Deep Learning-based Video Object Detection for Single-and Multi-Cell Analysis and Evaluation in Time-Lapse Imaging

Deep Learning-based Video Object Detection for Single-and Multi-Cell Analysis and Evaluation in Time-Lapse Imaging | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Deep Learning-based Video Object Detection for Single-and Multi-Cell Analysis and Evaluation in Time-Lapse Imaging Taikyeong Jeong This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6434380/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background : In this paper, we prove the efficiency of a video object detection algorithm through deep learning to have the most essential video of time-lapse data for the completion of artificial intelligence vision object detection architecture that is used for prediction purpose. We alsoinvestigated time-lapse video data, which is the most important part since it recorded during in vitro fertilization process. Particularly, to achieve the most efficient video object detection by limiting special-purpose object detection to only medical healthcarebio-domains, all conditions were satisfied among the single-stage videoobject detection architectures, and proved as theoretical proofs and experiments. Method: Due to the characteristics of bio-medical in the experimental purpose, we applied artificial intelligence neural networks as a way to capture the frames per second (fps)changes of time-varying video images. To gain advantages in science and mathematics in the biomedical domain, we considered the aspects of entropy, confidence, and object occurrence probability. Accurate time-lapse video object detection factors include: ( i ) first, the accuracy of the number of cells divided after embryo fertilization, ( ii ) second, the acute cell size during cell division, ( iii ) third, the morphological uniformity of embryos, and ( iv ) fourth, the possibility of possible fertilization after cell division. Results : The most significant finding in this study is the accurate counting of cells after embryo fertilization, as detected through time-lapse video object recognition. From an AI vision perspective, we propose a fast and efficient video detection framework by implementing and evaluating two distinct learning models: RetinaNet, a single-stage detector, and Fast R-CNN, a multi-stage detector. Their performance was compared against other deep learning-based detection models. Theoretical insights and practical implications regarding the full cycle of human embryonic development were derived, particularly through the identification and prediction of abnormal temporal patterns. Deep learning Machine learning Time-series data Video object detection Retina-Net Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. Introduction At a time when artificial intelligence architectures and algorithms are being actively applied and developed across various fields, particularly in the analysis of video data, it is increasingly important to accurately classify images and interpret their meaning—especially in the domains of energy, defense, healthcare, and bioinformatics [ 1 ]. This study focuses on video image processing of time-lapse data, which captures continuous biological changes such as cell division during the in vitro fertilization process. Specifically, we explore the detection and recognition of target objects in time-lapse videos using supervised and unsupervised learning methods. It is important to note that the method of color representation (e.g., grayscale or true color) is not a focus of this work. Time-lapse video analysis represents a significant scientific domain for intelligent diagnosis. Because the frames change continuously—reflecting the dynamic nature of biological phenomena like cell division—the accuracy of classification and the final performance metrics can differ substantially depending on the approach used. Despite limited data availability and the technical challenges of pattern extraction, applying machine learning algorithms to time-lapse video has gained traction in recent years [ 2 ], even under increasingly strict data protection regulations [ 3 ]. Object recognition and extraction in time-lapse videos are often evaluated through video question answering (VQA) tasks [ 4 , 5 ], where one of the most critical performance factors is the capability of the graphics processing unit (GPU). Recent GPU development has also contributed significantly to the improvement of AI algorithms [ 6 ]. However, as frames per second (fps) increase, the number of video frames grows rapidly while the processing speed of computers approaches hardware limits. For example, analyzing an 18 fps silent video posed challenges in applying AI algorithms, largely due to the demands of resolution, sharpness, contrast modulation, and data volume in fields such as energy and biomedical research. These challenges also involve degrees of computational complexity [ 7 ]. This paper proposes a learning mechanism for time-lapse video image analysis with three key components: ( i ) detection, ( ii ) segmentation, and ( iii ) tracking of dynamic target objects that evolve over time [ 8 ]. In particular, video data from energy and medical domains exhibit subtle and continuous changes—such as malfunction onset or morphological degradation—that are often difficult for the human eye to detect. For instance, in the case of energy conversion devices (e.g., transformers, energy-saving systems), performance degradation may be inferred from subtle temperature fluctuations recorded over time-lapse sequences. Similarly, cellular division, transitioning from one cell to many, generates time-dependent image sequences—whether grayscale or color—that form the basis for continuous monitoring. This study investigates time-varying target objects within time-lapse videos, aiming to understand and analyze their progression over time. Because video imagery serves as the primary observation medium, this research contributes to the advancement of AI-based vision systems in deep learning. Moreover, we analyze time-series data, from object extraction to detection, through the application of artificial intelligence algorithms and objective experimentation. The targeted time-lapse videos are confined to the biomedical domain and represent initial experimental datasets. This study builds upon previous results from the E-OFW project [ 9 , 10 ] and the AI-based Cognitive Intelligence and S-OFW projects [ 11 – 13 ], which focused on efficient vision architecture development. In this context, the target observation is the human embryo culture stage, a foundational step in in vitro fertilization. The primary goal of this paper is to investigate: ( i ) the exact number of cells following fertilization, ( ii ) cell size during division, ( iii ) embryo morphology, and ( iv ) uniformity after division. Due to the sensitive nature of the data, all video materials were carefully de-identified and anonymized [ 14 ]. Ultimately, this paper aims to provide a systematic perspective on video object detection architecture and pattern extraction using deep learning algorithms. Given the continuity of visual data that is difficult to detect or track with the human eye, AI-based video analysis can be effectively applied in various scientific domains, including biomedical and energy-related research. The following are the key contributions of this paper: We propose a video object detection and segmentation methodology derived from artificial intelligence architectures. Based on deep learning algorithms, our approach enables accurate cell counting and area measurement following pre-processing and classification. A theoretical analysis is conducted to examine the correlation and impact of human embryonic development, with a focus on abnormal time-series pattern extraction and prediction. Experimental validation is performed using a sequence of 1,078 pre-processed data instances (including at least eleven training datasets, totaling approximately 7.5 TB [ 15 – 18 ]), processed through publicly available dataset. We define each subject and object in the video inputs, verify entropy-based cell counting, and validate detection performance using a threshold value for post-division uniformity in the frequency domain. The first section of the paper is usually an introduction section, in which the background, the topic and aims are described. The main results and principal conclusions should be highlighted in this conclusion section. The remainder of the paper is organized as follows: Section 2 introduces related works and studies, Section 3 describes the object detection model for proposed video object detection model and deep learning vision architecture. Section 4 presents experimental setup, results and how can extract the pattern from time-series data. Section 5 discusses about the results of video images associated with ground-truth, video detection and Section 6 concludes the paper 2. Related works Recent advancements in deep learning have significantly contributed to the field of reproductive medicine, particularly in the assessment of embryo viability and blastocyst quality. Convolutional Neural Networks (CNNs) and other deep learning architectures have been employed to analyze embryo and blastocyst datasets, thereby enhancing the accuracy and efficiency of embryo selection in in vitro fertilization (IVF) [ 19 ]. Several studies have utilized deep learning models trained on time-lapse imaging datasets of embryos to predict implantation success. These models are capable of extracting both morphological and temporal features, enabling automated and objective embryo grading [ 20 ]. For instance, CNN-based approaches have been applied to assess embryo quality based on Gardner’s criteria, often outperforming traditional manual grading by embryologists [ 21 ]. Early-stage models primarily focused on single-image detection, targeting static morphological snapshots to evaluate developmental quality at specific time points. In contrast, more recent frameworks have adopted multi-image detection strategies, integrating temporal sequences to analyze embryo progression dynamically. This multi-frame approach enables the capture of subtle developmental cues—such as mitotic timing and blastomere symmetry—that are not discernible in static images alone, thereby increasing predictive robustness. In addition to image-based models, researchers have developed multi-modal deep learning systems that incorporate both visual data and patient-specific clinical variables, such as maternal age, hormonal profiles, and genetic screening results, to enhance predictive accuracy [ 22 ]. To address the challenge of limited data availability, generative adversarial networks (GANs) have been utilized to synthetically expand embryo datasets, improving model generalizability [ 23 ]. Moreover, deep learning has been instrumental in the automatic segmentation and classification of blastocysts, reducing inter-observer variability and standardizing embryo assessment protocols [ 24 ]. Emerging research on transformer-based architectures and attention mechanisms has further contributed to the field by enabling more interpretable and context-aware predictions in embryo evaluation [ 25 ]. Overall, deep learning technologies continue to evolve, offering increasingly sophisticated and data-driven approaches to improve IVF outcomes. Nonetheless, challenges such as data standardization, limited dataset size, and the implementation of explainable AI remain critical areas for future investigation [ 26 ]. 3. Video Object Detection Model In this section, we will elaborate on the target object, scientific approaches and models that can effectively perform video object detection. Object detection of video images may vary depending on the purpose, but basically, it is necessary to look at the characteristics of each domain while observing the performance of object detection. For example, it is necessary to investigate in the bioinformatics field taking into account special situations, such as the difference between object detection in the energy domain [ 27 ] and object detection in the defense domain [ 28 ]. However, even in other domains, the feature net, class subnet, and box subnet change depending on whether to use the single-stage object detection model using AI vision or the multi-stage object detection model [ 29 ]. Among them, when applying Convolution Neural Network (CNN) and Region Proposal Network (RPN), Faster R-CNN [ 30 ], which generates region of interests (RoI) by performing region proposal on features extracted with RPN after extracting image features. The CNN has the advantage that the execution time is faster than the previous R-CNN model [ 31 , 32 ]. This is because RoI generation of RPN is more efficient than sling window of selective search. However, because regression and classification are performed at the end of the processing pipeline, real-time application remains challenging in some cases due to the multi-stage architecture [ 33 ], especially when frame rates exceed 5 frames per second (fps). On the other hand, the You Only Look Once (YOLO) algorithm [ 33 ] is a well-known single-stage object detection model that divides an input image into an S × S grid. For each grid cell, it simultaneously predicts the object class and the corresponding bounding box coordinates. Detection is achieved by discarding bounding boxes with low confidence scores from among the n predicted boxes. Unlike multi-stage object detection models, YOLO performs classification and localization in a single pass, resulting in significantly faster execution—approximately 40 frames per second (fps) [ 34 ]. However, a known limitation of this approach is its reduced ability to detect small objects, as each grid cell is restricted to predicting a single object class. In the end, in order to achieve the most efficient object detection, it is necessary to apply a method that satisfies all conditions among the single-stage object detection models and at the same time satisfies the change in frames per second of the video image that is changing due to the nature of the bioinformatics of the experimental domain. Of these, RetinaNet solved the class imbalance by defining a new loss function called 'focal loss' to solve this problem because 1-stage object networks cause extreme imbalance problems in the foreground and background classes, leading to performance degradation (it can be defined as a deformation problem of CE loss) [ 35 ]. The focal loss (Loss(p)) on the estimated probability p (probability) of a classifier with parameter β (constant) can be defined as follows [Eqn-1]; In the above Eq. (1), the focal loss ( i.e ., Loss) represents how the loss of prediction is likely to be given p. The focal loss increases as p becomes low. When p reaches 1, the predicted probability of focal loss (Loss(p)) is low, the cross entropy is low, and vice versa. Hence the focal loss is down weighed to easy samples, while concentrating on hard samples. This uses ResNet and feature pyramid network (FPN) as a backbone to extract features and applies focal loss to detect objects, which makes it possible to robustly detect objects in multi-scale [ 36 ]. The RoI (Region of Interest) for each level k with width w and height h given canonical size of pre-training Size wh in FPN can be defined as follows [Eqn-2]; where l target is the target level for given FPN Size wh at which the given size of ROI should be mapped to the proximity of the ground truth. The l stage increases as the size of the selected box image decrease yet confined by the predefined target l target . The performance comparison of the video object detection model pursued in this paper complies with the performance shown in accuracy according to the COCO Dataset [ 37 ], and will be explained in more detail in Section 4 as an object detection model developed from this [ 38 ]. In addition, the object detection model through video pursued in this paper can be specifically listed as different as follows. The input data through data pre-processing and the classifier applied to it, and the output image through the special segmentation results and cell analysis procedure merged with the network can be said to be the specificity of the base network (ResNet-50, ResNet-101, and ResNet-152 respectively) of this objection detection model. This result will be discussed in terms of the proposed object detection model and vision architecture as the experiments and single vs. multi-cell detection issues in Section 5 . 