TWI687937B - Establishing method of chromosome abnormality detection model, chromosome abnormality detection system, and chromosome abnormality detection method - Google Patents

Abstract

The present disclosure provides a chromosome abnormality detection system including an image capturing unit and a non-transitory machine readable medium. The image capturing unit is for obtaining a target metaphase chromosomes image of a subject. The non-transitory machine-readable medium is for storing a program for assessing whether the subject has a chromosome abnormality when executed by a processing unit. Therefore, the chromosome abnormality detection system of the present disclosure can effectively improve the accuracy and sensitivity of the chromosome abnormality detection, and the time for assessing whether the subject has the chromosome abnormality can be further shorten.

Description

Establishment method of chromosome abnormality detection model, chromosome abnormality detection system and infection Color body abnormality detection method

The invention relates to a medical information analysis model, system and method, in particular to a chromosome abnormality detection model, chromosome abnormality detection system and chromosome abnormality detection method.

Chromosomal abnormality examinations are mostly used for genetic disease screening or cancer cell mutation detection such as blood cancer and lymphoma. Among them, genetic disease screening is mainly for pregnant women to receive relevant tests during pregnancy. Because the fetus carries the chromosomes of the father's sperm cells and the mother's egg cells through meiosis, so each time the embryonic life occurs It is possible to produce chromosomal mutations in the embryo. It is necessary to confirm the health status of the fetus by detecting the abnormality of the fetal chromosome.

Chromosomal abnormalities can generally be divided into abnormal chromosome number, abnormal chromosome structure and abnormal patchwork chromosome. Among them, the abnormal chromosome number is the meiosis of the germ cells. If a nondisjunction phenomenon occurs, the chromosome number of sperm or egg cells will be caused. The purpose is abnormal. After conception, it becomes an embryo with a haploid or polyploid chromosome number, and a malformed fetus is born. Common chromosome abnormalities include trisomy 21 (Down's disease), trisomy 18 (Edward's disease), and monochromosome X (Terner's disease). Abnormal chromosome structure is caused by one or more defects or abnormal combination of chromosome structure. The more common chromosomal patchwork abnormalities are 46, XX/47, XX, +21 Down syndrome patchwork, 45, X/46, XX, 45, X/46, XY or 45, X/46, X, i(Xq) is a patchwork of Turner's disease. Generally speaking, it is a patchwork of cells with some normal chromosomes, and the symptoms are usually lighter than a single pure chromosome abnormality.

The conventional chromosomal abnormality detection method is to take a chromosomal cell division mid-phase image, and the tester manually arranges the chromosome karyotype map, and then judges whether the chromosome is haploid or polyploid to determine whether the chromosome number is abnormal. And whether the chromosomes are out of shape, circular, inverted or misaligned to determine whether there is a chromosomal structural abnormality, the evaluation results of chromosomal abnormality detection vary greatly among different inspectors, and the process is also tedious and time-consuming. Therefore, how to develop a chromosomal abnormality detection system with high accuracy and rapid detection is actually a technical topic with commercial value.

In view of this, one object of the present invention is to provide a chromosome abnormality detection model, a chromosome abnormality detection method, and a chromosome abnormality detection system, which can objectively and accurately determine whether a subject has a chromosome abnormality condition, and can thereby carry out a disease Classification and risk assessment.

One aspect of the present invention is to provide a chromosome abnormality detection model, which includes the following building steps: obtaining a reference database, performing image conversion steps, performing preliminary classification steps, performing feature selection steps, and performing training steps. The reference database contains a plurality of metaphase images of reference chromosome cells. In the image conversion step, an unsupervised learning method classifier is used to arrange 23 pairs of chromosomes in the reference chromosome cell division metaphase image to obtain a plurality of reference chromosome karyotype images. The preliminary classification step is based on the number of chromosomes in the reference karyotype image. If the number of chromosomes is 46, the number of chromosomes is normal; if the number of chromosomes is greater than or less than 46, the classification is classified as The number of chromosomes is abnormal. The feature selection step is to use the feature selection module to analyze the reference karyotype image to obtain at least one image feature value. The training step is to train the at least one image feature value through a convolutional neural network learning classifier to achieve convergence to obtain the chromosome abnormality detection model, wherein the chromosome abnormality detection model is used to judge the subject Does the person have abnormal chromosomal structure or abnormal patchwork?

According to the aforementioned chromosome abnormality detection model, the unsupervised learning classifier can be a Generative Adversarial Network (GAN).

According to the aforementioned chromosome abnormality detection model, at least one image feature value may include chromosome size, chromosome position or chromosome shape.

According to the aforementioned chromosome abnormality detection model, the convolutional neural network learning classifier may be an Inception-ResNet-v2 convolutional neural network or an Inception V3 convolutional neural network.

