WO2023151578A1

WO2023151578A1 - Embryo selection process

Info

Publication number: WO2023151578A1
Application number: PCT/CN2023/074963
Authority: WO
Inventors: Pengwei HU; Qiao HUA; Siyu Zeng
Original assignee: Merck Patent Gmbh
Priority date: 2022-02-10
Filing date: 2023-02-08
Publication date: 2023-08-17

Abstract

Method for aiding to select a suitable embryo (3) for implantation in an in-vitro fertilization process whereupon at least one two-dimensional digital picture (5) of an in-vitro fertilized embryo (3) is taken, transferred to a computer (6) performing a software (8) with access to clinical data of a patient (20) where the embryo (3) is to be implanted, the software (8) then evaluates embryo morphology features (11) extracted from the digital pictures (5) and predicts the pregnancy success rate (12) using the evaluated embryo morphology features (11) and the clinical patient data (20), wherein the software (8) implements several trained first Artificial Intelligent (Al) models (10) to evaluate the embryo morphology features (11) and one different trained second Al model (9) to predict the pregnancy success rate (12) based on outputs of the first trained Al models (10) and the clinical data of the patient (20).

Description

Embryo selection process

Technical Field

The invention deals with the technological area of assisted reproductive technologies, in particular a method and a system to select suitable embryos for implantation in an in-vitro fertilization process using Al-based mechanisms to predict the pregnancy success rate.

Background

With the increasing number of infertility patients, doctors in various reproductive centers are under increasing pressure to work. Doctors need to evaluate the development of embryos by observing the development of embryos at different times, combining embryo morphology and embryo development dynamics and referring to the consensus of experts at home and abroad, and then choose high-quality embryos for implantation, so as to improve the pregnancy success rate of patients. However, the scoring of embryo morphology depends on the doctor′sown long-term experience of embryo observation. For inexperienced doctors, it is difficult to give accurate embryo scoring results, and there are some differences in the way different doctors are scoring the embryos due to their different subjective knowledge, experience and attitude. In recent years, some experts have begun to study the application of artificial intelligence in the field of artificial reproduction, and have made some preliminary achievements. Artificial intelligence technology can help doctors to quickly and accurately identify embryo morphology at different stages, automatically obtain some important morphology and kinetic parameters of embryo development process, and intelligently analyze the quality of embryos and predict their pregnancy success rates, saving doctors a lot of time and energy. In the process of embryo development, experienced doctors screen and implant embryos according to various morphology parameters of embryos, in which the change of blastomeres during the cleavage phase, embryo fragments and blastocyst morphology are of important roles.

The change of blastomeres during the cleavage phase is one of the important bases for doctors to score the quality of embryos. The change rate of the blastomeres of the embryo must be within a specific range to keep the embryo development in a conducive stage. Normally developed embryos can reach 7 to 9 cells on the third day. Medical studies have shown that the probability of chromosomal abnormalities in stagnant embryos, stunted growing or too fast-growing embryos is significantly higher than that in normal embryos. Therefore, it is very important to identify the number of blastomeres in the third day (Day 3) embryo to score the embryo.

The production of embryo fragments is a common phenomenon in the process of embryo in vitro culture. Fragmentation ratios range from a small amount of fragmentation of 5%to 100%. Doctors usually grade the fragments according to their proportion in the embryo. Analysis of the relationship between cell fragments and embryonic development potential reveals that the fragmentation ratio affects development potential of the embryo. The greater the fragmentation ratio, the worse the development potential of the embryo, so the proportion of fragments can be an important factor in determining the quality of the embryo.

At present, the basic consensus among embryo experts at home and abroad is that the observation time of the blastocyst should be five to six days after fertilization. The development of the blastocyst is divided into six periods according to the size of the blastocyst cavity and whether the blastocyst is hatched. There is a certain relationship between the blastocyst morphology and pregnancy rate or implantation rate, and by evaluating the development speed of blastocyst and blastocyst morphology, embryo quality may be evaluated, and thus higher quality embryos can be screened and implanted to improve pregnancy success rate.

Summary

The task of this patent application is therefore to help doctors to quickly and accurately complete the selection of high-quality embryos based on both morphology features identified and pregnancy success rate predicted results for each embryo.