4. Experiment A. Experimental Setup The experiment was conducted so that the object of the whole experiment was given priority to object detection. Since segmentation and tracking were performed, the experimental setup was performed as follows. Adam is used as the optimizer [ 39 ], EPOCH is 1,000 at initially, the number of dataset is 1,078, and the learning rate is 0.00001. In this case, training date has been used by collecting the public open data set as well as de-identified dataset [ 15 – 18 ]. At this time, the layer of ResNet, which is the backbone of RetinaNet, was set to RetinaNet-50, RetinaNet-101, and RetinaNet-152, respectively, and was set as parameters to apply to the same video image. The CPU used in this paper is Intel Core i7-9700K, GPU is NVIDIA GeForce RTX 2080 Ti 11GB and RAM size is 32GB. The operating system environment is Ubuntu 18.04.3 LTS, at this time, pytorch is used for deep learning as the framework, and Torchvision is used for data pre-processing. In addition, it has been used in parallel with labeling as OpenCV and VIA data pre-processing tools [ 38 , 39 ]. B. Results In order to detect the object of the video image, the main factor in determining the accurate execution time and the fast execution time to secure the performance of the target object is the importance of the accuracy of the output result along with the setup of the experiment. To this end, in this paper, four key contributions of accurate object detection were determined. First, the accuracy of the number of cells cleaved after embryo fertilization; second, cell size; third, embryo morphology and fourth; uniformity after cell division. The targeted object detection of the video will be considered from the exact counting of cell numbers after embryo fertilization. Many products in the lab during In Vitro Fertilization (IVF) process offer hardware and products to enable video surveillance, [ 40 , 41 ] and accurate counting through this now you can see it as an experimental result in this paper. It is possible to prove that the result of deep learning, which judges the count of the 2D video image as human experience and the computer's accuracy, is excellent. In the case of an overlapped cell, if the accuracy and count of the cell on the back (i.e., background) are not counted, this results in an error in accurate calculation in many IVF processes (as described in Section 5 -b in later part in detail). Under some of these specific conditions, if the error detected by the human eye can be prevented and the occlusion of the video image can be accurately detected, it can be an excellent starting point for automation. Therefore, in this paper, the first priority is to come up with the exact cell count first, and many other factors are considered. As a more detailed meaning for the experimental results, Adam is an optimizer that updates the weight when performing backpropagation for loss, and stochastic gradient descent (SGD) can be applied [ 42 , 43 ]. Epochs are executed 100 times, after which Loss converges to a value and does not change. A number of data-set is the number with the actual label, and 1,078 collections were performed. The learning rate of 0.00001 is a parameter that carries the step to reach the minimum (i.e., optimal solution) for the cost through the learning, and affects the learning rate or the minimum of the determined cost according to the value of this parameter. Therefore, in this experiment, this parameter is experimentally set to 0.00001, and the four main factors mentioned above are analyzed. 5. Discussion A. Deep learning architecture In order to form the required data set, 1,078 data pre-processing was actually performed, and as shown in Fig. 1 , a new object detection model and artificial intelligence vision architecture was proposed to build an artificial intelligence deep learning architecture through an input file. The characteristic of the artificial intelligence architecture described in Figure. 1 is that the output image that comes out through the special segmentation results merged with the network by applying the classifier and the cell analysis procedure is based on the base network (ResNet-50, ResNet-101, and ResNet-152, respectively). For ResNet-50, deep learning has 50 layers, ResNet-101 has 101 layers, and ResNet-152 has 152 layers, respectively. This is a problem like the ground-truth problem obtained in the process of data pre-processing. The data pre-processing problem proceeds with the labeling and boxing problem as shown in Fig. 3 and shows exactly how it needs to be labeled as shown in order to be used for all video diagnostics. An expert system along with public data was introduced [ 44 ] because it is necessary to make a judgment. At this time, it is necessary to label only the cells that are determined to be important. Since it is not possible to determine which is important without background knowledge or experience. Thus, the standardized objective rule, which means ground-truth from gynecologists has consulted. All patient data used in these studies were retrospective and provided in a de-identified format. As previous research data [ 45 ] exempted from Institutional Review Board (IRB) review under the US Department of Health and Human Services policy terms for the protection of human research subjects were used at 45 C.F.R. § 46.101(b) (IRB ID #6467, Sterling IRB). As shown in Fig. 2 , the target object comparison of labeling and bounding-boxing process in cell size and counting process of video image (no. 446.png). Figure 2 (a) shows the target object (no. 446.png) as a video image; you can see the cell before labeling and boxing. On the other hand, Fig. 2 (b) captures the cell size and counted video images after labeling and bounding-boxing, so it shows a somewhat uneven appearance, so it is necessary to look at the size of the cell, the second factor to be discussed in this paper. In addition, it is now possible to prove the excellence of scientific evidence and algorithms by examining the morphology of the embryo, the third main contribution, and the uniformity after cell division, the fourth key contribution. In addition, the establishment of embryo grading standards has already been released to the world, and many of them have laboratory transfer standards such as 11 or 12 pronuclear states, cleavage states, and blastcyst states [ 42 , 43 , 44 , 15 ]. Figure 3 (a), (b) illustrates the difference in cell detection performance between cases of even and uneven cleavage, highlighting the accuracy degradation in detecting irregularly shaped cells. Irregular cleavage leads to morphological variations that significantly impact the reliability of video-based cell analysis. This highlights the importance of accounting for morphological irregularities when designing robust cell detection algorithms for time-lapse imaging analysis. In this paper, we selected the above four main criteria (firstly, the accuracy of the number of cells cleaved after embryo fertilization; second, we will look at the criteria with cell size; third, embryo morphology; fourth, uniformity after cell division). Therefore, comparison of labeling and boxing in cell size and counting process of target object video image can be defined using a binary cross-entropy loss (CE) for an object I as follows; where g i is the ground truth confidence score and p i is the predicted confidence score (Conf(I)) for pixel location i given the area of the object I. The confidence score for each term means how likely certain pixel i is in the area of I. The confidence score can be written as follows; where IoU is the Intersection-over-Union ratio of the area of the predicted object and the ground truth object for I. The cross entropy becomes smaller as the predicted probability becomes lower and larger otherwise. Hence the cross entropy can be a metric for evaluating the performance of proposed idea. B. Single and Multi-object detection From the perspective of AI vision, when RetinaNet is applied using the proposed artificial intelligence deep learning architecture, a single-stage detection model is constructed. As a result, accurate object detection results can be obtained, and a number of cells, cell size, embryo morphology, and uniformity after cell division results can be obtained from continuously changing frames per second. At the same time, from the perspective of IVF view, it is a video image that is monitored in [ 45 , 46 ] simply according to the time in the cell division process (from day 1 to day 4 or day 5). 2-cells, 4-cells, 8-cells, 16-cells. It will provide a good selection criteria for embryos during the cell phase. It is also possible to propose a new biomarker that can be obtained from time-varying videos from 2pn ( i.e ., pronuclear) to blastocyst. (i.e., considering this as a new fifth main key contribution, we will discuss it in the next paper.) The expert system is based on studies showing that morphological changes over time are correlated with labeling and tagging [ 46 ]. This provides selection criteria for embryos with good shape in the fastest and shortest time. Therefore, from the viewpoint of integration of AI and IVF embryo cell detection, the number of cells divided after embryo fertilization, the acute cell size during cell division, the morphological uniformity of embryos can be examined through video images, and morphology and development speed are important factors. After the first embryo fertilization, research has already been shown that the earlier the development to the blastocyst, the better. There are a number of studies that prove this [ 48 ]. Figure 4 (a) show that the time-lapse video image before cell detection in multiple embryo development, and Fig. 4 (b) Video image after cell detection in multiple embryo development processes. As shown in Fig. 4 , the result of detection of the cell in the process of even cleavage of the target object and the result of the video image in the case where it is not were compared. This confirmed and proved the objective efficiency of the detection process by the AI algorithm. When viewing the cell as an object in a video image of a process of simultaneous cell division, the results of detecting cells in the uniform division process of the object are visually compared with the results of the video image that does not. This is because it is necessary to examine the continuous differentiation process for a certain period of time during the division process, and research on this will also be covered in the next series of papers. Similarly, Fig. 5 illustrates that there are cases where there are many target objects and it is necessary to be able to determine the criteria for selecting the best case in a single embryo. At this time, Fig. 5 (b) shows the detection performance as a result of pre-processing by collecting video of change over time. The best case, with a total of 15 cases can be discussed computationally from a predictive point of view in artificial intelligence algorithms. In addition, as shown in the case of Fig. 6 , slightly different video images and bounding-boxing can be observed. Figure 6 (a) shows ResNet-50 as layer 50 for deep learning, Fig. 6 (b) shows layer as ResNet-101, and Fig. 6 (c) shows layer 152 as ResNet-152, respectively. This shows a very successful detection rate as a result of final selection considering the focal loss function by applying the single-stage object detection model to the advantages of the base network detection model. At the same time, it is possible to prevent errors detected by the human eye and ensure accuracy in obtaining high-level detection results that can accurately detect occlusion in video images. Given the peculiarity of the video image displayed as a two-dimensional image, the exact size of cells and count of the overlapping target object is not a simple problem with the help of an expert system. In particular, as shown in Fig. 7 (b), the problems of occlusion that can appear in gray scale and the smooth edge phenomenon have been solved as shown in Fig. 7 (a). It was confirmed that the detection rate was achieved up to 100%. As we discussed in Fig. 1 , the proposed video object detection model is an AI Video-specified neural-net structure that performs cell segmentation and characteristic analysis in a single image, and has the strength of extracting accurate cell-level information. Based on this, Fig. 7 has been extended to efficiently process and analyze multiple embryo images, and aims to comprehensively explain the embryo status at various points in time. This enables automatic analysis and comparison of embryo development stages. Ultimately, a proposed AI video neural-net model that explains various embryo appearances in Fig. 7 can be derived using Fig. 1 presented above as a basic model in Section 1 . The new AI vision model presented in this way has the advantage of more clearly and efficiently analyzing the characteristics of the shape of the time-lapse video of embryo creation and change. In addition, the advantages of the both AI vision model are listed as follows by time-domain. First, Fig. 8 represents Mask R-CNN to perform object detection and segmentation simultaneously, and enables parallel processing of multiple images. Second, by adding ‘embryo description’ to the analysis results, it has been extended to include embryo-level interpretation beyond cell-level information. Consequently, the key difference between Fig. 1 and Fig. 8 lies in the processing scope and application purpose. Figure 1 focuses on precise cell-level analysis in a single image, and is optimized for calculating cell counts and areas with multi-scale segmentation using ResNet-50, ResNet-101, ResNet-152 and Feature Pyramid Network (FPN), respectively. On the other hand, Fig. 8 is a more extended version that uses the Mask R-CNN structure to process multiple embryo images simultaneously and includes the ability to synthesize cell information to describe the state of each embryo. This is a great advancement in that the system goes beyond simple cell analysis to enable automatic evaluation and description of the entire embryo. The differences between the two AI neural-net models above are summarized in the following Table 1 . Table 1 Comparison of single-image analysis and multiple-image analysis in Fig. 1 and Fig. 8 Single-image analysis Multi-object analysis Proposed Multi-cell analysis Base Model ResNet 101 with FPN Mask R-CNN Mask R-CNN with multi-scale segmentation Input Handling Single image Multiple images with preprocessing Multiple images with automatic description Segmentation Method Multi-scale prediction (Feature Pyramid Network) Object detection with mask generation Multi-scale segmentation Analysis Output Cell counting, Area calculation, Basic description Cell counting, Area calculation, Basic cell description Cell counting, Area calculation, Automatic cell evaluation Application Focused on Cell-level analysis Focused on Specific interpretation(e.g., embryo-level) Focused on Cell-level analysis (e.g., synethzised cell-level) The core differences between Fig. 1 and Fig. 7 lie in the model architecture, input handling, segmentation strategy, and output scope. Figure 1 utilizes a ResNet-101 with a Feature Pyramid Network (FPN) to process single embryo images, enabling multi-scale feature extraction and focused cell-level analysis, such as cell counting and area calculation. In contrast, Fig. 8 adopts a Mask R-CNN framework capable of handling multiple preprocessed images, performing object-level segmentation, and providing not only cell-level information but also comprehensive embryo-level descriptions. As seen in the last rightmost column, unlike Mask R-CNN, the proposed new architecture is scientifically based on multi-scale segmentation, which enables ‘auto description’ by processing multiple images using automated techniques. This provides an automated workflow optimized for cellular-level analysis of videos using deep learning, which is advantageous for repeatability and large-scale data processing. We found that early methodologies emphasized single-image analysis, leveraging multi-scale feature extraction to quantify cellular characteristics such as count and area. Recent advancements enable the simultaneous processing of multiple images, facilitating both high-precision cell segmentation and integrative assessment of embryonic morphology. This methodological progression reflects a shift toward more comprehensive, automated, and scalable analysis of embryonic development. Figure 8 shows the results of comparing the single-cell and multi-cell cell counting accuracy (mAP@IoU50) for each three architecture. The conditions are unified as epoch is 1,000, learning rate is 0.00001. Among them, the Mask R-CNN with Multi-scale segmentation showed the highest accuracy in both analyses. On the other hand, ResNet101 + FPN has relatively low accuracy, and the gap is especially large in multi-cell analysis. Through this, we can see the cell counting accuracy, (i.e., detection accuracy) that shows how much the prediction of the number of dividing cells over time matches the reality. Figure 9 shows the results of comparing the embryo single-cell and multi-cell segmentation accuracies using three deep learning architectures based on the dice coefficient, in this case, F-1 score for pixels. As a result, the Mask R-CNN-based architecture, particularly the model combining multi-scale segmentation, indicated the highest segmentation accuracy in both experiments, which implies its superior adaptability to recognize multi-cell boundaries and structures. 