Chromosome image segmentation method based on artificial intelligence

Abstract

The invention discloses a chromosome image segmentation method based on artificial intelligence, which can solve the problems that the contour of chromosomes in a high-power lens image depends on manual annotation, the annotation workload is large, and the accuracy and consistency of manual annotation are difficult to guarantee, realizes the training of a diversified chromosome image sample set, guarantees the effectiveness of a chromosome image segmentation model, and improves the image segmentation efficiency. And the generalization ability and accuracy of the model are improved. The method comprises the following steps: acquiring a target chromosome image, and preprocessing the target chromosome image through morphological operation; acquiring contour data of overlapped chromosomes in the target chromosome image; constructing a diversified chromosome image training sample set, and training a chromosome image segmentation model by adopting a plurality of chromosome image training samples to obtain a trained chromosome image segmentation model; and inputting the preprocessed target chromosome image and the contour data of the overlapped chromosomes in the target chromosome image into the trained chromosome image segmentation model for processing to obtain a segmentation result of the overlapped chromosomes.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing and artificial intelligence, and in particular to a chromosome image segmentation method based on artificial intelligence. Background Art

[0002] Currently, due to the softness of chromosomes and the discreteness of samples, there are often various overlapping and adhesion phenomena in chromosome samples, such as Figure 1 As shown, this makes chromosome segmentation difficult.

[0003] To address this issue, existing techniques typically involve manually marking chromosome outlines and obtaining mask images of overlapping chromosomes. This is followed by building and training an intelligent segmentation network for overlapping chromosomes to form a segmentation model. This model is then used to segment the overlapping chromosome images and generate mask images of individual chromosomes.

[0004] However, although this method can achieve the segmentation of overlapping chromosomes, it relies on manual annotation of the chromosome outlines in high-magnification images. Therefore, it may lead to a large amount of annotation work, and the accuracy and consistency of manual annotation are difficult to guarantee.

[0005] Secondly, the intelligent segmentation network used in this method is limited to specific datasets and image quality. Once the dataset is not diverse enough or the chromosome image quality is not high, the generalization ability and accuracy of the model may be affected. Summary of the Invention

[0006] In order to overcome the defects of the existing technology, the technical problem to be solved by the present invention is to provide a chromosome image segmentation method based on artificial intelligence, which can solve the problem of relying on manual annotation of the outlines of chromosomes in high-magnification images, resulting in a large amount of annotation workload and difficulty in ensuring the accuracy and consistency of manual annotation. It can realize the training of diverse chromosome image sample sets, ensure the effectiveness of the chromosome image segmentation model, and improve the generalization ability and accuracy of the model.

[0007] The technical solution of the present invention is: a chromosome image segmentation method based on artificial intelligence, which comprises the following steps:

[0008] (1) Obtaining the target chromosome image and preprocessing the target chromosome image through morphological operations;

[0009] (2) Identify whether there is an overlapping region on the target chromosome based on the preprocessed target chromosome image. If yes, execute step (3); otherwise, execute step (6);

[0010] (3) obtaining the outline data of the overlapping chromosomes in the target chromosome image;

[0011] (4) constructing a diverse chromosome image training sample set, and using multiple chromosome image training samples to train a chromosome image segmentation model to obtain a trained chromosome image segmentation model;

[0012] (5) inputting the preprocessed target chromosome image and the outline data of the overlapping chromosomes in the target chromosome image into the trained chromosome image segmentation model for processing to obtain the segmentation results of the overlapping chromosomes;

[0013] (6) End;

[0014] The chromosome image segmentation model in step (4) includes: encoder, decoder,

[0015] BConvLSTM layer, Swin Transformer module;

[0016] The encoder consists of four stages, each of which consists of two separable convolutional layers using 3×3 filters, followed by a 2×2 max pooling layer and a rectified linear unit (ReLU) activation function. For each subsequent stage, the number of filters doubles compared to the previous stage. By gradually increasing the layer dimension, the encoder path gradually extracts visual features and generates high-level semantic information based on the high-dimensional image representation in the last layer; the corresponding decoder performs output calculations;

