Chromosome structure anomaly detection method and device

Abstract

The invention provides a chromosome structure anomaly detection method and device, and solves the problems of insufficient local feature representation of chromosomes, neglect of an inherent sequential relationship in a chromosome image and limited detection range in the prior art. Comprising the following steps: S1, acquiring image data of chromosomes; s2, carrying out segmentation processing on the image data of the chromosomes to form image blocks; s3, encoding the image blocks according to a sequence form to form context representation of the image blocks; and S4, detecting whether the chromosome is abnormal or not according to the context representation of the image block, thereby realizing efficient and accurate detection of the local abnormality of the chromosome structure.

Description

Technical Field

[0001] The present invention relates to the field of detection technologies, and particularly to a method and device for detecting chromosomal structural abnormalities. Background Art

[0002] The detection of chromosomal structural abnormalities is crucial for diagnosing genetic diseases and advancing biomedical research. Different from numerical abnormalities that can be detected by simply counting chromosomes, structural abnormalities usually manifest as subtle local changes in chromosomal band patterns. These changes require more complex processing methods and the ability to reason about the internal temporal relationships of chromosomes. With the rapid development of Deep Learning (DL) technology, the application of DL to the detection of chromosomal structural abnormalities has attracted increasing attention from researchers. However, despite some progress, the accuracy of existing methods in general structural abnormality detection is only about 80%, highlighting the urgent need for more advanced and robust detection methods. Existing methods face the following three key limitations in practical applications:

[0003] (1) Insufficient local feature representation:

[0004] Structural abnormalities usually manifest as local changes in certain specific regions of chromosomes. Existing methods mainly focus on instance-level features, that is, the entire chromosome is used as a whole for global representation. Although this method helps to capture macroscopic patterns, it often fails to identify key subtle local changes, which are crucial for detecting structural abnormalities. Local changes reflect the fine-grained changes in chromosomal sub-regions, such as the offset of band patterns or the disruption of specific regions, and these details are often ignored in global representations. The gap between instance-level abstraction and the need for fine-grained local analysis significantly reduces the detection accuracy, highlighting the necessity of effectively modeling and integrating the information of both.

[0005] (2) Lack of sequence modeling:

[0006] Existing methods regard chromosomes as static images and ignore their inherent temporal relationships. However, cytogeneticists regard chromosomes as sequences of band patterns arranged along their longitudinal axes. Failing to model these spatio-temporal relationships significantly limits the ability of deep learning models to capture the rich sequence dependencies for structural abnormality detection.

[0007] (3) Limited detection range:

[0008] Many existing methods only detect specific types of structural anomalies and mainly rely on predefined classification tasks. However, this approach faces significant challenges in clinical scenarios because the manifestations of structural anomalies are diverse and complex. These anomalies often appear in unpredictable and heterogeneous forms, making rigid classification difficult to meet the needs of precise detection. In addition, classification-based methods usually lack the flexibility to span a wide range of anomalies, limiting their practicality in real-world applications. This reflects an urgent need for a more adaptable and comprehensive framework that can address anomaly types deviating from traditional classification, thereby enhancing its practical value in clinical diagnosis. Summary of the Invention

[0009] The present invention provides a method and device for detecting chromosomal structural anomalies to solve the problems of insufficient representation of local features of chromosomes, ignoring the inherent temporal relationship in chromosome images, and limited detection range in the prior art.

[0010] In a first aspect, the present invention provides a method for detecting chromosomal structural anomalies, specifically including the following steps:

[0011] Step S1: Obtain the image data of the chromosome;

[0012] Step S2: Segment the image data of the chromosome to form image patches;

[0013] Step S3: Encode the image patches in the form of a sequence to form the context representation of the image patches;

[0014] Step S4: Detect whether the chromosome has anomalies according to the context representation of the image patches.

[0015] Preferably, in step S2, segmenting the image data of the chromosome to form image patches specifically includes the following steps:

[0016] Step S201: Generate the longitudinal axis of the chromosome image according to the chromosome image;

[0017] Step S202: Segment the chromosome image into continuous image patches according to the longitudinal axis of the chromosome image.

[0018] Preferably, in step S201, generating the longitudinal axis of the chromosome image according to the chromosome image specifically includes the following steps:

[0019] Step S201a: Skeletonize the chromosome image through a thinning algorithm to form the longitudinal axis of the chromosome image;

[0020] Step S202b: Perform branch pruning on the longitudinal axis of the chromosome image to form the longitudinal axis of the chromosome image after branch pruning.

[0021] Step S203c: Smooth the longitudinal axis of the chromosome image after branch pruning to form the refined longitudinal axis of the chromosome image.

[0022] Preferably, in step S202, the chromosome image is segmented into continuous image blocks, satisfying the following rules: 1) Sample pixels at fixed intervals along the longitudinal axis of the chromosome image, and use each sampling point as the center to form an image block; 2) There is an overlapping area between adjacent image blocks to maintain continuity and prevent information loss.

[0023] Preferably, in step S3, the image blocks are encoded in the form of a sequence to form the context representation of the image blocks, specifically including the following steps:

[0024] Step S301: Obtain the image block, and extract the spatial features of the image block through the backbone network f enc Extract the spatial features of the image block.

[0025] Step S302: According to the local spatial features of the image block, add position information to the image block through the autoregressive model g ar Add position information to the image block.

[0026] Step S303: Process the image block through the autoregressive model g ar To form the context representation of the image block.

[0027] Preferably, in step S302, the autoregressive model g ar Combines position encoding and Transformer layers to add position information to the image block; wherein, the position encoding PE is calculated according to the index of each image block in the sequence, and is specifically expressed as follows:

[0028]

[0029]

[0030] Among them, i represents the index of the image block, i is a positive integer; k represents the dimension index of the feature vector, k is a positive integer; d represents the total dimension of the feature vector, d is a positive integer.