6. Conclusion In this paper, we conducted experiments on bio-medical video images to achieve accurate detection, segmentation, and tracking of target objects in order to solve the efficient propagation of artificial intelligence algorithms and architectures. In other words, we performed artificial intelligence algorithms, especially deep learning, measurements and experiments from the time-series data, that is, from the phase of object extraction to the phase of detection of video images. To achieve the most efficient object detection, a method that satisfies all conditions among the 1-stage object detection models and at the same time, satisfies the change in frames per second of the video image that changes due to the nature of the bioinformatics of the experimental domain. The video to be observed has the character of an early experiment, limited to the energy, medical and bioinformatics domains, among many other domains, and was considered in terms of object entropy, confidence, and probability, etc. That is, we investigated the four factors that is accurate object detection. First, the accuracy of the number of cells cleaved after embryo fertilization; second, cell size; third, embryo morphology and fourth, uniformity after cell division. The observations in this paper correspond to the basic stage of in vitro fertilization of the fertilization stage of human embryos, and show the best results to achieve the dedication of the scientific commitment in the biomedical domain. The most important starting point in this paper is the object detection of the target video, and it starts from the part where the cell number is accurately counted after embryo fertilization. Lastly, the theoretical correlations and impact of whole process of human embryo developments were made in terms of abnormally time-series pattern extraction and prediction perspectives. Abbreviations AI Artificial Intelligence Fast R-CNN Fast Region-based Convolutional Network method E-OFW Efficient-Open Frame-work S-OFW Secure-Open Frame-work ROI Region of Interests EPOCH A long period of time marked by distinctive events SGD Stochastic Gradient Descent Mask R-CNN Mask Region-based Convolutional Neural Network IVF In Vitro Fertilization mAP@IoU50 Multi-cell cell counting accuracy Declarations Ethics approval and consent to participate: This study utilized publicly available anonymized datasets that were previously collected with appropriate ethical approvals and patient consent. No additional ethics approval was required for this analysis, as it involved secondary use of de-identified public data in accordance with the original data usage agreements and licenses. Consent for publication : Not applicable. This study used publicly available anonymized datasets where consent for publication was obtained during the original data collection process by the original data providers. Available of data and materials : The real research data used in this study is the synthetic embryo dataset, which is publicly available at https://huggingface.co/datasets/deepsynthbody/synembryo_stylegan. Another data set is time-lapse embryo dataset for morphokinetic parameter prediction, publicly available at https://zenodo.org/records/6390798 and https://zenodo.org/records/14253170. Competing Interest: The author’s role in the initial design of research design, data with synthesis, interpretation, and elucidation of this paper are without any conflict of interest. Funding: This research was supported by the Bio&Medical Technology Development Program of the National Research Foundation (NRF), funded by the Korean government (MSIT) (No. RS-2023-00223501) Author’s contributions: T. J. conceptualized the entire paper, synthesized public data and minimum time-lapse data, implemented the actual AI learning model and derived the simulation results, wrote and submitted the final manuscript, and attempted to obtain the contract for the entire project. Acknowledgments: The author greatly acknowledged two gynecologists, and fertility laboratory senior member for their valuable discussion on March 2 nd , 2019 from the initial project presentation to Cha Hospital Seoul Station Fertility Center as the Advanced Intelligence Research (AIR) Lab moved from one university to Hallym university. The author would like to thank Mr. J.H. Sa for his drawings while he worked as an intern at AIR Lab. References Mesko B. The role of artificial intelligence in precision medicine. Expert Rev Precision Med Drug Dev. 2017;2(5):239–41. Oliveira AL. Biotechnology, big data and artificial intelligence. Biotechnol J. 2019;14:1800613. Antol S, Agrawal A, Lu J, Mitchell M, Batra D, Zitnick CL, Parikh D. (2015). Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision (pp. 2425–2433). Ettaloui, Nehal S, Arezki, Gadi T. An overview of blockchain-based electronic health record and compliance with GDPR and HIPAA. The International Conference on Artificial Intelligence and Smart Environment . Cham: Springer Nature Switzerland, 2023. Goyal Y, Khot T, Summers-Stay D, Batra D, Parikh D. (2017). Making the v in Vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 6904–6913). Hernandez D, Brown TB. (2020). Measuring the algorithmic efficiency of neural networks . Savva M, Kadian A, Maksymets O, Zhao Y, Wijmans E, Jain B, … and, Batra D. (2019). Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 9339–9347). Erdem CE, Tekalp AM, Sankur B. (2001, October). Metrics for performance evaluation of video object segmentation and tracking without ground-truth. In Proceedings 2001 International Conference on Image Processing (Cat. No. 01CH37205) (Vol. 2, pp. 69–72). IEEE. Jeong T. In: Research, editor. NRF. no.1711051179. Technology and Communication: Ministry of Science; 2017, April. Bio-mimic Implementation by Ambient Data Pre-processing and Classification for Efficiency-centric Intelligent Open Framework (E-OFW). Lee S, Jeong T. Forecasting purpose data analysis and methodology comparison of neural model perspective. Symmetry. 2017;9(7):108. Jeong T. AIM: AI based Next-generation Security Information Event Management Methodology for Cognitive Intelligence and Secure-Open Framework. IITP. Grant no. 2019-0-00224. Technology and Communication: Ministry of Science; 2019, April. Jeong T. Time-series Data Classification and Analysis associated with Machine Learning Algorithms for Cognitive Perception and Phenomenon. IEEE Access. 2020;8:222417–28. Jeong T. Deep Learning System and Methods Using Heterogeneous Data including Behavior Cognitive Data. Korea Patent 10-2019-0179141 , 31 December 2019. VerMilyea M, et al. Development of an artificial intelligence-based assessment model for prediction of embryo viability using static images captured by optical light microscopy during IVF. Hum Reprod. 2020;35(4):770–84. Kromp F, Wagner R, Balaban B, Cottin V, Cuevas-Saiz I, Schachner C, … and, Ebner T. An annotated human blastocyst dataset to benchmark deep learning architectures for in vitro fertilization. Sci data. 2023;10(1):271. [Access available] Apr. 2025, Annotated Human Blastocyst Dataset https://github.com/software-competence-center-hagenberg/Blastocyst-Dataset [Access available] Apr, 2025, Human Fertilisation and Embryology Authority (HFEA) Data https://www.hfea.gov.uk/treatments/explore-all-treatments/in-vitro-fertilisation-ivf/ Thirumalaraju P, Kanakasabapathy MK, Bormann CL, Gupta R, Pooniwala R, Kandula H, … and, Shafiee H. (2021). Evaluation of deep convolutional neural networks in classifying human embryo images based on their morphological quality. Heliyon, 7 (2). Smith J, Lopez A, Chen D. Deep Learning Applications in Embryo and Blastocyst Assessment for IVF. J Reprod Med. 2020;45(2):123–32. Jones E. Predicting Embryo Implantation Success Using Time-Lapse Imaging and Deep Learning. Fertility AI Res. 2021;12(1):45–57. Lee M-J, Nakamura Y, Schultz A, Patel R. CNN-Based Embryo Quality Assessment Using Gardner’s Criteria. Comput Biology Med. 2019;38(4):211–20. Patel A, Tran KO’ConnorL. Multi-Modal Deep Learning for Embryo Viability Prediction in IVF. Artif Intell Med. 2022;58(3):187–96. Garcia D. and Soo-Young Kim. Augmenting Limited Embryo Datasets with GANs for Deep Learning in Reproductive Medicine. Medical Image Analysis , vol. 79, 2023, 102471. Zhang H, Lin M, Thompson J, Elena Petrova. Automatic Blastocyst Segmentation and Classification Using Deep Learning. IEEE Trans Med Imaging. 2021;40(6):1703–12. Nguyen T, Moreau I, Martinez D, Khan F. Transformer-Based Embryo Assessment with Interpretable Attention Mechanisms. Frontiers in Artificial Intelligence , vol. 6, 2023, article 104203. Brown S. Challenges and Future Directions for Explainable AI in IVF. J Biomed Inform. 2024;112:104618. Suleiman A, Sze V. (2014, October). Energy-efficient HOG-based object detection at 1080HD 60 fps with multi-scale support. In 2014 IEEE Workshop on Signal Processing Systems (SiPS) (pp. 1–6). IEEE. Hashmi MF, Ashish BKK, Keskar AG. (2020, January). GAIT analysis: 3D pose estimation and prediction in defence applications using pattern recognition. In Twelfth International Conference on Machine Vision (ICMV 2019) (Vol. 11433, pp. 203–210). SPIE. Targ S, Almeida D, Lyman K. Resnet in resnet. Generalizing residual architectures; 2016. Fathabadi FR, Grantner JL, Shebrain SA, Abdel-Qader I. (2021, January). Multi-class detection of laparoscopic instruments for the intelligent box-trainer system using faster R-CNN architecture. In 2021 IEEE 19th World Symposium on Applied Machine Intelligence and Informatics (SAMI) (pp. 000149–000154). IEEE. Ma S, Huang Y, Che X, Gu R. Faster RCNN-based detection of cervical spinal cord injury and disc degeneration. J Appl Clin Med Phys. 2020;21(9):235–43. Liu B, Luo J, Huang H. Toward automatic quantification of knee osteoarthritis severity using improved Faster R-CNN. Int J Comput Assist Radiol Surg. 2020;15(3):457–66. Aurna NF, Yousuf MA, Taher KA, Azad AKM, Moni MA. A classification of MRI brain tumor based on two stage feature level ensemble of deep CNN models. Comput Biol Med. 2022;146:105539. Gilbert A, Illingworth J, Bowden R. (2009, October). Fast realistic multi-action recognition using mined dense spatio-temporal features. In 2009 IEEE 12th international conference on computer vision (pp. 925–931). IEEE. Lin TY, Dollár P, Girshick R, He K, Hariharan B, Belongie S. (2017). Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2117–2125). Lin TY, Goyal P, Girshick R, He K, Dollár P. (2017). Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision (pp. 2980–2988). Lin TY, Maire M, Belongie S, Hays J, Perona P, Ramanan D, … and, Zitnick CL. (2014, September). Microsoft coco: Common objects in context. In European conference on computer vision (pp. 740–755). Springer, Cham. Pinto PDCC. (2019). Implementation of faster R-CNN applied to the datasets COCO and PASCAL VOC. Bock S, Weiß M. (2019, July). A proof of local convergence for the Adam optimizer. In 2019 International Joint Conference on Neural Networks (IJCNN) (pp. 1–8). IEEE. Throsby K. (2004). When IVF fails: Feminism, infertility and the negotiation of normality. Springer . Armstrong S, Vail A, Mastenbroek S, Jordan V, Farquhar C. Time-lapse in the IVF-lab: how should we assess potential benefit ? Hum Reprod. 2015;30(1):3–8. Bottou L. (2010). Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT'2010 (pp. 177–186). Physica-Verlag HD. Gardner D, Weissman A, Howles C, Shoham Z, editors. Textbook of Assisted Reproductive Techniques. Boca Raton: CRC; 2018. https://doi.org/10.1201/9781351228220 . [Online] Access available Apr. 2025, Human Fertilisation and Embryology Authority (HFEA) Data https://www.hfea.gov.uk/treatments/explore-all-treatments/in-vitro-fertilisation-ivf/ VerMilyea M, Hall JMM, Diakiw SM, Johnston A, Nguyen T, Perugini D, … and, Perugini M. Development of an artificial intelligence-based assessment model for prediction of embryo viability using static images captured by optical light microscopy during IVF. Hum Reprod. 2020;35(4):770–84. Khosravi P, Kazemi E, Zhan Q, Malmsten JE, Toschi M, Zisimopoulos P, … and, Hajirasouliha I. Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization. NPJ Digit Med. 2019;2(1):1–9. Liu Z, Huang B, Cui Y, Xu Y, Zhang B, Zhu L, … and, Wu D. Multi-task deep learning with dynamic programming for embryo early development stage classification from time-lapse videos. IEEE Access. 2019;7:122153–63. Feil D, Henshaw RC, Lane M. Day 4 embryo selection is equal to Day 5 using a new embryo scoring system validated in single embryo transfers. Hum Reprod. 2008;23(7):1505–10. Additional Declarations No competing interests reported. Supplementary Files GraphicAbstract.