This task has been solved by aiding to select a suitable embryo for implantation in an in-vitro fertilization process whereupon at least one two-dimensional digital picture of an in-vitro fertilized embryo is created, transferred to a computer performing a software with access to clinical data of a patient where the embryo is to be implanted, the software then evaluates embryo morphology features extracted from the digital pictures and predicts the pregnancy success rate using the evaluated embryo morphology features and the clinical patient data, wherein the software implements trained first Artificial Intelligent (Al) models to evaluate the embryo morphology features and a different trained second Al model to predict the pregnancy success rate based on outputs of the first trained Al models and the clinical data of the patient. To do so the software uses several models which comprise Al based algorithms to analyze pictures which are taken from the embryo cell culture to determine the condition of the embryo by extracting and evaluating specific parameters from those pictures. The results of these evaluations are then used by a different Al algorithm which also includes specific data about the patient, like age, blood type etc but also clinical data like hormone levels and so on, to whom the embryo might be implanted, in its prediction. The prediction contains the probability of a successful pregnancy success rate when using the evaluated embryo for the in-vitro fertilization and can be used by the responsible physician, the embryologist, to choose the most suitable available embryo. The pregnancy success rate defines the probability that the evaluated embryo, if implanted to a patent, will lead to a successful pregnancy as measured by whether fetal heart can be detected in the patient.

Advantageous and therefore preferred further developments of this invention emerge from the associated subclaims and from the description and the associated drawings.

One of those preferred further developments of the disclosed method comprise that two or three of the embryos with a respective ranking score, derived from the evaluated embryo morphology features and the predicted pregnancy success rate, are used as a reference to a following authority which selects at least one of those embryos for implantation to the patient. This authority is in most cases a human embryologist who makes the final decision which embryo (s) is/are implanted. With the information provided the embryologist can choose the most promising of those two or three embryos for in-vitro fertilization. This approach also ensures that a human physician is participating in the whole process who can use his personal knowledge and experience to combine it with the results of the data driven software analyzes. Alternatively, it is also possible that the authority is another automatic system and/or software which double-checks embryo selection.

Another one of those preferred further developments of the disclosed method comprise that a Deep Learning Object Detection Algorithm is used for the models which evaluate the individual parameters of the embryo extracted from the digital pictures and a Machine Learning Boosting Algorithm is used for the model which predicts the pregnancy success rate from the results of the different Al based algorithms for the individual parameters and the patient information. Those are the two preferred types of algorithms, but the invention is not limited to those two types. One example for Deep Learning Object Detection Algorithm is CenterNet, while for the Machine Learning Boosting Algorithm CatBoost could be used.

Another one of those preferred further developments of all disclosed methods comprise that the embryo morphology features extracted from the at least one digital picture by the first Al models comprise of number of blastomeres, fragmentation ratio, blastocyte stage, symmetry, inner cell mass, trophectoderm and other suitable parameters. Especially the first three parameters of the blastomeres, fragmentation ratio, blastocyte stage are important to determine and evaluate the status of the embryo.

Another one of those preferred further developments of all disclosed methods comprise that the first Al models which obtain the embryo morphology features of the number of blastomeres of the embryo and the fragmentation ratio of the embryo are extracting it in a first stage, while the first Al model which obtain the blastocyte stage of the embryo are extracting it in a second stage later than the first stage. Using those specific dated pictures grants accurate training and work results. Therefore preferably those pictures should be used in the described invention. Of course, pictures from other states/days of embryo development can be used as well.

Another one of those preferred further developments of all disclosed methods comprise that the first stage is within the third day of the embryo development while the second stage is within the fifth day of the embryo development. For an adequate evaluation of the blastocytes an observation time of the blastocyst of at least five days after fertilization is required, so 5-day pictures are used. For the other parameters and the pregnancy success rate prediction 3-day pictures are more suitable.

Another one of those preferred further developments of all disclosed methods comprise that respective mAP values, comprising of Recall and Precision parameters for the models, are calculated and respective ROC curves, which reflect sensitivity and specificity of the models, are determined for each model and used to evaluate and improve their respective prediction accuracy. Both parameters for the mAP values are calculated using the [TP] and [FN, respectively FP] . The ROC curve on the other hand is a graph that reflects the continuous variables of sensitivity and specificity. The ordinate shows sensitivity. The lower the sensitivity value for it, the higher the accuracy of the model prediction. The abscissa shows the 1-specificity. Here accounts that the lower it indicates the lower is the misjudgment rate of the model.