[0017] The BConvLSTM layer extends the

[0018] ConvLSTM, the input data is processed in two separate paths, forward and backward. Each path has a ConvLSTM layer for processing data in sequence. The forward path processes the input data in its natural order, from the first image to the last image, and the reverse path processes the data in reverse from the last image to the first image. The ConvLSTM layer includes a storage unit Output gate φ t , input gate i t and forget the door f t ,These gates are used as the control mechanism of the ConvLSTM layer, ,where the input gate, output gate, and forget gate control the access, ,update, and clearing of the memory cell respectively;

[0019] The Swin Transformer module divides the input into non-overlapping patches, each patch encapsulating a 4×4 pixel area, a patch token whose associated feature is a combination of RGB pixel values; these features of the original value are projected onto the selected dimension using a linear embedding layer LE; a series of transformer blocks equipped with a modified self-attention calculation are applied to these patch tokens. Block 1 consists of a transformer block and LE, retaining h / 4×w / 4, where h is the height and w is the width. The layers are combined to create a hierarchical representation by reducing the number of tokens. Transformers are deployed to change the features while keeping the resolution at h / 4×w / 8. The beginning stage of patch merging and feature conversion is designated as block 2.

[0020] The beneficial technical effects of the present invention are as follows:

[0021] 1. The preprocessed target chromosome image and the outline data of the overlapping chromosomes in the target chromosome image (including local curvature, pixel intensity gradient, radius from the outline point to the chromosome center of gravity, change in radius of adjacent points, and outline point set information) are input into a trained chromosome image segmentation model for processing to obtain the segmentation results of the overlapping chromosomes, thereby achieving overlapping chromosome image segmentation. This method can solve the problem that the existing technology relies on manual annotation of chromosome outlines in high-magnification microscopy images, resulting in a large annotation workload and difficulty in ensuring the accuracy and consistency of manual annotation.

[0022] 2. Construct a diverse chromosome image training sample set to ensure the effectiveness of the chromosome image segmentation model. This can avoid the problem in existing technologies where insufficient diversity in the dataset leads to low generalization and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A sample diagram of an actual chromosome.

[0024] Figure 2 Schematic diagram of the chromosome image segmentation method based on artificial intelligence according to the present invention used in a slide scanning imaging system.

[0025] Figure 3 The figure is a flow chart of the chromosome image segmentation method based on artificial intelligence according to the present invention.

[0026] Figure 4 : is a network structure diagram of the chromosome image segmentation model according to the present invention.

[0027] Figure 5 Schematic diagram of overlapping chromosome images before and after segmentation using the method of the present invention. DETAILED DESCRIPTION

[0028] like Figure 3 As shown, a chromosome image segmentation method based on artificial intelligence includes the following steps:

[0029] (1) Obtaining the target chromosome image and preprocessing the target chromosome image through morphological operations;

[0030] (2) Identify whether there is an overlapping region on the target chromosome based on the preprocessed target chromosome image. If yes, execute step (3); otherwise, execute step (6);

[0031] (3) obtaining the outline data of the overlapping chromosomes in the target chromosome image;

[0032] (4) constructing a diverse chromosome image training sample set, and using multiple chromosome image training samples to train a chromosome image segmentation model to obtain a trained chromosome image segmentation model;

[0033] (5) inputting the preprocessed target chromosome image and the outline data of the overlapping chromosomes in the target chromosome image into the trained chromosome image segmentation model for processing to obtain the segmentation results of the overlapping chromosomes;

[0034] (6) End;

[0035] like Figure 4 As shown, the chromosome image segmentation model in step (4) includes: encoder, decoder, BConvLSTM layer, Swin Transformer module;

[0036] The encoder consists of four stages, each of which consists of two separable convolutional layers using 3×3 filters, followed by a 2×2 max pooling layer and a rectified linear unit (ReLU) activation function. For each subsequent stage, the number of filters doubles compared to the previous stage. By gradually increasing the layer dimension, the encoder path gradually extracts visual features and generates high-level semantic information based on the high-dimensional image representation in the last layer; the corresponding decoder performs output calculations;