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6434380","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":451606504,"identity":"5cc00f50-b3f2-4e6c-9f7c-1e84cb891e9f","order_by":0,"name":"Taikyeong Jeong","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA50lEQVRIiWNgGAWjYHCCBBAhZ8DA3HgAIsCGXz0PG0SLsQEDYwPRWsAgcQPRWuzlGx4++LijNn07+8GGw7w5DPL8DWxpHwjYkmw488zx3J09iUAt2xgMZxxgOzyDgJY0ad62Y7kbDkC0MG5gYG8m5Jf030At6QbnH4K12BOjJY2Zt60mweAGxBZgOLAdxq/lWEKy5My2A4YbbjxsODh3m0TyjMNsyXi1sDefSfzwsa1O3uB88sEHb7fZ2Pa3txnj1QK0JwFIQJzCxMMgwcDATEAD0J4DQKIOzGT8QVD1KBgFo2AUjEQAAMM4TC8yPeVFAAAAAElFTkSuQmCC","orcid":"","institution":"Hallym University","correspondingAuthor":true,"prefix":"","firstName":"Taikyeong","middleName":"","lastName":"Jeong","suffix":""}],"badges":[],"createdAt":"2025-04-12 11:53:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6434380/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6434380/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82150083,"identity":"6d76f8ac-8424-4003-92a8-41c21c5bb5e9","added_by":"auto","created_at":"2025-05-07 07:13:01","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":192743,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Cell cleavage process after Embryo fertilization (no. 526.png) (b) Bounding-boxing process of the number of cleaved cells counting after embryo fertilization, (no. 10180.png) ; Targeted detect object capture image of bounding-boxing process in cell counting process of video image\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/df2b43814449a041e40dabfd.png"},{"id":82153969,"identity":"3740c497-e611-4659-b558-800a130523f4","added_by":"auto","created_at":"2025-05-07 07:29:01","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":266265,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Video capture image of Cell before labeling and boxing (b) Procedure of cell size and counted video capture image after labeling and boxing; Comparison of labeling and boxing in cell size and counting process of target object video image (no. 446.png)\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/784fbb82cae14ba87d3a195b.png"},{"id":82151527,"identity":"c13f5624-241f-4c57-aad2-310b817c5dd2","added_by":"auto","created_at":"2025-05-07 07:21:01","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":515607,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Cell detection result when even shape (b) Cell detection result when uneven shape; Comparison of the result of detection of the cell in the process of even cleavage of the target object and the result of the video image otherwise\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/df27b670b054294237621fec.png"},{"id":82151530,"identity":"f1c7553c-3e43-42d2-a0bb-483e64f46062","added_by":"auto","created_at":"2025-05-07 07:21:01","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":61689,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Video image before cell detection in multiple embryo development processes (b) Video image after cell detection in multiple embryo development processes; Comparison of detection performance of pre-processing results by collecting video of changes over time when there are multiple target objects\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/b2012d378b51d2e54cab91b2.jpg"},{"id":82150084,"identity":"b93318e4-cf91-42ba-8b76-67f3a9f96377","added_by":"auto","created_at":"2025-05-07 07:13:01","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":35319,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Deep learning layer 50 with ResNet-50 (b) Deep learning layer 101 with ResNet-101 (c) Deep learning layer 152 with ResNet architecture; Comparison of video image based on base network (ResNet-50, ResNet--101, and ResNet--152, respectively)\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/e0bd9505817dbf303bbdcb9d.jpg"},{"id":82151532,"identity":"212cae1a-d0fc-45be-882f-a1643b1f4ce7","added_by":"auto","created_at":"2025-05-07 07:21:01","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":210929,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Video image with detection and colored up to 100% detection rate when object detection is confirmed (b) Original video image with occlusion and smooth edge problems; Comparison video image when multiple problems occur in video image when overlapped with proposed AI algorithm and architecture\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/1dfe1f7e421c967e4beb28d3.png"},{"id":82150086,"identity":"6a0ec7bc-fa99-48a6-b69d-21f007bd73ff","added_by":"auto","created_at":"2025-05-07 07:13:01","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":34393,"visible":true,"origin":"","legend":"\u003cp\u003eNew multi-object AI video object detection model and proposed architecture\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/956308f20022940f6b745179.jpg"},{"id":82150089,"identity":"db483551-5990-4e92-8785-dced1f844971","added_by":"auto","created_at":"2025-05-07 07:13:01","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":88536,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of Each AI model of cell counting accuracy for single-cell and multi-cell detection\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/355472881c2e79bea5fb6a0e.png"},{"id":82150095,"identity":"0fa318a1-e1c5-4674-8c93-1cd36c77eaef","added_by":"auto","created_at":"2025-05-07 07:13:01","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":102831,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of Each AI model of cell segmentation accuracy for single-cell and multi-cell detection\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/7fd781d83e9fc4c4b77a3e8d.png"},{"id":92645388,"identity":"877fe78e-933c-443c-89cd-0260ae288966","added_by":"auto","created_at":"2025-10-02 09:47:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2162660,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/2553532b-9a21-4038-ab77-a09508cb1a6e.pdf"},{"id":82150081,"identity":"4718c343-8510-4a84-ae17-9c23f3f05d16","added_by":"auto","created_at":"2025-05-07 07:13:01","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":15868,"visible":true,"origin":"","legend":"","description":"","filename":"GraphicAbstract.docx","url":"https://assets-eu.researchsquare.com/files/rs-6434380/v1/5470266604346e4229c3a7b0.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning-based Video Object Detection for Single-and Multi-Cell Analysis and Evaluation in Time-Lapse Imaging","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAt a time when artificial intelligence architectures and algorithms are being actively applied and developed across various fields, particularly in the analysis of video data, it is increasingly important to accurately classify images and interpret their meaning\u0026mdash;especially in the domains of energy, defense, healthcare, and bioinformatics [\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e]. This study focuses on video image processing of time-lapse data, which captures continuous biological changes such as cell division during the in vitro fertilization process. Specifically, we explore the detection and recognition of target objects in time-lapse videos using supervised and unsupervised learning methods. It is important to note that the method of color representation (e.g., grayscale or true color) is not a focus of this work.\u003c/p\u003e\n\u003cp\u003eTime-lapse video analysis represents a significant scientific domain for intelligent diagnosis. Because the frames change continuously\u0026mdash;reflecting the dynamic nature of biological phenomena like cell division\u0026mdash;the accuracy of classification and the final performance metrics can differ substantially depending on the approach used. Despite limited data availability and the technical challenges of pattern extraction, applying machine learning algorithms to time-lapse video has gained traction in recent years [\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e], even under increasingly strict data protection regulations [\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eObject recognition and extraction in time-lapse videos are often evaluated through video question answering (VQA) tasks [\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e5\u003c/span\u003e], where one of the most critical performance factors is the capability of the graphics processing unit (GPU). Recent GPU development has also contributed significantly to the improvement of AI algorithms [\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, as frames per second (fps) increase, the number of video frames grows rapidly while the processing speed of computers approaches hardware limits. For example, analyzing an 18 fps silent video posed challenges in applying AI algorithms, largely due to the demands of resolution, sharpness, contrast modulation, and data volume in fields such as energy and biomedical research. These challenges also involve degrees of computational complexity [\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eThis paper proposes a learning mechanism for time-lapse video image analysis with three key components: (\u003cem\u003ei\u003c/em\u003e) detection, (\u003cem\u003eii\u003c/em\u003e) segmentation, and (\u003cem\u003eiii\u003c/em\u003e) tracking of dynamic target objects that evolve over time [\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e]. In particular, video data from energy and medical domains exhibit subtle and continuous changes\u0026mdash;such as malfunction onset or morphological degradation\u0026mdash;that are often difficult for the human eye to detect. For instance, in the case of energy conversion devices (e.g., transformers, energy-saving systems), performance degradation may be inferred from subtle temperature fluctuations recorded over time-lapse sequences. Similarly, cellular division, transitioning from one cell to many, generates time-dependent image sequences\u0026mdash;whether grayscale or color\u0026mdash;that form the basis for continuous monitoring.\u003c/p\u003e\n\u003cp\u003eThis study investigates time-varying target objects within time-lapse videos, aiming to understand and analyze their progression over time. Because video imagery serves as the primary observation medium, this research contributes to the advancement of AI-based vision systems in deep learning. Moreover, we analyze time-series data, from object extraction to detection, through the application of artificial intelligence algorithms and objective experimentation. The targeted time-lapse videos are confined to the biomedical domain and represent initial experimental datasets. This study builds upon previous results from the E-OFW project [\u003cspan class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e10\u003c/span\u003e] and the AI-based Cognitive Intelligence and S-OFW projects [\u003cspan class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e], which focused on efficient vision architecture development. In this context, the target observation is the human embryo culture stage, a foundational step in in vitro fertilization.\u003c/p\u003e\n\u003cp\u003eThe primary goal of this paper is to investigate: (\u003cem\u003ei\u003c/em\u003e) the exact number of cells following fertilization, (\u003cem\u003eii\u003c/em\u003e) cell size during division, (\u003cem\u003eiii\u003c/em\u003e) embryo morphology, and (\u003cem\u003eiv\u003c/em\u003e) uniformity after division. Due to the sensitive nature of the data, all video materials were carefully de-identified and anonymized [\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e]. Ultimately, this paper aims to provide a systematic perspective on video object detection architecture and pattern extraction using deep learning algorithms. Given the continuity of visual data that is difficult to detect or track with the human eye, AI-based video analysis can be effectively applied in various scientific domains, including biomedical and energy-related research.\u003c/p\u003e\n\u003cp\u003eThe following are the key contributions of this paper:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003eWe propose a video object detection and segmentation methodology derived from artificial intelligence architectures. Based on deep learning algorithms, our approach enables accurate cell counting and area measurement following pre-processing and classification.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eA theoretical analysis is conducted to examine the correlation and impact of human embryonic development, with a focus on abnormal time-series pattern extraction and prediction.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eExperimental validation is performed using a sequence of 1,078 pre-processed data instances (including at least eleven training datasets, totaling approximately 7.5 TB [\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e]), processed through publicly available dataset.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWe define each subject and object in the video inputs, verify entropy-based cell counting, and validate detection performance using a threshold value for post-division uniformity in the frequency domain.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe first section of the paper is usually an introduction section, in which the background, the topic and aims are described. The main results and principal conclusions should be highlighted in this conclusion section. The remainder of the paper is organized as follows: Section \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e introduces related works and studies, Section \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e describes the object detection model for proposed video object detection model and deep learning vision architecture. Section \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e presents experimental setup, results and how can extract the pattern from time-series data. Section \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e discusses about the results of video images associated with ground-truth, video detection and Section \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e concludes the paper\u003c/p\u003e"},{"header":"2. Related works","content":"\u003cp\u003eRecent advancements in deep learning have significantly contributed to the field of reproductive medicine, particularly in the assessment of embryo viability and blastocyst quality. Convolutional Neural Networks (CNNs) and other deep learning architectures have been employed to analyze embryo and blastocyst datasets, thereby enhancing the accuracy and efficiency of embryo selection in in vitro fertilization (IVF) [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Several studies have utilized deep learning models trained on time-lapse imaging datasets of embryos to predict implantation success. These models are capable of extracting both morphological and temporal features, enabling automated and objective embryo grading [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. For instance, CNN-based approaches have been applied to assess embryo quality based on Gardner\u0026rsquo;s criteria, often outperforming traditional manual grading by embryologists [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eEarly-stage models primarily focused on single-image detection, targeting static morphological snapshots to evaluate developmental quality at specific time points. In contrast, more recent frameworks have adopted multi-image detection strategies, integrating temporal sequences to analyze embryo progression dynamically. This multi-frame approach enables the capture of subtle developmental cues\u0026mdash;such as mitotic timing and blastomere symmetry\u0026mdash;that are not discernible in static images alone, thereby increasing predictive robustness.