyst	0.987	1.476	1.102	296.1	443.0	330.7	100	100	100

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0.

. All models achieved 100% accuracy across all developmental stages when using the ResNet18, DenseNet121, and EfficientNet B0 CNN architectures. This demonstrates that these pretrained CNN models are highly effective for identifying individual developmental stages relative to non-developing embryos.

Although all CNNs achieved perfect accuracy, the processing times for training, evaluation, and per-image inference varied significantly. In the SOF model, DenseNet121 required the most time for training and evaluation (200.3–219.1 s), compared to ResNet18 (167.3–187.2 s) and EfficientNet B0 (168.3–177.2 s). A similar trend was observed in the Gx-TL™ model, where DenseNet121 again had the highest processing time (203.0–209.1 s), followed by ResNet18 (163.7–185.5 s) and EfficientNet B0 (167.6–183.0 s).

3.2.2. Nine-Class Model

SOF model 

Table 7. Classification accuracy and processing time for nine-class classification of bovine embryos using the SOF model.

Developmental Stage	Average Inference Time/Image (S)			Total Evaluation Time (S)			Total Training Time (S)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	0.021	0.019	0.020	14.19	13.41	13.73	766.0	1022	768.0	100	100	100
T3										100	100	100
T4										100	100	100
T5										98	98	100
T6										98	98	96
T7										100	97	98
T8										94	98	94
Morula										100	100	100
Blastocyst										100	100	100
Overall accuracy										99.1	99.2	98.8

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0.

presents the classification accuracy and processing time for the nine-class analysis using the SOF model. All CNN architectures correctly identified the T2, T3, T4, morula, and blastocyst stages.

However, some misclassifications occurred in other developmental stages. Specifically, confusion was observed between T5 and T6, and between T7 and T8 across all CNNs, which could be due to the similarity in images of those stages (

Table 8. Confusion matrix for nine-class classification of bovine embryos using the SOF model.

ResNet18	DenseNet121	EfficientNet B0

). The confusion matrices depict the percentage distributions of test samples across the predicted developmental stages. Diagonal values represent correctly classified samples, whereas off-diagonal values represent misclassified samples and the misclassified stages. Overall classification accuracy was used as the primary evaluation metric in the present study.

Gx-TL™ model In the Gx-TL™ model, the overall classification accuracy for the nine-class analysis was 100% for both ResNet18 and DenseNet121, and 97.7% for EfficientNet B0 (

Table 9. Classification accuracy and processing time for nine-class classification of bovine embryos using Gx-TL™ model.

Developmental Stage	Average Inference Time/Image (S)			Total Evaluation Time (S)			Total Training Time (S)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	0.019	0.022	0.020	13.10	15.07	13.54	727.8	1687	747.3	100	100	96
T3										100	100	98
T4										100	100	98
T5										100	100	93
T6										100	100	98
T7										100	100	98
T8										100	100	97
Morula										100	100	98
Blastocyst										100	100	100
Overall accuracy										100	100	97.7

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0.

). The confusion matrix for EfficientNet B0, shown in

Table 10. Confusion matrix for nine-class classification of bovine embryos using Gx-TL™ model.

EfficientNet B0

, indicates that the T4 stage was frequently misclassified as other developmental stages.

Combined model (SOF + Gx-TL™) 

Table 11. Classification accuracy and processing time for nine-class classification of bovine embryos using combined model (SOF + Gx-TL™).

Developmental Stage	Average Inference Time/Image (S)			Total Evaluation Time (S)			Total Training Time (S)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	0.017	0.022	0.019	24.14	30.18	26.46	1487	2012	1559	100	100	98
T3										100	98	99
T4										100	100	98
T5										100	80	95
T6										99	99	97
T7										99	98	98
T8										98	85	96
Morula										100	100	99
Blastocyst										100	100	100
Overall accuracy										99.5	95.5	97.7

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0.

shows the classification accuracy and processing time for the nine-class analysis using the combined model. The classification accuracies were 99.5% for ResNet18, 95.5% for DenseNet121, and 97.7% for EfficientNet B0. In DenseNet121, several developmental stages were misclassified as T6 and T7 (

Table 12. Confusion matrix for nine-class classification of bovine embryos using combined model (SOF + Gx-TL™).

ResNet18	DenseNet121	EfficientNet B0

). In EfficientNet B0, all developmental stages except T2 were misclassified as at least one other stage.

3.2.3. Ten-Class Model

SOF model The classification accuracy of the three CNN architectures ranged from 98.4% to 98.8% for embryos cultured in SOF medium (

Table 13. Classification accuracy and processing time for ten-class classification of bovine embryos using the SOF model.

