RU2846177C1 - Method and system for determining probability of presence of chromosomal abnormality in foetus - Google Patents

Abstract

FIELD: bioinformatics.

SUBSTANCE: group of inventions relates to bioinformatics, specifically to methods for early detection of abnormalities in a foetus during pregnancy using machine learning. Method for determining the probability of the presence of a chromosome anomaly in the foetus is performed by a computing device and comprises steps of: at the preliminary stage, labelled data are obtained in the form of an image of foetal chromosomes; the data are processed: for each image, selecting regions of the image containing points of a given colour, finding the coordinates of all the pixels of the point, calculating the median value of the coordinate X of each point, calculating the distance from the origin of coordinates to each of the pixels, median distance from the origin of coordinates to each pixel of said point is calculated for each point; the obtained data are stored in the form of a numerical vector associated with said image and a machine learning model is trained to judge on the probability of presence of a chromosomal anomaly in the foetus. At the working stage: obtaining data, performing processing and supplying the processed data to the input of a trained machine learning model, obtaining a verdict and transmitting information on the probability of having a chromosomal abnormality in the foetus to medical personnel.

EFFECT: higher probability of detecting the chromosomal anomaly in the foetus.

11 cl, 5 dwg

Description

Field of technology to which the invention relates

The group of inventions relates to the field of bioinformatics, namely to methods of machine learning, processing and analysis of data aimed at early detection of abnormalities in the fetus during pregnancy.

In modern medicine and bioinformatics, where data flows are becoming increasingly extensive and complex, the use of machine learning methods for data processing and analysis is of particular importance. One of the important tasks in the field of healthcare is the early detection of fetal abnormalities during pregnancy, which allows timely measures to prevent or minimize possible complications for the health of the child and mother.

In order to determine the presence of chromosomal abnormalities in the fetus, a non-invasive prenatal test (NIPT) is performed, the results of which form data containing a variety of information. The use of machine learning methods allows us to effectively process this data, identifying hidden patterns that may further indicate the presence of abnormalities in the fetus. This approach helps to create models that can predict the likelihood of developing specific abnormalities, which provides doctors and parents with additional information for decision-making.

During the NIPT process, Halos (software for automated bioinformatics processing of results obtained from the sequencer) builds and evaluates a chromosomal profile graph and provides a prognosis. However, this prognosis is not always perfectly accurate. In this regard, there is a need to develop tools that will help improve the accuracy of screening and reduce the likelihood of erroneous results.

With the development of machine learning technologies and data analysis algorithms, new opportunities arise to create such tools that can effectively process and interpret NIPT results. Machine learning methods can be used to create models that take into account a variety of factors, which can allow for more accurate prediction of fetal abnormalities.

The development of additional tools based on machine learning methods for analyzing NIPT results is a pressing issue in modern medical science. Improving diagnostic accuracy will allow timely detection of fetal abnormalities and taking appropriate measures to protect the health of both the mother and the unborn child.

State of the art

A technology for performing non-invasive prenatal testing using an artificial intelligence model is known (US 2022343178 A1, “Method and system for performing non-invasive genetic testing using an artificial intelligence (AI) model” PRESAGEN PTY LTD, 27.10.2022). An artificial intelligence (AI)-based computing system is used to non-invasively assess the presence of a range of aneuploidies and mosaicism in an embryo before implantation through image analysis. Aneuploidies and mosaicisms with similar risks of adverse outcomes are grouped, and training images are labeled with their group. Separate AI models are trained for each group using the same training dataset, and then the individual models are combined, for example, using an ensemble or distillation approach, to develop a model that can identify a wide range of aneuploidy and mosaicism risks. The AI model for the group is generated by training multiple models, including binary models, hierarchical multi-level models, and a multi-class model. Specifically, hierarchical multi-level models are generated by assigning quality labels to images. At each level, the training set is split into the best quality images and other images. The model at that level is trained on the best quality images, and the other images are passed to the next level, and the process is repeated (so that the remaining images are split into the next best quality images and other images). The final model can then be used to non-invasively detect aneuploidy and mosaicism, and their associated risk of adverse outcomes, using pre-implantation embryo imaging.

The disadvantage of this technology, compared to the one described below, is the use of a separate model for each group of images (united by common deficiencies, pathologies). Thus, several different training models are used, which significantly increases the required computing power and analysis time. At the same time, as is known, each model has its own accuracy of verdict, and combining several models into one can lead to an increase in the number of false verdicts.

A system for detecting the probability of a chromosomal abnormality in a fetus is known from the prior art (US 12020492 B2, "Cell analysis method, cell analysis device, and cell analysis system", SYSMEX CORP, 06/25/2024). The present invention proposes a cell analysis method, a cell analysis device, a cell analysis system, and a cell analysis program that facilitate the analysis of a plurality of analysis elements. The cell analysis method generates data for analyzing cells contained in a sample, selects an artificial intelligence algorithm from a plurality of artificial intelligence algorithms, inputs the generated data for analysis into the selected artificial intelligence algorithm, and generates, using the selected artificial intelligence algorithm, data indicating the properties of cells based on the generated data for analysis.

This solution mainly considers only the analysis of cells, and the result of the study based on artificial intelligence models does not result in a verdict on the presence of any chromosomal pathologies and defects.