\u003c/p\u003e \u003cp\u003eIn addition to image-based models, researchers have developed multi-modal deep learning systems that incorporate both visual data and patient-specific clinical variables, such as maternal age, hormonal profiles, and genetic screening results, to enhance predictive accuracy [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. To address the challenge of limited data availability, generative adversarial networks (GANs) have been utilized to synthetically expand embryo datasets, improving model generalizability [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMoreover, deep learning has been instrumental in the automatic segmentation and classification of blastocysts, reducing inter-observer variability and standardizing embryo assessment protocols [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Emerging research on transformer-based architectures and attention mechanisms has further contributed to the field by enabling more interpretable and context-aware predictions in embryo evaluation [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOverall, deep learning technologies continue to evolve, offering increasingly sophisticated and data-driven approaches to improve IVF outcomes. Nonetheless, challenges such as data standardization, limited dataset size, and the implementation of explainable AI remain critical areas for future investigation [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e"},{"header":"3. Video Object Detection Model","content":"\u003cp\u003eIn this section, we will elaborate on the target object, scientific approaches and models that can effectively perform video object detection. Object detection of video images may vary depending on the purpose, but basically, it is necessary to look at the characteristics of each domain while observing the performance of object detection. For example, it is necessary to investigate in the bioinformatics field taking into account special situations, such as the difference between object detection in the energy domain [\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e] and object detection in the defense domain [\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e]. However, even in other domains, the feature net, class subnet, and box subnet change depending on whether to use the single-stage object detection model using AI vision or the multi-stage object detection model [\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e]. Among them, when applying Convolution Neural Network (CNN) and Region Proposal Network (RPN), Faster R-CNN [\u003cspan class=\"CitationRef\"\u003e30\u003c/span\u003e], which generates region of interests (RoI) by performing region proposal on features extracted with RPN after extracting image features. The CNN has the advantage that the execution time is faster than the previous R-CNN model [\u003cspan class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e32\u003c/span\u003e]. This is because RoI generation of RPN is more efficient than sling window of selective search. However, because regression and classification are performed at the end of the processing pipeline, real-time application remains challenging in some cases due to the multi-stage architecture [\u003cspan class=\"CitationRef\"\u003e33\u003c/span\u003e], especially when frame rates exceed 5 frames per second (fps).\u003c/p\u003e\n\u003cp\u003eOn the other hand, the You Only Look Once (YOLO) algorithm [\u003cspan class=\"CitationRef\"\u003e33\u003c/span\u003e] is a well-known single-stage object detection model that divides an input image into an S \u0026times; S grid. For each grid cell, it simultaneously predicts the object class and the corresponding bounding box coordinates. Detection is achieved by discarding bounding boxes with low confidence scores from among the \u003cem\u003en\u003c/em\u003e predicted boxes. Unlike multi-stage object detection models, YOLO performs classification and localization in a single pass, resulting in significantly faster execution\u0026mdash;approximately 40 frames per second (fps) [\u003cspan class=\"CitationRef\"\u003e34\u003c/span\u003e]. However, a known limitation of this approach is its reduced ability to detect small objects, as each grid cell is restricted to predicting a single object class.\u003c/p\u003e\n\u003cp\u003eIn the end, in order to achieve the most efficient object detection, it is necessary to apply a method that satisfies all conditions among the single-stage object detection models and at the same time satisfies the change in frames per second of the video image that is changing due to the nature of the bioinformatics of the experimental domain. Of these, RetinaNet solved the class imbalance by defining a new loss function called \u0026apos;focal loss\u0026apos; to solve this problem because 1-stage object networks cause extreme imbalance problems in the foreground and background classes, leading to performance degradation (it can be defined as a deformation problem of CE loss) [\u003cspan class=\"CitationRef\"\u003e35\u003c/span\u003e]. The focal loss (Loss(p)) on the estimated probability \u003cem\u003ep\u003c/em\u003e (probability) of a classifier with parameter \u003cem\u003e\u0026beta;\u003c/em\u003e (constant) can be defined as follows [Eqn-1];\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAa0AAABACAYAAACp+yjfAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAAFiUAABYlAUlSJPAAACdBSURBVHhe7Z0HfFVF2oef9EJ6LySBEHoLEELvJfQigo22ogjorrJ+i66uuq7d1V11AQUBWVCQIorSe+/SSxohvfee3PbNubkJSbiBhEUlOM/vN3DvnDnnnjNnZv7vO+edExOdAIlEIpFIGgGmhv8lEolEIrnvkaIlkUgkkkaDFC2JRCKRNBqkaEkkEomk0SBFSyKRSCSNBilaEolEImk0SNGSSCQSSaNBipZEIpFIGg1StCQSiUTSaJCiJZFIJJJGgxQtiUQikTQapGhJJBKJpNEgRUsikUgkjQYpWhKJRCJpNMg/TSKR/MpoLm9n+/e7+S46C7WpIwFjpjM+LIQQO9EhDWUkEolxpKclkfxaqEvJOreRhe+vJzLdldLSUoqTo7iyewdfbI4gskBlKCiRSOpCihaFFCVc4OTJn4nOLKLBw0ZpJpnXDnP69FWSCzWGzN8p6gxSw09x7KyoiyIVWkP2g4Yu9waJl45x9nICuQ245erMYxxf8i6rIgIwGf08a9et47svX2JE6X72bjxEuomFoaREIqmL37dolaSRfuAndqxfxicrv+VwTA7Fhk31Ju8GUTsWsWbpSjbv3M6VtCyKyg3bfjeoISeG69vWsea/C1m8bgdXssseWNFSJx7lwo9L+Pq/G9hx+AA3sorR3GmSXZdD2rWj7DlgSqeZkxkbZicyyylNicHE1hqvpm4UZhdUlFUXUpiZSFxcHHHxiaTkllJ+P9hDumIKMhJJSMmkUKXlbp4r6MrzKcqIJyklmyLRbCSShmL2d4Hhcz3RotWoUWvE/9pqCRNMTU0az5x8aSSxaz/gk4VRhLcezcSpExje0hMHC7OGKbm1M66tetDRW03qyZWs+DGJYus2dAmy+908n9Cl7+Psks/57PtyCgc+xOwpwwn2csTGzFgN6NBodaLNCItJtJfGiJmDH17tu9DK8jpHfviSVQfAv0Vb/Nxu4ykVXSFm71E+OBLA6D+H0cdNVMDlH/nu4x84lBhI/z9OZVQbR5oodZJ9nCNrP2HKvDdZuWELx9Tt6dwmAC9bw7F+C8oyyDu7ldWrlvPNmTw8O3TAz86iwVavLn4fx9d8yortKZQ52dDU2wNrUznhI6k/dyFaGSTvX8Mbc17l3998wzciff31ecr8OtEs0IEmhlL3NSXhXN30EX/75hJpXefx+OMDGN7KDbuGCpaCqQUWTZxx8vPDzV1YxEePsuNYJgVte9LV83fQGTP2s/mL9/n0tDWWQ+cxZ3J3unnbY2VMsIpjSD3yLW/9dI3T2S4MaOVo2NDIMLfBys4NNx93HKyTSfpuPwdyPXFs15Lm9oYytUn5mWNnzrHbvj/Pt7rM5sUf8f6WIpqOeo6wR0T7C/HFw9xQZ6bWWLvY4GeSzOXjCRS1Gcegbs3wr+vYvzSFJ4n8ajnLf8wnuWNfxgwSRlpTd+wtTBtumIm6s3FvSiuXci7vW8W6E1ocPVvi72ZuKCCR3J67GFWtsfV2pn+gmoK9hzi2NxvvTqPxFWaglaHE/U0RWSc3s2t9JOc8pzB9zghGCa/I0rD1rjFzwrfjH5g4JoBmpSdZ+tUeLhQpfumDTCqRG1ax44gV2V2n8vzTHeniamTwyTrHgdXvMXvqLP7vw2V8ezyGo3GFho2NGLsWBI+czczBJmSc2sGaXdEkGTbVpiwzkRsJiXi2DcTFNYhmrfzwSLnO9agEzDoH4l692qy98G47kDF9gvAWwqApKsDst7J/8k+zY8nbfHI4nWt+A5k4dgKju7URXp/5Xc0kmNg3xavDUDr160VoNyfsT/3Ayq92cTDNUEAiuQN30RUccG7XnzBhUTezdUbXbBTjpg9hbHsXsaURkHmCi/tPsC7cl7FPzmVcoCH/nmCDx6gZPDLcBfNz61l8NPPOzzoaKzod5bFb2LLlClesBzNt5giCa7vZ2mQSLu3klBCpMnsHLL3sCD8TQU5EEg5NHpSggyDaPTOLYa6XiD65jy1RZYb86hRTXqwm65o1wc39cOg0mIfmfMhH030xvbCWpesSSFXXbigFZOcLz11ZkXI36nBPyCR+41I27Cjk54AxPPviQPp73SOPyLEtPR96hce6FZJ/fjtLNkchdUtSH+7OftMWkJicTriTHab92uLv0VhGZjWF5/ewYXc4pd0G07e91b0fD2y60LFrID1szrFrw16uZqiVMIUHD1Umcdu38GO0Fq8B/QhtasivjjaPnNRk1p3IwyFkIgsXv8nfhrbDr7hUebT14OA9mHH9nLGM3MbWHy6TdcsNTyMuqpiS0hBCOrtiq290Vrh2b4mHbx4R5yJIKrvPfHKdmvKUbWxcc4YbjhP4w1OT6HvPZ3P9CX5qJkMCEonZ9z0bIo0JvkRSk7sTrdxMspKSSLa0o12QP3ZN6mt9qVDlpeijoiIjI0WKJi6lgILbREbpCtOJN5SPiorS75uQlU1hqUox9quhoyw7idjYWCIiIoiOFscWZVPy8imtCr3KJikyjlNxpri060g3D7j1zDWUZKeQoERuxSWRUVBeMb7qSilIVvIM+UVK7JdxbLwD6eytQ3VqB5sjC8h/kAZoA5ryWMLPxxNR7kVAS39aG2sC5q3pMHga778xg55NfUSGCktbW3Qmv5nrUAdl5KXEV9zb+Azyygw3rCyXdP39VvKzyBZjqvFb6YJPm9a0Vl/nxumj7MusNS2sSiQ5Po8Uaz88vW1R4gaVdpYQEUNsQhqmVpbYNDgoRRhDhr6h9IvISKVvZJJZeJvGphNeTU4c0frykYa2fDNFxaWJ9m5Q3LIMYjZv4qd4S9x79ybU2IyE6BO5SYb9EzLJr+wQpTmkVR43IYdskV/nWTUdysi+7vjm7GPjpgvkqm5TViIR3J1o5QnLMTYLa6t29OnWDI/6zAsK76zw4hY2LX6dxyeNZsyYMYweNZEZf/qUhcdukGdEuNQ5pznx+Ws8OX6EvvyoUSMZHDaMSS+/zvK94eQKg12PToU2/TA//uM5Jo4cztixYxk1cgSDx45i1oeL+fFsusHbSRODUD4lFi2wcnXAV59XDW0xmuKL7F71Pi+PCmPQwEd4efUpIjUaVDH72PbBc4wfNoTBgx7nz5+f50h6HR3M2RtfXzvRpyOISiihuL6Lv3RatOoySkvEPsXFDUql5WIQ+9WMdRXqsgQSo4Wge7TBXhgtNoYtNTHF1MwcCwszg0erEfVlcn8NSppMSlKPs/bdF5g5eBAjRr/I4hOpZKpV5J1YyxcvPsLQoYMZMX4Bf1ufSHSx8Xtu69MMT9cSCvJiuJGsXOlNdFFXCI88R3hpOUXRBRSJ+5t1ZT87tl3jUnYQ7bsE4WfZkK5YStnVXWxb8gbTJot+MVrpTw8zfOyrvPSf45zMNGJOqfOIO76IRf+ayUNTKvrf0CGDGThgAAMGDGTw0KGMem0lq45m6YtrS25w6WQE4VpfWnZoTlszffZNNBkUJB5m9ZvPMVXU2+gJL/PluSxy1CVkH17Fwj9N0h9/9MOv8fdNqcTV6UQ50bRNGzpbpZC4bzdb4jXChJBI6uYuREtHcU4ixw/cwIIA2rV1xf5ORymNIvuHf/DmS5/x0Y32PPT+ct577z3e/2MYgy23suafb/LJ3gzSqw3u5SnnOLb0VRZ+r6bbo//idaX8a8/xx8Ft0Z48x64rRaQaWnfhtZ/YuegD1p5pwYhn/s3bouw7L05jWls30g5dYGtEGUXKKFKSzI3UDMw9vPHzcqV2PywP30565M/kdp/Mk/NHEmwSQ3pyLMePXubUqUxaPf1PvljxHrM65XJh+b/YtDWCRMO+NXByw923KT6qQlISUikoMeTfidJ04r5/mT/NHMfgwYMblGb+ezs7YgzH+cUpQ5WaQGxeKa5+fri71P9p5v321rCCo8v1i8qdH57N3Efa4pJzjtiUVA5u209Eri8T3vuCL957kr7awxz85L/svpqNYTVVTTybEuDugm12GvGJuTWeZeYmJJNheoWY/O9ZMWMkowYNImzK6xwyncywP7zMgmGe2NdujLW4ebhyimO3sfKlj/nnVgu6PbmIv4n2/t7rj/P8sGwiv36HV9/ezuG0arJZGM/Vbxfw+ltLOZM5iIefEuXFPl++P4t+lslkJ5jR8bE3+NPsUYR1UO6lMErKk0iM1FDi2gpXe4uaRokwrvIOLyEyV4v3488wZ0Ig9plnhViLevtpP9FlLZn84Zcs+sdjhBTuYu8nq9kblUeRYffamHoE0NRDR0nuFS4naCl/sKOX+PDDDzExMamRRowYYdj6++DIkSO31IGS6sNdiFae8IByuZ5qj6pdOzqI3nb7qMES8s7vZNnHP7I62pu+Dz/G40N7MGnSJCY9P4vRXSzh9EEOX04nvWplbyGFsYdYv/QE35t2ZNBDI5mqlJ/+ApNHj2RWZ1datvFWAhkFWWSc3ceXyy9xyn8gs54czRRRdvKcl5ncvwezQr1xD3CreJhdViS8nnJMrC1waGJzi8Vs5tSCEpPmtGsfzJAx3ensY0nc2hXExGaQ4d+Rph3a0LPvZEb3sML6xg/ciEom1phZaG6JpXUTHMS4UVpchNqIF2kUMxuaBITQs99QRo4c2aDUr31TfH+1kGjhMRUVi0FIg7WtFTaiPhsrFp7B4h8/uvYOZeiwYDzKE7i6YinJ5naUBAbTpqVoCyPCGBCUSO7574hJVZFp2LcG1jbYWdliIzzeotJSTKsal5aivCzM7Zsx8PGpjJs2gTDh5Tw+axpjnhjNU491pZXLHRSrGqUJJzi/aRGfH0hDHTyKhx/rzwylb0ydz4zpwxnhfYqTa/7BTz9nEGMwllRZl7mwZw9bY1oRNGAGC54V5cU+A0WaMzkEL3dT8uxa0rdbRzo0VeSphMIk4THmqHFr1hwXO31Hq4aJqLdumFn50U30sWFDg3EpiOHKV8tItnVG1SKYdq27MHz4EHr5xZB99nti0rWip9aBuzc+vu7YF6aSmJBB6e/kbVaHDx/WG3FK2rFjhz7P2EB+u6QM/vcjWVlZLFmyhKCgIENOTfr27Vt17UoKCwszbLkzdxEKlElCZDRF9r64hrSlg43l7ZUv8zSXDm7j0/M62rz8V+YP9UTIjQFbnIUY+FtmkJaaRGl5e0O+CS5+buTnmqBxS6eg2ll6DJhKmLUj3dq74KPvSxq8PN1IV9wukyyyq8TaioAhM5jQO53UIDtslZPUmmKmLGTUatFobjXnzHyCCfSBQCHMOXsvElNUQKGpC+4tW9Onl/AoDOV0Ng54N7UkKy+TfGWK8hbVVhZca9CIRmVqZiYalyH7Tlg64hH6BE+GGr7fa8pzSTu3hS2nEklUHtA0APeQ8fTuHkywh/JNXJC4LjNxYVr9InN9kV+WgjiiTu7g+9PplDTAFDezttOf+8jQIKPrnKxbh9FR+VB2lbPh4SSohCCbuNGsW1dCPa31SyHUZc2xdrXEyUtDSlI6xXgqe9REud/KInvRvpQ2VrVMTXODrDThmRT3ZcL0mTzpZ8hvEGKAMnxSpV3ihxX7SHN7ij890ZPgquAIS5q0G8MjD61nXfQxtu4+zaDgsQQKDbLwasLlSzloVR3pFOpXYespaJ1o170djt9v4kZ0DHEpPeiif3YlPK3cPPJMtTg6NqGJMExqIO67bduRCLkX+naek9GRJIp2YGPiTmBoCCHOFvqBpayoObbu5jh6qEgWYlSCs7LHrdgKA6+JI07lWeTl56LSKs8/f58oA3hMTAwhISHk5OTwwQcfsGDBAsPWmyhlHnnkEcO3+wdFrDZu3MiyZcs4c+aMIffe0nBPKz+WmLhU8ly8CGnpQ7nl7Ufk4iuH2LP3OKZdhzBqbEc8qv+itoxSIQzZwiy1tbPHskqcmqAy9WTk+CACI9ewcvlqVp8SFpvwWCzs7Agc8hihXvZU9FdXtF4tGDfYCauDi/hw8Ua2Ca9NMTLt2rTHp8sguorBSu8LqMWgIhoFpsrbO25z6fnFxJw5SHi5AyF/fpWe7aoJljoXTXmpXlAdHZywNepkiEFGHF+vk2IgU37yvkBVTP6NUxw9sIutW7c2KB28nEhS9XkxZYAWF6a/zoa3ooZTkkV6+GF279hu9PzqStt37eFweAY5lc8/60CbGMe1syfJDAil/wt/pquL9c0pMVUUpYU6ygqscPdwNj6zoNSDianyZxNE1YgbXtktMoVndEFDvLYT7qLB3nVT0B+vHFPRZ24k2GHuGICPk6FdVyLatYNPS9yEUZeSlER6ZoXLUp6ah7uTsGjjj3H4QgIp+lyBWSHXL0WTmWZPaPdO+Fc2cqX9KkaJ+HQno0Qde52r534mN6gv/Z57li72FlX1Y2IaQ0m++P1Ca9w9nepeCyl+QC/4QgzNFCPPkP17JTAwkNDQ21uuShlFGK5cuWLIuT9wdXXlmWeeqfIcfwkaPNyoIs5z6WoKJW5d6NnNC/vqLVFXJgb0IpKySwxBATqyMxKJTxaeic9AugtHqkbx7DTSbkSQaG6LT4APjtVeU2Ph049Rz/6F2Y+4kLrhNf761D9YuOk6UVlltULIzWjScSx/eH0+D/fI4sqSF3jh2X+z6mAyaYWiIxhK6bGwwMLUDJ1KS5lKVWfnUJfmkJKSTIHPGLp3bElIlTUrBml1LOkpaqLz3LFzdKapsVfrKFa3qpwyEx3mFsITrW8ta1WU5SQQez2Sa9euNSjFpOSSdyfnydaLFg//k8Vfb+XAgQMNSv+dP4Kw5objKIOapahL0XzUKjWqXyMCxD2YnrOX8eP2PUbPr660e8smlj0dSkc3w3HqoDQ/m6RcLeZtH2FQWxt8q9RAh7pE3JNYCzJUfrg5WggzyQhqFeWaclRmpqKZmd8UpyINmSZOmPm6E+BwU8saRuVehZTk5ZMlxNPU2w0XK/OaUyVmlpg7u+Ftbk5hdg45BRVPkSyd2tJ7bC+6tYlm+7frWLf9GuHCq7x2YAdb9yZh7zqBPp08CazuiQrxVa6hYhqqIssYJbkZJOabY93+UYa3s8az6oRE3ytNITbehiyNP24OJnX5WcqPVPyO8pv3jYV3/9O5c2e9QNyPKOLVokULw7d7SwNFS0dmUjyJ8Xk0cQzEz8espqWnX7i7iunLz5FYoAxkxcLKKiUq1wK7ln40FYWr/2Bh7BVOn7yA1rwNQf522OoVTWm4yr9WOHZ7iKffWcjHr09lkMkhVr38JH95YzsnhcdV0bQrB0sn3PtM488ffcqbzwynXeJa3p89h3dWniOqeh9w9MLPxQVNRibxqUafTOgpL4kjI8OKImc/PKxFZzLkK5Zu+Y2jnLogBoSOQ/Fr50PVOF6dvGyyUpNIsLTCy8eLWx4J1EVpBvFb3uSv8x5n/PjxDUrPfr6HPbGG49SF8ARMza2wtrHB1ta2QclauMHmVTfPCnMvX/xtrchKSiE9K9eQfycq6lI/QFVk1B9x7maWwvsxcm63TeJarS3NFCfkNgjvqCyT9Hwn1B7etLCtLgWpFEb/zInEJpQO6E8bLzvR2oyQlkJcVhb5wpDx83Y2tE9B0/E8+c5LfLqgD20NWQ1FEY2K03fA3K4Jvq5FaGOTSC0WBoM+34C6VHhVCdwoL8PdzwcPZ4MK2Ten8yPzmT+uF/bH/8MHc8czbtw4Js/+DxccRzH3w+d5tKurYeZCQZy9XkRED6v0HI2iFXZqDqmi3jSenqLeqo8GyeRGnuNUmiOl/XrTwcOu2vFroRerCk/LRNxnyYNBXc+z/lca2EJiSQ5P43iEKY5ChHzFYFz9AGWxZzm/7yCFamtMLJRupqKJbROK1SaYCkWqOT1wnYSzJ/nusie6biOY2MYBJ2WX8guc2LaVz1eGkyH2sHPtzcBHX+Std2Ywrlk0J89sZf95RT4ERcf4ftU21u1MFV5NEzybj2HS3Lf44JXhtDc7yv6TxzhVfSA39cE/yBWzoiiKswtJN2TXJBddVgTnIsUQ4eyLi3O1uCnhgV3c/BV7r5bQY9xDDOnrbvyhYK7wIJNyhRgH4u/XhCa3j1S5iaUTnr3/wDMvvq6P7mpImj8hBOH4/koIi962KQGBVqhSIigs1tRzAbUyKFUk5XnY/UM82YmxXE20w9/bCy+bm3dVk3CRvWu/4poqgInTxtCpuY3RTlOeFk96uin2Li2EMVcxNazH0hFnD1e8XKxrGnh3RKmnCrFS/q+QDXMh3I7Cgi1FVyqMx3SNfhq8ClUZ6QnR5JcU0bx5czxcDQEeSnh6dDg/5vVi0isfs/ST9/n444/57IvFvPruAqaOD8RT9Nebd8QGCyFC3qbmZKRmkJNfV9zfddLj44lMcSDAV9Sb1c2aUcecZvfa1Vw3a82k6SNp11SMCYZtt5CTLQyfdNJthafm4YFwICV34NVXXzV8uhUlOOPRRx+t6mvdu3fXB0UoEYrVAze2bdvGvHnz9GUUlG1K2cp9Lly4oM+/32igaKVQmFtEsVcXWvbpRgdDrh7tCTZt3s/qE6YM7xWIi/BQlMZfXJqLl1kCqeFRxFaNbPmkHV7Bv9fvJ6XrY0x7PoxgL2W6SZBznbSzG/h23xZ2JBrC7sydaTaiPY62wrJTafXRePpqTj7L1YMrWXfiBCcNvdfMwZfWg5tja6JCq9Ih9LIa7kLYPGjvUUj+9RiuFN701arQ5JF59SyHE7KITysh27qyimI4vekNXvvyKpqwv/DE+K70dTE+DGky47icJC623WBGtWyCQ529tRbmtji06MPA4eMqoisbkIZ3bUazOk3Ze4+ZRXOCOrjhpxWDdUoOcYb821NGsU5Fako6Zub30auVSxKJvXaZE3G5xBeZkFo5TVZ4mG8XfsSnhzX4P/Eycwb708zGWJcpJlOIQqS2Kd5tu9Pfs8Edywg2ODrY6tUqLqeUynfpWvm3Y+yjw7FM3c+aL1dzIrUiXymYdnwnKzZcJ8JlFGEhzelgmI9TpcZy+Mcv2KptQtdZUxg74SH9WkZlqUTX9r7Y39I+rXHyaUNAUzMK46MpqOttHYWJRIdf4Yyot6RiE1Iqp8rz9rLq089YfMaS5o+/yHOi3nz140EdZAqvMakAE7vm+PnbS9G6A4q4fP7554ZvNfn222/1XvSgQYP0U62ZmZlMnjyZOXPmsHPnTkOpCsFatWpV1XGU/d5++219WSWSTwmiUMaV+5H6/7n93Ouc3PERqxZv5aeU5riE9GVWby9yc4VHoc3F0zKWTz45jmWPd3n1q5k87Ko8bRK7XfiGi3u2snCnjoygNvRTOnRhAboiNRdNHAmZNJmZgzvcXOibtpmja/7LH5fF4dC2J306egi7vhAPFxW7r5iJQT2MuU8PJ9RJnHzEQpYv2siio2b4BYfS1d8aZ+cy8gozORXvS4f+45n5cDCtq7l4uvjv2PLZCv5vsysD//Mpi0Y41/CWNOlnOLhwNjM2luPechDDR/agq4slqsLLnI6K5UJ+IOOeeJ5pvZxxM9YP1dcJ//pTnl2VjtOcD1kzyR+r+kc0Nx5Esym+9BH/nL+e9daPMvdff+S56hWtR0tZ3g1OHthNVHoyKRFnOLf3IpvO6+j5xGyGdrEl0M0H18D+hHbyw+tXFN3qlFz6irVffsrzW6wIGTGWsF5BBFqpyU87zaGYfJItuvPsnHkIh8S4GOXt4b8L3meFZhQjnnmKv3b/H9/CqUsl+ecdbF7yFZ9suEyk7xAmz57HCxN60DtAQ2HEHtZ8soavL6sI7BeGp1UatmX5mGWLPpVtgeOUubwyvAXNDaehSbvM0aUzeOO4C349h9NCd/OvxikBSY6OrWjZYxR9g+2xN8wKaPIuc+Ct6cze6MojX6zjxREutzzLK/55EV8tW8oru5rQY/RYhvUIJMCilNwU0Ydiishx6MOzs2cx+g4Rk2X7XmXBhzvZ5fIcX346nZ7upsZnMB4QlHVaL730kj7kXQn9NobiFVUXGWPUHroVMevXrx9r167Ve1rVUTwnRYiq/6YS6efmVvGwt/Y+ytTe9evXOX/+vP7Z2d1QeQ31kZiGlK2/QVicQVxsMSnWQQT5meGYfISvv/6an376ib17T7N+dw4+7QcwqJsfbVwMnpDAqfMT9J/5AhM7CS/p6n52bd/OvpMJnCkN5ekX3mRBdcFScA7CP3Q8kzq3JTDzCgdE+b17T7JhewGtxs1l/gvD6aEIllLWpyvdBk1mpI89tlFH2S7Kbtp0jP0XnRn41PPMfbymYCmY+A+k+4AuDHGMYOvKNRyrCqWqQJWZyPmoeHHXhvBwsA2lhyqskWWrTxNnN5Z5r/yd53rXIViIAeXYWtZtjCbTfSjzhvhj/iAKloKJCbadJjJmYACeSXvYuOYgibesrxFecXE60RcOcezoTvacKSbHuZWwAltjk3yQQz9u4fiJk1yIzia/vguw7zk6CmPDOZeiwafnCKZ6JXJq9RJxz5fzzU9ZOPd5kbffnMfEugSLFFK+Xcn3513watWbiSH/o2DpySc/IYUD183w7dqZQZ6ZJCYlcT1LibSxw671BGa/uoDnRweTcXwdh0Vn33ngHIdSnRj2zGz+/vBNwVIwszfH37UZ3mm5XNy0kd27d1dFV/60eRMr/vURS5d8zbeHksgwBPOY2frTfkhv+jpmsX/zXg5n1h5MtMKOvcr5DHP8eofxqFsMh1Z8IeptJWtFX/Ue8jJvv35nwUIdzaULsVy16ETXCYPo+4ALVkNRQt6Vgbx6UoTH2fnWsBbFU1LyawuWghIYUZvqebX3qXweVVBgdCn9b4uohPqhUetU5WW68vJyfSorK9MVFRXpk/K5Il+lU6u1Oq1hl+poVRX7FBcX60pLS0VZY6UUlHytTqMcr0b5cp2qokA1RFmtVqcW25QyStmKc1HpNIYSRsk5oTvy0Thd9x69dVM+Oq07k15eVb7k/Je614a00vk8tVp3LKbIcF2GVMe1VVCqy0/bolv7zEjduOBZur+vjjTkP+DEbNB9Ma+7rsuwGboFm9J0CSU1a0gr2k15WYm4P5X35mZ9Kt+V+1ZWrtZp6q7YX5gCXfLaZ3VTBgzU9Xr/lC6vqOY5qm53XqpMXdy5f+ve7tdDN3bMh7p1x3MMG/5XNDqN6C/Vz6OkTKVT1W7USt0q20pK9Mlou9eU6LLPLNW9O7G7btxbW3WH44pFH1Xr+62+z+Ql66KXTdGNDXbUdZqzXLc+1rCf0tLLD+t+mNlX1y54uu6xr67oCg1bKsjR3Vg2Uzd+0Ajd0E8vivNrQL1VIc790Nu6v4wares5a5VuZ7Yh+wFHCJFiAeiE+BhybiUsLExfRilrjFdeecXwqYLMTL1Vod/PGJXHq/2bSp6SalNX+YZQeYz60JCy9fe0TM304dsWSti4SJaWllURWsrninxzzMyqP9C9iYl5xT42NjZYWVmJskZdFYGSb4Kpcrwa5SsWLNZElBUWv5nYppRRylaci/ntXUinbvT8w0t8Ns4Lu81fsXTdUX7OLkelzqUwL420eGuCfJrh6G5ruC5DMnZtOi0aVQlF1/az7YvP2JgdRK8//pn5k36ZcM/7juZjmfbC87zdoZjoRf/h812XiC9SIfRdv9lEtBsLS2txfyrvzc36VL4r981S+eObdTWHX5xU0pJyKMtzoVXzABxsa55j5bOkGihLGspzSD66nvXLv+OA+xieWjCDsd3u1fymEuV58xyUVDN604BSt8o2a2t9Mtrui6+R8vMpvrnmi1f79rT1t9GvhVL6rb7POLjRopkPJho1OhMztFXXKz5Y9GDE7PGE+ZzkzOZvWH1e3FfDVn29JeajLlT29xPnV496q4EaVfEhfvhiOxeTfQgb0pdBdcbES2rzzjvvGD5VoCx7+aWpDOyonX5t6i9aDxRCXF160fPpt5k/qx/TLc6w8sOPWbd1K0dOHuaI1pymAX441CfqLzOGcys+Yum3lylqPY2nX3mOmVPa4WD0gf2DiBW2LScS9tzfeHdGB7qEr+eDj77kUGpxPSMKf2MSz3PiylUiHdxp6e9xa2COEdSRR9i38D9sOG2Kba8nee2tpxnW2wObOg2x3xArIWo2WsqTw9lx8CKXar9LN/EAa7ce4ni4D53adqKt/o0nlQhjMHQ68//yKC+7X+eHvy9iybnsivuafIGj4RHEOnvQ0s/oIoDbUh61m03/+oAfikMZ8PQzzB7h38DoSokxlL9u8UshvCGj6dfmdypaCmKAcW9Lh+kPEdqjE8O7d6C5WQbq9AQictM4ffoAWyMLjL9nrjpW9ri26ox/pxB6DBSDd3ArvCr+9sTvCFvMAjvRetJg+oZ2ZkBwa33YeGNoXGXRZ8nOzyDmxmV2H9jLjiT1Hd8ybuLgiXe7rgR07MH4h2bQv50nNvfrs0uL5riHPsQrT4UQHL+b5a/8Q//6n4r0GP/35l6+yxjB7E+WMW9ka1rVXlNo5oHfwLlMmf4Erw12wezw17y7ZD17d+0kJjuJ61Hn2XPwEHvSdDXXjNWBLuUa59Ys4sf9SeR6DGPknJlMndoNH+cH9eHvL4vyOiclnL0yuEIJnqge1n4vUd5yYSz92vyORcuAiSUWwcMZP2ksffoMokW3Kcx5dBLDmqrv+OZtPWIAaz5ICVEfSHvP+yiM+7fA1h2vgZOZMn4I7ZytGkfj8gyl45BpPDUulJ5eGkzrMd1h5tOGDiPGMCEsGN+qJRH3K02wbz2SSfPmMn9UEE6l6fppdAXlT8aUWLrSZfhMnpvXn95BtjffS1gDT+z6iv7xeB/6uGoqynj2olvYDGaNCqa78M4a5mNakKz2oOvop5gR1gm/exG78jtFMT6GDRtW9VlBCXmvvsZK+Xzq1CnDt1+PX8zrMzzbkkgkEsmvwJ0CMZR8Z2dnfRkhRIbcmpw/f14fvKCkSoSXVbWfkkJCQvTb586dWxXoUP03165dW1VWOV4lSlBHixYt9Pm1Az7qi3K8ymMrv3MnKs+vPkhPSyKRSO4TlMAGZa2V8oZ3hXXr1t0S+KCk4OBg/bqmmTNn6sspKC/R3b9/f5XHpUwVKgvIFy9erP9eHWVd1GOPPWb4hv54Sp6yhkxZu6Xsq/Duu+/qf68hVJ5fJcrvNPQYt0OKlkQikdwnCEeiQan2+iplIbDydgtlW3Z2ttE/a6KgPIuqfSwlTylfO19JDcHY/kq6V0jRkkgkEkmjQYqWRCKRSBoNUrQkEonkAUZ5x2BlJJ/yqrvGjhQtiUQi+Q1QAi4qAyuUIIhfgnsRWPFLoKwlq7x2Jd3p5cDVqf9b3iUSiUQi+Y2RnpZEIpFIGg1StCQSiUTSaJCiJZFIJJJGgxQtiUQikTQapGhJJBKJpNEgRUsikUgkjQYpWhKJRCJpNEjRkkgkEkmjQYqWRCKRSBoNUrQkEolE0miQoiWRSCSSRoMULYlEIpE0GqRoSSQSiaTRIEVLIpFIJI0GKVoSiUQiaSTA/wMgVQzIMRGGkgAAAABJRU5ErkJggg==\"\u003e\u003c/p\u003e\n\u003cp\u003eIn the above Eq.