Another aspect of the present invention is to provide a chromosome abnormality detection method, which includes the following steps. Provide a chromosomal abnormality detection model as described in the previous paragraph. Provide the subject's target chromosomal cell division mid-phase image. The unsupervised learning method classifier is used to arrange 23 pairs of chromosomes in the target chromosome cell division metaphase image to obtain a target chromosome karyotype image. The aforementioned chromosomal abnormality detection model is used to analyze the aforementioned target karyotype image to determine whether the subject has a chromosomal abnormality.

According to the aforementioned chromosomal abnormality detection method, the chromosomal abnormality may include abnormality of chromosome number, abnormality of chromosome structure or abnormality of chromosome patchwork. Preferably, the abnormal chromosome number may include the subject's target chromosome as a haploid or polyploid, and the abnormal chromosome structure may include the subject's target chromosome as a chromosome deletion, circular chromosome, chromosome translocation, chromosome inversion or Chromosome duplication.

Another aspect of the present invention is to provide a chromosome abnormality detection system including an image capturing unit and a non-transitory machine-readable medium. The image capturing unit is used to obtain the target chromosomal cell division metaphase image of the subject. The non-transitory machine-readable medium signal is connected to the image capturing unit, wherein the non-transitory machine-readable medium is used to store a program, and when the aforementioned program is executed by a processing unit, it is used to determine whether the subject has a chromosomal abnormality, And the aforementioned program includes reference database acquisition module, reference image conversion module, reference preliminary classification module, reference feature selection module, training module, target image conversion module, target preliminary classification module, target feature selection module Group and comparison module. The reference database obtaining module is used to obtain a reference database, and the aforementioned reference database is created by a plurality of reference chromosome cell division metaphase images. Ginseng According to the image conversion module, it uses an unsupervised learning method classifier to arrange 23 pairs of chromosomes in the reference chromosome cell division metaphase image to obtain a plurality of reference chromosome karyotype images. Refer to the preliminary classification module to classify the reference chromosome karyotype image according to the number of reference chromosomes. If the number of reference chromosomes is 46, the classification is normal. If the number of reference chromosomes is greater than or less than 46, then Classified as abnormal chromosome number. The reference feature selection module is used to analyze the reference chromosome karyotype image to obtain at least one reference image feature value. The training module is used to train at least one reference image feature value through the convolutional neural network to learn the classifier to achieve convergence, so as to obtain a chromosome abnormality detection model. The target image conversion module uses an unsupervised learning method classifier to arrange 23 pairs of chromosomes in the target chromosome cell division metaphase image to obtain the target chromosome karyotype image. The target preliminary classification module is used to classify the target chromosome karyotype image according to the number of target chromosomes. If the number of target chromosomes is 46, the classification is normal. If the number of target chromosomes is greater than or less than 46, the classification The number of chromosomes is abnormal. The target feature selection module is used to analyze the target karyotype image to obtain at least one target image feature value. The comparison module is used to analyze the target image feature value with the chromosome abnormality detection model to obtain a target image feature value weight data, and determine whether the subject has abnormal chromosome structure or chromosome patching according to the target image feature value weight data Type is abnormal.

According to the aforementioned chromosome abnormality detection system, the unsupervised learning classifier can be a Generative Adversarial Network (GAN).

According to the aforementioned chromosome abnormality detection system, at least one reference image feature value may include chromosome size, chromosome position or chromosome shape, and at least one target image feature value may include chromosome size, chromosome position or chromosome shape.

According to the aforementioned chromosome abnormality detection system, the convolutional neural network learning classifier may be an Inception-ResNet-v2 convolutional neural network or an Inception V3 convolutional neural network.

According to the aforementioned chromosomal abnormality detection system, the non-transitory machine-readable medium may further include an evaluation module for calculating the risk value of the chromosomal abnormality of the subject based on the target image feature value weighting data.

In this way, the chromosomal abnormality detection model, chromosomal abnormality detection system and chromosomal abnormality detection method of the present invention automatically convert the target chromosomal cell division metaphase image into the target karyotype image, and use the target feature selection module to analyze the target chromosome karyotype Obtaining at least one target image feature value after the image can effectively reduce the error caused by the subjective consciousness of different judges in the detection of chromosome abnormality. Furthermore, the chromosome abnormality detection model with deep neural network learning function can not only effectively improve the accuracy and sensitivity of chromosome abnormality detection, but also greatly shorten the judgment time of chromosome abnormality, making the chromosome abnormality detection model and chromosome of the present invention Abnormal detection systems and chromosomal abnormality detection methods are more efficient in detecting chromosomal abnormalities.

The above summary of the invention aims to provide a simplified summary of the present disclosure so that the reader can have a basic understanding of the present disclosure. Summary of this invention It is not a complete overview of the disclosure, and it is not intended to point out important/critical elements of embodiments of the invention or to define the scope of the invention.