A method for detecting the probability of a chromosomal abnormality in a fetus is known from the prior art (CN 110265087 A, "Chromosome abnormality detection model" UNIV MEDICAL HOSPITAL CHINA, 09/20/2019). This method is the closest analogue of the developed method and can be selected as a prototype. The invention is a chromosomal abnormality detection system. The chromosomal abnormality detection system consists of an image acquisition unit and a non-transient machine-readable medium. The image acquisition unit is used to obtain an image of the division metaphase of the target chromosome of the subject's cell. The non-transient machine-readable medium is used to store the program, and when executed by the processor, the program is used to determine whether the subject has chromosomal abnormalities. Thus, the chromosomal abnormality detection system can effectively improve the accuracy and sensitivity of chromosomal abnormality detection and reduce the time for assessing the presence or absence of a chromosomal abnormality in a subject.

However, this technical solution does not disclose the pre-processing of the images obtained as a result of testing, and as is known and will be reflected below, real images contain a large amount of information, which during subsequent data processing can affect the accuracy of the verdict. Thus, for direct training of the model, an important stage is pre-processing, which allows increasing the accuracy by excluding data that do not contribute to the characteristics of the chromosome.

The proposed solution is free from the listed shortcomings identified during the analysis of the state of the art. It is aimed at solving the problem of creating a tool for more accurate processing of results obtained in the form of profile graphs of chromosomes to determine the presence of chromosomal abnormalities in the fetus based on marked graphic data.

Revealing the essence

The technical result of the developed method and system is an increase in the accuracy of detecting chromosomal abnormalities in the fetus based on the results of non-invasive prenatal testing.

The technical result is achieved by using a method for identifying the probability of the presence of a chromosomal abnormality in a fetus and a system for identifying the probability of the presence of a chromosomal abnormality in a fetus, capable of implementing the claimed method.

One of the variants of the method for detecting the probability of the presence of a chromosomal abnormality in the fetus, performed by a computing device, is implemented by performing the following steps, in which:

at the preliminary stage:

• receive labeled data, where the data are images of fetal chromosomes,

• perform processing of the marked data, and for each image:

- select areas of the image containing dots of a given color,

- find the coordinates of all pixels that make up the points of a given color,

- calculate the median value of the X coordinate of each point of a given color,

- calculate the distance from the origin to each of the pixels that make up the points of a given color,

- calculate for each point of a given color the median value of the distances from the origin to each of the pixels that make up the given point,

- save the obtained data in the form of a numerical vector, assigned to the given image,

• train at least one machine learning model on the processed data to make a verdict on the probability of the presence of a chromosomal abnormality in the fetus;

at the working stage:

• receive data, where the data are images of fetal chromosomes,

• perform data processing,

• feed the processed data into the trained machine learning model,

• receive the verdict of the trained machine learning model,

• transmit the information contained in the verdict of the trained machine learning model about the probability of the presence of a chromosomal abnormality in the fetus to authorized medical personnel.

In another version of the method for identifying the probability of the presence of a chromosomal abnormality in a fetus, performed by a computing device, the following steps are implemented:

at the preliminary stage:

• receive labeled data, where the data are images of fetal chromosomes,

• perform processing of the marked data, and for each image:

- perform cropping, as a result of which only the area of the graphs remains on the image,

- replace all pixels that are black or gray with white pixels,

- save for each remaining pixel its coordinates along the X and Y axes, as well as the color in the RGB color space,

- cluster all pixels, assigning each of them, based on its color in the RGB color space, to one of four classes:

• pixels whose color matches the pre-known color of the T-score line,

• pixels whose color matches the previously known color of the line and Segment,

• pixels whose color matches the color resulting from mixing the color of the T-score line and the color of the Segment line,

• all other pixels,

- calculate the value of Tsmed as the average deviation from zero of the Y coordinate for all points that make up the T-score line,

- calculate the value of Tsso as the standard deviation of the Y coordinate for all points that make up the T-score line,

- the result of the Dickey-Fuller test is calculated at a significance level of 5% in relation to the points that make up the T-score line,

- calculate the value of Sgmed as the average deviation from zero of the Y coordinate for all points that make up the Segment line,

- calculate the value of Sgso as the standard deviation of the Y coordinate for all points that make up the Segment line,

- calculate the modal value Sgmd, that is, the value of the Y coordinate that is most often found at the points that make up the Segment line,

- calculate the proportion Sgd of points that make up the Segment line, for which the Y value is within plus or minus 5 pixels of the modal value Sgmd,

- save the obtained data in the form of a numerical vector, assigned to the given image,

• train at least one machine learning model on the processed data to make a verdict on the probability of the presence of a chromosomal abnormality in the fetus;

at the working stage:

• receive data, where the data are images of fetal chromosomes,

• perform data processing,

• feed the processed data into the trained machine learning model,

• receive the verdict of the trained machine learning model,

• transmit the information contained in the verdict of the trained machine learning model about the probability of the presence of a chromosomal abnormality in the fetus to authorized medical personnel.