A further solution to the given task comprises of an apparatus for aiding to select a suitable embryo for implantation in an in-vitro fertilization process, comprising a first obtaining component, configured for obtaining adigital picture of an in-vitro fertilized embryo in a first stage, and clinical data of a patient where the embryo is to be implanted; a second obtaining component, configured for obtaining embryo morphology parameters from the image via trained first Artificial Intelligent (Al) models; a third obtaining component, configured for obtaining pregnancy outcome from the embryo morphology parameters and the clinical data of the patient via a trained second Al model. The first obtaining component could be a digital camera which is set up to create digital pictures of the embryo. Or it could be a component which scans printed pictures of embryos and digitalizes them. The second obtaining component could be a computer which runs a software which performs the trained first Al models, while the third obtaining component could be a computer with a software that performs the second Al model. The computer and the software for the second obtaining component could be the same as for the third obtaining component, which are simply performing the different models, or it could be different computers and software which uses a data connection to forward their respective output data to its destiny. Furthermore, it comprises a computer device comprising a computer readable memory, a processor and a computer program which is stored on the memory and may be run on the processor, wherein when executing the computer program, the processor implements the previously described methods. The solution also includes a non-transitory computer readable storage medium storing a computer program for execution by a computing device having a processor, wherein the program, when executed by the processor, cause the computing device to perform the described methods. The invention therefore also covers the whole system used to execute the invented method, comprising not only the mentioned software components, but also the required hardware.

Additional parts of the solution for the given task include a method to train models with Al based algorithms used by a software as previously described, wherein the models with the first Al models to evaluate the embryo morphology features and the second Al model for the pregnancy success rate prediction are in a first step trained with a training data set, the trained Al models are then in a second step evaluated with a validation data set and in a third step the trained and evaluated Al models are tested with a test data set, which all comprise of suitable two-dimensional digital pictures of an in-vitro fertilized embryo. That describes the preferred way how to basically train the used models, but other ways are possible as well. Additional features like specific evaluation methods to determine if the model has already reached the required accuracy can be added. Also, the training steps usually need to be iterated, either individually or the whole training with all steps. Typical iteration numbers are 50 iterations per training.

Another improvement of these parts include that the respective data sets each comprise of at least several hundred two-dimensional digital pictures, that the two-dimensional digital pictures of the respective data sets are taken from a joint collection of at least several thousand two-dimensional digital pictures which is distributed in a ratio of 8∶1∶1 between the training data set, the validation data set and the test data and that for the second evaluation step a k-fold cross-validation method is used. Of course, also fewer pictures could be sufficient sometimes. That depends on the specific chosen algorithm and the embryo pictures. But experience has shown that at least several hundred, better even thousands of different pictures should be used to result in adequately trained models. The ratio is considered most advantageous since the first step of training the models takes the most effort while validation and final test can be done by using fewer digital pictures. The standard k-fold cross-validation is adequate to evaluate how accurately the trained model will perform in practice. But other types of cross-validation methods can be used.

Detailed description of the invention

The method and system according to the invention and functionally advantageous developments of those are described in more detail below with reference to the associated drawings using at least one preferred exemplary embodiment. In the drawings, elements that correspond to one another are provided with the same reference numerals.

The drawings show:

Figure 1a illustrates an abnormal picture.

Figure 1b illustrates another abnormal picture.

Figure 1c illustrates a third abnormal picture.

Figure 1d illustrates a fourth abnormal picture.

Figure 2 illustrates predicted pregnancy success rate from the system according to embodiments of the present disclosure.

Figure 3 illustrates an example of an ROC curve diagram according to embodiments of the present disclosure.

Figure 4 illustrates a schematic overview about the method to predict the pregnancy success rate according to embodiments of the present disclosure.

Figure 5 illustrates a schematic overview about the method to train system according to embodiments of the present disclosure.