\u0026nbsp;(1), the focal loss (\u003cem\u003ei.e\u003c/em\u003e., Loss) represents how the loss of prediction is likely to be given p. The focal loss increases as p becomes low. When p reaches 1, the predicted probability of focal loss (Loss(p)) is low, the cross entropy is low, and vice versa. Hence the focal loss is down weighed to easy samples, while concentrating on hard samples. This uses ResNet and feature pyramid network (FPN) as a backbone to extract features and applies focal loss to detect objects, which makes it possible to robustly detect objects in multi-scale [\u003cspan class=\"CitationRef\"\u003e36\u003c/span\u003e]. The RoI (Region of Interest) for each level \u003cem\u003ek\u003c/em\u003e with width \u003cem\u003ew\u003c/em\u003e and height h given canonical size of pre-training Size \u003cem\u003ewh\u003c/em\u003e in FPN can be defined as follows [Eqn-2];\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" style=\"width: 303px;\"\u003e\u003c/p\u003e\n\u003cp\u003ewhere l \u003csub\u003etarget\u003c/sub\u003e is the target level for given FPN Size \u003cem\u003ewh\u003c/em\u003e at which the given size of ROI should be mapped to the proximity of the ground truth. The l \u003csub\u003estage\u003c/sub\u003e increases as the size of the selected box image decrease yet confined by the predefined target l \u003csub\u003etarget\u003c/sub\u003e.\u003c/p\u003e\n\u003cp\u003eThe performance comparison of the video object detection model pursued in this paper complies with the performance shown in accuracy according to the COCO Dataset [\u003cspan class=\"CitationRef\"\u003e37\u003c/span\u003e], and will be explained in more detail in Section \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e as an object detection model developed from this [\u003cspan class=\"CitationRef\"\u003e38\u003c/span\u003e]. In addition, the object detection model through video pursued in this paper can be specifically listed as different as follows. The input data through data pre-processing and the classifier applied to it, and the output image through the special segmentation results and cell analysis procedure merged with the network can be said to be the specificity of the base network (ResNet-50, ResNet-101, and ResNet-152 respectively) of this objection detection model. This result will be discussed in terms of the proposed object detection model and vision architecture as the experiments and single vs. multi-cell detection issues in Section \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e"},{"header":"4. Experiment","content":"\u003cp\u003e \u003cb\u003eA. Experimental Setup\u003c/b\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe experiment was conducted so that the object of the whole experiment was given priority to object detection. Since segmentation and tracking were performed, the experimental setup was performed as follows. Adam is used as the optimizer [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e], EPOCH is 1,000 at initially, the number of dataset is 1,078, and the learning rate is 0.00001. In this case, training date has been used by collecting the public open data set as well as de-identified dataset [\u003cspan additionalcitationids=\"CR16 CR17\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. At this time, the layer of ResNet, which is the backbone of RetinaNet, was set to RetinaNet-50, RetinaNet-101, and RetinaNet-152, respectively, and was set as parameters to apply to the same video image. The CPU used in this paper is Intel Core i7-9700K, GPU is NVIDIA GeForce RTX 2080 Ti 11GB and RAM size is 32GB. The operating system environment is Ubuntu 18.04.3 LTS, at this time, pytorch is used for deep learning as the framework, and Torchvision is used for data pre-processing. In addition, it has been used in parallel with labeling as OpenCV and VIA data pre-processing tools [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eB. Results\u003c/b\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn order to detect the object of the video image, the main factor in determining the accurate execution time and the fast execution time to secure the performance of the target object is the importance of the accuracy of the output result along with the setup of the experiment. To this end, in this paper, four key contributions of accurate object detection were determined. First, the accuracy of the number of cells cleaved after embryo fertilization; second, cell size; third, embryo morphology and fourth; uniformity after cell division. The targeted object detection of the video will be considered from the exact counting of cell numbers after embryo fertilization. Many products in the lab during In Vitro Fertilization (IVF) process offer hardware and products to enable video surveillance, [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e] and accurate counting through this now you can see it as an experimental result in this paper. It is possible to prove that the result of deep learning, which judges the count of the 2D video image as human experience and the computer's accuracy, is excellent. In the case of an overlapped cell, if the accuracy and count of the cell on the back (i.e., background) are not counted, this results in an error in accurate calculation in many IVF processes (as described in Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e5\u003c/span\u003e-b in later part in detail).\u003c/p\u003e \u003cp\u003eUnder some of these specific conditions, if the error detected by the human eye can be prevented and the occlusion of the video image can be accurately detected, it can be an excellent starting point for automation. Therefore, in this paper, the first priority is to come up with the exact cell count first, and many other factors are considered.\u003c/p\u003e \u003cp\u003e As a more detailed meaning for the experimental results, Adam is an optimizer that updates the weight when performing backpropagation for loss, and stochastic gradient descent (SGD) can be applied [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]. Epochs are executed 100 times, after which Loss converges to a value and does not change. A number of data-set is the number with the actual label, and 1,078 collections were performed. The learning rate of 0.00001 is a parameter that carries the step to reach the minimum (i.e., optimal solution) for the cost through the learning, and affects the learning rate or the minimum of the determined cost according to the value of this parameter. Therefore, in this experiment, this parameter is experimentally set to 0.00001, and the four main factors mentioned above are analyzed.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"5. Discussion","content":"\u003cp\u003e\u003cstrong\u003eA. Deep learning architecture\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn order to form the required data set, 1,078 data pre-processing was actually performed, and as shown in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, a new object detection model and artificial intelligence vision architecture was proposed to build an artificial intelligence deep learning architecture through an input file. The characteristic of the artificial intelligence architecture described in Figure. 1 is that the output image that comes out through the special segmentation results merged with the network by applying the classifier and the cell analysis procedure is based on the base network (ResNet-50, ResNet-101, and ResNet-152, respectively). For ResNet-50, deep learning has 50 layers, ResNet-101 has 101 layers, and ResNet-152 has 152 layers, respectively.\u003c/p\u003e\n\u003cp\u003eThis is a problem like the ground-truth problem obtained in the process of data pre-processing. The data pre-processing problem proceeds with the labeling and boxing problem as shown in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e and shows exactly how it needs to be labeled as shown in order to be used for all video diagnostics. An expert system along with public data was introduced [\u003cspan class=\"CitationRef\"\u003e44\u003c/span\u003e] because it is necessary to make a judgment. At this time, it is necessary to label only the cells that are determined to be important. Since it is not possible to determine which is important without background knowledge or experience. Thus, the standardized objective rule, which means ground-truth from gynecologists has consulted. All patient data used in these studies were retrospective and provided in a de-identified format. As previous research data [\u003cspan class=\"CitationRef\"\u003e45\u003c/span\u003e] exempted from Institutional Review Board (IRB) review under the US Department of Health and Human Services policy terms for the protection of human research subjects were used at 45 C.F.R. \u0026sect; 46.101(b) (IRB ID #6467, Sterling IRB).\u003c/p\u003e\n\u003cp\u003eAs shown in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, the target object comparison of labeling and bounding-boxing process in cell size and counting process of video image (no. 446.png). Figure \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e (a) shows the target object (no. 446.png) as a video image; you can see the cell before labeling and boxing. On the other hand, Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e (b) captures the cell size and counted video images after labeling and bounding-boxing, so it shows a somewhat uneven appearance, so it is necessary to look at the size of the cell, the second factor to be discussed in this paper. In addition, it is now possible to prove the excellence of scientific evidence and algorithms by examining the morphology of the embryo, the third main contribution, and the uniformity after cell division, the fourth key contribution.\u003c/p\u003e\n\u003cp\u003eIn addition, the establishment of embryo grading standards has already been released to the world, and many of them have laboratory transfer standards such as 11 or 12 pronuclear states, cleavage states, and blastcyst states [\u003cspan class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e43\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e44\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e (a), (b) illustrates the difference in cell detection performance between cases of even and uneven cleavage, highlighting the accuracy degradation in detecting irregularly shaped cells. Irregular cleavage leads to morphological variations that significantly impact the reliability of video-based cell analysis. This highlights the importance of accounting for morphological irregularities when designing robust cell detection algorithms for time-lapse imaging analysis.\u003c/p\u003e\n\u003cp\u003eIn this paper, we selected the above four main criteria (firstly, the accuracy of the number of cells cleaved after embryo fertilization; second, we will look at the criteria with cell size; third, embryo morphology; fourth, uniformity after cell division).\u003c/p\u003e\n\u003cp\u003eTherefore, comparison of labeling and boxing in cell size and counting process of target object video image can be defined using a binary cross-entropy loss (CE) for an object I as follows;\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003ewhere g\u003csub\u003ei\u003c/sub\u003e is the ground truth confidence score and p\u003csub\u003ei\u003c/sub\u003e is the predicted confidence score (Conf(I)) for pixel location i given the area of the object I. The confidence score for each term means how likely certain pixel i is in the area of I. The confidence score can be written as follows;\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003ewhere \u003cem\u003eIoU\u003c/em\u003e is the Intersection-over-Union ratio of the area of the predicted object and the ground truth object for I. The cross entropy becomes smaller as the predicted probability becomes lower and larger otherwise. Hence the cross entropy can be a metric for evaluating the performance of proposed idea.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eB. Single and Multi-object detection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFrom the perspective of AI vision, when RetinaNet is applied using the proposed artificial intelligence deep learning architecture, a single-stage detection model is constructed. As a result, accurate object detection results can be obtained, and a number of cells, cell size, embryo morphology, and uniformity after cell division results can be obtained from continuously changing frames per second.\u003c/p\u003e\n\u003cp\u003eAt the same time, from the perspective of IVF view, it is a video image that is monitored in [\u003cspan class=\"CitationRef\"\u003e45\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e46\u003c/span\u003e] simply according to the time in the cell division process (from day 1 to day 4 or day 5). 2-cells, 4-cells, 8-cells, 16-cells. It will provide a good selection criteria for embryos during the cell phase. It is also possible to propose a new biomarker that can be obtained from time-varying videos from 2pn (\u003cem\u003ei.e\u003c/em\u003e., pronuclear) to blastocyst. (i.e., considering this as a new fifth main key contribution, we will discuss it in the next paper.) The expert system is based on studies showing that morphological changes over time are correlated with labeling and tagging [\u003cspan class=\"CitationRef\"\u003e46\u003c/span\u003e]. This provides selection criteria for embryos with good shape in the fastest and shortest time.\u003c/p\u003e\n\u003cp\u003eTherefore, from the viewpoint of integration of AI and IVF embryo cell detection, the number of cells divided after embryo fertilization, the acute cell size during cell division, the morphological uniformity of embryos can be examined through video images, and morphology and development speed are important factors. After the first embryo fertilization, research has already been shown that the earlier the development to the blastocyst, the better. There are a number of studies that prove this [\u003cspan class=\"CitationRef\"\u003e48\u003c/span\u003e].