[0031] Preferably, before the autoregressive model g ar Performs processing, it also includes guiding the autoregressive model g ar To perform pre-training by constructing a spatio-temporal loss function; wherein, the spatio-temporal loss function The specific representation is as follows:

[0032]

[0033] Among them, λ 1 and λ 2 are hyperparameters used to control the relative importance of the two parts of the loss; represents the spatial alignment loss, The specific representation is as follows:

[0034]

[0035] Among them, y i represents the one-hot encoded label of the starting position of k subsequence blocks (when the prediction is correct, y i is 1, and when the prediction is incorrect, y i is 0), p i represents the probability that the predicted starting position is the correct position i, which is calculated through the softmax function; n is the total number of image blocks in the chromosome sequence, and n is a positive integer;

[0036] Among them, represents the time prediction loss, The specific representation is as follows:

[0037]

[0038] Among them, x k+m represents the (k + m)-th image block; the autoregressor g ar summarizes the local representations z≤k of all the previous k image blocks into the context latent representation c k = g ar (z≤k) (modeling the image blocks as a sequence, so each image block has a dependency on the previous image block. In the representation space (latent space) of the model, each image block has its own representation (latent representation). To emphasize that the representation learned by this method contains the dependency between the front and the back, this application uses "context" to modify this latent representation); among them, f k (x k+m , c k ) represents the density ratio model for maintaining the mutual information between x k+m and c k , and the specific representation is as follows:

[0039]

[0040] Among them, ∝ represents direct proportion.

[0041] Preferably, in step S4, according to the context representation of the image block, detecting whether the chromosome is abnormal specifically includes the following steps:

[0042] Step S401, obtain the context representation of homologous chromosomes (humans normally have 23 types of chromosomes, and each type of chromosome appears in pairs. Each pair of the above-mentioned chromosomes is called a homologous chromosome), and calculate the cosine similarity between the corresponding context representations of two homologous chromosomes (in this application, the "corresponding" means that homologous chromosome A includes two chromosomes a1 and a2. If the context representation formed by taking the 6th image block of chromosome a1 is taken, then its corresponding context representation is the context representation formed by taking the 6th image block of chromosome a2);

[0043] Step S402, calculate the cosine similarity of the context representation formed by each image block in each chromosome to form a threshold matrix;

[0044] Step S403, judge whether the chromosome is abnormal according to the type of the chromosome and the threshold matrix.

[0045] Preferably, in step S402, the threshold matrix T is specifically represented as follows:

[0046]

[0047] Among them, cls represents the chromosome category, cls ∈ [1, 22], cls is an integer; t represents the threshold coefficient, which is used to enhance the adaptability between different data sets; N represents the number of samples of each type of chromosome; Indicates that the result is rounded to the nearest integer.

[0048] Preferably, the CS-Net network model is used to implement the functions described in steps S2 - S3. The CS-Net network model includes a C-Patcher network unit and an S-Encoder network unit. The C-Patcher network unit and the S-Encoder network unit are connected in sequence. The C-Patcher network unit is used to implement the functions described in step S2, and the S-Encoder network unit is used to implement the functions described in step S3. The S-Encoder network unit includes a backbone network and an autoregressive model. In the S-Encoder network unit, the backbone network and the autoregressive model are connected in sequence; AB-Detector is used to implement the functions described in step S4.

[0049] In a second aspect, the present invention also provides a chromosome structural abnormality detection device, which specifically includes the following modules:

[0050] An image data acquisition module for acquiring image data of chromosomes;

[0051] An image segmentation module for segmenting the image data of the chromosomes to form image blocks;

[0052] A context representation module for encoding the image blocks in the form of a sequence to form a context representation of the image blocks;

[0053] An anomaly detection module for detecting whether there are anomalies in the chromosomes according to the context representation of the image blocks.

[0054] Preferably, the image segmentation module specifically includes the following sub-modules:

[0055] The first sub-module of image segmentation for generating a longitudinal axis of the chromosome image according to the chromosome image;

[0056] The second sub-module of image segmentation for segmenting the chromosome image into continuous image blocks according to the longitudinal axis of the chromosome image.

[0057] Preferably, the first sub-module of image segmentation specifically includes the following grandchild-modules:

[0058] The first grandchild-module for skeletonizing the chromosome image through a thinning algorithm to form a longitudinal axis of the chromosome image;

[0059] The second grandchild-module for pruning branches of the longitudinal axis of the chromosome image to form a longitudinally axis of the chromosome image after branch pruning;

[0060] The third grandchild-module for smoothing the longitudinally axis of the chromosome image after branch pruning to form a refined longitudinally axis of the chromosome image.

[0061] Preferably, in the second sub-module of image segmentation, when segmenting the chromosome image into continuous image blocks, the following rules are satisfied: 1) Sampling pixels at a fixed interval along the longitudinal axis of the chromosome image, and taking each sampling point as the center to form an image block; 2) There is an overlapping area between adjacent image blocks to maintain continuity and prevent information loss.

[0062] Preferably, the context representation module specifically includes the following sub-modules:

[0063] The first sub-module of context representation for obtaining the image block and extracting the spatial features of the image block through the backbone network f enc ;

[0064] The second sub-module of context representation for, according to the local spatial features of the image block, through the autoregressive gar Add position information to the image block;

[0065] The context represents the third sub-module, which is used to pass through the autoregressive model g ar Process the image block to form a context representation of the image block.

[0066] Preferably, in the context representation second sub-module, the autoregressive model g ar Combines position encoding and Transformer layers to add position information to the image block; where the position encoding PE is calculated according to the index of each image block in the sequence, and is specifically expressed as follows:

[0067]

[0068]

[0069] where i represents the index of the image block, i is a positive integer; k represents the dimension index of the feature vector, k is a positive integer; d represents the total dimension of the feature vector, d is a positive integer.

[0070] Preferably, the autoregressive model g ar Before processing, it also includes guiding the autoregressive model g by constructing a spatio-temporal loss function ar for pre-training; where the spatio-temporal loss function is specifically expressed as follows:

[0071]

[0072] where λ 1 and λ 2 are hyperparameters used to control the relative importance of the two parts of the loss; represents the spatial alignment loss, which is specifically expressed as follows:

[0073]

[0074] where y i represents the one-hot encoded label at the starting position of the k-block subsequence (y i is 1 for a correct prediction and y i is 0 for an incorrect prediction), p i represents the probability that the predicted starting position is the correct position i, calculated through the softmax function; n is the total number of image blocks in the chromosome sequence, n is a positive integer;

[0075] where, represents the time prediction loss, which is specifically expressed as follows:

[0076]

[0077] where x k+m represents the (k + m)-th image patch; the autoregressor summarizes the local representations z≤k of all the previous k image patches into a context latent representation c k = g ar (z≤k); where f k (x k+m , c k ) represents a density ratio model for maintaining the mutual information between x k+m and c k , which is specifically expressed as follows:

[0078]

[0079] where ∝ represents direct proportion.