100‧‧‧ Steps for establishing chromosome abnormality detection model

110, 120, 130, 140, 150‧‧‧ steps

200‧‧‧ Chromosome abnormality detection method

210, 220, 230, 240 ‧‧‧ steps

300‧‧‧ Chromosome abnormality detection system

400‧‧‧Image capture unit

500‧‧‧non-transitory machine-readable media

510‧‧‧Refer to the database to obtain the module

520‧‧‧Reference image conversion module

530‧‧‧Refer to the preliminary classification module

540‧‧‧Reference feature selection module

550‧‧‧Training module

560‧‧‧Target image conversion module

570‧‧‧Preliminary classification module

580‧‧‧ target feature selection module

590‧‧‧ Comparison module

610‧‧‧The image of metaphase of target chromosome cell division

620‧‧‧ Target karyotype image

700, 800 ‧‧‧ Convolutional Neural Network Learning Classifier

701, 801‧‧‧ target image feature value weight data

In order to make the above and other objects, features, advantages and embodiments of the present invention more obvious and understandable, the drawings are described as follows: Figure 1 illustrates the establishment of a chromosome abnormality detection model according to one embodiment of the present invention Step flow chart; Figure 2 shows a step flow chart of a chromosome abnormality detection method according to another embodiment of the present invention; Figure 3 shows a block diagram of a chromosome abnormality detection system according to yet another embodiment of the present invention; 4 is a diagram showing a result of conversion of a target chromosomal cell division metaphase image into a target karyotype image; FIG. 5 is a convolutional neural network learning classifier of a chromosome abnormality detection model according to an embodiment of the invention FIG. 6 is a schematic diagram of a convolutional neural network learning classifier of a chromosome abnormality detection model according to another example of an embodiment of the present invention; and FIG. 7 is a chromosome abnormality detection model of the present invention. Confusion matrix used to judge the chromosomal abnormality of the subject.

The embodiments of the present invention will be discussed in more detail below. However, this embodiment can be an application of various inventive concepts and can be specifically implemented in various specific ranges. The specific embodiments are for illustrative purposes only, and are not limited to the scope of disclosure.

Please refer to FIG. 1, which shows a flowchart of a step 100 for establishing a chromosome abnormality detection model according to an embodiment of the present invention. The step 100 of establishing a chromosome abnormality detection model of the present invention includes steps 110, 120, 130, 140 and 150. The established chromosome abnormality detection model can be used to determine whether the subject has abnormal chromosome number, abnormal chromosome structure or Abnormal chromosome patchwork.

Step 110 is to obtain a reference database, which includes a plurality of reference chromosome cell division metaphase images. In the non-dividing phase, the chromatin is mostly distributed in the nucleus from 30nm to 300nm. When the cell enters the mitotic phase, the chromosomes will begin to be gradually arranged closely. In the metaphase of the cell, the nuclear membrane of the cell disappeared completely, and the spindle filament became clear. The centromere on each chromosome is attached to the spindle filament (or star ray) respectively. The centromere begins to move up and down under the tension of its two poles, and finally the tension of the two poles reaches equilibrium. The centromeres are arranged on the equatorial plate in the center of the cell. Chromosome intelligibility reached the highest point. Therefore, before obtaining the image of the reference chromosome cell division, the reference subject's cells are entered into the cell division phase by administering hormones, and then the specific cells of the reference subject are extracted, and the reference is obtained by staining and microscopic observation. Chromosomal metaphase imaging.

Step 120 is an image conversion step. An unsupervised learning method classifier is used to arrange 23 pairs of chromosomes in the reference chromosome cell division metaphase image to obtain a plurality of reference chromosome karyotype images. The reference chromosome karyotype image system is obtained by analyzing, comparing, sorting and numbering chromosomes according to the characteristics of chromosome length, centromere position, ratio of long and short arms, presence or absence of satellite, etc. Image. The unsupervised learning method classifier may be a Generative Adversarial Network (GAN).

Step 130 is a preliminary classification step, which is based on the number of chromosomes in the reference karyotype image. If the number of chromosomes is 46, the number of chromosomes is normal; if the number of chromosomes is greater than or less than 46, then Classified as abnormal chromosome number.

Step 140 is a feature selection step. The feature selection module analyzes the reference karyotype image to obtain at least one image feature value. The at least one image feature value may include chromosome size, chromosome position or chromosome shape.

Step 150 is a training step, which is to train the at least one image feature value through a convolutional neural network learning classifier to achieve convergence, so as to obtain the chromosome abnormality detection model. The convolutional neural network learning classifier may be an Inception-ResNet-v2 convolutional neural network or an Inception V3 convolutional neural network.