In particular cases of the implementation of the above-mentioned variants of methods for identifying the probability of the presence of a chromosomal abnormality in the fetus, performed by a computing device:

images of fetal chromosomes are generated using bioinformatics software;

the marked data contain both the results of examination of fetuses without chromosomal abnormalities and the results of examination of fetuses with chromosomal abnormalities;

The data are PNG images obtained as a result of non-invasive prenatal testing;

during data processing additionally

• convert an image from RGB color space to HSV color space,

• exclude from the set of found pixels those pixels that are outside the zone that has pre-set coordinates along the x-axis and y-axis;

at least one machine learning model is at least one of the following models:

• decision trees,

• decision trees using gradient boosting, CatBoost,

• decision trees using gradient boosting, LightGBM,

• linear regression,

• logistic regression,

• K Nearest Neighbors Classifier,

• Multilayer Perceptron, MLP Classifier,

• Naive Bayes classifier, Gaussian NB,

• support vector machines, SVC,

• decision trees with Ada correction algorithm, DTC with Ada,

• random forest, Random Forest Classifier;

As a metric for the quality of training, the model uses one of the following values:

• sensitivity,

• True Positive Rate,

• recall score,

• overall accuracy;

Model training is considered successful if the overall accuracy exceeds 85%;

The information contained in the verdict of the trained machine learning model about the probability of the presence of a fetal anomaly is transmitted to authorized medical personnel in one of the following ways:

• by displaying a message in the user interface of the system implementing this method,

• by sending an email message,

• by sending SMS or MMS,

• by sending a push notification,

• through an API that implements the exchange of information between the system implementing this method and other information systems.

A system for identifying the probability of the presence of a chromosomal abnormality in a fetus has been created, containing at least:

• processor,

• RAM,

• data bus,

• input/output means,

• input/output interfaces,

• read-only memory,

• a device storing machine-readable instructions, the execution of which by the processor results in the implementation of one of the methods described above.

Brief description of the drawings

The accompanying drawings, which are included to provide further understanding of the solution, show embodiments of the invention and serve to explain the principles of the solution.

The stated solution is illustrated by the following drawings, which show:

Fig. 1 illustrates a block diagram of the algorithm of one of the variants of implementing the preparatory stage of the method for identifying the probability of the presence of a chromosomal abnormality in the fetus.

Fig. 2A, 2B illustrate examples of images of fetal chromosomes without anomalies and with anomalies, respectively.

Fig. 2B, 2G illustrate examples of processed images of fetal chromosomes.

Fig. 3 illustrates a block diagram of the algorithm of another possible embodiment of the preparatory stage of the method for identifying the probability of the presence of a chromosomal abnormality in the fetus.

Fig. 4 illustrates the algorithm of the working stage of the described method.

Fig. 5 illustrates an example of a general diagram of a computing device that provides the data processing necessary for implementing the claimed solution.

Implementation of the invention

Below is a description of exemplary embodiments of the claimed solution.

The objects and features of the present solution, the methods for achieving these objects and features will become obvious by reference to exemplary embodiments. However, the present solution is not limited to the exemplary embodiments disclosed below, it can be embodied in various forms. The essence given in the description is nothing more than specific details provided to help a person skilled in the art in a comprehensive understanding of the invention, and the present solution is defined only in the scope of the appended claims.

The method for identifying the probability of a chromosomal abnormality in a fetus begins with a preliminary stage. According to the method described, the preliminary stage can be performed in either of two equivalent, independent variants.

In one possible embodiment, a preliminary step 100 is performed, which will be described below with reference to the drawing Fig. 1.

In another possible embodiment, a preliminary step 300 is performed, which will be described below with reference to the drawing Fig. 3.

As shown in Fig. 1, the preliminary step 100 begins with step 110, in which labeled data is obtained, wherein these data are images of fetal chromosomes. Examples of such an image are shown in Fig. 2A, Fig. 2B.

The labeled data is obtained in any well-known way.

For example, the system implementing the method may be designed to receive data over a computer network, such as the Internet. The labeled data in step 110 may be received by the system implementing the method by sending via e-mail, automatically downloading from a remote web server that functions as a file storage, or by any other method known from the prior art.

The input data in the described method may be images obtained from a sequencer, such as a BGISEQ-500 sequencer or a DNBSEQ-G400 sequencer, operating in conjunction with the Halos software. This equipment and software are named as a non-limiting example of implementation; in the claimed method, it is possible to use data obtained from any other functionally similar software and hardware complex.

The images themselves can be of any known computer format, for example, PNG.

It should be noted that the data obtained in step 110 is labeled. This means that the resulting data set contains both images of fetal chromosomes that correspond to developmental norms (i.e., these are images that do not contain chromosomal abnormalities) and images of fetal chromosomes that indicate any developmental pathologies (i.e., these are images that contain chromosomal abnormalities).

The resulting labeled data is saved in any generally known manner, and step 110 is completed.

The method proceeds to step 120, where the obtained labeled data is processed by sequentially taking one image after another from the set of labeled data.

On each image, as shown in Fig. 1, at step 122, areas are selected that contain dots of a predetermined color, for example, blue (similar dots and the nature of their arrangement on the original images are shown in Fig. 2A, Fig. 2B). A specific color, for example, blue (RGB 0, 0, 255) is selected in advance, when designing a system implementing the method, depending on the colors that a specific model of the sequencer and the software used by it uses to form images that serve as the original data of the described method. Where further in the text the colored dots on the analyzed images are called blue, it should be understood that this is done only for simplicity of description; this is a non-limiting example, and the specific color can be any of those present in the RGB color space.

