Embryo development stage onset detection by time lapse

monitoring based on deep learning

Wided Souid Miled1,2 , Sana Chtourou3 , Nozha Chakroun3 and Khadija Kacem Berjeb3
1 LIMTIC Laboratory, Higher Institute of Computer Science, University of Tunis El-Manar, Ariana, Tunisia.
2 National Institute of Applied Science and Technology, University of Carthage, Centre Urbain Nord, Tunisia
3 University of Medicine of Tunis, Laboratory of Reproductive Biology and Cytogenetic, Aziza Othmana Hospital, Tunisia
[email protected], sana [email protected], [email protected], [email protected]

Keywords: IVF, Pronuclei Detection, Embryo Selection, Computer Vision, Classification, Deep Learning, Sequential
Models.

Abstract: In Vitro Fertilisation (IVF) is a procedure used to overcome a range of fertility issues, giving many couples the
chance of having a baby. Accurate selection of embryos with the highest implantation potentials is a necessary
step toward enhancing the effectiveness of IVF. The detection and determination of pronuclei number during
the early stages of embryo development in IVF treatments help embryologists with decision-making regarding
valuable embryo selection for implantation. Current manual visual assessment is prone to observer subjectivity
and is a long and difficult process. In this study, we build a CNN-LSTM deep learning model to automatically
detect pronuclear-stage in IVF embryos, based on Time-Lapse Images (TLI) of their early development stages.
The experimental results proved possible the automation of pronuclei determination as the proposed deep
learning based method achieved a high accuracy of 85% in the detection of pronuclear-stage embryo.