Please refer to FIG. 2, which illustrates a flowchart of a method for detecting a chromosome abnormality 200 according to another embodiment of the present invention. Dyeing of the invention The body abnormality detection method 200 includes step 210, step 220, step 230, and step 240.

Step 210 is to provide a chromosome abnormality detection model, and the chromosome abnormality detection model is established through the foregoing steps 110 to 150.

Step 220 is to provide the image of the subject's target chromosomal cell division. Before obtaining the target chromosomal cell division image, the subject's cells are entered into the cell division phase by hormone administration, and then the subject's specific image is extracted. Cells, and through the staining and microscope observation to obtain the target chromosome cell division in the middle image.

Step 230 uses an unsupervised learning method classifier to arrange 23 pairs of chromosomes in the target chromosome cell division metaphase image to obtain the target chromosome karyotype image. The target chromosome karyotype imaging system analyzes, compares, sorts and numbers chromosomes according to the characteristics of chromosome length, centromere position, ratio of long and short arms, presence or absence of satellite, etc. The resulting image. The unsupervised learning method classifier may be a Generative Adversarial Network (GAN).

Step 240 is to analyze the target chromosome karyotype image using a chromosome abnormality detection model to determine whether the subject has a chromosome abnormality. The chromosomal abnormalities may include abnormal chromosome number, abnormal chromosome structure or abnormal patchwork type. Preferably, the abnormal chromosome number may include the subject's target chromosome as a haploid or polyploid, and the abnormal chromosome structure may include the subject's target chromosome as a chromosome deletion, circular chromosome, chromosome translocation, chromosome inversion or Chromosome duplication.

In this way, the chromosomal abnormality detection model and chromosomal abnormality detection method of the present invention can automatically convert the target chromosomal cell division metaphase image into the target karyotype image, and analyze the target karyotype image using the feature selection module to obtain at least The method of image feature value can effectively reduce the error caused by the subjective consciousness of different judges in the detection of chromosome abnormality. Furthermore, the chromosome abnormality detection model with deep neural network learning function can not only effectively improve the accuracy and sensitivity of chromosome abnormality detection, but also greatly shorten the judgment time of chromosome abnormality, making the chromosome abnormality detection model and chromosome of the present invention Anomaly detection methods are more efficient in detecting chromosome abnormalities.

Please refer to FIGS. 3 and 4 again. FIG. 3 shows a block diagram of a chromosome abnormality detection system 300 according to yet another embodiment of the present invention. FIG. 4 shows the conversion of the target chromosome cell metaphase image 610 into the target chromosome Karst image 620 result graph. The chromosome abnormality detection system 300 of the present invention includes an image capture unit 400 and a non-transitory machine-readable medium 500. The chromosomal abnormality detection system 300 can be used to determine whether the subject has an abnormal number of chromosomes, abnormal structure of chromosomes, or abnormality of chromosome patchwork.

The image capturing unit 400 is used to obtain the target chromosomal cell division mid-phase image 610 of the subject. The image capturing unit can be an imaging device matched with a microscope, which is used to capture the chromosome image observed by the microscope.

The non-transitory machine-readable medium 500 is connected to the image capturing unit 400. The non-transitory machine-readable medium is used to store a program. When the aforementioned program is executed by a processing unit, it is used to evaluate whether the subject has a chromosome Abnormal, where the aforementioned program includes the reference database acquisition module 510, Reference image conversion module 520, reference preliminary classification module 530, reference feature selection module 540, training module 550, target image conversion module 560, target preliminary classification module 570, target feature selection module 580, and comparison module Group 590.

The reference database obtaining module 510 is used to obtain a reference database, and the aforementioned reference database is created from a plurality of reference chromosome cell division mid-phase images.

The reference image conversion module 520 uses an unsupervised learning method classifier to arrange 23 pairs of chromosomes in the reference chromosome cell division metaphase image to obtain a plurality of reference chromosome karyotype images. The unsupervised learning method classifier may be a generative adversarial neural network.

The reference preliminary classification module 530 is used to classify the reference chromosome karyotype image according to the number of reference chromosomes. If the number of reference chromosomes is 46, the number of chromosomes is classified as normal; if the number of reference chromosomes is greater or less than 46, the number of chromosomes is classified as abnormal. Preferably, the abnormal chromosome number may include whether the subject's target chromosome is haploid or polyploid.

The reference feature selection module 540 is used to analyze the reference karyotype image to obtain at least one reference image feature value. The at least one reference image feature value may include chromosome size, chromosome position, or chromosome shape.

The training module 550 is used to train at least one reference image feature value through the convolutional neural network to learn the classifier to achieve convergence, so as to obtain a chromosome abnormality detection model. The convolutional neural network learning classifier may be an Inception-ResNet-v2 convolutional neural network or an Inception V3 convolutional neural network.