To select areas in step 122 by any generally known method, for example, by a previously prepared mask based on the range of blue color in the color space of the image, select areas containing blue pixels. It should be noted that in some cases, this step may additionally require converting the analyzed image, for example, from the RGB color space to the HSV color space. Such conversion is performed, if necessary, by any generally known method, for example, by calling the RGB-to-HUE function of the DirectX system utility with a previously prepared script.

The selected areas of the image containing the blue dots are saved, and this completes step 122.

The method proceeds to step 124, where the coordinates of all pixels that make up the blue dots found in the previous step are found.

At step 124, the coordinates of the pixels found at the previous step using the blue mask are found. Thus, at this step, we are dealing with pixels that are obviously of the desired color, and the task is only to determine their coordinates.

This can be done in any well-known way, for example, by means of a pre-prepared script that sequentially goes through all the available pixels of the image and records the coordinates of those that are blue.

It is possible to implement the described method in such a way that at this step the selected pixels are additionally filtered by coordinates in any generally known way in order to exclude blue pixels located outside the graph. Such filtering includes setting threshold values for the X values and along the Y axis, as well as removing pixels whose coordinates are within the set thresholds. Filtering of coordinate data can be carried out in any generally known way from the state of the art.

After storing all the obtained blue pixel coordinates, step 124 is completed.

The method then moves to step 125, where the median value of the X coordinate is calculated for each blue point. To do this, the X coordinate values are ordered in ascending order in any generally known manner, and a value is selected that is greater than approximately half of the values and also less than approximately half of the values. The calculated median values of the X coordinates for all blue points found are saved, after which step 125 is completed.

The graph 200 shown in Fig. 2B shows a graphical representation of the result of performing steps 122-125, which has the form of a set of points 201 in the drawing.

After step 125, the method proceeds to step 126, where the distances from the origin (0,0) to each of the pixels that make up the blue dots are calculated. For clarity, step 126 is illustrated by the drawing Fig. 2G, which shows a graph 210. A set of straight line segments 211, each of which connects the origin (0;0) and the current (x;y) coordinates of the next pixel, is drawn along the shortest path, i.e., each segment from the set of segments 211 is a graphical representation of the distance from the origin to a specific pixel.

The calculation of distances D _i is carried out in any well-known way according to the well-known formula

D _i = ,

where X _i , Y _i are the coordinates of the next pixel.

The calculated distances are saved, and this completes step 126.

The method then moves to step 127, where for each point of a given, for example, blue color, the median value of the distances from the origin to each of the pixels that make up the given point is calculated. To do this, in any generally known way, the distances D _i calculated in the previous step are ordered in ascending order for each point, and a value is selected that is greater than approximately half of the values and also less than approximately half of the values.

After step 127, step 128 is performed, where the obtained data is saved as a numerical vector, and the obtained vector is put in correspondence with the processed image. In this way, a training sample is formed, where the calculated median distances are used as features for classifying the presence or absence of anomalies in the chromosomes of the fetus.

Next, the preliminary step 100 proceeds to step 129, where it is determined whether the processed image was the last one from the set of received data. In response to the fact that the processed image was not the last one, i.e., the processing of the set of received data is not yet complete, they return to step 120, where the next image is selected for processing. In response to the fact that the processed image was the last one, they proceed to step 130.

In step 130, at least one machine learning model is trained on the processed data to make a verdict on the probability of the presence of a chromosomal abnormality in the fetus.

The machine learning model is trained in step 130 on the training sample obtained during the above-described image preprocessing. The criterion for maximizing classification accuracy is used, as well as cross-validation checks. The training is considered successful if the accuracy of predictions on the test sample exceeds a predetermined value. In one non-limiting example, the threshold value for prediction accuracy may be chosen to be 92%. In another non-limiting example, this value may be chosen to be 85%.

The actual training of the model is carried out in any generally known way.

In various possible implementation options of the described solution, the following can be used as a training model: decision trees, decision trees using gradient boosting, CatBoost, decision trees using gradient boosting, LightGBM, linear regression, logistic regression, K Nearest Neighbors Classifier, KNeighbors Classifier, multilayer perceptron, MLP Classifier, naive Bayes classifier, Gaussian NB, support vectors, SVC, decision trees with Ada correction algorithm, DTC with Ada, random forest, Random Forest Classifier.

Once the machine learning model has been trained to make a verdict on the probability of a fetus having a chromosomal abnormality, the trained model is saved. This completes step 130 and pre-step 100.

Below, with reference to Fig. 3, a preliminary step 300 will be described, which represents an alternative embodiment of pre-processing images obtained in a set of pre-labeled data.

The preliminary stage 300 is also performed in order to obtain a training sample and train the machine learning model. The difference between the preliminary stage 300 and the preliminary stage 100 described above is that in this case, the features used are not the coordinates of the points of a given color, but the behavioral features of the T-score and Segment curves on the images obtained as the initial data of the method.

The T-score (standard deviation) and Segment curves are constructed on the original images by sequencers. The algorithm for constructing these curves is a commercial secret of the sequencer manufacturers, but within the framework of the described method this circumstance is insignificant, since the features to be extracted and further analyzed are contained not in the construction methods, but in the form and location of the already constructed curves on the image.