1 INTRODUCTION In recent years, new advanced IVF incubators en-

tered the market with Time Lapse Imaging (TLI) tech-
Statistically, almost 10% to 15% of couples suffer nology (Dolinko et al., 2017). These TLI incubators
from infertility in the world. Multiple infertility treat- make it possible to monitor embryonic development
ments have been developed over the years, collec- continuously. They take photographs of each embryo
tively referred to as Assisted Reproductive Technol- at regular intervals and compile them in a time-lapse
ogy (ART). In Vitro Fertilization (IVF) has prevailed video, giving dynamic insight into embryonic devel-
as the most effective and commonly used type of opment in vitro without disturbing the stable culture
ART. To undergo an IVF cycle, patients should have conditions. These incubators, often accompanied by
an ovarian stimulation in order to collect multiple a dedicated annotation software, have provided both
oocytes which will be incubated with selected motile biologists and clinicians with a new set of data re-
sperm from a semen collection. The intra cytoplasmic garding embryonic behaviour during preimplantation
sperm injection is a more advanced technique where development and its association with embryo quality.
every spermatozoa is injected in a mature oocyte. The
As detailed in academic guidelines (Ciray et al.,
resulting embryos are kept in an incubator for three to
2014), the human embryo undergoes different de-
five days where their development is observed con-
velopment stages, from a fertilized egg (zygote) to
tinuously by embryologists, on an x400 microscopic
a transferable blastocyst. The main developmental
scale, to extract their morphokinetic parameters. Mor-
events are polar body appearance (pPB2), pronuclei
phokinetics comprise the timing and morphological
appearance and fading (pPNa and pPNf), cleavage
changes of embryo as it grows and passes through
or cell divisions (p2 to p9+), compaction or Morula
a series of sequential developmental stages defined
(phase pM), and Blastocyst formation and expansion
in academic guidelines (Ciray et al., 2014). Based
(pB and pEB). Figure. 1 illutrates some of these em-
on these observations, embryologists decide whether
bryo development phases.
to transfer the developed embryo for implantation,
freeze it for later use, or discard it if it doesn’t show a Typically, the pronuclear stage occurs within
good implantation potential. about 16-18 hours, after the sperm is combined with
 more consistent results. Recent AI and DL advance-
ments in the embryology laboratory are summarized
in the review of Dimitratis et al. (I. Dimitriadis and
Bormann, 2022).
(a) (b) (c) (d) In this work, we are concerned with the problem
Figure 1: Embryonic development stages. (a) Pronuclear of automatic detection of pronuclei in the early stages
(b) First cleavage (c) Morula (d) Blastocyst. of IVF embryos development. We aim to develop a
proof of concept (PoC) computer vision solution to
the egg. At this stage, a male and female pronuclei automatically grade the quality of pronuclei in fer-
(2PN) appear containing the genetic material from the tilized embryos, based on time-lapse images of their
sperm and the egg, respectively. The two pronuclei of early development stages.
a normal fertilization are generally equal in size and The main contributions of this work are as follows:
centrally located. Indeed, several studies have shown • We build a supervised data collected from TLI
that the morphology of the embryo at the pronuclear IVF incubators making a dataset of 250 anno-
stage is a valuable parameter in the process of evaluat- tated time-lapse sequences of unique embryos
ing embryo quality and developmental potential. Cur- framed each into 20 annotated images. The an-
rently, embryologists do the assessment visually, in a notations refer to critical embryo development in-
manual process, leaning on their visual experience. stants, namely tPB2, tPNa, tPNf, and t2. We infer
This poses several challenges including: the selection from these annotations the tPN assessment, which
is prone to human perception error, which can lead to confirms successful fertilization.
the loss of promising embryos, or to failed pregnan-
cies; the process is highly subjective as it is difficult • We create a deep learning model based on a CNN-
to agree on quality assessment between embryologists LSTM network with a pre-trained VGG16 back-
(Adolfsson and Andershed, 2018). Manual assess- bone.
ment is also a difficult and time-consuming process. • Hyperparameter selection and comparative exper-
Apart from having to take out the embryo from the in- iments are conducted to optimize and evaluate the
cubators thus disturbing its culture conditions. These proposed CNN-LSTM model
challenges suggest that an automated evaluation so- • To our knowledge, this work represents the first
lution leveraging computer vision and artificial intel- attempt at automatic video annotation of human
ligence would provide a more reliable and accurate embryos from an ART center in north Africa.
solution that helps embryologists and supports their
decision-making with embryo selection.
Artificial intelligence (AI) is a field whose goal is 2 RELATED WORK
to create machines capable of learning and improving
themselves in an autonomous way. This technology According to the literature review by Louis et al.