The target image conversion module 560 uses an unsupervised learning method classifier to arrange the 23 pairs of chromosomes in the target chromosome cell division metaphase image 610 to obtain the target chromosome karyotype image 620. The unsupervised learning method classifier may be a generative adversarial neural network.

The target preliminary classification module 570 is used to classify the target chromosome karyotype image according to the number of target chromosomes. If the number of target chromosomes is 46, the number of chromosomes is classified as normal; if the number of target chromosomes is greater or less than 46, the number of chromosomes is classified as abnormal. Preferably, the abnormal chromosome number may include whether the subject's target chromosome is haploid or polyploid.

The target feature selection module 580 is used to analyze the target karyotype image to obtain at least one target image feature value. The at least one target image feature value may include chromosome size, chromosome position, or chromosome shape.

The comparison module 590 is used to analyze the target image feature value with the chromosomal abnormality detection model to obtain a target image feature value weighting data, and determine whether the subject has abnormal chromosome structure or chromosome according to the target image feature value weighting data Abnormal patchwork. Preferably, the abnormal chromosomal structure may include that the target chromosome of the subject is a chromosomal deletion, a circular chromosome, a chromosomal translocation, a chromosomal inversion, or a chromosomal duplication.

In addition, the non-transitory machine-readable medium 500 may further include an evaluation module (not shown) for further calculating the risk value of the subject having a chromosomal abnormality based on the target image feature value weighting data.

According to the above-mentioned embodiment, specific test examples are presented below and explained in detail in conjunction with the drawings.

<Test example> 1. Reference database

The reference database used in the present invention is the retrospective delinked clinical content of the subject collected by China Medical University Hospital (CMUH), which was approved by the Research Ethics Committee of China Medical University and Affiliated Hospital The clinical trial plan is numbered CMUH107-REC3-151. The aforementioned reference database contains 30,000 images of the reference chromosome metaphase cell division, and the gender of the subject to which the aforementioned reference chromosome cell division metaphase image belongs is not particularly limited, and there is no specific age range.

2. The chromosome abnormality detection model of the present invention

After obtaining the reference database in the chromosome abnormality detection model of the present invention, each reference chromosome cell division mid-phase image will use a reference image conversion module to convert each reference chromosome cell division mid-phase image by an unsupervised learning method classifier to classify each reference chromosome In the mid-phase cell division image, 23 pairs of chromosomes are arranged to obtain a plurality of reference chromosome karyotype images.

In detail, because the current deep neural network model requires a large amount of training data (Training Data, that is, the mid-phase image of each reference chromosome cell division of the chromosome abnormality detection model of the present invention) to achieve stable convergence and a high degree of classification criteria To be sure, if the number of training data is not sufficient, the deep neural network will cause overfitting (Overfitting) and the error value of the judgment result will be too high, resulting in a low reliability of the deep neural network model. In order to solve the aforementioned problems, the chromosome abnormality detection model of the present invention further includes an image pre-processing step, which is to carry out each reference chromosome karyotype image. Perform black-and-white contrast correction and normalize the image value so that the image value is between 0 and 1.

A preliminary classification step is first performed to determine whether the subject has an abnormal number of chromosomes, which is classified according to the number of chromosomes in each reference chromosome karyotype image. If the number of chromosomes is 46, the number of chromosomes is classified as normal; if the number of chromosomes is greater or less than 46, the number of chromosomes is classified as abnormal.

Then, each reference chromosome karyotype image will be analyzed with a feature selection module to obtain at least one image feature value. In detail, the feature selection module can further distinguish image feature values of chromosome size, chromosome position or chromosome shape in each reference chromosome karyotype image.

Next, the aforementioned image feature values will be trained through a convolutional neural network learning classifier to achieve convergence, so as to obtain the chromosome abnormality detection model of the present invention. In this test example, the chromosome abnormality detection model will be used to determine whether the subject has abnormal chromosome number, abnormal chromosome structure, or abnormal patchwork type. The convolutional neural network learning classifier can be Inception-ResNet-v2 convolutional neural network or Inception V3 convolutional neural network.

Please refer to FIG. 5, which is a schematic diagram of the convolutional neural network learning classifier 700 of the chromosome abnormality detection model of the present invention. In the experimental example of FIG. 5, the convolutional neural network learning classifier 700 is an Inception-ResNet-v2 convolutional neural network, which includes a plurality of convolution layers (Convolution), a plurality of maximum pooling layers (MaxPool), Plural Averaging pool (AvgPool) and multiple cascade layers (Concat) to train and analyze image feature values.