As shown in Fig. 3, the preliminary step 300 begins with step 310, in which labeled data is obtained, wherein the data are images of the chromosomes of the fetus. This step has no special features and is performed completely analogously to the step 110 described above.

After completing step 310, step 300 proceeds to step 320, in which the obtained labeled data is processed by sequentially taking one image after another from the set of labeled data.

On each image, as shown in Fig. 3, at step 321 cropping is performed, as a result of which only the graph area remains on the image. Each image contains a graph plotting area with a legend (the inner area in a rectangular frame) and axis labels. The key information on each image is limited to the graph plotting area, so each image is cropped, leaving only the graph area. At step 321 cropping is performed, as a result of which only the graph area remains on the image. This process is performed in any well-known way, for example, by calling a pre-prepared Python script. The cropping itself, i.e. cropping the image by specified coordinates, can be implemented, for example, by calling the corresponding function of the open OpenCV library with arguments representing the X and Y coordinates of the opposite corners of the rectangular fragment of the image that should remain as a result of the described operation.

The cropped images are saved and the preliminary step 300 proceeds to step 322. In step 322, all pixels that are black or gray are replaced with white pixels. Thus, as a result of step 322, the legend area of each graph and the grid lines (pixels from dark gray to black shades) will be filled with white. After the described processing, each image will contain only information about the T-score and Segment lines.

The replacement is performed in any well-known way, for example, by means of a previously prepared script that alternately goes through the pixels of the image and changes these values for all pixels related to dark gray or black colors, i.e. having RGB component values in the range from (0, 0, 0) to, for example, (30, 30, 30) to those corresponding to white color, i.e. (255, 255, 255). A specific shade of dark gray color, in the given example corresponding to RGB (30, 30, 30), can be selected in advance depending on the parameters of the original images and be some other, for example, (50, 50, 50) or some other.

The modified images are saved and the preliminary step 300 proceeds to step 323. In step 323, the coordinates of each non-white pixel along the X and Y axes, as well as its color in the RGB color space, are saved. This can be done in any well-known way, for example, by means of a script prepared in advance, successively going through all available pixels of the image and recording the coordinates and values of the color components R, G, B for each pixel for which these values differ from the RGB components of white color (255, 255, 255).

The preliminary step 300 then proceeds to step 324, where the image pixels are clustered according to their color. It should be noted that all colors in the T-score and Segment graphs after all the processing steps described above have been performed can belong to one of four classes:

1. Pixel colors that are related exclusively to the T-score line.

2. Pixel colors that are specific to the Segment line.

3. Pixel colors related to the intersection of the T-score and Segment lines.

4. Pixel colors that do not belong to any of the T-score and Segment lines.

Clustering is performed using a pre-trained machine learning model. To do this, a training sample consisting of images similar to those used as input in the described method is created in advance, before the method starts working, and the training sample is manually marked.

Manual labeling of the training sample consists of the operator evaluating all pixels related to any curves, evaluating their RGB values, and manually assigning one of the above-mentioned four classes to each such pixel. In this way, a data array is formed that contains the colors of the pixels in the RGB system and the class of each pixel. This array of labeled data is used to train the model to recognize the type (class) of a pixel by its color in RGB coordinates. Any of the following can be used as a training model:

• Logistic regression, in any known implementation,

• Decision trees, in any known implementation,

• Random forest, in any known implementation,

• LightGBM (decision trees using gradient boosting), in any known implementation,

• CatBoost (decision trees using gradient boosting), in any known implementation.

The overall accuracy (the proportion of correctly identified classes) is used as a metric for the quality of the model. The cross-validation approach is used for training. Cross-validation is a method for assessing the performance of a machine learning model, used to test its ability to generalize data to new, previously unseen cases. The essence of the method is to divide the existing data set into two or more parts. One part is used to train the model, and the other is used to test its quality. This process is repeated several times using different data partitions to obtain a more reliable assessment of the model's performance. In one possible implementation of the described method, the original array was divided into 5 equal parts for cross-validation.

At step 324, the analyzed image is transferred to a model previously trained in the specified manner. The model interprets the color coordinates (RGB components) of the pixels and assigns each of them to one of the above-described classes, thus performing classification. It should be noted that a number of possible classification errors will not significantly affect the final result of the graph digitization (in particular, this concerns errors when a pixel located at the intersection of the T-score and Segment lines is assigned exclusively to one of these lines). Technically, this means that at this step, the coordinates of the pixels belonging to each of the classes, based on the results of step 340, are saved separately, for example, in a section of the database dedicated to this. Thus, upon completion of this step, the coordinates of the pixels, each of which belongs to one of the above-mentioned classes, will be obtained and saved.

In other words, step 324 results in lists of pixels that make up the T-score and Segment lines, as well as the X and Y coordinates of each of them.

Then, the preliminary step 300 proceeds to step 325. In step 325, the following quantities are calculated: the average deviation from zero of the Y coordinate of the points that make up the T-score line, the standard deviation of the Y coordinate of the points that make up the T-score line, the result of the Dickey-Fuller test at a significance level of 5% with respect to the Y coordinates of the points that make up the T-Score line, the average deviation from zero of the Y coordinate for all the points that make up the Segment line, the standard deviation of the Y coordinate for all the points that make up the Segment line, the value of the Y coordinate that occurs most frequently among the points that make up the Segment line, and the proportion of the points that make up the Segment line whose Y value is within plus or minus 5 pixels of the modal value.