is proving to be useful in all intellectual tasks. The (Louis et al., 2021), existing research employing
concept of (IA) has been extended to encompass sev- computer vision and deep learning techniques for
eral subfields, including image classification, which IVF embryo selection focuses on the following main
has made considerable progress in recent years (Ya- tasks: automatic embryo stage development anno-
dav and Sawale, 2023). This progress is due to numer- tation (Gomez et al., 2022), (V. Raudonis, 2019)
ous works in this field and to the availability of public cell counting and detection during cleavage (Rad and
datasets that have allowed researchers to report the ex- Havelock, 2019), blastocyst quality grading accord-
ecution of their approaches. This direction of research ing to Gardner’s grading system (Gardner and School-
has resulted in the emergence and evolution of Deep craft, 1999), (L. Lockhart and Havelock, 2019),
Learning (DL), with the advent of Convolutional Neu- (G. Vaidya and Banker, 2021), (M. F. Kragh and
ral Networks (CNN), a particular type of neural net- Karstoft, 2019) and implantation outcome prediction.
work whose architecture of connections is inspired
by that of the visual cortex. In the same trend, the Leahy et al. (Leahy et al., 2020) created a pipeline
use of artificial intelligence (AI) techniques is being of five CNNs for automated measurements of key
intensively researched in the field of IVF. Many au- morphological features of human embryos for IVF.
tomated systems based on artificial intelligence have A Mask R-CNN network with a ResNet50 backbone
been proposed to improve IVF success rates by as- was proposed for pronucleus object instance segmen-
sisting embryologists with their decision and ensuring tation. The model detects pronuclei by outputting
an object mask and a confidence score from 0 to 1
for each frame of a TLI embryo sequence, cropped main unchanged), is only valid from 0PN to either
around the embryo region of interest. Another in- 1PN or 2PN and from 1PN to 2PN. This integration
sightful research that uses deep learning for automat- of time-series information resulted in improvement of
ing assessment of human embryos in IVF treatment performance in sensitivity, however the accuracy re-
is reported in (Lockhart, 2018). Three tasks were the mains relatively low. To the best of our knowledge,
focus of this work: blastocyst grading, cell detection this workb (Fukunaga et al., 2020) is the only exist-
and counting, and embryo stage classification and on- ing reference that deals with detecting and determin-
set detection. For the latter task, the proposed model ing pronuclei number in IVF embryos.
incorporates temporal learning over the TLI sequence In this work, we aim to automate the annotation
and automatically detects three classes, namely cleav- process of the early stages of embryonic development,
age, morula, and blastocyst stage onsets. In order to from Polar Body appearance (tPB2) to just before
detect stage transitions, two image sequence batches the first cell division (t2). We create a deep learn-
are fed in parallel, in pairwise learning, through two ing model that analyzes the TLI incubator’s sequences
separate CNNs, which are based on VGG16 architec- of embryonic development and annotates tPN, de-
tures pre-trained on the ImageNet dataset with three fined as the time at which fertilization status is con-
final convolution layers fine-tuned. Fully connected firmed, immediately before the time fading of pronu-
layers from each classifier are concatenated and used clei (tPNf) (Ciray et al., 2014).
to predict whether the input images fed through each
branch were at the same stage. Synergic loss from this
binary output is backpropagated through both classi- 3 METHODOLOGY
fier branches. Stage transitions predictions are then
refined using temporal context in an LSTM layer sep-
arately for each synergic branch. 3.1 Dataset
Gomez et al. (Gomez et al., 2022) worked on the
The dataset used in this work is a collection of 352
automatic annotation of the 16 embryo development
videos of unique embryos exported from a private TLI
phases. In addition to providing a fully annotated
IVF Incubator manufactured by Esco Medical®. The
dataset composed of 704 time-lapse videos, authors
frames of each video are time-lapse embryo images
applied ResNet, ResNet-LSTM and ResNet3D mod-
taken every five minutes, starting shortly after fertil-
els to automatically annotate the stage development
ization. Each video contains between 600 and 1400
phases. The evaluation results showning the superi-
frames in gray scale with a resolution of 1280 × 720
ority of ResNet-LSTM and ResNet-3D over ResNet,
pixels.
prove the importance of using the temporal informa-
An experienced biologist notes the start and end
tion in the automatic annotation process. However,
time of each phase of the embryo’s development.
predicting the 16 classes of embryonic development
Each image of each video has therefore a class, which
is prone to numerous challenges, primarily due to
corresponds to the phase seen in the image. The an-
the extensive computational requirements necessary
notations follow the same convention used by Gomez
for training DL models on more than 300k images,
et al. (Gomez et al., 2022) and academic guidelines
which demand high-performance GPUs. Fukunaga et
(Ciray et al., 2014). There are, in general, 16 annota-