[0080] Preferably, the anomaly detection module specifically includes the following sub-modules:

[0081] The first anomaly detection sub-module is used to obtain the context representation of homologous chromosomes and calculate the cosine similarity between the corresponding context representations of two homologous chromosomes;

[0082] The second anomaly detection sub-module is used to calculate the cosine similarity of the context representations formed by each image patch in each chromosome to form a threshold matrix;

[0083] The third anomaly detection sub-module is used to determine whether there is an anomaly in the chromosome according to the type of the chromosome and the threshold matrix.

[0084] Preferably, in the second anomaly detection sub-module, the threshold matrix T is specifically expressed as follows:

[0085]

[0086] where cls represents the chromosome category, cls ∈ [1, 22], cls is an integer; t represents the threshold coefficient, which is used to enhance the adaptability between different data sets; N represents the number of samples of each type of chromosome; means rounding the result to the nearest integer.

[0087] Preferably, the CS-Net network model is used to implement the functions described in the image segmentation module and the context representation module. The CS-Net network model includes a C-Patcher network unit and an S-Encoder network unit. The C-Patcher network unit and the S-Encoder network unit are connected in sequence. The C-Patcher network unit is used to implement the functions described in the image segmentation module. The S-Encoder network unit is used to implement the functions described in the context representation module. The S-Encoder network unit includes a backbone network and an autoregressive model. In the S-Encoder network unit, the backbone network and the autoregressive model are connected in sequence; the AB-Detector is used to implement the functions described in the anomaly detection module.

[0088] In a third aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the chromosome structure anomaly detection method according to any one of the first aspects of the present application.

[0089] In a fourth aspect, the present invention also provides an electronic device, which includes: a memory storing a computer program; a processor communicatively connected to the memory and executing the chromosome structure anomaly detection method according to any one of the first aspects of the present application when calling the computer program.

[0090] Compared with the prior art, the present invention has the following obvious prominent substantial features and remarkable advantages:

[0091] The present invention provides a chromosome structure anomaly detection method and device, which solves the problems of insufficient representation of local features of chromosomes, ignoring the inherent temporal relationship in chromosome images, and limited detection range in the prior art. At the same time, a CS-Net network model is proposed for chromosome structure anomaly detection. The network model includes three key components: C-Patcher, which is used to divide chromosomes into meaningful image blocks to retain sequence information; S-Encoder, which uses prediction-position loss to enhance representation learning through self-supervised pre-training; and AB-Detector, which combines domain-specific medical prior knowledge to achieve generalization in various structural anomalies. Through the CS-Net and the AB-Detector, efficient and accurate detection of local anomalies in chromosome structures is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] The accompanying drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0093] Figure 1 It is a flowchart of a method for detecting chromosomal structural abnormalities according to a preferred embodiment of the present invention.

[0094] Figure 2 It is a structural diagram of the CS-Net network model according to a preferred embodiment of the present invention.

[0095] Figure 3 It is a schematic diagram of extracting the context representation of homologous chromosomes through the pre-trained CS-Net network in a method for detecting chromosomal structural abnormalities according to a preferred embodiment of the present invention.

[0096] Figure 4 It is a schematic structural diagram of a device for detecting chromosomal structural abnormalities according to a preferred embodiment of the present invention. Detailed implementation manners

[0097] The present invention provides a method and a device for detecting chromosomal structural abnormalities. To make the objectives, technical solutions and effects of the present invention clearer and more definite, the following further describes the present invention in detail with reference to the accompanying drawings and by way of examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0098] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and do not necessarily need to describe a specific order or sequence. It should be understood that such data used can be interchanged under appropriate circumstances. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0099] Example 1:

[0100] As Figures 1 - 3 shown, a method for detecting chromosomal structural abnormalities described in this embodiment specifically includes the following steps:

[0101] Step S1, obtaining image data of chromosomes.

[0102] Step S2, performing segmentation processing on the image data of the chromosomes to form image blocks.

[0103] Optionally, step S2 specifically includes the following steps:

[0104] Step S201, generating a longitudinal axis of the chromosome image according to the chromosome image, as Figure 2 (B) shown.

[0105] Optionally, step S201 specifically includes the following steps:

[0106] Step S201a: Skeletonize the chromosome image through a thinning algorithm to form the longitudinal axis of the chromosome image.

[0107] Step S202b: Perform branch pruning on the longitudinal axis of the chromosome image to form the longitudinal axis of the chromosome image after branch pruning.

[0108] Step S203c: Smooth the longitudinal axis of the chromosome image after branch pruning to form the refined longitudinal axis of the chromosome image.

[0109] Step S202: Segment the chromosome image into continuous image blocks according to the longitudinal axis of the chromosome image.

[0110] Preferably, in step S202, segmenting the chromosome image into continuous image blocks satisfies the following rules: 1) Sample pixels at a fixed interval along the longitudinal axis of the chromosome image, and form an image block with each sampling point as the center; 2) There is an overlapping area between adjacent image blocks to maintain continuity and prevent information loss.

[0111] Step S3: Encode the image blocks in the form of a sequence to form the context representation of the image blocks, as Figure 3 shown.

[0112] Optionally, step S3 specifically includes the following steps:

[0113] Step S301: Obtain the image block, and extract the spatial features of the image block through the backbone network f enc Extract the spatial features of the image block.

[0114] Step S302: Add position information to the image block according to the local spatial features of the image block through the autoregressive model g ar Add position information to the image block.

[0115] Among them, the autoregressive model g ar Combines position encoding and Transformer layers to add position information to the image block; among them, the position encoding PE is calculated according to the index of each image block in the sequence, and is specifically expressed as follows:

[0116]

[0117]

[0118] Among them, i represents the index of the image block, and i is a positive integer; k represents the dimension index of the feature vector, and k is a positive integer; d represents the total dimension of the feature vector, and d is a positive integer.