In detail, the Inception-ResNet-v2 convolutional neural network is a large-scale visual recognition convolutional neural network based on the ImageNet visualization data database, and the image data in the ImageNet visualization data database are two-dimensional color images. Therefore, the conventional GoogLeNet convolutional neural network model has RGB three-channel filters in its first convolutional layer. However, the original image files of each reference chromosome karyotype image are three-dimensional grayscale images. The chromosome abnormality detection model of the present invention further uses the GoogLeNet convolutional neural network model including RGB three-channel filters through arithmetic averaging Convert to a single channel and apply Stochastic Gradient Descent (SGD) to the pre-trained model neural network of the chromosome abnormality detection model of the present invention to optimize its training process. The number of training sessions can be 100 (Epochs) and the gradient descent method using 96Mini-Batch Size, and by adjusting the initial learning rate (Learning Rates) to adjust, where the learning rate is the weight and bias of the neural network training The important parameter of change is that the chromosomal abnormality detection model of the present invention can further ensure the stable convergence of the loss function by adjusting the value of the learning rate.

In the process of training the image feature values of the chromosome abnormality detection model of the present invention, the image feature values of each reference chromosome karyotype image are processed by two layers of convolution layers and a layer of maximum pooling (MaxPool) to extract the extracted images Eigenvalues are used for maximum output, and after repeating the output of the above two convolutional layers and one maximum pooling layer again, multiple convolutional layers are used for parallel towers (parallel towers) training to complete the initial training of image feature values (Inception).

After completing the aforementioned primary training, the image feature values of each reference karyotype image will be performed 10 times (10×), 20 times (20×) and 10 times (10×) at different depths, different levels and different appearances Residual module training to train the image feature values of each reference chromosome karyotype image and achieve convergence. In detail, since the Inception-ResNet convolutional neural network undergoes the weight calculation of multiple layers, because each residual module performs different calculations and judgments on the image feature values of each standardized medial foot X-ray image data , Resulting in accumulation of errors, so the training of the Inception-ResNet convolutional neural network will pull the operation value of the node of a specific level back to the input of the level to perform the operation again to prevent the convolutional neural network from learning the classifier 700 to the aforementioned The image eigenvalues are degraded after the multi-layer weight operation training, and avoid the loss of information caused by error accumulation, and can effectively improve the training efficiency of the convolutional neural network learning classifier 700.

After completing the deep and repeated training of the residual module, a convolutional layer, an average pooling layer, a global average pooling layer (GloAvePool2D) and a linear rectification unit training layer (Rectified Linear Unit (ReLU) performs final training and processing on the converged image feature values to judge the chromosomal abnormality of the subject. Among them, the average pooling layer can first calculate the image feature values of the residual module training to obtain the average value of each image feature value. Instead of the global average pooling layer, the classifier 700 can be learned from the convolutional neural network. whole The volume network architecture is regularized to prevent the convolutional neural network learning classifier 700 from overfitting in the training mode that pursues low errors, which leads to excessively high error values in the judgment results. Finally, linear rectification The unit training layer further activates the image feature values after completion of training, and outputs a target image feature value weight data 701 for subsequent comparison and analysis. The aforementioned linear rectification unit training layer can prevent the target image feature value weight data 701 output by the foot deformity detection model from approaching zero or approaching infinity, so as to facilitate subsequent comparison steps, and thereby improve the chromosome abnormality detection of the present invention The judgment accuracy of the model.

Next, the results of the aforementioned subjects’ chromosomal abnormality judgment will be further integrated into the reference database to optimize the chromosomal abnormality detection model of the present invention, thereby further improving the training effect and judgment accuracy of the chromosomal abnormality detection model of the present invention. .

Please refer to FIG. 6 again, which shows a schematic diagram of the architecture of the convolutional neural network learning classifier 800 of the chromosome abnormality detection model of the present invention. In the experimental example of FIG. 6, the convolutional neural network learning classifier 800 is an Inception V3 convolutional neural network, which includes a plurality of convolution layers (Convolution), a plurality of average pooling layers (AvgPool), and a plurality of maximum Pooling layer (MaxPool) and multiple cascade layers (Concat), and use the dropout layer (Dropout), fully connected layer (Fully connected) and normalization layer (Softmax) to solve the problem of overfitting on machine learning, to Train and analyze image feature values.

A single-layer neural network will cause over-fitting in machine learning due to too many parameters. Inception V3 convolutional neural network is based on Large filter size decomposes the factorization of the convolutional network and reduces the order with parallel parameters, which can not only solve the problem of overfitting, but also increase the number of parameters by increasing the depth of the network to more closely approximate the original mathematical model.