Developmental Stage	Average Inference Time/Image (S)			Total Evaluation Time (S)			Total Training Time (S)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	0.020	0.019	0.021	15.62	14.39	15.87	802.4	1028	861.7	100	100	100
T3										100	100	100
T4										100	100	100
T5										100	100	100
T6										96	96	96
T7										98	98	96
T8										100	100	100
Morula										93	93	93
Blastocyst										100	100	100
NDE **										100	100	98
Overall accuracy										98.8	98.8	98.4

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0. ** NDE-non-developing embryo.

). In this approach, both ResNet18 and DenseNet121 achieved 100% accuracy in identifying non-developing embryos (NDE), demonstrating their potential for use in real-time applications. EfficientNet B0, however, did not reach perfect accuracy for NDE classification.

None of the CNN models achieved 100% accuracy in classifying the T6, T7, and morula stages. All three architectures exhibited misclassifications between the T6 and T7 stages (

Table 14. Confusion matrix for ten-class classification of bovine embryos using the SOF model.

ResNet18	DenseNet121	EfficientNet B0

). In particular, EfficientNet B0 misclassified morula-stage embryos as T8, blastocyst, or NDE.

The average inference time per image ranged from 0.019 to 0.021 s. Model training times were 802 s for ResNet18, 862 s for EfficientNet B0, and 1028 s for DenseNet121.

Gx-TL™ model All developmental stages and NDE cultured in the Gx medium were accurately classified (100%) by all three tested CNN architectures (

Table 15. Classification accuracy and processing time for ten-class classification of bovine embryos using the Gx-TL™ model.

Developmental Stage	Average Inference Time/Image (S)			Total Evaluation Time (S)			Total Training Time (S)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	0.020	0.020	0.020	15.32	15.40	15.66	784.5	1026	815.5	100	100	100
T3										100	100	100
T4										100	100	100
T5										100	100	100
T6										100	100	100
T7										100	100	100
T8										100	100	100
Morula										100	100	100
Blastocyst										100	100	100
NDE **										100	100	100
Overall accuracy										100	100	100

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0. ** NDE-non-developing embryo.

). The average inference time per image was 0.02 s across all CNNs. Model training times were 785 s for ResNet18, 816 s for EfficientNet B0, and 1026 s for DenseNet121.

Combined model (SOF + Gx-TL™) The overall classification accuracy of the combined media model ranged from 99.7% to 99.9% (

Table 16. Classification accuracy and processing time for ten-class classification of bovine embryos using the combined model (SOF + Gx-TL™).

Developmental Stage	Average Inference Time/Image (S)			Total Evaluation Time (S)			Total Training Time (S)			Classification Accuracy (%)
	Res. *	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.	Res.	Dense.	Eff.
T2	0.018	0.020	0.020	27.47	31.20	30.05	2019	2207	1721	100	100	100
T3										100	100	100
T4										100	100	100
T5										100	100	100
T6										98	99	100
T7										100	100	100
T8										98	100	100
Morula										100	99	99
Blastocyst										100	100	100
NDE **										100	100	100
Overall accuracy										99.7	99.8	99.9

* Res-ResNet18; Dense-DenseNet12; Eff-Efficient NetB0. ** NDE-non-developing embryo.

). EfficientNet B0 achieved a perfect classification accuracy of 100% accuracy across all developmental stages (except for the morula stage). ResNet18 and DenseNet121 showed misclassifications between the T6 and T7 stages, as well as between T8 and morula (

Table 17. Confusion matrix for ten-class classification of bovine embryos using combined model (SOF + Gx-TL™).

ResNet18	DenseNet121	EfficientNet B0

).

Figure 2 and Figure 3 show the typical MIRI images of the developmental stages of the bovine embryos cultured in SOF and Gx-TL™ media. In general, the top surface of the embryos cultured in Gx-TL™ was darker and more definite than that of the SOF medium. This might have contributed to the better performance of the Gx-TL™ ten-class model with all three architectures.

A potential limitation of this study is the variation in image appearance between embryos cultured in SOF and Gx-TL™. Under the imaging conditions applied, Gx-TL™ images appeared darker and more sharply defined. As a result, the convolutional neural networks (CNNs) may have learned medium-associated differences in image intensity, contrast, or background, in addition to embryo morphology. The high classification performance observed should be regarded as proof of concept under the specific MIRI imaging and culture conditions employed, rather than as evidence of generalizability across different culture media, imaging systems, or laboratory settings. Future research should assess intensity normalization and conduct independent cross-medium and external validations using larger numbers of biologically independent embryos.

