A chromosome abnormality detection system and method based on a denoising diffusion probability model

Abstract

The application discloses a chromosome abnormality detection system and method based on a denoising diffusion probability model, a forward diffusion model of which adds noise to an input chromosome image as an original image along a Markov process for t times of superposition until a pure noise image is obtained; a reverse diffusion model of which traverses backward along a Markov chain to perform t times of denoising operation on the pure noise image to obtain a reconstructed image; and an analysis model of which calculates mean square errors of the original image and the reconstructed image, and compares the obtained mean square errors with a preset threshold value, and determines that the original image is abnormal if the mean square error is greater than the threshold value, and is normal otherwise. The application has the advantages of high efficiency and accuracy compared with a traditional detection mode.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis, and in particular to a chromosomal abnormality detection system and method based on a denoising diffusion probability model. Background Technology

[0002] Chromosomal structural abnormalities refer to chromosomal aberrations and chromatid aberrations that occur through breakage-replacement or exchange mechanisms of chromosomes or chromatids. These include deletions (resulting in a reduction in the number of genes on the chromosome); duplications (resulting in an increase in the number of genes on the chromosome); 180° inversions (resulting in the reverse order of gene sequences on a segment of the chromosome, without changing the number of genes); and translocations (including reciprocal and non-reciprocal translocations) between non-homologous chromosomes.

[0003] In genetics, most congenital malformations, intellectual disabilities, and infertility are caused by chromosomal abnormalities. Furthermore, for hematology and tumors, cytogenetics and the detection of chromosomal structural abnormalities play a crucial role in diagnosis, prognosis, treatment screening, and monitoring.

[0004] Conventional techniques mostly involve counting chromosomes based on karyotype analysis, which is generally suitable for detecting triploid syndromes. However, there is a lack of accurate methods for detecting chromosomal structural abnormalities.

[0005] For the detection of chromosomal structural abnormalities, most existing methods focus on the detection process. That is, they perform binary classification prediction on the input single chromosome to determine whether there is an abnormality. However, the identification of abnormal categories and abnormal regions cannot be automated and must rely on annotation by professional physicians, which is very inefficient.

[0006] The technical problem this invention aims to solve is to provide an efficient and accurate detection system and method for chromosomal structural abnormalities such as fragment deletions. Summary of the Invention

[0007] The purpose of this invention is to provide an accurate and efficient method for detecting chromosomal abnormalities based on a denoised diffusion probability model.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A chromosome abnormality detection system based on a denoising diffusion probability model includes a forward diffusion model, a reverse diffusion model, and an analysis model. The forward diffusion model adds noise to the input chromosome image as the original image by performing t superpositions along a Markov process until a pure noise image is obtained. The reverse diffusion model traverses the Markov chain backward to perform t denoising operations on the pure noise image to obtain a reconstructed image, where t is a natural number greater than 0. The analysis model compares the similarity between the original image and the reconstructed image to determine whether the original image is abnormal or normal. The forward diffusion model, reverse diffusion model, and analysis model are all trained separately according to the chromosome type chr1, chr2...chr22, chrX, chrY to obtain a single model for that chromosome type.

[0010] Preferably, the noise is simplex noise, Gaussian noise, or white noise.

[0011] Preferably, during training, the value of t ranges from 1000 to 1500; during detection, the value of t ranges from 25 to 50.

[0012] This invention also provides a method for detecting chromosomal abnormalities based on a denoised diffusion probability model, which includes the aforementioned chromosomal abnormality detection system based on a denoised diffusion probability model, comprising the following steps:

[0013] S1. Use different types of chromosome images as the original images. The input is fed into the corresponding type of forward diffusion model to add simplex noise by stacking it t times along a Markov process, resulting in a purely noisy image; where the T-step of the Markov chain is used to formulate the forward diffusion model, i.e., given data points... At each step of a Markov chain Add a simplex noise to each to generate a new latent variable. Among them, Simplex noise The distribution formula is as follows: t is the sequence number of each operation. It is an image. This is the image at step t. This is the image at step t-1. It is the original input image. It is a Markov process at step t, where T represents the maximum number of steps and T >> 1000. N All of these are hyperparameters. Represents the initial startup frequency, and N represents the noise metric. Noise attenuation rate;

[0014] S2. The reverse diffusion model traverses backwards along the Markov chain to perform t denoising operations on the pure noise image until a reconstructed image is obtained. ;

[0015] S3, the analysis model for the original image and reconstructed images A similarity comparison is performed, and the difference obtained is compared with a pre-trained threshold c. If the difference is greater than the threshold c, the original image is discarded. If it is deemed abnormal, it is deemed normal.