al. (Fukunaga et al., 2020) proposed an automated
tions corresponding to 16 different instants of embryo
pronuclei determination system based on few amount
evolution. Here, as we are only interested in detect-
of supervised data. In their paper, authors proposed
ing two key instants, namely tPB2 and tPN, we only
a framework of four stages. First, images are pre-
consider the following phases:
processed to detect and focus on the embryo area us-
ing a circular Hough mask. Then, images are passed • tPB2: time of appearance of second polar body
for main processing to two CNNs, both composed of • tPNa: time of pronuclei appearance
two convolution layers and two fully connected lay- • tPNf: time of pronuclei fading
ers. The first model detects the outline around pronu-
• t2: time of first cell division marks the end of
clei and passes these outline images to the second
pronuclear phase
CNN, which gives a probability distribution of the
number of pronuclei (0PN, 1PN, 2PN). Finally, pre- The stage tPN, which is defined as the time at which
dictions are postprocessed through a Hidden Markov fertilization status is confirmed, is calculated from
model, while setting conditions for the change in the tPNa and tPNf (Ciray et al., 2014). We received the
number of pronuclei over time. Thus, the change of annotation in Excel sheets generated by the software
the number of pronuclei, if occurred (the state can re- of the TLI incubator, which we had to parse to extract
useful information.
3.2 Data Preprocessing
Before feeding the sequence of images to the pro-
posed deep learning based model, we made some
preprocessing treatments. First, as we reviewed the
dataset, we observed that some videos suffered from
excessive lighting changes and motion blur. Other
images were taken from a bad angle where the em-
bryo was not entirely visible. Some other videos did
not cover some critical stages of the embryo’s devel-
opment. After discarding these unusable videos, we
obtained 250 annotated videos of unique embryos.
Then, as the embryo cell presents only a small part
of the image, we cropped them, reducing the frame
size to 360 × 360 and gaining in memory efficiency. Figure 3: Proposed input convolution flows.
To achieve this, we applied the Hough transform to
detect circular shapes in the images, then we cropped chitecture.
the detected circles with a fixed size of 360 × 360 pix- It should be noted that a sequence of images must not
els. be fed to a single convolution. If we take a common
After choosing the frames and preprocessing sequential network, each entry is connected to all the
them, each frame has been labeled based on the ex- neurons in the first layer. With multiple images as
pert’s annotations. We repeated the same process for batch entries to the CNN network, all the pixels of all
every video in the dataset. For each video, we ended images are merged and sent to the first layer. Con-
up with 20 images, sampled over the first 18 hours sequently, their distinctive features and the temporal
of embryo development. The retained frames are an- information will be lost. To overcome this problem,
notated as 0 (neither tPB2 nor tPN occured), 1 (tPB2 as illustrated in Figure 3, we need to share the network
occured), 2 (tPN occured). We can see examples of layers across the video frames to reduce the number
the three classes in Figure. 2. of tensors, thus having filters for each image input,
not for the whole stack of frames.
With this adopted architecture, each image has got its
own convolution flows. If we separately train each
convolution flow, we will have several unwanted be-
haviors:
• We will need long training time because several
convolution flows need to be trained (one per in-
class 0 (no event) class 1 (tPB2) class 2 (tPN) put image).
Figure 2: Examples of labelled images from the dataset. • Some convolution flows will not detect what other
flows could detect.
3.3 Proposed Model • Each convolution flow, for one sequence, can have
several different weights, and so we get different
In this work, we are concerned with a sequence clas- detection features that are not linked.
sification problem, which implies that the model’s In order to make sure that all the convolution flows
input is not a series of independent images to be can extract the same features, we propose to add a
classified as categorical targets, but rather a time- time distributed layer which applies the same convo-
dependent sequence of images to be predicted accord- lution layer to several inputs. This allows to apply
ing to a certain order. Sequence classification is a the layer operation on each timestamp. Otherwise,
challenging problem because the sequences can vary when we flatten the data all the image instances will
in length, contain a very large vocabulary of input be combined and the time dimension will be lost.
symbols, and may require the model to learn the long- As shown in Figure 4, the proposed model has two
term context or dependencies between symbols in the main parts: a CNN and an LSTM network, linked by
input sequence. The solution to this sequentially- a time distributed layer. Each layer that is time dis-
classified problem is to use a combination of the two tributed will share the same weights, saving calcula-
approaches: the LSTM architecture, and the CNN ar- tion and computation time.
 Figure 4: The proposed architecture integrating a time distributed layer.