Figure 6 illustrates a schematic overview of the environment where the method according to embodiments of the present disclosure is applied.

Figure 7 illustrates a block diagram showing the interaction between the system components to perform the invented method.

To apply the invented method a suitable system 1 comprising a computer 6, a software 8 performed on the computer 6, acell culture 3 containing an in-vitro fertilized embryo 3 and a digital camera 4 connected to the computer 6 configured to take digital pictures 5 of the embryo 3 is required which is disclosed in more detail by Figure 6. The computer 6, preferably realized as a form of a control unit 6, further provides a user interface 7 which the user 2 can use to input the settings or control the system 1 in any suitable way. In a preferred embodiment the computer 6 furthermore hosts the software 8 which runs the different Al models 9, 10 in form of the three deep learning models 10 for analyzing the parameters of blastomeres, fragmentation ratio and blastocytes and a machine learning Al 9 model for the pregnancy success rate prediction 12. The software 8 itself can be a stand-alone software used for instance in a hospital by the embryologist. Alternatively, it could also be part of another software, be it a software module for this other software or directly integrated into this software -whatever deems most suitable.

The invented method as shown summarized in Figures 4 and 5 further comprises several different aspects which are explained in the following chapters.

1. Data acquisition

In order to ensure the reasonableness and compliance of the data, the project solicited the recommendations of the Fertility Center and adopted the certification of the ethics committees of five reproductive centers. Based on the content and requirements of embodiments of the present disclosure, embryo images from the five reproductive centers taken under a microscope from day 0 to day 7 were collected, totalling approximately 250,000, as well as basic clinical information and embryo image annotation results for patients with pregnancy success rates.

2. Data cleansing

Because of subjective or objective factors, data samples have common problems such as image blur, image anomaly, value missing, a large difference in range of feature values of different dimensions, outlier, etc. In order to obtain more data of higher quality, data preprocessing of the original samples is required. Common data pre-processing methods include image screening, missing value processing, normalization processing, outlier processing, and feature quantification. Through a series of data pre-processing, data collected by multiple reproductive centers is built into a unified standard database.

In the process of image data collection, there may be abnormal images, such as when embryos 3 do not appear in the image 5, as shown in Figure 1a, embryos 3 in the image blurs, as shown in Figure 1b, individual morphology parameters of embryos 11 can not be accurately identified, as shown in Figure 1c where the embryo 3 shrinks and in Figure 1d where the embryo 3 does not appear in the center, etc. . In order to ensure the robustness of the algorithm, embodiments of the present disclose only retains of the blurred images while neglecting the other abnormal images which are relatively complex to take into account.

In the data collection process, some subjective or objective factors may result in missing and vacancy of feature values in some dimensions in a table of collected clinical data 20. There are many cases for missing of feature values, and corresponding pre-treatment methods need to be selected according to the distribution characteristics of variables and the importance analysis of variables.

If feature values of a certain dimension are highly missing (more than 80%) , feature coverage and importance of the dimension are low, and feature values of this dimension may be deleted.

If the missing rate of a variable is low (no more than 5%) and of some importance, missing values can be filled according to statistics of existing data. When data distribution is relatively uniform, the missing values can be filled, meaning interpolated, with averages, and when there is a tilt distribution of the data, the missing values can be filled with median values.

There are different interpolation methods which work through the analysis of existing data, to complete the missing value complement. Common methods include hot card interpolation, fitting interpolation and multi-interpolation. Hot card interpolation finds an object most similar to the complete data, and then the object′sdata is used for filling. Fitting interpolation method are used to predict missing values using labeled supervised learning methods, such as regression algorithms, nearest neighbor algorithms, random forests, support vector machines, and other models. Multi-interpolation combines Bayesian estimation method to estimate the value to be interpolated, which in combination with different noises, forms multiple candidate sets of interpolation values. The most suitable values are selected according to the distribution of the values to be interpolated.

Features of different dimensions have different ranges of values, and at the same time, values of different hospitals may have certain differences even in the same dimension, which may affect accuracy of machine learning model 9 prediction results, especially for features with extreme outliers, when the difference in the range of values between the two is large. For example, a binary feature has a range of values of [0, 1] , and a distance feature has a range of values of [0, ∞) , which will lead the model to prefer features with a larger range of values. In order to balance the problem, data normalization processing needs to be introduced. The method commonly used is the minimum maximum normalization method, calculated as follows:

Where x_min and x_max represents the minimum and maximum values of the feature values respectively, x′represents a normalized feature value.