[0037] The BConvLSTM layer extends the

[0038] ConvLSTM, the input data is processed in two separate paths, forward and backward. Each path has a ConvLSTM layer for processing data in sequence. The forward path processes the input data in its natural order, from the first image to the last image, and the reverse path processes the data in reverse from the last image to the first image. The ConvLSTM layer includes a storage unit Output gate φ t , input gate i t and forget the door f t,These gates are used as the control mechanism of the ConvLSTM layer, ,where the input gate, output gate, and forget gate control the access, ,update, and clearing of the memory cell respectively;

[0039] The Swin Transformer module divides the input into non-overlapping patches, each patch encapsulating a 4×4 pixel area, a patch token whose associated feature is a combination of RGB pixel values; these features of the original value are projected onto the selected dimension using a linear embedding layer LE; a series of transformer blocks equipped with a modified self-attention calculation are applied to these patch tokens. Block 1 consists of a transformer block and LE, retaining h / 4×w / 4, where h is the height and w is the width. The layers are combined to create a hierarchical representation by reducing the number of tokens. Transformers are deployed to change the features while keeping the resolution at h / 4×w / 8. The beginning stage of patch merging and feature conversion is designated as block 2.

[0040] The beneficial technical effects of the present invention are as follows:

[0041] 1. The preprocessed target chromosome image and the outline data of the overlapping chromosomes in the target chromosome image (including local curvature, pixel intensity gradient, radius from the outline point to the chromosome center of gravity, change in radius of adjacent points, and outline point set information) are input into a trained chromosome image segmentation model for processing to obtain the segmentation results of the overlapping chromosomes, thereby achieving overlapping chromosome image segmentation. This method can solve the problem that the existing technology relies on manual annotation of chromosome outlines in high-magnification microscopy images, resulting in a large annotation workload and difficulty in ensuring the accuracy and consistency of manual annotation.

[0042] 2. Construct a diverse chromosome image training sample set to ensure the effectiveness of the chromosome image segmentation model. This can avoid the problem in existing technologies where insufficient diversity in the dataset leads to low generalization and accuracy of the model.

[0043] Preferably, in step (1), the target chromosome image is converted into binary using a preset threshold; an erosion operation is performed on the binary image, using a 3x3 square structure element to reduce the pixels of the chromosome in the image for better segmentation; objects with more than 800 pixels are removed using connected components; the chromosome is restored to its original size by enlarging using the same 3x3 square structure element; and a bitwise AND operation is performed on the enlarged image and the original image to preserve the original size and shape of the chromosome.

[0044] Preferably, in step (4), constructing a chromosome image training sample set includes the following steps:

[0045] (4.1.1) Extracting the outer contours of overlapping chromosomes in the target chromosome image using an edge detection algorithm;

[0046] (4.1.2) Based on the outer contour information of the overlapping chromosomes, calculate the main axis direction and center point of each chromosome to form the initial contour data;

[0047] (4.1.3) Based on the outer contour and initial contour data, calculate the local curvature of each point on the contour:

[0048]

[0049] x′(p), y′(p) represent the tangent direction of point p, x″(p), y″(p) represent the curvature change rate of point p, and K(p) represents the local curvature of point p on the contour;

[0050] (4.1.4) For each point on the contour, calculate its pixel intensity gradient:

[0051]

[0052] Indicates the rate of change of pixel intensity along the x and y directions;

[0053] (4.1.5) Determine the curvature enhancement gradient based on the pixel intensity gradient and the local curvature:

[0054]

[0055] G c (p) represents the curvature enhancement gradient, α represents the weight value of the local curvature, and β represents the weight value of the pixel intensity gradient;

[0056] (4.1.6) Further calculate the radius from the contour point to the chromosome center of gravity, and calculate the adjacent

[0057] The change in point radius;

[0058] (4.1.7) Combine the radius change and curvature enhancement gradient to calculate the union of each contour point