[0016] Preferably, the original image in step S1 After pretreatment, the processing steps include:

[0017] S0-1. Perform chromosome segmentation on the initial image containing multiple chromosomes using an automatic segmentation network to obtain an image containing only one chromosome and a mask of the chromosome's outline, and extract the chromosome from it.

[0018] S0-2. Arrange the extracted chromosomes with the short arm on top and the long arm on the bottom, and fill the blank areas with white pixels to obtain a rectangular single chromosome image.

[0019] S0-3. Identify the chromosomes in the single chromosome image to determine their types.

[0020] Preferably, step S2 includes the following steps:

[0021] S2-1, Calculate the original image using the Zhang-Suen algorithm. The chromosome skeleton in;

[0022] S2-2, Traverse the Markov chain backward to perform t denoising operations on the pure noise image until a reconstructed image is obtained. Each noise reduction operation is constrained by the chromosome backbone.

[0023] The training method for the inverse diffusion model is as follows:

[0024] S2-11. Building a Deep Learning Model Using U-Net++ Network , Refers to deep learning models;

[0025] S2-12. Traverse the Markov chain backward to perform t denoising operations on the pure noise image; wherein, each denoising operation samples one noise from the Simplex noise. Then, the input image is denoised and its diffusion probability is reconstructed; among which, the deep learning model Each layer of the network embeds the chromosome backbone obtained in step S2-1, enabling the network to learn different morphological features and structural information at different levels. The U-Net++ network introduces SE modules to incorporate an attention mechanism, the implementation steps of which are as follows:

[0026] S2-12-1. Compress each channel of the U-Net++ network using global average pooling to obtain the global feature vector for each channel.

[0027] S2-12-2. Perform a linear transformation on the global feature vector of each channel and activate it using an activation function to obtain the weight of each channel;

[0028] In steps S2-12, each denoising operation multiplies the channel weights by the input feature map to obtain the output feature map reconstructed by the denoising diffusion probability after being weighted by the attention mechanism. The input feature map is the output feature map of the previous step.

[0029] The U-Net++ network is trained through pruning, and the implementation steps are as follows:

[0030] 1) Train the complete U-Net++ network and record the gradient values ​​and Hessian matrix of each neuron or weight;

[0031] 2) Based on the gradient values ​​and the indices of the Hessian matrix, calculate the degree of influence of each neuron or weight on the output of the U-Net++ network. :

[0032]

[0033] in, This represents the loss function for the i-th neuron. This represents the i-th neuron or its weight;

[0034] 3) Prune neurons or weights according to a preset contribution threshold;

[0035] S2-13. Each noise reduction operation calculates the noise distribution using a loss function. The result of deep learning prediction at step t loss function between Its expression is:

[0036]

[0037] in, Represents the first diffusion process in the diffusion probability model step, Representative at the The loss function value during step diffusion. It is Simplex noise generated by random sampling, and its input is and , These are manually set hyperparameters. This indicates the calculation of the expected value of a mathematical expression. Calculate the gradient, train t times, until the loss is calculated. Until convergence.

[0038] Preferably, in step S3, the mean squared error (MSE) or the cross-entropy loss function is used to process the original image. and reconstructed images Perform a similarity comparison;

[0039] The mean square error (MSE) is calculated as follows:

[0040] Where I is the reconstructed image K is the original image ; It is the first in the test image The value of each pixel. It is the first one in the reference image The value of each pixel; original image The number of pixels is m×n; the mean square error (MSE) less than the threshold c is considered normal, and the MSE greater than or equal to the threshold c is considered abnormal.