It should be noted that the list of methods for calculating the named quantities given below should not be considered as an indication of the order in which the calculations are performed. The order of calculation can be arbitrary, including a number of quantities can be calculated in parallel.

The calculations are performed as follows:

The value of Tsmed, that is, the average deviation from zero of the Y coordinate for all points that make up the T-score line, a parameter that takes into account the deviation of the T-score from zero, is calculated using the formula

Where Y coordinate for the i-th point related to the T-score line

points related to the T-score line in the analyzed image.

The value of Tsso, that is, the standard deviation of the Y coordinate for all points that make up the T-score line, a parameter that takes into account the degree of scatter of T-score values, is calculated using the formula:

The result of the Dickey-Fuller test for the points that make up the T-score line is a binary value: 1 if the series is stationary, and 0 if the series is not stationary. Generally speaking, in econometrics, stationarity is a property of a time series that means that its mean and standard deviation do not change over time. In this case, the data are not a time series. However, stationarity in this context can be interpreted as the constancy of the mean and standard deviation of the T-score relative to the chromosomes.

Obtaining the Dickey-Fuller test result for a series of Y-coordinate data points related to the T-score line consists of the following substeps:

1. Based on a series of data on the Y-coordinates of points related to the T-score line, the parameter is estimated using the least squares method. in a regression of the form:

random error.

2. Using the meaning , we obtain a number of predicted values:

and estimate a series of squared errors of the model:

.

3. Using the obtained values, calculate the test statistics. :

Where total number of points in the T-score curve,

– the Y coordinate of a specific point on the pre-processed graph,

e – model error.

4. The resulting value are compared with the critical value from the special tables of Dickey and Fuller. If greater than the table value at a significance level of 5%, then the series is stationary, i.e. the value of the feature is equal to 1. Otherwise, the series is not stationary and the value of the feature is taken to be equal to 0.

The value of Sgmed, the average deviation from zero of the Y coordinate for all points that make up the Segment line, is calculated using the formula

Y coordinate for the i-th point related to the line Segment,

points related to the Segment line on the analyzed image

The value of Sgso, the standard deviation of the Y coordinate for all points that make up the Segment line, is calculated using the formula:

The modal value of Sgmd, that is, the Y coordinate value that occurs most often in the points that make up the Segment line, is calculated in any well-known way by counting the number of points of the Segment line that have the same Y coordinate and selecting the coordinate value that corresponds to the largest number of points. Since the Segment line is parallel to the X axis with possible delta breaks, the modal value allows us to understand how much the longest continuous section of the Segment line deviates from zero.

The proportion Sgd of points making up the line Segment, for which the value of Y is within plus or minus 5 pixels of the previously calculated modal value Sgmd, is calculated by any well-known method, by counting the number of points Nm satisfying the said condition, and then calculating the ratio of this number to the total number of points N:

Sgd = Nm/N.

This metric is designed to account for delta breaks in the Segment series. The larger the proportion of values in the vicinity of the modal value, the smaller the number of breaks and/or the shorter the sections with these breaks. The +/- 5 pixel range is necessary because a straight Segment line is typically 4–5 pixels wide, i.e. 4–5 rows.

At this point, step 325 is completed and the preliminary stage 300 method proceeds to step 326, where the obtained data is stored in the form of a numerical vector associated with the given image.

Next, the preliminary step 300 proceeds to step 327, where it is determined whether the processed image was the last one from the set of received data. In response to the fact that the processed image was not the last one, i.e., the processing of the set of received data is not yet complete, it is returned to step 320, where the next image is selected for processing. In response to the fact that the processed image was the last one, it is proceeded to step 330.

In step 330, at least one machine learning model is trained on the processed data to make a verdict on the probability of the presence of a chromosomal abnormality in the fetus. This is done in a completely analogous manner to that described above in relation to step 130.

Regardless of whether the described method implements the preliminary stage 100 or the preliminary stage 300, after training at least one machine learning model, the preliminary stage is completed.

Below, with reference to the drawing Fig. 4, the working step 400 of the method for detecting the probability of the presence of a chromosomal abnormality in a fetus will be described.

As shown in Fig. 4, the working step 400 begins with step 410, in which data is obtained, wherein these data are images of fetal chromosomes. An example of such data is illustrated in Fig. 2B. This is performed in a completely analogous manner to the steps 110 and 310 described above.

The next step 420 is data preprocessing.

Depending on whether the preliminary step 100 or the preliminary step 300 was implemented, in step 420 the data is pre-processed either in accordance with the algorithm of the preliminary step 100 described above or in accordance with the algorithm of the preliminary step 300 described above.

In any of the possible embodiments of the method, as a result of step 420, numerical vectors are obtained and stored, each of which corresponds to one of the images of the fetal chromosomes obtained in step 410.

At this point, step 420 is completed and the method proceeds to step 430, where the numerical vectors obtained in step 420 are fed to the input of the machine learning model trained in the preliminary stage.