[0059] score:

[0060] S(p)=γ·G c (p)+δ·Δr(p)

[0061] S(p) represents the joint score of the contour points, Δr(p) represents the change in the radius of adjacent points, γ represents the weight of the curvature enhancement gradient in the selection of segmentation points, and δ represents the weight of the radius change in the selection of segmentation points;

[0062] (4.1.8) Use density clustering algorithm to cluster high joint score points, and each cluster center point represents a segmentation point group;

[0063] (4.1.9) Based on the clustering results, a segmentation line from one segmentation point group to another segmentation point group is generated through a dynamic elastic deformation function model;

[0064] (4.1.10) Perform morphological operations on the generated segmentation lines to remove pseudo-boundaries and ensure that the segmentation lines are closed. Combined with the biological characteristics of chromosome length, width, and aspect ratio, the segmentation results are verified and adjusted to eliminate abnormal segmentation results; the normal segmentation results are used as the first chromosome image sample set.

[0065] Preferably, the step (4.1.9) includes the following:

[0066] E=E elastic +E boundary +E curvature

[0067] E represents the optimization objective function, and the goal is to optimize the objective function to make the generated segmentation line smooth and accurately fit the boundary characteristics of the chromosome;

[0068]

[0069] E elastic It is the second-order derivative integral of the dividing line, which indicates the curvature or smoothness of the dividing line in space. represents the second-order derivative of the dividing line, indicating the degree of bending of the dividing line;

[0070]

[0071] E boundary Indicates the matching degree between the segmentation line and the target boundary point, P represents the contour point, and I(p) represents the image boundary point corresponding to the segmentation line at the contour point P;

[0072]

[0073] E curvature It indicates the matching degree between the segmentation line and the high curvature area, ensuring that the segmentation line passes through the key points of the chromosome overlapping area. K(p) is the curvature value of the high curvature point, which indicates the intensity of the local mutation or concave of the contour of the high curvature point.

[0074] Alternatively, in step (4), constructing the chromosome image training sample set includes the following steps:

[0075] (4.2.1) Convert the input chromosome contour into a reference shape that matches the reference shape, and use the Procrustes transform to perform a first deformation operation on the input shape so that the key points of the two are aligned;

[0076] (4.2.2) Through the second Procrustes transformation, the initial deformed reference shape is guided back to the coordinate system of the input contour point set;

[0077] (4.2.3) Generate segmentation lines based on the optimized shape point set to obtain the second chromosome image sample set.

[0078] Preferably, the specific formula of step (4.2.2) is as follows:

[0079]

[0080] Among them, pts is the shape point set of the input chromosome, representing the input target shape, c is the geometric center of the input shape, which is used to adjust the input shape point set to align with the geometric center of the reference shape to ensure that the transformed shape is centered on the reference shape, T -1 is the inverse of the rotation and reflection matrix, which reflects the angle and direction adjustment of the input shape to the reference shape to ensure the consistency of the direction between the input shape and the reference shape. b is the scaling component, which is used to adjust the size of the input shape to make it consistent with the size ratio of the reference shape. A correction parameter ∈ is added to prevent the denominator from being zero to prevent numerical overflow. global is the global weight, which controls the influence of the global transformation and is related to the global matching degree of the overall shape of the chromosome. Its value is dynamically adjusted according to the overall difference between the input shape and the reference shape. local is the local weight, used to enhance the local characteristics of the overlapping area, Δ local It is the local error feedback used to correct the local transformation deviation and is calculated as follows:

[0081]

[0082] Among them, pts i is the i-th point in the input point set, ref i is the position of the corresponding point in the reference shape.

[0083] Preferably, the step (4.2.3) comprises the following steps:

[0084] (4.2.3.1) Use density clustering algorithm to cluster high joint score points, and each cluster center point represents a segmentation point group;

[0085] (4.2.3.2) Based on the clustering results, a segmentation line from one segmentation point group to another segmentation point group is generated using a dynamic elastic deformation function model;

[0086] (4.2.3.3) Perform morphological operations on the generated segmentation lines to remove pseudo-boundaries and ensure that the segmentation lines are closed. Combined with the biological characteristics of chromosome length, width, and aspect ratio, verify and adjust the segmentation results to eliminate abnormal segmentation results; the normal segmentation results are used as the first chromosome image sample set.