[0041] The calculation method for the cross-entropy loss function is as follows:

[0042] I is the reconstructed image K is the original image ; For the original image and reconstructed images The pixel values ​​are normalized and denoted as p and q respectively;

[0043] by It is the first in the original image The value of each pixel. It is the first in the reconstructed image The value of each pixel; original image The number of pixels is m×n;

[0044] Calculate the cross-entropy loss function :

[0045]

[0046] if Values ​​less than the threshold c are considered normal. Data greater than or equal to a threshold c is considered abnormal; the threshold c is obtained through pre-training.

[0047] Before step S3, the original image is first aligned. and reconstructed images The steps for alignment are as follows:

[0048] a) From the original image and reconstructed images Several feature points were extracted from the original image using the SIFT algorithm, followed by the FLANN algorithm. and reconstructed images Matching feature points in the data;

[0049] b) Remove incorrect matching points by using the RANSAC algorithm to remove incorrectly matched feature points;

[0050] c) Use the fundamental matrix or homography matrix to obtain the original image. and reconstructed images Perform geometric transformations to obtain the transformation matrix; and transform the original image based on the feature points of the matching points. and reconstructed images Any image in the image can be aligned with another image.

[0051] Preferably, the steps for solving the threshold c are as follows:

[0052] S4-1. Take N chromosome images that have been manually labeled as normal and N chromosome images that have been manually labeled as abnormal, and input them into the denoising diffusion probability model to obtain the mean square error (MSE) of the original image and the reconstructed image. Use different thresholds to determine whether the input chromosome image is normal or abnormal.

[0053] S4-2, Solving for the true positive rate: ;

[0054] S4-3, Solve for the false positive rate: ;

[0055] S4-4, Solving for sensitivity: ;

[0056] S4-5, Solving for specificity: ;

[0057] S4-6. Calculate the Youden Index J: The point with the largest J value is selected as the optimal threshold c, where TP is the number of cases where both the judgment result and the labeling result are positive, FN is the number of cases where the judgment result is negative but the labeling result is positive, TN is the number of cases where both the judgment result and the labeling result are negative, and FP is the number of cases where the judgment result is positive but the labeling result is negative.

[0058] Preferably, the analysis model calculates the pixel difference matrix D:

[0059]

[0060] in, This represents the value of the pixel in the i-th row and j-th column in the difference image. and These represent the values ​​of the pixels in the i-th row and j-th column of the original image and the reconstructed image, respectively. The difference image is an m×n matrix obtained by calculating the difference between the original image and the reconstructed image for each pixel, where m and n are the length and width of the original image, respectively. The analysis model visualizes the difference image using a heatmap.

[0061] The analytical model calculates the coordinates of abnormal chromosome regions through the following steps:

[0062] S3-1. For the difference matrix D, set a judgment threshold. ;extract The area is denoted as The threshold It was obtained by optimizing and adjusting based on the results;

[0063] S3-2, To Find the geometric center, denoted as . ;

[0064] S3-3, For Perform connected component calculation to obtain the connected component it belongs to, denoted as . ;

[0065] S3-4, to The process involves expansion followed by corrosion to fill the voids; the resulting area is denoted as [missing information]. This refers to the abnormal chromosomal region;

[0066] S3-5. Perform skeleton extraction processing on the input chromosome image to obtain its skeleton coordinate vector S;

[0067] S3-6, Calculate abnormal chromosome regions upper and lower boundaries and ,calculate and Intersection point with skeleton coordinate vector S and Then calculate and The distance from the starting end of the position in the skeleton S and And the total length l of S;

[0068] S3-7. Based on the human chromosome coordinate information published by the international nomenclature system for human cell genomics, according to and The percentage of l is used to determine the coordinates of the corresponding chromosomal abnormal region.

[0069] The beneficial effects of this invention are: it utilizes denoising diffusion probabilistic models to destroy the image and reconstruct an approximate healthy chromosome image. The reconstructed chromosome image is compared with the original chromosome image, extracting regions with pixel differences. Heatmaps are generated for these difference regions, and areas potentially exhibiting structural abnormalities are labeled and their coordinates calculated. Compared to traditional detection methods, this invention offers advantages such as high efficiency and accuracy. Attached Figure Description

[0070] Figure 1 This is the original image of the present invention. A diagram illustrating the preparation process;

[0071] Figure 2 This is a flowchart illustrating the forward diffusion model and the reverse diffusion model of the present invention, wherein the network structure diagram represents the U-Net++ network used;

[0072] Figure 3 This is a schematic diagram illustrating the principle of a chromosome abnormality detection method based on a denoised diffusion probability model.