[0119] Optionally, the autoregressive model g ar Before processing, it further includes guiding the autoregressive model g by constructing a spatio-temporal loss function ar for pre-training, as shown in Figure 2 (C); among them, the spatio-temporal loss function is specifically expressed as follows:

[0120]

[0121] Among them, λ 1 and λ 2 are hyperparameters used to control the relative importance of the two parts of the loss; represents the spatial alignment loss, which is specifically expressed as follows:

[0122]

[0123] Among them, y i represents the one-hot encoded label of the starting position of the k-block subsequence (y i is 1 for a correct prediction and y i is 0 for an incorrect prediction), p i represents the probability that the predicted starting position is the correct position i, calculated through the softmax function; n is the total number of image blocks in the chromosome sequence, and n is a positive integer;

[0124] Among them, represents the time prediction loss, which is specifically expressed as follows:

[0125]

[0126] Among them, x k+m represents the (k + m)-th image block; the autoregressive model summarizes the local representations z≤k of all the previous k image blocks in the latent space as the context latent representation c k = g ar (z≤k); among them, f k (x k+m , c k ) represents the density ratio model for maintaining the mutual information between x k+m and c k , and is specifically expressed as follows:

[0127]

[0128] Among them, ∝ represents direct proportionality.

[0129] Step S303: Process the image block through the autoregressive model g ar to form a context representation of the image block.

[0130] Step S4: Detect whether the chromosome is abnormal according to the context representation of the image block.

[0131] Optionally, step S4 specifically includes the following steps:

[0132] Step S401: Obtain the context representations of homologous chromosomes, and calculate the cosine similarity s i between the corresponding context representations of two homologous chromosomes, which is specifically expressed as follows:

[0133]

[0134] where and respectively represent the context representations of two homologous chromosomes.

[0135] Step S402: Calculate the cosine similarities of the context representations formed by each image block in each chromosome to form a threshold matrix.

[0136] Among them, the threshold matrix T is specifically expressed as follows:

[0137]

[0138] where cls represents the chromosome category, cls ∈ [1, 22], cls is an integer; t represents the threshold coefficient, which is used to enhance the adaptability between different datasets; N represents the number of samples of each type of chromosome; means rounding the result to the nearest integer.

[0139] Step S403: Judge whether the chromosome is abnormal according to the type of the chromosome and the threshold matrix.

[0140] If Figure 2As shown in (A), it is a schematic diagram of the CS-Net network model. The CS-Net network model is used to implement the functions described in steps S2 - S3. The CS-Net network model includes a C-Patcher network unit and an S-Encoder network unit. The C-Patcher network unit and the S-Encoder network unit are connected in sequence. The C-Patcher network unit is used to implement the functions described in step S2, and the S-Encoder network unit is used to implement the functions described in step S3. The S-Encoder network unit includes a backbone network and an autoregressive model. In the S-Encoder network unit, the backbone network and the autoregressive model are connected in sequence; the AB-Detector is used to implement the functions described in step S4.

[0141] Example 2:

[0142] As Figure 4 shown, a chromosome structure abnormality detection device according to this embodiment specifically includes an image data acquisition module, an image segmentation module, a context representation module, and an abnormality detection module.

[0143] The image data acquisition module is used to acquire the image data of chromosomes.

[0144] The image segmentation module is used to perform segmentation processing on the image data of the chromosomes to form image blocks.

[0145] Among them, the image segmentation module specifically includes a first sub-module for image segmentation and a second sub-module for image segmentation:

[0146] The first sub-module for image segmentation is used to generate the longitudinal axis of the chromosome image according to the chromosome image.

[0147] Among them, the first sub-module for image segmentation specifically includes the following grandson modules:

[0148] The first grandson module is used to perform skeletonization processing on the chromosome image through a thinning algorithm to form the longitudinal axis of the chromosome image.

[0149] The second grandson module is used to perform branch pruning processing on the longitudinal axis of the chromosome image to form the longitudinal axis of the chromosome image after branch pruning.

[0150] The third grandson module is used to perform smoothing processing on the longitudinal axis of the chromosome image after branch pruning to form a refined longitudinal axis of the chromosome image.

[0151] The second sub-module for image segmentation is used to segment the chromosome image into continuous image blocks according to the longitudinal axis of the chromosome image.

[0152] Among them, the chromosome image is segmented into continuous image blocks, satisfying the following rules: 1) Sampling pixels at fixed intervals along the longitudinal axis of the chromosome image, and taking each sampling point as the center to form an image block; 2) There is an overlapping area between adjacent image blocks to maintain continuity and prevent information loss.

[0153] A context representation module, configured to perform encoding processing on the image blocks in the form of a sequence to form a context representation of the image blocks;

[0154] Among them, the context representation module specifically includes a first context representation sub-module, a second context representation sub-module, and a third context representation sub-module:

[0155] The first context representation sub-module is configured to obtain the image block and extract the spatial features of the image block through the backbone network f enc Extract the spatial features of the image block;

[0156] The second context representation sub-module is configured to add position information to the image block according to the local spatial features of the image block through the autoregressive model g ar Add position information to the image block.

[0157] Among them, the autoregressive model g ar Combines positional encoding and Transformer layers to add position information to the image block; among them, the positional encoding PE is calculated according to the index of each image block in the sequence, and is specifically expressed as follows:

[0158]

[0159]

[0160] Among them, i represents the index of the image block, i is a positive integer; k represents the dimension index of the feature vector, k is a positive integer; d represents the total dimension of the feature vector, d is a positive integer.