For the CNN backbone, in the field of medical images into RGB. Furthermore, as each pixel value
image analysis, it is common to use a deep learn- can vary from 0 to 255, representing the color inten-
ing model pre-trained on a large and challenging im- sity, feeding an image directly to the neural network
age classification task, such as the ImageNet classi- will result in complex computations and a slow train-
fication competition. The research organizations that ing process. To address this problem, we normalize
develop models for these competitions often release the high numeric values to range from 0 to 1 by di-
their final models under a permissive license for reuse. viding all pixel values by 255. Then, we labeled the
These models can take days or weeks to train on mod- dataset marking images in the tPB2 phase as class 1,
ern hardware. But, we can directly use them pre- those attaining the stage tPN as class 2, and the re-
trained employing transfer learning technique for a maining images where no event occurs into class 0.
target specific task. In this work, we opted for a Finally, we split the dataset, conventionally, into 80%
VGG16 model pre-trained on the ImageNet compe- training data and 20% test data.
tition dataset.
4.2 Models Implementation

4 EXPERIMENTAL RESULTS Since the backbone pre-trained CNN model wasn’t

designed to annotate pronuclei stage development
phases in embryo image datasets, we have to make
4.1 Dataset it more specific to our needs, taking advantage of the
transfer learning technique and using the ImageNet
In this section, the performance of the proposed deep
pre-tuned weights. We chose to train only the last
learning model for the task of early stage human em-
four layers and reduce the number of outputs using
bryo detection is discussed. Our dataset contains 250
the last pooling layer with a maximum operation ap-
annotated videos of unique embryos augmented five
plied to the convolutional values. First, we specify
times (Horizontal flip, vertical flip, transpose, and
the top layers by the VGG implementation, taking our
transpose horizontal flip). We further resized the im-
custom input 180 × 180 × 3 images. We then link the
ages from 360 × 360 to 180 × 180 resolution. Since
time distributed layer with the VGG16 output layer
the number of frames can be very large, it is imprac-
via a sequential mode, which will fully connects each
tical to feed all of them to the model, as this would
neuron from both sides. The next layers are the LSTM
slow the training and reduce the performance. Our
layers, followed by five dense layers, separated with
strategy was to choose 20 frames between the start
50% dropout layers to prevent over-fitting. We use
of the video and the instant tPNf (which denotes the
the ReLu activation function and Softmax as a final
fading of the pronuclei) and feed them to the model,
activation function, which will output the correspond-
since this range covers all the phases we are interested
ing class probabilities. As an optimization algorithm,
in. We chose our frames in a way where the number
we opted for Adam (Adaptive Moment Estimation),
of frames between two consecutive chosen frames is
as it is straightforward to implement, is computation-
constant. Every sequence is therefore framed into 20
ally efficient, has little memory requirements and is
(180 × 180 × 3) images. As the VGG model requires
well suited for problems with large data and/or pa-
3-channels input images, we converted our grayscale
rameters (Kingma and Ba, 2014). We fix the learning
rate at a value of 0.01, to converge the learning in a
faster, more efficient way, and to avoid the problem of
vanishing gradients. We choose the categorical cross-
entropy as the loss function since we are dealing with
a multi-class dataset.
After experimenting with the transfer learning
technique, we decided to build a custom CNN model
using six convolutional layers. We also introduced
a batch normalization to reduce the inter-variance of
the layer inputs. This technique stabilizes the learn-