[0073] Figure 4 This is a flowchart illustrating the coordinate calculation of abnormal chromosomal regions, where each text marker corresponds to the detection target in steps S3-5 to S3-7. Detailed Implementation

[0074] The technical solution of this patent will be further described in detail below with reference to specific embodiments.

[0075] In the description of this invention, it should be noted that the terms "inner", "outer", "upper", "lower", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0076] like Figures 1 to 4 As shown, the chromosome abnormality detection system based on a denoised diffusion probability model of the present invention includes a forward diffusion model, a reverse diffusion model and an analysis model.

[0077] The forward diffusion model adds noise to the input chromosome image as the original image by stacking it t times along a Markov process until a pure noise image is obtained. The backward diffusion model traverses the Markov chain backward to perform denoising operations on the pure noise image t times to obtain the reconstructed image, where t is a natural number greater than 0. The analysis model compares the similarity between the original image and the reconstructed image to determine whether the original image is abnormal or normal. The forward diffusion model, backward diffusion model and analysis model are all trained separately according to the chromosome type chr1, chr2...chr22, chrX, chrY to obtain a single model for that chromosome type.

[0078] The noise can be simplex noise, Gaussian noise, or white noise, with simplex noise being preferred.

[0079] During training, the value of t ranges from 1000 to 1500; during detection, the value of t ranges from 25 to 50. This reduces the computational load and increases the analysis speed.

[0080] This invention also provides a method for detecting chromosomal abnormalities based on a denoised diffusion probability model, which includes a chromosomal abnormality detection system based on a denoised diffusion probability model, comprising the following steps:

[0081] S1. Use different types of chromosome images as the original images. The input is fed into the corresponding type of forward diffusion model to add simplex noise by stacking it t times along a Markov process, resulting in a purely noisy image; where the T-step of the Markov chain is used to formulate the forward diffusion model, i.e., given data points... At each step of a Markov chain Add a simplex noise to each to generate a new latent variable. Among them, Simplex noise The distribution formula is as follows: t is the sequence number of each operation. It is an image. This is the image at step t. This is the image at step t-1. It is the original input image. It is a Markov process at step t, where T represents the maximum number of steps and T >> 1000. N All of these are hyperparameters. Represents the initial starting frequency, and N represents the noise metric octave. T is the noise decay rate; in practice, T is taken as 1200.

[0082] S2. The inverse diffusion model traverses the Markov chain backwards to perform t denoising operations on the pure noise image until a reconstructed image is obtained. The noise reduction operation refers to sampling a noise source from the simplex noise. Then, the input image is subjected to denoising diffusion probability reconstruction; that is, the inverse process of process q, denoted by p, and the training objective is to obtain a good probability distribution of the reconstructed image: By traversing backward along the Markov chain, new data can be regenerated. .

[0083] S3, Analysis Model on Original Image and reconstructed images A similarity comparison is performed, and the difference obtained is compared with a pre-trained threshold c. If the difference is greater than the threshold c, the original image is discarded. If it is deemed abnormal, it is deemed normal.

[0084] Further, see Figure 1 The original image in step S1 After pretreatment, the processing steps include:

[0085] S0-1. Perform chromosome segmentation on the initial image (i.e., the original metaphase image) containing multiple chromosomes using an automatic segmentation network or other automatic or manual segmentation software (such as Ikaros, CytoVision, etc.) to obtain an image containing only one chromosome and a mask of the chromosome outline, and extract the chromosome from it.

[0086] S0-2. Arrange the extracted chromosomes with the short arm on top and the long arm on the bottom, and fill the blank area with white pixels as the background to obtain a rectangular single chromosome image.

[0087] S0-3. Identify the chromosomes in the single chromosome image to determine their type.

[0088] This step allows the input chromosomes to be fed into specific types of forward diffusion models, reverse diffusion models, and analysis models for detection, thereby achieving a higher recognition rate.