[0161] Optionally, before the autoregressive model g ar Performs processing, it further includes guiding the autoregressive model g to perform pre-training by constructing a spatio-temporal loss function; among them, the spatio-temporal loss function ar Is specifically expressed as follows: Is specifically expressed as follows:

[0162]

[0163] Among them, λ 1 And λ 2 Are hyperparameters used to control the relative importance of the two parts of the loss; Represents the spatial alignment loss, Is specifically expressed as follows:

[0164]

[0165] Among them, y i represents the one-hot encoded label of the starting position of k subsequences (when the prediction is correct, y i is 1, and when the prediction is incorrect, y i is 0), and p i represents the probability that the predicted starting position is the correct position i, which is calculated through the softmax function; n is the total number of image patches of the chromosome sequence, and n is a positive integer;

[0166] Among them, represents the time prediction loss, which is specifically shown as follows:

[0167]

[0168] Among them, x k+m represents the (k + m)-th image patch; the autoregressor summarizes the local representations z≤k of all the previous k image patches in the latent space as the context latent representation c k = g ar (z≤k); among them, f k (x k+m , c k ) represents the density ratio model for maintaining the mutual information between x k+m and c k , which is specifically shown as follows:

[0169]

[0170] Among them, ∝ represents direct proportion.

[0171] The context representation third sub-module is used to process the image patch through the autoregressor g ar to form the context representation of the image patch.

[0172] The anomaly detection module is used to detect whether there is an anomaly in the chromosome according to the context representation of the image patch.

[0173] Among them, the anomaly detection module specifically includes an anomaly detection first sub-module, an anomaly detection second sub-module, and an anomaly detection third sub-module:

[0174] The anomaly detection first sub-module is used to obtain the context representations of homologous chromosomes and calculate the cosine similarity between the corresponding context representations of two homologous chromosomes.

[0175] The second sub-module for anomaly detection is used to calculate the cosine similarity of the context representations formed by each image patch in each chromosome, and form a threshold matrix.

[0176] Among them, the threshold matrix T is specifically represented as follows:

[0177]

[0178] Among them, cls represents the chromosome category, cls ∈ [1, 22], and cls is an integer; t represents the threshold coefficient, which is used to enhance the adaptability between different data sets; N represents the number of samples of each type of chromosome; It means that the result is rounded to the nearest integer.

[0179] The third sub-module for anomaly detection is used to determine whether there is an anomaly in the chromosome according to the type of the chromosome and the threshold matrix.

[0180] Preferably, the CS-Net network model is used to implement the functions described in the image segmentation module and the context representation module. The CS-Net network model includes a C-Patcher network unit and an S-Encoder network unit. The C-Patcher network unit and the S-Encoder network unit are connected in sequence. The C-Patcher network unit is used to implement the functions described in the image segmentation module, and the S-Encoder network unit is used to implement the functions described in the context representation module. The S-Encoder network unit includes a backbone network and an autoregressive model. In the S-Encoder network unit, the backbone network and the autoregressive model are connected in sequence; AB-Detector is used to implement the functions described in the anomaly detection module.

[0181] Example 3:

[0182] To further verify the performance of a chromosome structural anomaly detection method proposed by the present invention, this embodiment is verified through experiments, and the specific situation is as follows:

[0183] A) Data set

[0184] The data set used in this experiment comes from the Medical Genetics and Prenatal Diagnosis Center of a certain hospital, and includes the following two parts: (1) Normal data set: a set containing normal chromosomes; (2) Abnormal data set: a set containing chromosomally structurally abnormal chromosomes.

[0185] The normal data set contains a total of 24 types of chromosomes, including 22 autosomes and X and Y chromosomes. All samples are paired homologous chromosomes. The data set contains approximately 250,000 chromosome images, comprehensively covering the normal chromosome structure.

[0186] The abnormal dataset focuses on the structural abnormalities of chromosomes, covering 22 autosomes and involving 76 types of structural abnormalities. Similar to the normal dataset, all samples are paired homologous chromosomes. The abnormal dataset totals approximately 550 pairs of images, reflecting the inherent data imbalance between normal and abnormal samples.

[0187] Under the guidance of the proposed chromosome spatio-temporal loss, all samples in the normal dataset are used to pre-train the S-Encoder in a self-supervised manner. Subsequently, only the autosome samples from the normal dataset are used to calculate the chromosome similarity metric required for threshold matrix calculation. Finally, the normal and abnormal datasets are used to comprehensively evaluate the effectiveness of the model in detecting chromosome structural abnormalities.

[0188] B) Implementation details

[0189] 1. C-Patcher configuration

[0190] The C-Patcher module is designed to segment chromosomes into a sequence of consecutive image patches. The number of patches n is set to 13, i.e., blank image patches with a fixed length of 13 are padded in the chromosome image patch sequence. The size of each small patch is 32×32×3, i.e., p = 32.

[0191] 2. S-Encoder and pre-training process

[0192] The S-Encoder is based on the Pixel-CNN architecture for sequence modeling. To enhance its ability to capture spatial and temporal dependencies between image patches, positional encoding and Transformer layers are jointly introduced in Pixel-CNN. This hybrid design enables the S-Encoder to effectively integrate local and global relationships in the chromosome image patch sequence.

[0193] In the S-Encoder, each context representation is defined as a 2048-dimensional vector. The chromosome spatio-temporal loss is configured with k = 5, and the loss balance coefficients are empirically set as λ 1 = 0.1 and λ 2 = 1.

[0194] In the self-supervised pre-training of the S-Encoder, the model is trained for 300 epochs with a batch size of 256. The initial learning rate is set to 10 -4 and decays by a factor of 5 every 120 epochs. The training process is carried out on a GeForce RTX 3090 graphics card.

[0195] In addition, the threshold matrix for anomaly detection is calculated based on the autosome samples in the normal dataset. Both the normal and abnormal datasets are used for evaluation to ensure that the results can reflect the generalization ability of the system under various abnormal conditions.

[0196] C) Evaluation of Learned Representations

[0197] 1. Setup

[0198] To evaluate the quality of the representations learned by the S-Encoder, the Linear Probing Accuracy (L.P.ACC) is used as the main metric. The normal dataset is divided into a training set and a test set in an 8:2 ratio. ResNet14, ResNet34, ResNet50, and ViT-Base (ViT-B) are used as the encoder backbone networks of the S-Encoder to evaluate the impact of different architectures.

[0199] For fair comparison, CPCv2 and CVAE-GAN are reproduced under the same conditions and ResNet50 is used as the encoder. Additionally, ResNet34 and ResNet50 directly trained using cross-entropy loss are used as supervised learning baseline models.