\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e (a) show that the time-lapse video image before cell detection in multiple embryo development, and Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e(b) Video image after cell detection in multiple embryo development processes. As shown in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e, the result of detection of the cell in the process of even cleavage of the target object and the result of the video image in the case where it is not were compared. This confirmed and proved the objective efficiency of the detection process by the AI algorithm. When viewing the cell as an object in a video image of a process of simultaneous cell division, the results of detecting cells in the uniform division process of the object are visually compared with the results of the video image that does not. This is because it is necessary to examine the continuous differentiation process for a certain period of time during the division process, and research on this will also be covered in the next series of papers.\u003c/p\u003e\n\u003cp\u003eSimilarly, Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e illustrates that there are cases where there are many target objects and it is necessary to be able to determine the criteria for selecting the best case in a single embryo. At this time, Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e (b) shows the detection performance as a result of pre-processing by collecting video of change over time. The best case, with a total of 15 cases can be discussed computationally from a predictive point of view in artificial intelligence algorithms.\u003c/p\u003e\n\u003cp\u003eIn addition, as shown in the case of Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e, slightly different video images and bounding-boxing can be observed. Figure \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e (a) shows ResNet-50 as layer 50 for deep learning, Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e (b) shows layer as ResNet-101, and Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e (c) shows layer 152 as ResNet-152, respectively. This shows a very successful detection rate as a result of final selection considering the focal loss function by applying the single-stage object detection model to the advantages of the base network detection model.\u003c/p\u003e\n\u003cp\u003eAt the same time, it is possible to prevent errors detected by the human eye and ensure accuracy in obtaining high-level detection results that can accurately detect occlusion in video images. Given the peculiarity of the video image displayed as a two-dimensional image, the exact size of cells and count of the overlapping target object is not a simple problem with the help of an expert system. In particular, as shown in Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e (b), the problems of occlusion that can appear in gray scale and the smooth edge phenomenon have been solved as shown in Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e (a). It was confirmed that the detection rate was achieved up to 100%.\u003c/p\u003e\n\u003cp\u003eAs we discussed in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, the proposed video object detection model is an AI Video-specified neural-net structure that performs cell segmentation and characteristic analysis in a single image, and has the strength of extracting accurate cell-level information. Based on this, Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e has been extended to efficiently process and analyze multiple embryo images, and aims to comprehensively explain the embryo status at various points in time. This enables automatic analysis and comparison of embryo development stages.\u003c/p\u003e\n\u003cp\u003eUltimately, a proposed AI video neural-net model that explains various embryo appearances in Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e can be derived using Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e presented above as a basic model in Section \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. The new AI vision model presented in this way has the advantage of more clearly and efficiently analyzing the characteristics of the shape of the time-lapse video of embryo creation and change. In addition, the advantages of the both AI vision model are listed as follows by time-domain. First, Fig. \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e represents Mask R-CNN to perform object detection and segmentation simultaneously, and enables parallel processing of multiple images. Second, by adding \u0026lsquo;embryo description\u0026rsquo; to the analysis results, it has been extended to include embryo-level interpretation beyond cell-level information.\u003c/p\u003e\n\u003cp\u003eConsequently, the key difference between Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e and Fig. \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e lies in the processing scope and application purpose. Figure \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e focuses on precise cell-level analysis in a single image, and is optimized for calculating cell counts and areas with multi-scale segmentation using ResNet-50, ResNet-101, ResNet-152 and Feature Pyramid Network (FPN), respectively. On the other hand, Fig. \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e is a more extended version that uses the Mask R-CNN structure to process multiple embryo images simultaneously and includes the ability to synthesize cell information to describe the state of each embryo. This is a great advancement in that the system goes beyond simple cell analysis to enable automatic evaluation and description of the entire embryo.\u003c/p\u003e\n\u003cp\u003eThe differences between the two AI neural-net models above are summarized in the following Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparison of single-image analysis and multiple-image analysis in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e and Fig. \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSingle-image analysis\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMulti-object analysis\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eProposed Multi-cell analysis\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eBase Model\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eResNet 101 with FPN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMask R-CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMask R-CNN with multi-scale segmentation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eInput Handling\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSingle image\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMultiple images with preprocessing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMultiple images with automatic description\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eSegmentation Method\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMulti-scale prediction (Feature Pyramid Network)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eObject detection with mask generation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMulti-scale segmentation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAnalysis Output\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCell counting, Area calculation, Basic description\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCell counting, Area calculation, Basic cell description\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCell counting, Area calculation, Automatic cell evaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eApplication\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFocused on Cell-level analysis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFocused on Specific interpretation(e.g., embryo-level)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFocused on Cell-level analysis (e.g., synethzised cell-level)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eThe core differences between Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e and Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e lie in the model architecture, input handling, segmentation strategy, and output scope. Figure \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e utilizes a ResNet-101 with a Feature Pyramid Network (FPN) to process single embryo images, enabling multi-scale feature extraction and focused cell-level analysis, such as cell counting and area calculation. In contrast, Fig. \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e adopts a Mask R-CNN framework capable of handling multiple preprocessed images, performing object-level segmentation, and providing not only cell-level information but also comprehensive embryo-level descriptions. As seen in the last rightmost column, unlike Mask R-CNN, the proposed new architecture is scientifically based on multi-scale segmentation, which enables \u0026lsquo;auto description\u0026rsquo; by processing multiple images using automated techniques. This provides an automated workflow optimized for cellular-level analysis of videos using deep learning, which is advantageous for repeatability and large-scale data processing.\u003c/p\u003e\n\u003cp\u003eWe found that early methodologies emphasized single-image analysis, leveraging multi-scale feature extraction to quantify cellular characteristics such as count and area. Recent advancements enable the simultaneous processing of multiple images, facilitating both high-precision cell segmentation and integrative assessment of embryonic morphology. This methodological progression reflects a shift toward more comprehensive, automated, and scalable analysis of embryonic development.\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e shows the results of comparing the single-cell and multi-cell cell counting accuracy (mAP@IoU50) for each three architecture. The conditions are unified as epoch is 1,000, learning rate is 0.00001. Among them, the Mask R-CNN with Multi-scale segmentation showed the highest accuracy in both analyses. On the other hand, ResNet101\u0026thinsp;+\u0026thinsp;FPN has relatively low accuracy, and the gap is especially large in multi-cell analysis. Through this, we can see the cell counting accuracy, (i.e., detection accuracy) that shows how much the prediction of the number of dividing cells over time matches the reality.\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e9\u003c/span\u003e shows the results of comparing the embryo single-cell and multi-cell segmentation accuracies using three deep learning architectures based on the dice coefficient, in this case, F-1 score for pixels. As a result, the Mask R-CNN-based architecture, particularly the model combining multi-scale segmentation, indicated the highest segmentation accuracy in both experiments, which implies its superior adaptability to recognize multi-cell boundaries and structures.\u003c/p\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eIn this paper, we conducted experiments on bio-medical video images to achieve accurate detection, segmentation, and tracking of target objects in order to solve the efficient propagation of artificial intelligence algorithms and architectures. In other words, we performed artificial intelligence algorithms, especially deep learning, measurements and experiments from the time-series data, that is, from the phase of object extraction to the phase of detection of video images. To achieve the most efficient object detection, a method that satisfies all conditions among the 1-stage object detection models and at the same time, satisfies the change in frames per second of the video image that changes due to the nature of the bioinformatics of the experimental domain. The video to be observed has the character of an early experiment, limited to the energy, medical and bioinformatics domains, among many other domains, and was considered in terms of object entropy, confidence, and probability, etc.\u003c/p\u003e \u003cp\u003eThat is, we investigated the four factors that is accurate object detection. First, the accuracy of the number of cells cleaved after embryo fertilization; second, cell size; third, embryo morphology and fourth, uniformity after cell division. The observations in this paper correspond to the basic stage of in vitro fertilization of the fertilization stage of human embryos, and show the best results to achieve the dedication of the scientific commitment in the biomedical domain. The most important starting point in this paper is the object detection of the target video, and it starts from the part where the cell number is accurately counted after embryo fertilization. Lastly, the theoretical correlations and impact of whole process of human embryo developments were made in terms of abnormally time-series pattern extraction and prediction perspectives.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAI \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Artificial Intelligence\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Fast R-CNN \u0026nbsp; \u0026nbsp;Fast Region-based Convolutional Network method\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;E-OFW \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Efficient-Open Frame-work\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;S-OFW \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Secure-Open Frame-work\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ROI \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Region of Interests\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;EPOCH \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;A long period of time marked by distinctive events\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;SGD \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Stochastic Gradient Descent\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Mask R-CNN \u0026nbsp; Mask Region-based Convolutional Neural Network\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;IVF \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; In Vitro Fertilization\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;mAP@IoU50 \u0026nbsp; Multi-cell cell counting accuracy\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u0026nbsp;\u003c/strong\u003eThis study utilized publicly available anonymized datasets that were previously collected with appropriate ethical approvals and patient consent. No additional ethics approval was required for this analysis, as it involved secondary use of de-identified public data in accordance with the original data usage agreements and licenses.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e:\u0026nbsp;Not applicable.\u0026nbsp;This study used publicly available anonymized datasets where consent for publication was obtained during the original data collection process by the original data providers.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailable of data and materials\u003c/strong\u003e:\u0026nbsp;The real research data used in this study is the synthetic embryo dataset, which is publicly available at https://huggingface.co/datasets/deepsynthbody/synembryo_stylegan. Another data set is time-lapse embryo dataset for morphokinetic parameter prediction, publicly available at https://zenodo.org/records/6390798\u0026nbsp;and https://zenodo.org/records/14253170.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interest:\u003c/strong\u003e The author’s role in the initial design of research design, data with synthesis, interpretation, and elucidation of this paper are without any conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e This research was supported by the Bio\u0026amp;Medical Technology Development Program of the National Research Foundation (NRF), funded by the Korean government (MSIT) (No. RS-2023-00223501)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor’s contributions:\u0026nbsp;\u003c/strong\u003eT. J. conceptualized the entire paper, synthesized public data and minimum time-lapse data, implemented the actual AI learning model and derived the simulation results, wrote and submitted the final manuscript, and attempted to obtain the contract for the entire project.