[0089] Further, step S2 includes the following steps:

[0090] S2-1, Calculate the original image using the Zhang-Suen algorithm. The chromosome skeleton in;

[0091] S2-2, Traverse the Markov chain backwards to perform t denoising operations on the pure noise image until the reconstructed image is obtained. Each noise reduction operation is constrained by the chromosome backbone; because t and the forward diffusion model are the same, the output of each step has a desired image, but it cannot be guaranteed to be the same each time;

[0092] The training method for the inverse diffusion model is as follows:

[0093] S2-11. Building a Deep Learning Model Using U-Net++ Network , Refers to deep learning models;

[0094] S2-12. Traverse the Markov chain backward to perform t denoising operations on the pure noise image; where each denoising operation samples one noise from the Simplex noise. Then, the input image is denoised and reconstructed using diffusion probability. During diffusion probability reconstruction, the probability model needs to be trained to obtain model parameters, which is typically achieved through maximum likelihood estimation, where the goal is to maximize the probability of the observed noisy image under the model. To achieve this, a reverse process can be used, i.e., generating noise iteratively and updating the model parameters by comparing the generated noise with the observed noise. The deep learning model... Each layer of the network embeds the chromosome skeleton obtained in step S2-1, enabling the network to learn different morphological features and structural information at different levels, thereby gaining a more comprehensive understanding of chromosome morphology and structure. The U-Net++ network introduces an SE module (Squeeze-and-Excitation: a novel image recognition structure announced by the autonomous driving company Momenta in 2017, which improves accuracy by modeling the correlation between feature channels and enhancing important features) to incorporate an attention mechanism. The implementation steps are as follows:

[0095] S2-12-1. Use global average pooling to compress each channel of the U-Net++ network to obtain the global feature vector for each channel.

[0096] S2-12-2. Perform a linear transformation on the global feature vector of each channel and activate it using an activation function to obtain the weight of each channel;

[0097] In steps S2-12, each denoising operation multiplies the channel weights by the input feature map to obtain the output feature map reconstructed by the denoising diffusion probability after being weighted by the attention mechanism. The input feature map is the output feature map of the previous step.

[0098] The U-Net++ network is trained after pruning, and the implementation steps are as follows:

[0099] 1) Train the complete U-Net++ network and record the gradient value of each neuron or weight using the Hessian matrix.

[0100] 2) Based on the gradient values ​​and the indices of the Hessian matrix, calculate the degree of influence of each neuron or weight on the output of the U-Net++ network. :

[0101]

[0102] in, This represents the loss function for the i-th neuron. This represents the i-th neuron or its weight;

[0103] 3) Prune neurons or weights according to a preset contribution threshold;

[0104] S2-13. Each noise reduction operation calculates the noise distribution using a loss function. The result predicted by deep learning at step t loss function between Its expression is:

[0105]

[0106] in, Represents the first diffusion process in the diffusion probability model step, Representative at the The loss function value during step diffusion. It is simplex noise generated by random sampling. This represents the noise at the network output, and its input is... and , These are manually set hyperparameters. This indicates the calculation of the expected value of a mathematical expression. Calculate the gradient, train t times, until the loss is calculated. Until convergence.

[0107] In step S3, the mean squared error (MSE) or the cross-entropy loss function is used to process the original image. and reconstructed images Perform a similarity comparison;

[0108] The mean square error (MSE) is calculated as follows:

[0109] Where I is the reconstructed image K is the original image ; It is the first in the test image The value of each pixel, K It is the first one in the reference image The value of each pixel; original image The number of pixels is m×n; the mean square error (MSE) less than the threshold c is considered normal, and the MSE greater than or equal to the threshold c is considered abnormal.

[0110] The calculation method for the cross-entropy loss function is as follows:

[0111] I is the reconstructed image K is the original image ; For the original image and reconstructed images The pixel values ​​are normalized and denoted as p and q respectively;

[0112] by It is the first in the original image The value of each pixel. It is the first in the reconstructed image The value of each pixel; original image The number of pixels is m×n;

[0113] Calculate the cross-entropy loss function :

[0114]

[0115] if Values ​​less than the threshold c are considered normal. Data greater than or equal to the threshold c is considered abnormal; the threshold c is obtained through pre-training.