[0087] Preferably, in step (4), is the encoding feature χ in The output of the i-th encoding block, where i = 1, 2, 3, is expressed as:

[0088]

[0089] represents the maximum pooling operation,

[0090]

[0091] l *×* Represents the result of batch normalization operation, represents depthwise separable convolution, (*×*) represents kernel size, represents the ReLU activation function, β N Indicates batch normalization operation; there are three encoding blocks in the chromosome image segmentation model, and the spatial input dimension gradually becomes smaller. is W×H×C, for for After the encoding block, three densely connected depthwise separable convolutional blocks Used with spatial input dimensions The output of the first dense block is calculated by applying two consecutive operations l 3 ×3 and the ReU activation on the last encoding block Then STB3(S ViT ) and connect it with itself to get

[0092]

[0093] Indicates concatenation operation;

[0094] The output of the second dense block is obtained by applying two consecutive l 3×3 operation and ReLU activation and connect and two consecutive operations l 3×3 Followed by ReLU activation Get

[0095]

[0096] The output of the third dense block is obtained by applying two consecutive l 3×3 operation and ReLU activation and connect and two consecutive operations l 3×3 Followed by ReLU activation Get

[0097]

[0098] Correspondingly, the output of the decoder block is calculated as follows:

[0099]

[0100] Among them, T c is the transposed convolution operation of the i-th encoding block, S ViT Represents a lightweight Swin TransformerBlock, represents BConvLSTM;

[0101] The final output of the chromosome image segmentation model is χ out It is obtained by the following calculation method:

[0102]

[0103] Output of the batch normalization step It is fed into the ConvLSTM layer. The structure and operation of ConvLSTM are as follows:

[0104]

[0105] and Represent Hadamard and convolutional operations respectively, and the input tensor and hidden tensor are expressed as and The 2D convolution masks of hidden and input states are represented by The bias terms of storage unit, output, input and forget gate are represented by β φ ,β i and β f express.

[0106] Preferably, the output of BConvLSTM in step (4) is calculated as follows:

[0107]

[0108] and The hidden state tensors representing the forward and backward states respectively, β is the bias component, and the hyperbolic tangent function tanh is used to nonlinearly combine the outputs of the forward and backward states in BConvLSTM.

[0109] Preferably, the training of the chromosome image segmentation model in step (4) comprises the following steps:

[0110] (4.3.1) Inputting the first chromosome image sample set, the second chromosome image sample set, and the contour data involved in generating the first chromosome image sample set and the second chromosome image sample set into the chromosome image segmentation model to obtain a model segmentation result;

[0111] (4.3.2) Compare the model segmentation results with the actual segmentation results and calculate the loss difference between the two;

[0112] (4.3.3) The chromosome image segmentation model is iteratively trained using the loss difference until the loss difference of the chromosome image segmentation model meets the preset threshold. The iterative update is stopped and the final model is used as the trained chromosome image segmentation model.

[0113] Figure 2 A schematic diagram of the AI-based chromosome image segmentation method according to the present invention used in a slide scanning imaging system. The intelligent slide scanning imaging system based on the present invention can provide powerful AI-assisted evaluation capabilities for chromosome technology, enabling fully intelligent karyotype analysis and chromosome image segmentation through AI. Figure 5 The diagram shows overlapping chromosome images before and after segmentation using the method of the present invention. The method can achieve a chromosome segmentation accuracy of ≥99.5%.

[0114] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of protection of the technical solution of the present invention.