Step 440 results in a verdict from the trained machine learning model on the probable presence or absence of fetal developmental abnormalities based on the examined images of fetal chromosomes. The verdict may be, for example, a number between 0 and 1, with the closer this value to 1, the higher the probability of the presence of a chromosomal abnormality in the fetus whose chromosome images were obtained in step 410.

The resulting verdict is saved in any generally known manner.

At this point, step 440 is completed and the method 400 proceeds to step 450, where the obtained verdict of the trained machine learning model on the probability of the presence of a chromosomal abnormality in the fetus is transmitted to the authorized medical personnel. The transmission of the verdict can be performed by any generally known method. For example, the verdict can be transmitted by displaying a message in the user interface of the system implementing this method, by sending a message by e-mail, by sending an SMS or MMS, by sending a push notification, or by means of an API implementing the exchange of information between the system implementing this method and other information systems.

This completes step 450 and method 400.

Thus, the method for identifying the probability of the presence of a chromosomal abnormality in a fetus includes two stages: preliminary and operational, wherein the preliminary stage can be performed in any of the alternative variants 100 and 300 described above with reference to Fig. 1, Fig. 3. Regardless of which of the alternative variants 100 and 300 is selected, at the preliminary stage, the machine learning model is directly trained on numerical vectors obtained from a pre-labeled sample of fetal chromosome images, from which images, in various ways, the data that will allow the trained model to assess the presence or absence of abnormalities in the fetus are extracted and processed. At the operational stage 400, this trained machine learning model is used for directly identifying and providing information containing information on the probability of the presence of fetal abnormalities.

Fig. 5 shows an example of a general diagram of a computing device (500) that provides data processing necessary for implementing the claimed solution.

In this context, devices similar to those shown in the general diagram (500) are understood to mean any computing devices built on the basis of software and hardware, such as personal computers, servers, smartphones, laptops, tablets, etc.

A data processing device may be a processor, microprocessor, computer (electronic computer), PLC (programmable logic controller) or integrated circuit, configured to execute specific commands (instructions, programs) for data processing. The processor may be multi-core, for parallel data processing.

Memory devices may include, but are not limited to, hard drives (HDD), flash memory, ROM (read-only memory), solid-state drives (SSD), etc.

It should be noted that the said device may also include any other devices known in the art, such as sensors, data input/output devices, display devices (displays), etc. The data input/output device may be, but is not limited to, a mouse, keyboard, touchpad, stylus, joystick, trackpad, etc.

In general, the device (500) comprises such components as one or more processors (501), at least one random access memory or memory (502), data storage means (503), input/output interfaces (504), I/O means (505), network interaction means or, which is the same, data transmission means (506).

The processor (501) of the device performs the basic computing operations necessary for the functioning of the device (500) or the functionality of one or more of its components. The processor (501) executes the necessary machine-readable instructions contained in the RAM (502).

Memory (502) is typically in the form of RAM and contains the necessary software logic to provide the required functionality.

The data storage device (503) can be implemented in the form of HDD, SSD disks, RAID array, network storage, flash memory, optical storage devices (CD, DVD, MD, Blue-Ray disks), etc.

Interfaces (504) are standard means for connecting and working with the server part, for example, USB, RS232, RJ45, LPT, COM, HDMI, PS/2, Lightning, FireWire, etc. The choice of interfaces (504) depends on the specific design of the device (500), which can be a personal computer, mainframe, server cluster, thin client, smartphone, laptop, etc.

The following may be used as I/O data means (505): keyboard, joystick, display (touch display), projector, touchpad, mouse, trackball, light pen, speakers, microphone, etc.

Network interaction means (506) are selected from devices that ensure the reception and transmission of data over a network, for example, an Ethernet card, a WLAN/Wi-Fi module, a Bluetooth module, a BLE module, an NFC module, an IrDa module, an RFID module, a GSM modem, etc. With the help of the means (506), the organization of data exchange over a wired or wireless data transmission channel is ensured, for example, WAN, PAN, LAN, Intranet, Internet, WLAN, WMAN or GSM.

The components of the device (500) are connected via a common data transmission bus (510).

The systems and methods illustrated in this document may be described in terms of functional block components. It should be appreciated that such functional blocks may be implemented by any number of hardware and/or software components configured to perform the specified functions. For example, the system may use various integrated circuit components, such as memory elements, processing elements, logic elements, lookup tables, etc., that may perform a variety of functions under the control of one or more microprocessors or other control devices. Similarly, the software elements of the system may be implemented using any programming or scripting language, such as C, C++, C#, Java, JavaScript, VBScript, Macromedia Cold Fusion, COBOL, Microsoft Active Server Pages, assembly language, Perl, PHP, AWK, Python, Visual Basic, SQL stored procedures, PL/SQL, any UNIX shell scripts, and Extensible Markup Language (XML), with the various algorithms being implemented with any combination of data structures, objects, processes, procedures, or other software elements.

It should also be noted that the system implementing the described method can use any number of known technologies for data transmission, transmission of service signals, data processing, network management, etc.

In conclusion, it should be noted that the information provided in the description is only examples that do not limit the scope of the present invention described by the claims. A person skilled in the art will understand that there may be other embodiments of the present invention consistent with the essence and scope of the present invention.

The present application materials present a preferred disclosure of the implementation of the claimed technical solution, which should not be used as limiting other, particular embodiments of its implementation that do not go beyond the scope of the requested scope of legal protection and are obvious to specialists in the relevant field of technology.