ing process and dramatically reduces the number of
epochs required for training. The batch normaliza-
tion momentum uses the moving average of the sam-
ple mean and variance in a mini-batch for training.
By adjusting a dynamic momentum parameter, the Figure 5: Accuracy and loss graphs of the CNN-LSTM pro-
noise level in the estimated mean and variance can posed model
be well controlled. We fixed the momentum value
to 0.9. We kept the Adam optimizer and categori- lied on the circles in the centre of the embryo, which
cal cross-entropy loss function. We set the learning correspond to the two pronucleus. Thus, the Grad-
rate to 0.001, which is considerably lower than the Cam method makes it possible the visualisation of the
one used with the VGG16 architecture. The model is areas that contributed the most to the prediction of the
built from scratch, so the gradients are initially ran- specific tPN class.
domized, and to reach a similar accuracy, the weights
need to be adjusted carefully. In a second experiment, we trained the custom-
built CNN model, along with a LSTM network, for
a total of 50 epochs and a batch size of 8. We no-
4.3 Evaluation ticed that this second model has taken more time, and
more failed attempts to reach the threshold accuracy.
The metrics we used for the performance evaluation We can visibly conclude this from all the fluctuations
of proposed DL models are accuracy and sensitivity. in the accuracy per epoch graph in Fig. 7, where the
Accuracy is defined as the ratio of correctly classified validation accuracy doesn’t exceed 60%. This was ex-
instances by the total amount of instances. Sensitivity pected since the pre-trained model has already learned
is defined as the number of correctly classified pos- high-level features, is assigned pre-trained weights
itive samples divided by the number of all positive and only needs fine-tuning to fit the training dataset
samples. on the target task, while the custom-built CNN model
We conduct a first experiment where we trained starts with randomized weights.
the CNN-LSTM model based on a pre-trained
VGG16 backbone for a total of 90 epochs and a batch 4.4 Comparison with state of the art
size of 16. The accuracy and loss graphs for training
and validation are shown in Figure 5. The accuracy For state-of-the-art comparison, as there is no bench-
curve represents few variations and is up to 0.86%. In mark available in the literature, we reported the re-
addition, the loss curve is almost stable and the vali- sults of Fukunaga et al. (Fukunaga et al., 2020) and
dation and training curves are almost similar, showing those of Gomez et al. (Gomez et al., 2022) given
that the model is well fitted. in their corresponding papers and conducted on their
own datasets. Comparative results in terms of accu-
In order to make the proposed classification model
racy and sensitivity metrics are reported in Table 1.
interpretable, we implemented the Grad-Cam method
that exploits the features map from the last convolu- The common aspects between our work and the
tion layers to calculate the gradients of the features work of Fukunaga et al. (Fukunaga et al., 2020) are
map against the class score to identify the most im- the limited amount of supervised data available, and
portant filters. Figure 6 shows the generated heatmaps the classification task. However, the main difference
on the tPN stage prediction, where red pixels indicate is the methodology of the detection systems: we pro-
highest contribution towards stage prediction and no posed a CNN network linked to an LSTM layer while
colour represents no contribution, As seen in this fig- they developed a 2-CNN architecture, with no deploy-
ure, for tPN stage prediction, the network mainly re- ment of a sequential model that would deal with time
 Model Dataset LSTM Accuracy. Sensitivity Classes
usage
Proposed Model 250 videos Yes 85% 96% 3
Fukunaga et al. (Fukunaga et al., 2020) 300 videos No 40% 82% 3
Gomez et al. (Gomez et al., 2022) 873 videos Yes 73% 96% 16
Table 1: Comparison with state-of-the-art methods.