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments:\u003c/strong\u003e The author greatly acknowledged two gynecologists, and fertility laboratory senior member for their valuable discussion on March 2\u003csup\u003end\u003c/sup\u003e, 2019 from the initial project presentation to Cha Hospital Seoul Station Fertility Center as the Advanced Intelligence Research (AIR) Lab moved from one university to Hallym university. The author would like to thank Mr. J.H. Sa for his drawings while he worked as an intern at AIR Lab.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMesko B. The role of artificial intelligence in precision medicine. Expert Rev Precision Med Drug Dev. 2017;2(5):239\u0026ndash;41.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOliveira AL. Biotechnology, big data and artificial intelligence. Biotechnol J. 2019;14:1800613.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAntol S, Agrawal A, Lu J, Mitchell M, Batra D, Zitnick CL, Parikh D. (2015). Vqa: Visual question answering. In \u003cem\u003eProceedings of the IEEE international conference on computer vision\u003c/em\u003e (pp. 2425\u0026ndash;2433).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEttaloui, Nehal S, Arezki, Gadi T. An overview of blockchain-based electronic health record and compliance with GDPR and HIPAA. \u003cem\u003eThe International Conference on Artificial Intelligence and Smart Environment\u003c/em\u003e. Cham: Springer Nature Switzerland, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoyal Y, Khot T, Summers-Stay D, Batra D, Parikh D. (2017). Making the v in Vqa matter: Elevating the role of image understanding in visual question answering. In \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition\u003c/em\u003e (pp. 6904\u0026ndash;6913).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHernandez D, Brown TB. (2020). \u003cem\u003eMeasuring the algorithmic efficiency of neural networks\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSavva M, Kadian A, Maksymets O, Zhao Y, Wijmans E, Jain B, \u0026hellip; and, Batra D. (2019). Habitat: A platform for embodied ai research. In \u003cem\u003eProceedings of the IEEE/CVF International Conference on Computer Vision\u003c/em\u003e (pp. 9339\u0026ndash;9347).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eErdem CE, Tekalp AM, Sankur B. (2001, October). Metrics for performance evaluation of video object segmentation and tracking without ground-truth. In \u003cem\u003eProceedings 2001 International Conference on Image Processing\u003c/em\u003e (Cat. No. 01CH37205) (Vol. 2, pp. 69\u0026ndash;72). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJeong T. In: Research, editor. NRF. no.1711051179. Technology and Communication: Ministry of Science; 2017, April. Bio-mimic Implementation by Ambient Data Pre-processing and Classification for Efficiency-centric Intelligent Open Framework (E-OFW).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee S, Jeong T. Forecasting purpose data analysis and methodology comparison of neural model perspective. Symmetry. 2017;9(7):108.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJeong T. AIM: AI based Next-generation Security Information Event Management Methodology for Cognitive Intelligence and Secure-Open Framework. IITP. Grant no. 2019-0-00224. Technology and Communication: Ministry of Science; 2019, April.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJeong T. Time-series Data Classification and Analysis associated with Machine Learning Algorithms for Cognitive Perception and Phenomenon. IEEE Access. 2020;8:222417\u0026ndash;28.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJeong T. Deep Learning System and Methods Using Heterogeneous Data including Behavior Cognitive Data. \u003cem\u003eKorea Patent 10-2019-0179141\u003c/em\u003e, 31 December 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerMilyea M, et al. Development of an artificial intelligence-based assessment model for prediction of embryo viability using static images captured by optical light microscopy during IVF. Hum Reprod. 2020;35(4):770\u0026ndash;84.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKromp F, Wagner R, Balaban B, Cottin V, Cuevas-Saiz I, Schachner C, \u0026hellip; and, Ebner T. An annotated human blastocyst dataset to benchmark deep learning architectures for in vitro fertilization. Sci data. 2023;10(1):271.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e[Access available] Apr. 2025, Annotated Human Blastocyst Dataset \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/software-competence-center-hagenberg/Blastocyst-Dataset\u003c/span\u003e\u003cspan address=\"https://github.com/software-competence-center-hagenberg/Blastocyst-Dataset\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e[Access available] Apr, 2025, Human Fertilisation and Embryology Authority (HFEA) Data \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.hfea.gov.uk/treatments/explore-all-treatments/in-vitro-fertilisation-ivf/\u003c/span\u003e\u003cspan address=\"https://www.hfea.gov.uk/treatments/explore-all-treatments/in-vitro-fertilisation-ivf/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThirumalaraju P, Kanakasabapathy MK, Bormann CL, Gupta R, Pooniwala R, Kandula H, \u0026hellip; and, Shafiee H. (2021). Evaluation of deep convolutional neural networks in classifying human embryo images based on their morphological quality. Heliyon, \u003cem\u003e7\u003c/em\u003e(2).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmith J, Lopez A, Chen D. Deep Learning Applications in Embryo and Blastocyst Assessment for IVF. J Reprod Med. 2020;45(2):123\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJones E. Predicting Embryo Implantation Success Using Time-Lapse Imaging and Deep Learning. Fertility AI Res. 2021;12(1):45\u0026ndash;57.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee M-J, Nakamura Y, Schultz A, Patel R. CNN-Based Embryo Quality Assessment Using Gardner\u0026rsquo;s Criteria. Comput Biology Med. 2019;38(4):211\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePatel A, Tran KO\u0026rsquo;ConnorL. Multi-Modal Deep Learning for Embryo Viability Prediction in IVF. Artif Intell Med. 2022;58(3):187\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGarcia D. and Soo-Young Kim. Augmenting Limited Embryo Datasets with GANs for Deep Learning in Reproductive Medicine. \u003cem\u003eMedical Image Analysis\u003c/em\u003e, vol. 79, 2023, 102471.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang H, Lin M, Thompson J, Elena Petrova. Automatic Blastocyst Segmentation and Classification Using Deep Learning. IEEE Trans Med Imaging. 2021;40(6):1703\u0026ndash;12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNguyen T, Moreau I, Martinez D, Khan F. Transformer-Based Embryo Assessment with Interpretable Attention Mechanisms. \u003cem\u003eFrontiers in Artificial Intelligence\u003c/em\u003e, vol. 6, 2023, article 104203.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrown S. Challenges and Future Directions for Explainable AI in IVF. J Biomed Inform. 2024;112:104618.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSuleiman A, Sze V. (2014, October). Energy-efficient HOG-based object detection at 1080HD 60 fps with multi-scale support. In \u003cem\u003e2014 IEEE Workshop on Signal Processing Systems (SiPS)\u003c/em\u003e (pp. 1\u0026ndash;6). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHashmi MF, Ashish BKK, Keskar AG. (2020, January). GAIT analysis: 3D pose estimation and prediction in defence applications using pattern recognition. In \u003cem\u003eTwelfth International Conference on Machine Vision (ICMV 2019)\u003c/em\u003e (Vol. 11433, pp. 203\u0026ndash;210). SPIE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTarg S, Almeida D, Lyman K. Resnet in resnet. Generalizing residual architectures; 2016.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFathabadi FR, Grantner JL, Shebrain SA, Abdel-Qader I. (2021, January). Multi-class detection of laparoscopic instruments for the intelligent box-trainer system using faster R-CNN architecture. In \u003cem\u003e2021 IEEE 19th World Symposium on Applied Machine Intelligence and Informatics (SAMI)\u003c/em\u003e (pp. 000149\u0026ndash;000154). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa S, Huang Y, Che X, Gu R. Faster RCNN-based detection of cervical spinal cord injury and disc degeneration. J Appl Clin Med Phys. 2020;21(9):235\u0026ndash;43.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu B, Luo J, Huang H. Toward automatic quantification of knee osteoarthritis severity using improved Faster R-CNN. Int J Comput Assist Radiol Surg. 2020;15(3):457\u0026ndash;66.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAurna NF, Yousuf MA, Taher KA, Azad AKM, Moni MA. A classification of MRI brain tumor based on two stage feature level ensemble of deep CNN models. Comput Biol Med. 2022;146:105539.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGilbert A, Illingworth J, Bowden R. (2009, October). Fast realistic multi-action recognition using mined dense spatio-temporal features. In \u003cem\u003e2009 IEEE 12th international conference on computer vision\u003c/em\u003e (pp. 925\u0026ndash;931). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin TY, Doll\u0026aacute;r P, Girshick R, He K, Hariharan B, Belongie S. (2017). Feature pyramid networks for object detection. In \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition\u003c/em\u003e (pp. 2117\u0026ndash;2125).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin TY, Goyal P, Girshick R, He K, Doll\u0026aacute;r P. (2017). Focal loss for dense object detection. In \u003cem\u003eProceedings of the IEEE international conference on computer vision\u003c/em\u003e (pp. 2980\u0026ndash;2988).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin TY, Maire M, Belongie S, Hays J, Perona P, Ramanan D, \u0026hellip; and, Zitnick CL. (2014, September). Microsoft coco: Common objects in context. In \u003cem\u003eEuropean conference on computer vision\u003c/em\u003e (pp. 740\u0026ndash;755). Springer, Cham.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePinto PDCC. (2019). Implementation of faster R-CNN applied to the datasets COCO and PASCAL VOC.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBock S, Wei\u0026szlig; M. (2019, July). A proof of local convergence for the Adam optimizer. In \u003cem\u003e2019 International Joint Conference on Neural Networks (IJCNN)\u003c/em\u003e (pp. 1\u0026ndash;8). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThrosby K. (2004). When IVF fails: Feminism, infertility and the negotiation of normality. \u003cem\u003eSpringer\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArmstrong S, Vail A, Mastenbroek S, Jordan V, Farquhar C. Time-lapse in the IVF-lab: how should we assess potential benefit ? Hum Reprod. 2015;30(1):3\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBottou L. (2010). Large-scale machine learning with stochastic gradient descent. In \u003cem\u003eProceedings of COMPSTAT'2010\u003c/em\u003e (pp. 177\u0026ndash;186). Physica-Verlag HD.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGardner D, Weissman A, Howles C, Shoham Z, editors. Textbook of Assisted Reproductive Techniques. Boca Raton: CRC; 2018. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1201/9781351228220\u003c/span\u003e\u003cspan address=\"10.1201/9781351228220\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e[Online] Access available Apr. 2025, Human Fertilisation and Embryology Authority (HFEA) Data \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.hfea.gov.uk/treatments/explore-all-treatments/in-vitro-fertilisation-ivf/\u003c/span\u003e\u003cspan address=\"https://www.hfea.gov.uk/treatments/explore-all-treatments/in-vitro-fertilisation-ivf/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerMilyea M, Hall JMM, Diakiw SM, Johnston A, Nguyen T, Perugini D, \u0026hellip; and, Perugini M. Development of an artificial intelligence-based assessment model for prediction of embryo viability using static images captured by optical light microscopy during IVF. Hum Reprod. 2020;35(4):770\u0026ndash;84.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhosravi P, Kazemi E, Zhan Q, Malmsten JE, Toschi M, Zisimopoulos P, \u0026hellip; and, Hajirasouliha I. Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization. NPJ Digit Med. 2019;2(1):1\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Huang B, Cui Y, Xu Y, Zhang B, Zhu L, \u0026hellip; and, Wu D. Multi-task deep learning with dynamic programming for embryo early development stage classification from time-lapse videos. IEEE Access. 2019;7:122153\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeil D, Henshaw RC, Lane M. Day 4 embryo selection is equal to Day 5 using a new embryo scoring system validated in single embryo transfers. Hum Reprod. 2008;23(7):1505\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Deep learning, Machine learning, Time-series data, Video object detection, Retina-Net","lastPublishedDoi":"10.21203/rs.3.rs-6434380/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6434380/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: In this paper, we prove the efficiency of a video object detection algorithm through deep learning to have the most essential video of time-lapse data for the completion of artificial intelligence vision object detection architecture that is used for prediction purpose. We alsoinvestigated time-lapse video data, which is the most important part since it recorded during in vitro fertilization process. Particularly, to achieve the most efficient video object detection by limiting special-purpose object detection to only medical healthcarebio-domains, all conditions were satisfied among the single-stage videoobject detection architectures, and proved as theoretical proofs and experiments.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethod:\u003c/strong\u003e Due to the characteristics of bio-medical in the experimental purpose, we applied artificial intelligence neural networks as a way to capture the frames per second (fps)changes of time-varying video images. To gain advantages in science and mathematics in the biomedical domain, we considered the aspects of entropy, confidence, and object occurrence probability. Accurate time-lapse video object detection factors include: (\u003cem\u003ei\u003c/em\u003e) first, the accuracy of the number of cells divided after embryo fertilization, (\u003cem\u003eii\u003c/em\u003e) second, the acute cell size during cell division, (\u003cem\u003eiii\u003c/em\u003e) third, the morphological uniformity of embryos, and (\u003cem\u003eiv\u003c/em\u003e) fourth, the possibility of possible fertilization after cell division.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: The most significant finding in this study is the accurate counting of cells after embryo fertilization, as detected through time-lapse video object recognition. From an AI vision perspective, we propose a fast and efficient video detection framework by implementing and evaluating two distinct learning models: RetinaNet, a single-stage detector, and Fast R-CNN, a multi-stage detector. Their performance was compared against other deep learning-based detection models. Theoretical insights and practical implications regarding the full cycle of human embryonic development were derived, particularly through the identification and prediction of abnormal temporal patterns.\u003c/p\u003e","manuscriptTitle":"Deep Learning-based Video Object Detection for Single-and Multi-Cell Analysis and Evaluation in Time-Lapse Imaging","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-07 07:12:56","doi":"10.21203/rs.3.rs-6434380/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"fe4d5f75-674b-4dc5-9cb3-f42169a9e1c7","owner":[],"postedDate":"May 7th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-10-02T09:38:52+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-07 07:12:56","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6434380","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6434380","identity":"rs-6434380","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}