In the process of data collection, due to sensor failure, manual entry error or abnormal accident, outliers may appear in features of some dimensions, which deviates significantly from reasonable range of clinical data values, and has a great impact on accuracy of the applied model 9, 10. Data screening is required via outliers pre-processing methods. Commonly used outlier detection methods mainly include distance-based methods, density-based methods, model detection based methods and cluster-based methods.

There are several common ways to process detected outliers as discussed below. A first way is to delete data record comprising the outlier directly. Although this method is simple and easy to use, it is not preferred in case of insufficient data. A second way is take the detected outliers as missing values and use missing value handling methods based on existing variable information. A third way is to use an average of two observed values to correct an outlier, especially when data sample quantity is not too big. A fourth way is replace values less than (μ-3σ) %and greater than (μ+3σ) %with values of (μ-3σ) %and (μ+3σ) %respectively, where μ is an average of the feature values, σ is the standard deviation of the feature values, as it can generally be assumed that feature values are almost entirely concentrated in the (μ-3σ, μ+3σ) interval. In an example, (μ-3σ) %is approximately 1%and (μ+3σ) %is approximately 99%.

3. Algorithm design

Based on the embryologist′sprocess of embryo scoring according to Figures 4 and 5, embodiments of the present disclosure propose a system 1 using Al technology to identify the morphology features of embryo images 5 under microscopes and to predict pregnancy success rates 12 based on the identified morphology features and patient clinical data, a structure of which is shown in Figures 6 and 7. The system can help an embryologist user 2 to fulfill tasks such as morphology identification of the number of blastomeres and fragmentation ratio of Day 3 embryos 5a, stage identification of blastocysts of Day 5 embryos 5b, and prediction of pregnancy success rates of Day 3 embryos 5a based on above mentioned morphology identification in combination with basic clinical data of patients 20. According to embryonic development, embodiments of the present disclosure divide the entire system 1 mainly for Day 3 and Day 5 identification or prediction tasks, comprising a total of three deep learning models 10 and a machine learning model 9, with each model 9, 10 responsible for fulfilling a different identification or prediction task. As shown in Figures 6 and 7, a first deep learning model 10a is configured for identifying number of blastomeres in an embryo image 5a in particular of a Day 3 embryo, a second deep learning model 10b is configured for identifying fragmentation ratio in an embryo image 5a in particular of a Day 3 embryo, a third deep learning model 10c is configured for identifying a stage of blastocyte in an image 5b in particular of the Day 5 embryo developed from the Day 3 embryo, and a machine learning model 9 is configured for pregnancy success rate prediction to obtain a predicted pregnancy success rate 12 based on the first deep learning model 10a, the second deep learning model 10b and clinical data of a patient 20.

After collecting and cleansing data obtained from multiple reproductive centers, and building uniformed standard data, the first deep learning model 10a, the second deep learning model 10b, the third deep learning model 10c and the machine learning model 9 shown in Figure 6 and 7 are trained for intelligently fulfilling respective tasks mentioned above.

In some preferred embodiments of the present disclosure, a centerNet algorithm is applied for the first deep learning model 10a, the second deep learning model 10b, and the third deep learning model 10c, and a Catboost algorithm is applied for the machine learning model 9.

Centernet is an end-to-end deep learning algorithm 10 and a single-stage target detection algorithm based on no anchor points. The algorithm filters directly on the hot spot map calculated by ResNet basic feature extraction network, removes the time-consuming Non-MaximumSuppression (NMS) post-processing operation, and thus can further improve the running speed of the algorithm, which makes it easy for users such as embryologists to obtain pregnancy success rate prediction 12 and stage of blastocyte in time after inputting the embryo image 5.

Catboost is a conventional machine learning algorithm 9, a GBDT framework with fewer parameters, support for category variables, and high accuracy implemented with symmetrical decision trees as a base learner. Catboost′smain features lie in its support for category variables, without any requirement to preprocess non-numeric features, and without any requirement for tuning parameters a lot. Default parameters may work for high model quality and very good prediction success rate, reducing time spent on tuning parameters. During model training, the algorithm can quickly scale the model to a GPU version, and quickly complete the training with GPU-based gradient lift algorithms.