Figure 7: Accuracy graph of the custom CNN proposed

model

outperform algorithmic approaches to the automatic

annotation of embryo development phases. Further-
more, the compared models are detecting 16 classes
of 16 morphokinetic events, compared to 2 events in
our case. The three models they benchmarked con-
cluded a 73% accuracy score.

5 CONCLUSION
Continuous embryo monitoring with time-lapse imag-
ing enables time based development metrics along-
side visual features to assess an embryo’s quality be-
fore transfer and provides valuable information about
its likelihood of leading to a pregnancy. In this work,
we developed a deep learning based model to classify
a sequence of time-lapse Human embryo images with
the aim of helping embryologists with embryo selec-
tion for IVF implantations. The classification task
aims to detect tPB2 and tPN key instants from an in-
put sequence of images by predicting the class of each
image among three classes; denoting the appearance
of the second polar body (tPB2), the appearance of the
Figure 6: The heatmaps generated by the Grad-CAM
method on the tPN stage pronuclei (tPN), or none of the two events. The pro-
posed model is a combination of a pre-trained VGG16
dependency with a deep learning technique. Their backbone, and an LSTM network. It has proven to be
model’s sensitivity reached 82%, but with only a 40% powerful enough to fit the data as it achieved a high
accuracy rate, which makes our method more accurate training accuracy, In future work, our model can be
with 85% accuracy score and 96% sensitivity score. enhanced by being incorporated into a pipeline where
the second part detects the number of pronuclei as
Regarding Gomez et al. (Gomez et al., 2022), 0PN, 1PN, 2PN or more. This pipeline can then be
the used dataset is composed of 337 thousand images part of a whole automatic embryo assessment deep
from 873 annotated videos. This big ground-truth learning framework, integrating the work on blasto-
helped apply three approaches: ResNet, LSTM, and cyst segmentation and cell counting.
ResNet-3D architectures, and demonstrate that they
REFERENCES upsampling convolution. volume 7, pages 81945–
81955.
Adolfsson, E. and Andershed, A. (2018). Morphology vs V. Raudonis, A. Paulauskaite-Taraseviciene, K. S. e. a.
morphokinetics: A retrospective comparison of inter- (2019). Towards the automation of early-stage human
observer and intra-observer agreement between em- embryo development detection. volume 18.
bryologists on blastocysts with known implantation Yadav, S. and Sawale, M. D. (2023). A review on image
outcome. JBRA assisted reproduction, 22(3):228– classification using deep learning. volume 17.
237.
Ciray, H., Campbell, A., Agerholm, I., Aguilar, J.,
Chamayou, S., Esbert, M., and Sayed, S. (2014). Pro-
posed guidelines on the nomenclature and annotation
of dynamic human embryo monitoring by a time-lapse
user group. In Human Reproduction, volume 38,
pages 2650–660.
Dolinko, A. V., Farland, L. V., Kaser, D. J., and al. (2017).
National survey on use of time-lapse imaging systems
in ivf laboratories. Assisted Reproduction and Genet-
ics, 34(9):1167–1172.
Fukunaga, N., Sanami, S., Kitasaka, H., and al. (2020).
Development of an automated two pronuclei detec-
tion system on time-lapse embryo images using deep
learning techniques. Reprod Med Biol., 19(3):286–
294.
G. Vaidya, S. Chandrasekhar, R. G. N. G. D. P. and Banker,
M. (2021). Time series prediction of viable embryo
and automatic grading in ivf using deep learning. vol-
ume 15, pages 190–203.
Gardner, D. and Schoolcraft, W. (1999). In vitro culture of
human blastocyst. Towards Reproductive Certainty:
Fertility and Genetics Beyond 1999, page 378–388.
Gomez, T., Feyeux, M., and al. (2022). To-
wards deep learning-powered ivf: A large public
benchmark for morphokinetic parameter prediction.
https://arxiv.org/abs/2203.00531.
I. Dimitriadis, N. Zaninovic, A. C. B. and Bormann, C. L.
(2022). Artificial intelligence in the embryology lab-
oratory: a review. volume 44, pages 435–448.
Kingma, D. P. and Ba, J. (2014). Adam: A method for
stochastic optimization. In 3rd International Confer-
ence for Learning Representations. arXiv.
L. Lockhart, P. Saeedi, J. A. and Havelock, J. (2019). Multi-
label classification for automatic human blastocyst
grading with severely imbalanced data. pages 1–6.
Leahy, B., Jang, W., Yang, H., and al. (2020). Automated
measurements of key morphological features of hu-
man embryos for ivf. CoRR, abs/2006.00067.
Lockhart, L. (2018). Automating assessment of human em-
bryo images and time-lapse sequences for ivf treat-
ment.
Louis, C., Erwin, A., Handayani, N., and al. (2021). Review
of computer vision application in in vitro fertilization:
the application of deep learning-based computer vi-
sion technology in the world of ivf. Assist Reprod
Genet., 38(3):1627–1639.
M. F. Kragh, J. Rimestad, J. B. and Karstoft, H. (2019).
Automatic grading of human blastocysts from time-
lapse imaging. volume 115, page 103494.
Rad, P. Saeedi, J. A. and Havelock, J. (2019). Cell-net:
Embryonic cell counting and centroid localization via
residual incremental atrous pyramid and progressive