4. Discussion

4.1. Effect of Culture Medium on Embryo Development

Embryos cultured in SOF reached the two-cell stage faster (33.76 h) than those in Gx-TL™ (36.48 h), with similar trends observed in subsequent stages up to morula. However, embryos in Gx-TL™ reached the blastocyst stage slightly earlier (179.20 h) than those in SOF (190.71 h). For embryos that reached the blastocyst stage, those cultured in SOF demonstrated earlier cleavage at certain early stages, while the average time to reach the blastocyst stage was shorter in Gx-TL™. Because of the small number of blastocysts and the lack of consistent, statistically significant differences, these findings should be interpreted as descriptive trends rather than evidence that the culture medium affects developmental timing.

Somfai et al. [25] in 2010 were the first to study bovine embryo development using a commercial time-lapse culture system. The lengths of the first and second cell cycles were significantly shorter in viable embryos [25]. Abnormal division from the zygote stage to three or four blastomeres developed similarly to that in embryos developing with normal division [25]. These observations align with the results of this study, as the times taken to reach two and three blastomeres were shorter in embryos developing to the blastocyst stage. Similarly, many embryos reaching the blastocyst stage exhibited abnormal division. In 2016, Sugimura et al. [26] used microwell culture dishes (LinKID micro25) to observe bovine embryos in the time-lapse incubator and determine success post-implantation. A shorter time for first cleavage and further cleavages indicated higher rates of successful pregnancies in their study [26]. Additionally, the pregnancy rates of abnormally cleaving blastocysts were lower than those of normally cleaving blastocyts [26]. Although not directly comparable to this study, these findings provide a context for the potential relationship between early embryo development and subsequent developmental outcomes.

The observed association between earlier cleavage and subsequent developmental competence aligns with findings reported across multiple mammalian species. For example, in bovine embryos, Somfai et al. documented significantly shorter first- and second-cell cycles in embryos that progressed to the blastocyst stage, while Sugimura et al. identified the timing of first cleavage as a morphokinetic variable associated with pregnancy following bovine embryo transfer [25,26]. Similarly, Hillyear et al. [27] demonstrated that porcine embryos that later blastulated reached the two-cell and morula stages earlier than non-blastulating embryos, with the timing of first cleavage serving as a predictive marker for blastulation. Liu et al. [28] also reported delayed pre-blastulation developmental events in fishing cat–domestic cat iSCNT embryos compared to IVF controls, and observed reduced blastocyst development following abnormal direct cleavage [29]. These studies emphasize the biological significance of early cleavage kinetics. However, absolute developmental times should not be directly compared across species due to differences in embryo origin, culture medium, oxygen conditions, imaging intervals, and culture systems. Morphokinetic associations do not always translate into improved reproductive outcomes. For instance, in a human pilot randomized trial, Sacks et al. found no significant improvement in live birth or pregnancy outcomes resulting from time-lapse incubation with morphokinetic evaluation [30]. Therefore, the developmental timing patterns identified in this study should be regarded as indicators for further validation, rather than as direct predictors of bovine pregnancy success.

4.2. CNN-Based Classification

4.2.1. Two-Class Model

In the combined model, processing times increased due to the larger number of images per category. The same trend persisted; DenseNet121 took the longest (391.1–445.3 s), compared to ResNet18 (295.7–394.7 s) and EfficientNet B0 (330.7–349.9 s). This can clearly be attributed to the differences in architectural complexity and parameter usage. For example, ResNet-18 can allow certain input mappings to by-pass some convolutional layers, resulting in a hollow structure and the fastest inference times. EfficientNetB0 utilizes compound scaling, resulting in a competitive accuracy, a lightweight model and a moderate computational load. DenseNet 121, due to the input from the preceding densely connected layers and increased computational operations per layer, resulted in the highest processing and inference times among the three CNNs. The per-image inference time was also higher for DenseNet121 (1.303–1.484 s) compared to ResNet18 (0.985–1.315 s) and EfficientNet B0 (1.102–1.163 s) across the SOF, Gx-TL™, and combined models.

4.2.2. Nine-Class Model

Hachani et al. [31] recently introduced CLEmbryo, a supervised contrastive learning framework that integrates focal loss and a lightweight three-dimensional convolutional neural network (3D CNN) for the stage-wise classification of bovine embryos from time-lapse video microscopy. They used the Bovine Embryo Cell Stages dataset, which comprises 1221 video sequences and repeated evaluations; CLEmbryo achieved an overall classification accuracy of 81.25%. The greatest challenge was observed in classifying intermediate developmental stages, primarily due to class imbalance and morphological ambiguity [31]. In our present study, overall nine-class accuracy ranged from 95.5% to 100% across the SOF, Gx-TL™, and combined datasets, while the ten-class models achieved 98.4% to 100%. Misclassifications primarily occurred between morphologically adjacent stages, such as T5/T6, T6/T7, and T8/morula, which aligns with the difficulties reported by Hachani et al. in distinguishing intermediate stages. Nevertheless, the higher accuracies observed in the present study should not be interpreted as evidence of superior performance, as the datasets, classification tasks, imaging systems, and validation strategies differed. Furthermore, the augmentation-before-splitting procedure applied in the present study may have produced optimistic estimates of model performance.