Model training is divided into four main parts involving: number of blastomeres in a Day 3 embryo, fragmentation ratio of a Day 3 embryo, stage of blastocyte of the Day 5 embryo and pregnancy success rate prediction 12 for a Day 3 embryo.

A first training data set 13 is used to train the first deep learning model 10a and the training process is recoded as shown in table 1 below.

Table 1

A second training data 13 set is used to train the second deep learning model 10b and the training process is recoded as shown in table 2 below.

Table 2

A third training data 13 set is used to train the third deep learning model 10c and the training process is recoded as shown in table 3 below

Table 3

Predicted pregnancy success rate 12 from the machine learning model 9 is shown in Figure 2, where sensitivity tells us what proportion of the positive class got correctly classified. Specificity tells us what proportion of the negative class got correctly classified. A ROC curve 21 is a plot of the sensitivity in function of the specificity for different cut-off points of a parameter. The Area Under the ROC curve (AUC) 21 is a measure of how well a parameter can distinguish between two diagnostic groups. As analyzed from Figure 2, current overall accuracy of pregnancy success rate prediction 12 is 69.8%.

In the training process of the models 10, the trained models 14 are expected to not only to perform well on the training data set 13, but also perform well on new data set. Accordingly, obtained data is often divided in different ways. for the first, second and third deep learning models 10a, 10b, 10c as shown in Figure 6 and 7, obtained data is divided according to the ratio of 8∶1∶1 into training data set 13, validation data set 15 and test data set 17, to ensure that there is no intersection among the three sets. The deep learning model 9 is trained with images 5 in the training data set 13, the trained model 14 is then validated with images 5 in the validation data set 15 for evaluation, and the evaluated model 16 is then finally tested with the test data set 17. The tested 18 and therefore ready trained model 19 can then be used for an actual pregnancy success rate prediction 12. The identified results of each classification are counted to calculate mAP of the deep learning models 10. In the pregnancy success rate prediction 12, all models 10 in Figure 6 and 7 will be trained using a k-fold cross-validation method. During the data division phase, data set A containing n data records is randomly divided into k subsets of the same size A₁, A₂, ..., A_k, whereIn the course of training, A_i is taken as the test data set 17, and the rest of data in the data set A is taken as the training data set 13 to train the deep learning models 10 and this process iterates as i varies from 1 to 2, ..., k. According to statistics of test results of the deep learning models 10, the number of error results in each test set A_i is obtained, and error rate of the deep learning models 10 is calculated asThrough iterative training of the deep learning models 10, the error rate of the deep learning models 10 is reduced, and accuracy and generalization of the models are improved.

In order to better evaluate the accuracy of the deep learning models 10 in different phases, the two core indicators of mAP values and ROC curves 21 are used as the main measures.

Values calculated from formulas (1) and (2) below can be used to draw a PRcurve for a binary classification model, and the area below the PRcurve is approximated to the AP value of the binary classification model. For multi-classification models, for each classification, AP values are calculated according to formula (1) and formula (2) respectively, and the mAP value is obtained as an average AP values of each classification. The higher the value is, the better the model is, making it convenient to compare performance of multi-classification models.

The ROC curve 21 is a graph that reflects the continuous variables of sensitivity and specificity. As shown in Figure 3, the vertical coordinate indicates the sensitivity and the horizontal coordinate indicates the specificity. The lower the vertical coordicate is, the higher the accuracy of the deep learning model is. The lower the horizontal coordinate is, the lower the misjudgment rate of the deep learning models 10 is.

The straight line in Figure 3 represents the function y=x, and the dots on that straight line represent classification results for samples using a classifier with a random guessing strategy. The curve 21 indicates that the prediction ability of the deep learning models 10 is significantly better than that of the random guess strategy. On the other hand, when the curve 21 is closer to the upper left corner, the used model 10 is of a higher prediction ability. In the field of medicine, imbalance of class distribution is very common, and the ROC curve 21 is not affected by class distribution, so it is also very suitable for evaluating models 9, 10 in the field of medicine.