[0116] Before step S3, the original image is first aligned. and reconstructed images The steps for alignment are as follows:

[0117] a) From the original image and reconstructed images Several feature points were extracted using the SIFT (Scale Invariant Feature Transform) algorithm, followed by the FLANN (Fast Nearest Neighbor) algorithm on the original image. and reconstructed images Matching feature points in the data;

[0118] b) Remove incorrect matching points by using the RANSAC algorithm (RANSAC is short for Random Sample Consensus, which is an algorithm that calculates the mathematical model parameters of the data based on a set of sample datasets containing abnormal data to obtain valid sample data) to remove mismatched feature points.

[0119] c) Use the fundamental matrix or homography matrix to obtain the original image. and reconstructed images Perform geometric transformations to obtain the transformation matrix; and transform the original image based on the feature points of the matching points. and reconstructed images Any image in the image can be aligned with another image.

[0120] The steps to solve for the threshold c are as follows:

[0121] S4-1. Take N chromosome images that have been manually labeled as normal and N chromosome images that have been manually labeled as abnormal, and input them into a denoising diffusion probability model to obtain the mean square error (MSE) of the original image and the reconstructed image. Use different thresholds to determine whether the input chromosome image is normal or abnormal.

[0122] S4-2, Solving for the true positive rate: ;

[0123] S4-3, Solve for the false positive rate: ;

[0124] S4-4, Solving for sensitivity: ;

[0125] S4-5, Solving for specificity: ;

[0126] S4-6. Calculate the Youden Index J: The point with the largest J value is selected as the optimal threshold c, where TP is the number of cases where both the judgment result and the labeling result are positive, FN is the number of cases where the judgment result is negative but the labeling result is positive, TN is the number of cases where both the judgment result and the labeling result are negative, and FP is the number of cases where the judgment result is positive but the labeling result is negative.

[0127] In other words, for chromosome images that have been professionally labeled as normal or abnormal, abnormalities are defined as positive (denoted as P), and normalities as negative (denoted as N). Based on the original image... and reconstructed images The mean square error (MSE) is divided into different judgment thresholds c. If the MSE is less than c, it is judged as N, and if the MSE is greater than or equal to c, it is judged as P.

[0128] Then we have the confusion matrix:

[0129] Prediction / Reality P N P TP FP N FN TN

[0130] Plot an ROC curve with FPR on the x-axis and TPR on the y-axis. Calculate the Youden index (J).

[0131]

[0132] The point with the largest J value is selected as the optimal threshold to distinguish between normal and abnormal chromosome images.

[0133] Furthermore, the analysis model calculates the pixel difference matrix D:

[0134]

[0135] in, This represents the value of the pixel in the i-th row and j-th column in the difference image. and These represent the values ​​of the pixels in the i-th row and j-th column of the original image and the reconstructed image, respectively. The difference image is an m×n matrix obtained by calculating the difference between the original image and the reconstructed image for each pixel, where m and n are the length and width of the original image, respectively. The analysis model visualizes the difference image using a heatmap.

[0136] The analysis model calculates the coordinates of abnormal chromosomal regions through the following steps:

[0137] S3-1. For the difference matrix D, set a judgment threshold. ;extract The area is denoted as The threshold It was obtained by optimizing and adjusting based on the results;

[0138] S3-2, To Find the geometric center, denoted as . ;

[0139] S3-3, For Perform connected component calculation to obtain the connected component it belongs to, denoted as . ;

[0140] S3-4, to The process involves expansion followed by corrosion to fill the voids; the resulting area is denoted as [missing information]. This refers to the abnormal chromosomal region;

[0141] S3-5. Perform skeleton extraction processing on the input chromosome image to obtain its skeleton coordinate vector S;

[0142] S3-6, Calculate abnormal chromosome regions upper and lower boundaries and ,calculate and Intersection point with skeleton coordinate vector S and Then calculate and The distance from the starting end of the position in the skeleton S and And the total length l of S;

[0143] S3-7, Human chromosome coordinate information based on the international nomenclature system for human cell genomics (i.e., Figure 4 Biological coordinates in the middle), according to and The percentage of l is used to determine the coordinates of the corresponding chromosomal abnormal region.