[0200] 2. Results and Analysis

[0201] The evaluation results are summarized in Table 1, showing that the S-Encoder using the ResNet50 backbone network achieves the highest Linear Probing Accuracy (L.P.ACC) among all methods. Notably, its linear probing accuracy is better than that of the ResNet classifier directly trained in a supervised manner, indicating the effectiveness of the learned representations in capturing chromosome features.

[0202] Table 1 Evaluation of Learned Representations

[0203]

[0204] As the ResNet backbone expands from ResNet14 to ResNet50, the ability of the S-Encoder to represent chromosome features is correspondingly enhanced, reflecting the advantages of larger-scale architectures in representation learning. However, the experimental results show that when using ViT-B as the backbone of the S-Encoder, its representation ability decreases. We speculate that there are mainly two reasons for the performance degradation when using ViT-B as the backbone of the S-Encoder. 1) ViT-B is architecturally designed to process 16×16×3 image patches, while the input of the S-Encoder is 32×32×3 image patches. Larger image patches provide richer spatial information, but due to the limitations of the original design of ViT-B, it cannot effectively process this information. 2) The Transformer architecture (such as ViT) requires a significantly larger dataset to achieve effective representation learning. In the scenario of chromosome analysis where the data is relatively scarce, this poses a challenge to fully leveraging the potential of ViT. Therefore, the mismatch between the patch size and the encoder design, as well as the high data volume requirement of ViT, jointly lead to the performance degradation when using ViT-B as the backbone.

[0205] D) Anomaly Detection Performance Evaluation

[0206] 1. Evaluation Metrics and Settings

[0207] The performance of CS-Net in detecting chromosomal structural anomalies is mainly evaluated by the false positive rate (FPR) and the false negative rate (FNR), which are particularly crucial in anomaly detection scenarios. In addition, to compare with other methods, the accuracy (ACC) and F1-score (F1) are also included. The above metrics are defined as follows:

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214] In anomaly detection, the number of anomaly samples is usually much less than that of normal samples, resulting in severe class imbalance. In this case, a model that classifies all samples as normal may also obtain a high accuracy (ACC) as defined by the formula. However, this result is unacceptable in practice because it completely ignores the anomalies.

[0215] The FPR (False Positive Rate) is used to measure the false alarm rate, while the FNR (False Negative Rate) represents the frequency of missed detections. In a clinical setting, reducing the missed detection rate is particularly important as missed detections can have serious consequences for patient outcomes. In contrast, a slightly higher tolerance can be accepted for the false positive rate as false positives can typically be confirmed through subsequent testing. Therefore, the FPR and FNR are emphasized as more meaningful metrics for evaluating the performance of CS-Net.

[0216] To evaluate the effectiveness of the CS-Net framework in detecting chromosomal structural abnormalities, we conducted a comprehensive comparison with eight methods: including four traditional baselines (k-NN, LSTM, ResNet50, and Sequence Analysis) and four advanced representation learning methods (EfficientNet, AutoEncoder, CVAE-GAN, and CPCv2). These methods cover a variety of scenarios from general classification to specialized anomaly detection.

[0217] First, we experimented with the CVAE-GAN method, using its pre-trained encoder to encode chromosome pair images and detecting anomalies by comparing the similarity of the encoded representations. Second, we also experimented with the CPCv2 method, using its pre-trained encoder to generate representations of chromosome pair images and analyzing their similarity to identify anomalies. Next, we used the encoder of ResNet50, which is specialized for classification tasks, to process chromosome pair images and detect anomalies through the similarity of the encodings. Finally, we developed a non-deep learning sequence analysis method: extending both ends of the vertical axis of each chromosome and extracting the gray value sequence along the axis. Then, we applied Canonical Time Warping (CTW) to align and compare the gray value sequences of chromosome pairs to detect anomalies.

[0218] 2. Results and Analysis

[0219] Table 2 summarizes the comparison results of each method, including the ACC, F1, FPR, and FNR metrics, and details the performance of CS-Net under these evaluation criteria.

[0220] Table 2 Evaluation of Learned Representations

[0221]

[0222] Baseline methods (k-NN, LSTM, ResNet50, Sequence Analysis): These methods are difficult to effectively model spatial and temporal relationships, resulting in lower accuracy (ACC) and F1 score. Among them, the k-NN and sequence analysis methods have a high false positive rate and frequent false alarms, which are not practical in clinical use. Although ResNet50 has strong feature extraction capabilities, its performance is still poor due to its inability to combine time series relationships.

[0223] Advanced representation learning methods (EfficientNet, AutoEncoder, CVAE-GAN, CPCv2): EfficientNet outperforms AutoEncoder and CVAE-GAN due to its strong feature extraction capabilities, but it is still limited by the lack of time series modeling capabilities. CVAE-GAN and CPCv2 benefit from the representation learning framework and can capture data features, but due to the lack of explicit spatio-temporal modeling capabilities, the false negative rate is relatively high.

[0224] Performance of CS-Net: The experimental results show that CS-Net performs excellently in terms of false positive rate (0.116) and false negative rate (0.006), setting a new benchmark in the field of chromosome structural anomaly detection. This outstanding performance indicates that CS-Net can minimize the missed detection rate, which is particularly important in clinical applications, while maintaining a low false alarm rate. In addition, these results highlight the effectiveness of the proposed spatio-temporal modeling method, which significantly outperforms techniques that only consider space or time by simultaneously capturing the visual features of chromosome images and their inherent sequence information. This comprehensive modeling framework enables CS-Net to demonstrate excellent accuracy and robustness in detecting various structural anomalies.

[0225] 3. Comparison with existing advanced methods

[0226] To evaluate the performance of CS-Net in detecting chromosome structural anomalies, we conducted a comprehensive review of the current state-of-the-art methods. This review shows that there is a significant scarcity of general methods in this field, indicating a large gap in the field. Therefore, we also considered some specialized methods, although the number of relevant studies is still limited. Notably, the method proposed in

[18] is designed specifically for numerical anomaly detection and can be used as a reference point, reflecting the relative simplicity of numerical anomaly detection compared to structural anomaly detection. The comparison results are shown in Table 3.