List of references

1 Pregnancy Success Rate Prediction system

2 User

3 Embryo (s) (cell culture)

3b Selected Embryo (s)

4 Digital camera

5 Digital pictures

5a Digital picture -Embryo Day 3

5b Digital picture -Embryo Day 5

6 Computer /Control unit

7 User Interface

8 Prediction Software

9 Machine Learning Model

10 Deep Learning Models

10a First Deep Learning Model

10b Second Deep Learning Model

10c Third Deep Learning Model

11 Individual morphology parameters/features

12 Predicted pregnancy success rate

13 Training Data Set

14 Trained Models

15 Evaluation Data Set

16 Evaluated Models

17 Test Data Set

18 Tested Models

19 Ready trained models

20 Clinical Data from a patient

21 ROC Curve

Claims

Method for aiding to select a suitable embryo (3) for implantation in an in-vitro fertilization process whereupon at least one two-dimensional digital picture (5) of an in-vitro fertilized embryo (3) is created, transferred to a computer (6) performing a software (8) with access to clinical data of a patient (20) where the embryo (3) is to be implanted, the software (8) then evaluates embryo morphology features (11) extracted from the digital pictures (5) and predicts the pregnancy success rate (12) using the evaluated embryo morphology features (11) and the clinical data of the patient (20) , wherein

the software (8) implements several trained first Artificial Intelligent (Al) models (10) to evaluate the embryo morphology features (11) and one different trained second Al model (9) to predict the pregnancy success rate (12) based on outputs of the first trained Al models (10) and the clinical data of the patient (20) .
The method of claim 1, wherein

two or three of the embryos (3) with their respective ranking scores, derived from the evaluated embryo morphology features (11) and the predicted pregnancy success rate (12) , are used as a reference to a following authority which selects at least one of those embryos (3) for implantation into the patient.
The method of one of the previous claims, wherein

a Deep Learning Object Detection Algorithm is used for the first Al models (10) and a Machine Learning Boosting Algorithm is used for the second Al model (9) .
The method of one of the previous claims, wherein

the embryo morphology features (11) extracted from the at least one digital picture (5) by the first Al models (10) comprise one or more of the following: number of blastomeres, fragmentation ratio, blastocyte stage, symmetry, inner cell mass, trophectoderm and other suitable parameters.
The method of claim 4, wherein the first Al models (10) which obtain the embryo morphology features (11) of the number of blastomeres of the embryo and the fragmentation ratio of the embryo are extracting it in a first stage.
The method of claim 5, wherein the first Al models (10) which obtain the blastocyte stage of the embryo are extracting it in a second stage later than the first stage.
The method of claim 5 or 6, wherein the first stage is within the third day of embryo development.
The method of claim 6 to 7, wherein the second stage is within the fifth day of embryo development.
The method of one of the previous claims, wherein

respective mAP values, comprising of Recall and Precision parameters for the models, are calculated and respective ROC curves (21) , which reflect sensitivity and specificity of the models (10) , are determined for each model (10) and used to evaluate and improve their respective prediction accuracy.
An apparatus for aiding to select a suitable embryo (3, 3b) for implantation in an in-vitro fertilization process according to a method of one of the previous claims, comprising:

a first obtaining component (4) , configured for obtaining a digital picture (5) of an in-vitro fertilized embryo (3) in a first stage, and clinical data of a patient (20) where the selected embryo (3b) is to be implanted;

a second obtaining component (10) , configured for obtaining embryo morphology features (11 ) from the image (5) via trained first Artificial Intelligent (Al) models (19) ;

a third obtaining component (9) , configured for obtaining a pregnancy success rate (12) from the embryo morphology features (11) and the clinical data of the patient (20) via a trained second Al model (19) .
A computer device (6) comprising a computer readable memory, a processor and a computer program (8) which is stored on the memory and may be run on the processor, wherein when executing the computer program (8) , the processor implements the method as claimed in any of claims 1 to 9.
A non-transitory computer readable storage medium storing a computer program (8) for execution by a computing device (6) having a processor, wherein the computer program (8) , when executed by the processor, cause the computing device (6) to perform the method of any of claims 1 to 9.