[0144] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A chromosome abnormality detection method based on a denoised diffusion probability model, which is applied to a chromosome abnormality detection system based on a denoised diffusion probability model. The chromosome abnormality detection system based on the denoised diffusion probability model includes a forward diffusion model, a reverse diffusion model, and an analysis model. The forward diffusion model adds noise to the input chromosome image as the original image by performing T superpositions along a Markov process until a pure noise image is obtained. The backdiffusion model traverses backward along the Markov chain to perform T denoising operations on the pure noise image to obtain the reconstructed image; the analysis model compares the similarity between the original image and the reconstructed image to determine whether the original image is abnormal or normal; the forward diffusion model, backdiffusion model, and analysis model are all trained separately according to the chromosome type chr1, chr2...chr22, chrX, chrY to obtain a single model for that chromosome type; Its features are, Includes the following steps: S1. Use different types of chromosome images as the original images. The input is fed into the corresponding type of forward diffusion model to add simplex noise through T superpositions along a Markov process, resulting in a purely noisy image; where the T-step of the Markov chain is used to formulate the forward diffusion model, i.e., given data points... At each step of a Markov chain Add a simplex noise to each to generate a new latent variable. Among them, Simplex noise The distribution formula is as follows: t is the sequence number of each operation. It is an image. This is the image at step t. This is the image at step t-1. It is the original input image, q It is a Markov process at step t, where T represents the maximum number of steps and T > 1000. N All of these are hyperparameters. Represents the initial startup frequency, and N represents the noise metric. Noise attenuation rate; S2. The reverse diffusion model traverses backwards along the Markov chain to perform T denoising operations on the pure noise image until a reconstructed image is obtained. ; S3, the analysis model for the original image and reconstructed images A similarity comparison is performed, and the difference obtained is compared with a pre-trained threshold c. If the difference is greater than the threshold c, the original image is discarded. If it is deemed abnormal, then it is deemed normal; The original image in step S1 After pretreatment, the processing steps include: S0-1. Perform chromosome segmentation on the initial image containing multiple chromosomes using an automatic segmentation network to obtain an image containing only one chromosome and a mask of the chromosome's outline, and extract the chromosome from it. S0-2. Arrange the extracted chromosomes with the short arm on top and the long arm on the bottom, and fill the blank areas with white pixels to obtain a rectangular single chromosome image. S0-3. Identify the chromosomes in the single chromosome image to determine their types; Step S2 includes the following steps: S2-1, Calculate the original image using the Zhang-Suen algorithm. The chromosome skeleton in; S2-2, Traverse the Markov chain backward to perform t denoising operations on the pure noise image until a reconstructed image is obtained. Each noise reduction operation is constrained by the chromosome backbone. The specific method of step S2-2 is as follows: S2-11. The inverse diffusion model uses the U-Net++ network as a deep learning model. S2-12. Traverse the Markov chain backward to perform T denoising operations on the pure noise image; wherein, each denoising operation samples one noise from the Simplex noise. Then, the input image is denoised and its diffusion probability is reconstructed; among which, the deep learning model Each layer of the network embeds the chromosome backbone obtained in step S2-1, enabling the network to learn different morphological features and structural information at different levels. The U-Net++ network introduces an SE module to incorporate an attention mechanism, the implementation steps of which are as follows: S2-12-1. Compress each channel of the U-Net++ network using global average pooling to obtain the global feature vector for each channel. S2-12-2. Perform a linear transformation on the global feature vector of each channel and activate it using an activation function to obtain the weight of each channel; S2-12-3, each denoising operation multiplies the channel weights by the input feature map to obtain the output feature map reconstructed by the denoising diffusion probability after being weighted by the attention mechanism, wherein the input feature map is the output feature map of the previous step; The U-Net++ network is trained through pruning, and the implementation steps are as follows: 1) Train the complete U-Net++ network and record the gradient values ​​and Hessian matrix of each neuron or weight; 2) Based on the gradient values ​​and the indices of the Hessian matrix, calculate the degree of influence of each neuron or weight on the output of the U-Net++ network. : in, This represents the loss function for the i-th neuron. This represents the i-th neuron or its weight; 3) Prune neurons or weights according to the preset contribution threshold.