In the process of training the image feature values of the chromosome abnormality detection model of the present invention, the image feature values of each reference chromosome karyotype image are respectively subjected to one layer of average pooling layer and one layer of convolution layer; five layers of convolution layer; three layers of convolution layer; After the operation of a convolutional layer, the feature matrix values of each group of operations are stacked in cascade. Then repeat 2 times to perform one layer of average pooling layer and one layer of convolutional layer; five layers of convolutional layer; three layers of convolutional layer; one layer of convolutional layer after operation, and cascade the feature matrix values of each group of operations. Then perform a maximum pooling layer; three convolutional layers; one layer of convolutional layers, and then cascade the feature matrix values of each group of operations. Then repeat 4 times to perform an average pooling layer and a convolutional layer; five convolutional layers; three convolutional layers and one convolutional layer, and cascade the feature matrix values of each group of operations. Then perform an average pooling layer, two convolutional layers, a fully connected layer and a regularization layer operation. The calculated feature matrix values are repeated 2 times to perform an average pooling layer and a convolutional layer; three convolutional layers And a layer of cascading layers; two layers of convolutional layers and a layer of cascading layers; after the operation of a layer of convolutional layers, the feature matrix values of each group of operations are stacked in cascade. Finally, after performing an average pooling layer, a discarding layer, a fully connected layer, and a regularization layer, a target image feature value weight data 801 is output to obtain a trained chromosome abnormality detection model.

Next, the judgment results of the aforementioned subjects' chromosomal abnormalities will be further integrated into the reference database to detect the chromosomal abnormalities of the present invention The test model is optimized to further improve the training effect and judgment accuracy of the chromosome abnormality detection model of the present invention.

Please refer to FIG. 7 again for the confusion matrix of the chromosome abnormality detection model of the present invention for judging the chromosome abnormality of the subject. In the test example in Figure 7, the convolutional neural network learning classifier that establishes the chromosome abnormality detection model is the convolutional neural network learning classifier 800 shown in Figure 6 to determine whether the subject's chromosome is abnormal, and The results are divided into normal and abnormal. The horizontal axis is the predicted label, and the vertical axis is the actual label. The confusion matrix can be divided into True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (False) Negative, FN), and calculate the accuracy, sensitivity, specificity, positive predictive value and negative predictive value of the chromosome abnormality detection model of the present invention based on the data of TP, TN, FP and FN. The calculation method of correct rate is (TP+TN)/(TP+FP+TN+FN), the calculation method of sensitivity is TP/(TP+FN), and the calculation method of specificity is TN/(TN+FP), The calculation method for positive predictive value is TP/(TP+FP), and the calculation method for negative predictive value is TN/(FN+TN).

As shown in the results in Figure 7, the number of subjects in the TP block is 206, the number of subjects in the TN block is 201, the number of subjects in the FP block is 3, and the subjects in the FN block The number of people is 0 people. After calculation, the chromosomal abnormality detection model of the present invention is used to judge the prediction result of the chromosomal abnormality of the subject as shown in Table 1.

It is obvious from the above results that the chromosomal abnormality detection model of the present invention can be used to accurately determine whether a subject has a chromosomal abnormality condition, and the chromosomal abnormality condition may include abnormality in chromosome number, abnormality in chromosome structure, and abnormality in chromosome patchwork.

In this way, the chromosomal abnormality detection system of the present invention can effectively improve the accuracy and sensitivity of chromosomal abnormality detection, and can shorten the evaluation time of whether the subject has chromosomal abnormality. From the original image input to the interpretation result, the average is only 0.1- It can be completed in 1 second, making it more widely used.

Although the present invention has been disclosed as above by way of implementation, it is not intended to limit the present invention. Any person who is familiar with this art can make various modifications and retouching without departing from the spirit and scope of the present invention, so the protection of the present invention The scope shall be determined by the scope of the attached patent application.

300: Chromosome abnormality detection system

400: image capture unit

500: non-transitory machine-readable media

510: Refer to the database to obtain the module

520: Refer to image conversion module

530: Refer to the preliminary classification module

540: Reference feature selection module

550: Training module

560: Target image conversion module

570: Target Initial Classification Module

580: Target feature selection module

590: Comparison module

Claims (10)