Claims (82)

1. A method for determining the probability of the presence of a chromosomal abnormality in a fetus, performed by a computing device and comprising the steps of: at the preliminary stage: receive labeled data, where the data are images of fetal chromosomes, perform processing of the marked data, and for each image: - select areas of the image containing dots of a given color, - find the coordinates of all pixels that make up the points of a given color, - calculate the median value of the X coordinate of each point of a given color, - calculate the distance from the origin to each of the pixels that make up the points of a given color, - calculate for each point of a given color the median value of the distances from the origin to each of the pixels that make up the given point, - save the obtained data in the form of a numerical vector, assigned to the given image, train at least one machine learning model on the processed data to make a verdict on the probability of the presence of a chromosomal abnormality in the fetus; at the working stage: receive data, and the data are images of fetal chromosomes, perform data processing, feed the processed data into the trained machine learning model, receive the verdict of the trained machine learning model, transmit the information contained in the verdict of the trained machine learning model about the probability of the presence of a chromosomal abnormality in the fetus to authorized medical personnel. 2. A method for determining the probability of the presence of a chromosomal abnormality in a fetus, performed by a computing device and comprising the steps of: at the preliminary stage: receive labeled data, where the data are images of fetal chromosomes, perform processing of the marked data, and for each image: - perform cropping, as a result of which only the area of the graphs remains on the image, - replace all pixels that are black or gray with white pixels, - save for each remaining pixel its coordinates along the X and Y axes, as well as the color in the RGB color space, - cluster all pixels, assigning each of them, based on its color in the RGB color space, to one of four classes: pixels whose color matches the pre-known color of the T-score line, pixels whose color matches the previously known color of the line and Segment, pixels whose color matches the color resulting from mixing the color of the T-score line and the color of the Segment line, all other pixels, - calculate the value of Tsmed as the average deviation from zero of the Y coordinate for all points that make up the T-score line, - calculate the value of Tsso as the standard deviation of the Y coordinate for all points that make up the T-score line, - the result of the Dickey-Fuller test is calculated at a significance level of 5% in relation to the points that make up the T-score line, - calculate the value of Sgmed as the average deviation from zero of the Y coordinate for all points that make up the Segment line, - calculate the value of Sgso as the standard deviation of the Y coordinate for all points that make up the Segment line, - calculate the modal value Sgmd, that is, the value of the Y coordinate that is most often found at the points that make up the Segment line, - calculate the proportion Sgd of points that make up the Segment line, for which the Y value is within plus or minus 5 pixels of the modal value Sgmd, - save the obtained data in the form of a numerical vector, assigned to the given image, train at least one machine learning model on the processed data to make a verdict on the probability of the presence of a chromosomal abnormality in the fetus; at the working stage: receive data, and the data are images of fetal chromosomes, perform data processing, feed the processed data into the trained machine learning model, receive the verdict of the trained machine learning model, transmit the information contained in the verdict of the trained machine learning model about the probability of the presence of a chromosomal abnormality in the fetus to authorized medical personnel. 3. The method according to any one of paragraphs 1, 2, characterized in that the images of the fetal chromosomes are generated using bioinformatics software. 4. The method according to any of paragraphs 1, 2, characterized in that the marked data contain both the results of examination of fetuses without chromosomal abnormalities and the results of examination of fetuses with chromosomal abnormalities. 5. The method according to any one of paragraphs 1, 2, characterized in that the data are PNG format images obtained as a result of performing non-invasive prenatal testing. 6. The method according to any of paragraphs 1, 2, characterized in that during data processing additionally: convert an image from RGB color space to HSV color space, exclude from the set of found pixels those pixels that are outside the zone that has pre-set coordinates along the x-axis and the y-axis. 7. The method according to any one of paragraphs 1, 2, characterized in that at least one machine learning model is at least one of the following models: decision trees, decision trees using gradient boosting, CatBoost, decision trees using gradient boosting, LightGBM, linear regression, logistic regression, K Nearest Neighbors Classifier, Multilayer Perceptron, MLP Classifier, Naive Bayes classifier, Gaussian NB, support vector machines, SVC, decision trees with Ada correction algorithm, DTC with Ada, random forest, Random Forest Classifier. 8. The method according to any of paragraphs 1, 2, characterized in that one of the following values is used as a metric for the quality of training of the model: sensitivity, True Positive Rate, recall score, overall accuracy. 9. The method according to paragraph 8, characterized in that the training of the model is considered successful if the overall accuracy exceeds 85%. 10. The method according to any of paragraphs 1, 2, characterized in that the information contained in the verdict of the trained machine learning model about the probability of the presence of a fetal anomaly is transmitted to authorized medical personnel in one of the following ways: by displaying a message in the user interface of the system implementing this method, by sending an email message, by sending SMS or MMS, by sending a push notification, through an API that implements the exchange of information between the system implementing this method and other information systems. 11. A system for detecting the probability of the presence of a chromosomal abnormality in a fetus, comprising at least: CPU, RAM, data bus, input/output facilities, input/output interfaces, read-only memory, a device storing machine-readable instructions, the execution of which by the processor leads to the implementation of one of the methods according to paragraphs 1-10.