2. The method for detecting chromosome abnormalities based on a denoised diffusion probability model according to claim 1, characterized in that, In step S3, the mean squared error (MSE) or the cross-entropy loss function is calculated to evaluate the original image. and reconstructed images Perform a similarity comparison; The mean square error (MSE) is calculated as follows: Where I is the reconstructed image K is the original image ; It is the first in the reconstructed image The value of each pixel, K It is the first in the original image The value of each pixel; original image The number of pixels is m×n; if the mean square error (MSE) is greater than the threshold c, it is judged as abnormal, otherwise it is normal. The calculation method for the cross-entropy loss function is as follows: I is the reconstructed image K is the original image ; For the original image and reconstructed images The pixel values ​​are normalized and denoted as p and q respectively; by It is the first in the original image The value of each pixel. It is the first in the reconstructed image The value of each pixel; original image The number of pixels is m×n; Calculate the cross-entropy loss function : if If the value is greater than a threshold c, it is considered abnormal; otherwise, it is considered normal. The threshold c is obtained through pre-training. Before step S3, the original image is first aligned. and reconstructed images The steps for alignment are as follows: a) From the original image and reconstructed images Several feature points were extracted from the original image using the SIFT algorithm, followed by the FLANN algorithm. and reconstructed images Matching feature points in the data; b) Remove incorrect matching points by using the RANSAC algorithm to remove incorrectly matched feature points; c) Use a fundamental matrix or homography matrix to process the original image. and reconstructed images Perform geometric transformations to obtain the transformation matrix; and transform the original image based on the feature points of the matching points. and reconstructed images Any image in the image can be aligned with another image.

3. The method for detecting chromosome abnormalities based on a denoised diffusion probability model according to claim 1, characterized in that, The steps to solve for the threshold c are as follows: S4-1. Take N chromosome images that have been manually labeled as normal and N chromosome images that have been manually labeled as abnormal, and input them into the denoising diffusion probability model to obtain the mean square error (MSE) of the original image and the reconstructed image. Use different thresholds to determine whether the input chromosome image is normal or abnormal. S4-2, Solving for the true positive rate: ; S4-3, Solve for the false positive rate: ; S4-4, Solving for sensitivity: ; S4-5, Solving for specificity: ; S4-6. Calculate the Youden Index J: The point with the largest J value is selected as the optimal threshold c, where TP is the number of cases where both the judgment result and the labeling result are positive, FN is the number of cases where the judgment result is negative but the labeling result is positive, TN is the number of cases where both the judgment result and the labeling result are negative, and FP is the number of cases where the judgment result is positive but the labeling result is negative.

4. The method for detecting chromosome abnormalities based on a denoised diffusion probability model according to claim 1, characterized in that, The analysis model calculates the pixel difference matrix D: in, This represents the value of the pixel in the i-th row and j-th column in the difference image. and These represent the values ​​of the pixels in the i-th row and j-th column of the original image and the reconstructed image, respectively. The difference image is an m×n matrix obtained by calculating the difference between the original image and the reconstructed image for each pixel, where m and n are the length and width of the original image, respectively. The analysis model visualizes the difference image using a heatmap. The analytical model calculates the coordinates of abnormal chromosome regions through the following steps: S3-1. For the difference matrix D, set a judgment threshold. ;extract The area is denoted as The threshold It was obtained by optimizing and adjusting based on the results; S3-2, To Find the geometric center, denoted as . ; S3-3, For Perform connected component calculation to obtain the connected component it belongs to, denoted as . ; S3-4, to The process involves expansion followed by corrosion to fill the voids; the resulting area is denoted as [missing information]. This refers to the abnormal chromosomal region; S3-5. Perform skeleton extraction processing on the input chromosome image to obtain its skeleton coordinate vector S; S3-6, Calculate abnormal chromosome regions upper and lower boundaries and ,calculate and Intersection point with skeleton coordinate vector S and Then calculate and The distance from the starting end of the position in the skeleton S and And the total length l of S; S3-7. Based on the human chromosome coordinate information published by the international nomenclature system for human cell genomics, according to... and The percentage of l is used to determine the coordinates of the corresponding chromosomal abnormal region.

Images