A method for establishing a chromosome abnormality detection model includes the following steps: obtaining a reference database, wherein the reference database contains a plurality of reference chromosome cell division mid-phase images; performing an image conversion step, which uses an unsupervised learning method The classifier arranges the 23 pairs of chromosomes in the reference chromosome cell division metaphase images according to a chromosome length, a centromere position, a ratio of long and short arms and a satellite, to obtain a plurality of reference chromosome karyotype images, of which the non- The supervised learning classifier is a Generative Adversarial Network (GAN); a preliminary classification step is performed, which is based on the number of chromosomes in the reference chromosome karyotype images. If the number of chromosomes is 46, classified as normal number of chromosomes, if the number of chromosomes is greater than or less than 46, it is classified as abnormal number of chromosomes; to perform a feature selection step, it uses a feature selection module to analyze each reference chromosome karyotype After the image, at least one image feature value is obtained from each of the reference chromosome karyotype images to obtain a plurality of the image feature values; and a training step is performed, which is to learn the image feature values through a convolutional neural network The classifier is trained to achieve convergence to obtain the chromosome abnormality detection model, where the chromosome abnormality detection model is used to judge Whether a subject has a chromosomal structural abnormality or a chromosomal patchwork abnormality. The method for establishing a chromosome abnormality detection model as described in item 1 of the patent application scope, wherein the at least one image feature value includes a chromosome size, a chromosome position, or a chromosome shape. The method for establishing a chromosome abnormality detection model as described in item 1 of the patent scope, wherein the convolutional neural network learning classifier is Inception-ResNet-v2 convolutional neural network or Inception V3 convolutional neural network. A chromosomal abnormality detection system includes: an image capturing unit for acquiring a target chromosomal cell division metaphase image of a subject; and a non-transitory machine-readable medium to which a signal is connected to the image capturing unit, wherein the non- The transient machine-readable medium is used to store a program, which is used to determine whether the subject has a chromosomal abnormality when the program is executed by a processing unit, and the program includes: a reference database acquisition module for Obtain a reference database, and the reference database is created by multiple metaphase images of reference chromosome cells; A reference image conversion module, which uses an unsupervised learning method classifier to divide the 23 pairs of chromosomes in the metaphase image of these reference chromosomes according to a chromosome length, a centromere position, a long and short arm ratio and a satellite Whether to arrange to obtain a plurality of reference chromosome karyotype images, where the unsupervised learning method classifier is a Generative Adversarial Network (GAN); a preliminary classification module is used to refer to these reference chromosomes Karyotype images are classified according to the number of reference chromosomes. If the number of reference chromosomes is 46, the number of chromosomes is normal. If the number of reference chromosomes is greater than or less than 46, the number of chromosomes is abnormal. One reference A feature selection module for analyzing each reference chromosome karyotype image, and obtaining at least one reference image feature value in each reference chromosome karyotype image to obtain a plurality of the image feature values; a training module for The reference image feature values are trained through a convolutional neural network learning classifier to achieve convergence to obtain a chromosome abnormality detection model; a target image conversion module that uses the unsupervised learning method classifier to classify the target chromosome The 23 pairs of chromosomes in the mid-phase image of cell division are arranged according to the length of the chromosome, the location of the centromere, the ratio of the long and short arms, and the presence or absence of the satellite to obtain a target chromosome nuclei Type image, wherein the unsupervised learning method classifier is the generated adversarial neural network (Generative Adversarial Network, GAN); a target preliminary classification module, used to classify the target karyotype image according to a target chromosome number, If the target chromosome number is 46, the classification is normal. If the target chromosome number is greater than or less than 46, it is classified as abnormal chromosome number; a target feature selection module, which is used to analyze the target Chromosomal karyotype image to obtain at least one target image feature value; and a comparison module, which is used to analyze the target image feature value with the chromosome abnormality detection model to obtain a target image feature value weight data, and According to the target image feature value weighting data, it is determined whether the subject has a chromosome structural abnormality or a chromosome patchwork abnormality. The chromosome abnormality detection system as described in item 4 of the patent application scope, wherein the at least one reference image feature value includes a chromosome size, a chromosome position or a chromosome shape, and the at least one target image feature value includes a chromosome size and a chromosome Location or shape of a chromosome. The chromosome abnormality detection system as described in item 4 of the patent application scope, wherein the convolutional neural network learning classifier is Inception-ResNet-v2 convolutional neural network or Inception V3 convolutional neural network. The chromosomal abnormality detection system as described in item 4 of the patent application scope, wherein the program further includes: an evaluation module for calculating the risk value of the subject with the chromosomal abnormality based on the target image feature value weighting data. A chromosomal abnormality detection method, comprising: providing a chromosomal abnormality detection system as described in item 4 of the patent application scope; providing a metaphase image of a target chromosome cell of a subject; using the unsupervised learning classifier to classify the target The 23 pairs of chromosomes in the chromosome metaphase image are arranged according to the length of the chromosome, the position of the centromere, the ratio of the long and short arms, and the presence or absence of the satellite to obtain a target karyotype image, where the unsupervised learning classifier is The Generative Adversarial Network (GAN); and The chromosomal abnormality detection system is used to analyze the target chromosome karyotype image, and a confusion matrix is used to determine whether the subject has the chromosome number abnormality, the chromosome structure abnormality, or the chromosome patchwork type abnormality. The chromosomal abnormality detection method as described in item 8 of the patent application range, wherein the abnormal chromosome number includes that one of the subject's target chromosomes is a haploid or a polyploid. The chromosomal abnormality detection method as described in item 8 of the patent application scope, wherein the chromosomal abnormality includes that one of the subject's target chromosomes is a chromosomal deletion, a circular chromosome, a chromosomal translocation, a chromosomal inversion, or a chromosome repeat.

Images