Similar to the two-class model, DenseNet121 required a longer training time (1022 s) compared to ResNet18 (766 s) and EfficientNet B0 (768 s). In the Gx-TL™ model, the average inference time per image ranged from 0.019 to 0.021 s. Model training times were 728 s for ResNet18, 747 s for EfficientNet B0, and 1687 s for DenseNet121. In the Combined model (SOF + Gx-TL™), the average inference time per image ranged from 0.017 to 0.022 s, and total training time varied between 1487 and 2012 s—more than that of two-class model, which was probably due to the increase in the classification stages. The observed classification performance demonstrates the feasibility of the CNN-based classification of developmental stages within the present dataset; however, further validation using larger independent datasets is required to assess model generalizability.

4.2.3. Ten-Class Model

Total model training times ranged from 1721 to 2207 s (the most time-consuming among the three classification methods), while the average inference time per image varied between 0.018 and 0.02 s. Although an exact approach was not used, Rocha et al. in 2017 pursued a similar objective by using genetic algorithms (GA) and artificial neural networks (ANN) to assess blastocyst quality (grades 1–4) [14]. The GA evaluated different ANNs to determine the network with minimal errors in blastocyst classification [14,29]. The algorithm backpropogation was used to train 482 blastocyst images (70% for training, 15% for validation, 15% for testing) [14]. ANN accuracy was established based on variations between ANN values and embryologist evaluation. The optimal ANN using the GA method showed 76% accuracy in classifying bovine blastocyst quality [19].

Rocha et al. [14] reported about a 76% accuracy using artificial neural networks to classify bovine blastocyst quality, highlighting the challenges of automated morphological assessment in earlier research. More recently, Wells et al. [32] assessed machine learning approaches using bovine embryos collected during routine embryo-transfer procedures, and observed an 81.7% agreement with expert embryologists for the developmental stage, as well as an approximately 95% agreement for embryo transferability. Their findings also revealed considerable variability among human evaluators, with only 59.8% overall agreement regarding embryo stage. These results emphasize the potential value of automated assessment for enhancing standardization, while also indicating that performance should be evaluated using independent embryos under realistic laboratory or field conditions. In this context, the good stage-classification performance achieved in the present study is promising as a proof of concept. However, prospective validation using independent embryos and biologically relevant outcomes is required before direct comparisons with expert evaluation or embryo-transfer performance can be justified.

4.3. Practical and Research Implications

These findings have important practical and research implications for bovine embryo monitoring. Integrating time-lapse imaging with automated CNN-based stage recognition can reduce the time and subjectivity of manual assessments, and enable the standardized, continuous documentation of embryo development. The short inference times of the multiclass models support the technical feasibility of near-real-time monitoring. However, these models are a proof-of-concept and are not yet suitable for embryo selection or transfer decisions, as they require external validation and independent embryo-level testing. This study provides a basic framework for combining morphokinetic data with automated image analysis, and for comparing developmental patterns across culture conditions.

5. Current Limitations and Future Work

A limitation of the present study is that image augmentation was performed before splitting the dataset into training, validation, and test sets. Therefore, augmented images originating from the same embryo may have been distributed across different subsets. Hence, the reported classification performance may be optimistic because the models may have learned embryo-specific visual characteristics in addition to morphological features associated with developmental stage. Moreover, model performance was evaluated using a single 70:15:15 data split without repeated experiments or embryo-level cross-validation. Therefore, variability across different data partitions and generalizability to independent embryos could not be fully assessed.

Future studies should address these limitations by first dividing the original images at the embryo level and subsequently applying augmentation only to the training set. Embryo-level cross-validation (in which all images from an individual embryo are retained within the same fold) would provide a more rigorous assessment of model generalizability. Independent external larger datasets with embryo-level cross-validation should also be implemented to facilitate a more robust evaluation of model performance.

6. Conclusions

A classification accuracy of more than 90% was achieved in developmental stage identification by Resnet 18, Efficientnet B0, and Densenet 121 for all three model types. These findings demonstrate the feasibility and potential use of CNN-based approaches for the automated analysis of bovine embryo developmental stages. However, the relatively small dataset and lack of independent dataset validation limit the assessment of model robustness and generalizability. A larger collection of annotated images with diverse datasets for training more robust and generalizable models is required before the proposed approach can be considered for practical application in livestock reproduction.