[0227] Although we reviewed the research scope of existing methods, there are significant differences in the performance metrics reported in the literature for these methods. Among these studies, accuracy (ACC) and F1-score are the most commonly used evaluation metrics. Therefore, we adopted the ACC or F1-score reported in the original literature and provided these two metrics for CS-Net to ensure a fair and comprehensive evaluation.

[0228] In contrast, CS-Net achieved 0.896 and 0.656 in terms of ACC and F1-score respectively, outperforming the method published by Zhang et al. in the paper "Chromosome Classification and Straightening Based on an Interleaved and Multi-Task Network". Meanwhile, our validation dataset maintained a distribution of 90% normal samples and 10% abnormal samples, covering 76 different types of structural abnormalities, which is close to the real-world data distribution. If a balanced dataset with 50% normal samples and 50% abnormal samples is used, higher performance metrics (an ACC and F1-score of approximately 0.94) may be obtained, but this cannot reflect the applicability of the model in practice. Therefore, the performance metrics obtained from this experiment based on the real data distribution are shown in Table 3.

[0229] Meanwhile, the method published by Li et al. in the paper "Chromosomal Structural Abnormality Diagnosis by Homologous Similarity" achieved the highest F1-score of 0.975. However, this method only focuses on detecting five specific deletion abnormalities on chromosome 5, making it highly specialized. Deletion is a type of structural abnormality that manifests as the loss of a chromosomal segment, usually resulting in a significant shortening of the chromosome. This obvious length difference makes the identification of deletions relatively easy, and the simplicity of this detection may be the reason for the high F1-score of their method compared to abnormalities that do not involve significant changes in chromosome length.

[0230] In addition, the method proposed by Li et al. in the paper "CS-GANomaly: A Supervised Anomaly Detection Approach with Ancillary Classifier GANs for Chromosome Images" achieved the highest ACC (0.983) and F1-score (0.663) by synthesizing data for five specific chromosomal structural abnormalities. In contrast, CS-Net verified 76 different types of structural abnormalities in approximately 850 pairs of chromosomes, with an F1-score of 0.656. This highlights the versatility and robustness of our general method, whose performance is comparable to that of this method when dealing with a significantly more complex and diverse range of abnormalities.

[0231] In summary, the existing methods have limited performance in the field of chromosomal structural abnormality detection, and the method proposed in the present invention has obvious advantages in performing chromosomal structural abnormality detection.

[0232] E) Ablation experiments

[0233] To comprehensively evaluate the contributions of the key components and design choices in CS-Net, we conducted ablation experiments, focusing on three key aspects: the backbone architecture, the prediction location loss coefficients (λ 1 and λ 2 ), and the spatio-temporal modeling parameters (t and h). These experiments evaluated the effects of the backbone network configuration and pre-training strategy, as well as the impact of the threshold parameters on the downstream task of chromosomal structural abnormality detection. The experimental results are summarized in Tables 4 and 5.

[0234] 1. Effects of different backbone architectures and loss coefficients

[0235] The backbone network in the S-Encoder plays a crucial role in extracting meaningful features from chromosome patches. To evaluate its impact, we conducted experiments using three ResNet variants - ResNet14, ResNet34, and ResNet50 - as well as the advanced ViT-Base architecture. These experiments used different prediction location loss coefficients (λ 1 and λ 2 ) during the pre-training phase. Table 4 shows the results of the linear probing experiments, reporting the accuracy for each configuration.

[0236] Table 4 Ablation experiments on backbone network structure and spatio-temporal loss

[0237]

[0238]

[0239] In summary, among all the tested architectures, ResNet50 has consistently performed well and achieved the highest accuracy of 91.59 at λ 1 = 0.1 and λ 2 = 1. This result highlights the advantages of deep CNN architectures in capturing complex spatio-temporal relationships, especially when training on smaller datasets.

[0240] In contrast, ViT-Base performed poorly, probably because its architecture is less efficient in handling relatively small training datasets and the relatively large image patch size of 32×32 used in this study.

[0241] 2. Influence of threshold parameters (t and h)

[0242] In CS-Net, the threshold parameters t and h play a crucial role in determining the trade-off between sensitivity and specificity in anomaly detection. The threshold parameter t affects the calculation of the threshold matrix T and determines whether a patch is classified as an anomaly based on the similarity of homologous patches. The parameter h sets the minimum number of abnormal patches required to classify a chromosome as structurally abnormal. To evaluate the influence of the threshold parameters t and h, we evaluated the performance of CS-Net using different ResNet backbones. These experiments focused on anomaly detection metrics, especially the false positive rate (FPR) and false negative rate (FNR), details of which are shown in Table 5. It should be noted that ViT-B was excluded from these ablation experiments because of its poor performance in the early evaluations. The best pre-trained weights from previous ablation studies were used for each backbone architecture to ensure consistency and fairness in the comparison.

[0243] Table 5 Ablation experiments on threshold parameters (FPR / FNR)

[0244]

[0245]

[0246] As shown in Table 5, the threshold parameters t and h have a significant impact on the detection performance of CS-Net. A higher value of t corresponds to a more stringent threshold, which reduces the false negative rate (FNR) but increases the false positive rate (FPR) as fewer patches are labeled as anomalies. Conversely, a lower value of t relaxes the threshold, resulting in a lower FPR but at the cost of increasing the FNR. For example, there is a trade-off between FPR and FNR as t or h increases. For example, when using ResNet50 and h = 3, increasing t from 0.1 to 0.2 reduces the FNR from 0.039 to 0.011, but increases the FPR from 0.056 to 0.152. This highlights the trade-off that must be considered when choosing an appropriate value of t according to application requirements.

[0247] Similarly, the parameter h, which defines the minimum number of anomaly patches required to trigger anomaly detection, plays a crucial role in balancing FPR and FNR. For example, in ResNet50 and at t = 0.1, increasing h from 1 to 3 reduces the FPR from 0.379 to 0.056, but increases the FNR from 0.002 to 0.039. This shows that a higher h value helps reduce false positives by requiring stronger evidence for anomaly detection, but may overlook subtle structural anomalies.

[0248] Among all the tested configurations, ResNet50 achieved the most balanced performance at t = 0.1 and h = 2, with an FPR of 0.116 and an FNR of 0.006. This balance is crucial for practical applications where the goal is to minimize false positives and missed detections while maintaining strong detection accuracy.

[0249] The performance of a chromosome structural anomaly detection method described in the present invention was verified through the above-mentioned extensive experiments, demonstrating its ability to detect 76 different types of structural anomalies and outperforming existing dedicated and general methods in terms of performance metrics. At the same time, CS-Net achieved state-of-the-art accuracy while maintaining a low false positive rate and false negative rate. The above experimental results demonstrate the excellent performance of a chromosome structural anomaly detection method described in the present invention.

[0250] The specific embodiments of the present invention have been described in detail above, but they are only examples, and the present invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, all equivalent transformations and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.

Claims

1. A method for detecting chromosome structural abnormality, characterized in that: The specific steps include: Step S1, obtaining chromosome image data; Step S2, segmenting the chromosome image data to form image blocks; Step S3, encoding the image block in a sequence form to form a context representation of the image block; Step S4: Detect whether the chromosome is abnormal based on the context representation of the image block.

2. A method for detecting chromosome structural abnormality according to claim 1, characterized in that: In step S2, the chromosome image data is segmented to form image blocks, which specifically includes the following steps: Step S201, generating a longitudinal axis of the chromosome image according to the chromosome image; Step S202: Segment the chromosome image into continuous image blocks according to the longitudinal axis of the chromosome image.

3. A method for detecting chromosome structural abnormality according to claim 2, characterized in that: In step S201, the longitudinal axis of the chromosome image is generated according to the chromosome image, which specifically includes the following steps: Step S201a, skeletonizing the chromosome image by a thinning algorithm to form a longitudinal axis of the chromosome image; Step S202b, performing branch pruning processing on the longitudinal axis of the chromosome image to form the longitudinal axis of the chromosome image after branch pruning; Step S203c: smoothing the longitudinal axis of the chromosome image after the branch pruning to form a refined longitudinal axis of the chromosome image.

4. A method for detecting chromosome structural abnormality according to claim 2, characterized in that: In step S202, the chromosome image is segmented into continuous image blocks, satisfying the following rules: 1) pixels are sampled at fixed intervals along the longitudinal axis of the chromosome image, with each sampling point as the center to form an image block; 2) adjacent image blocks have overlapping areas to maintain continuity and prevent information loss.

5. A method for detecting chromosome structural abnormality according to claim 1, characterized in that: In step S3, the image block is encoded in a sequence to form a context representation of the image block, which specifically includes the following steps: Step S301: Obtain the image block and pass it through the backbone network f enc Extracting spatial features of the image block; Step S302: According to the local spatial features of the image block, the autoregressive g ar Adding position information to the image block; Step S303: using the autoregressor g ar The image block is processed to form a context representation of the image block.

6. A method for detecting chromosome structural abnormality according to claim 5, characterized in that: In step S302, the autoregressive g ar Combine the position encoding and the Transformer layer to add position information to the image block; wherein the position encoding PE is calculated according to the index of each image block in the sequence, and is specifically represented as follows: Among them, i represents the index of the image block, i is a positive integer; k represents the dimension index of the feature vector, k is a positive integer; d represents the total dimension of the feature vector, d is a positive integer.

7. A method for detecting chromosome structural abnormality according to claim 5, characterized in that: The autoregressor g ar Before processing, it also includes guiding the autoregressor g by constructing a space-time loss function ar Pre-training is performed; wherein the space-time loss function The specific representation is as follows: Among them, λ1 and λ2 are hyperparameters used to control the relative importance of the two parts of loss; represents the spatial alignment loss, The specific representation is as follows: Among them, y i The one-hot encoded label representing the starting position of the k-block subsequence (predicting the correct y i is 1, the prediction error y i is 0), p i It represents the probability that the predicted starting position is the correct position i, which is calculated by the softmax function; n is the total number of image blocks of the chromosome sequence, and n is a positive integer; in, represents the time prediction loss, The specific representation is as follows: Among them, x k+m represents the k+mth image patch; the autoregressive summarizes the local representations z≤k of all the first k image patches in the latent space into the contextual latent representation c k =g ar (z≤k); where f k (x k+m ,c k ) indicates that it is used to maintain x k+m and c k The density ratio model of the mutual information between is specifically expressed as follows: Here, ∝ indicates a direct proportion.

8. A method for detecting chromosome structural abnormality according to claim 1, characterized in that: In step S4, whether the chromosome is abnormal is detected according to the context representation of the image block, which specifically includes the following steps: Step S401, obtaining context representations of homologous chromosomes, and calculating the cosine similarity between the corresponding context representations of two homologous chromosomes; Step S402, calculating the cosine similarity of the context representation formed by each image block in each chromosome to form a threshold matrix; Step S403: judging whether the chromosome is abnormal according to the type of the chromosome and the threshold matrix; In step S402, the threshold matrix T is specifically represented as follows: Where cls represents the chromosome category, cls∈[1,22], cls is an integer; t represents the threshold coefficient, which is used to enhance the adaptability between different data sets; N represents the number of samples of each type of chromosome; Indicates that the result is rounded to the nearest integer.

9. A method for detecting chromosome structural abnormality according to claim 1, characterized in that: The CS-Net network model is used to implement the functions described in steps S2 to S3. The CS-Net network model includes a C-Patcher network unit and an S-Encoder network unit. The C-Patcher network unit and the S-Encoder network unit are connected in sequence. The C-Patcher network unit is used to implement the function described in step S2. The S-Encoder network unit is used to implement the function described in step S3. The S-Encoder network unit includes a backbone network and an autoregressive model. In the S-Encoder network unit, the backbone network and the autoregressive model are connected in sequence. AB-Detector is used to implement the function described in step S4.

10. A chromosome structural abnormality detection device, characterized in that: The modules include: An image data acquisition module, used for acquiring image data of chromosomes; An image segmentation module, used for segmenting the image data of the chromosome to form image blocks; A context representation module, used for encoding the image block in a sequence to form a context representation of the image block; The abnormality detection module is used to detect whether the chromosome has an abnormality according to the context representation of the image block.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, a method for detecting chromosome structural abnormalities according to any one of claims 1 to 9 is implemented.

12. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, a method for detecting chromosome structural abnormality as described in any one of claims 1 to 9 is implemented.

Images