Claims

1. A chromosome image segmentation method based on artificial intelligence, characterized by: It includes the following steps: (1) Obtaining the target chromosome image and preprocessing the target chromosome image through morphological operations; (2) Identify whether there is an overlapping region on the target chromosome based on the preprocessed target chromosome image. If yes, execute step (3); otherwise, execute step (6); (3) obtaining the outline data of the overlapping chromosomes in the target chromosome image; (4) constructing a diverse chromosome image training sample set, and using multiple chromosome image training samples to train a chromosome image segmentation model to obtain a trained chromosome image segmentation model; (5) inputting the preprocessed target chromosome image and the outline data of the overlapping chromosomes in the target chromosome image into the trained chromosome image segmentation model for processing to obtain the segmentation results of the overlapping chromosomes; (6) End; The chromosome image segmentation model in step (4) includes: encoder, decoder, BConvLSTM layer, Swin Transformer module; The encoder consists of four stages, each of which consists of two separable convolutional layers using 3×3 filters, followed by a 2×2 max pooling layer and a rectified linear unit (ReLU) activation function. For each subsequent stage, the number of filters doubles compared to the previous stage. By gradually increasing the layer dimension, the encoder path gradually extracts visual features and generates high-level semantic information based on the high-dimensional image representation in the last layer; the corresponding decoder performs output calculations; The BConvLSTM layer extends the ConvLSTM, the input data is processed in two separate paths, forward and backward. Each path has a ConvLSTM layer for processing data in sequence. The forward path processes the input data in its natural order, from the first image to the last image, and the reverse path processes the data in reverse from the last image to the first image. The ConvLSTM layer includes a storage unit Output gate φ t , input gate i t and forget the door f t ,These gates are used as the control mechanism of the ConvLSTM layer, ,where the input gate, output gate, and forget gate control the access, ,update, and clearing of the memory cell respectively; The Swin Transformer module divides the input into non-overlapping patches, each patch encapsulating a 4×4 pixel area, a patch token whose associated feature is a combination of RGB pixel values; these features of the original value are projected onto the selected dimension using a linear embedding layer LE; a series of transformer blocks equipped with a modified self-attention calculation are applied to these patch tokens. Block 1 consists of a transformer block and LE, retaining h / 4×w / 4, where h is the height and w is the width. The layers are combined to create a hierarchical representation by reducing the number of tokens. Transformers are deployed to change the features while keeping the resolution at h / 4×w / 8. The beginning stage of patch merging and feature conversion is designated as block 2.

2. The chromosome image segmentation method based on artificial intelligence according to claim 1, characterized in that: In the step (1), the target chromosome image is converted into binary using a preset threshold; Perform an erosion operation on the binary image, using a 3x3 square structure element to reduce the pixels of the chromosomes in the image for better segmentation; use connected components to remove objects with more than 800 pixels; restore the original size of the chromosomes by enlarging them using the same 3x3 square structure element; perform a bitwise AND operation on the enlarged image and the original image to preserve the original size and shape of the chromosomes.

3. The chromosome image segmentation method based on artificial intelligence according to claim 2, characterized in that: In step (4), constructing a chromosome image training sample set includes the following steps: (4.1.1) Extracting the outer contours of overlapping chromosomes in the target chromosome image using an edge detection algorithm; (4.1.2) Based on the outer contour information of the overlapping chromosomes, calculate the main axis direction and center point of each chromosome to form the initial contour data; (4.1.3) Based on the outer contour and initial contour data, calculate the local curvature of each point on the contour: x ′ (p),y ′ (p) represents the tangent direction of point p, x″(p), y″(p) represents the curvature change rate of point p, and K(p) represents the local curvature of point p on the contour; (4.1.4) For each point on the contour, calculate its pixel intensity gradient: Indicates the rate of change of pixel intensity along the x and y directions; (4.1.5) Determine the curvature enhancement gradient based on the pixel intensity gradient and the local curvature: G c (p) represents the curvature enhancement gradient, α represents the weight value of the local curvature, and β represents the weight value of the pixel intensity gradient; (4.1.6) Further calculate the radius from the contour point to the chromosome centroid, and calculate the change in radius of adjacent points; (4.1.7) Combine the radius change and the curvature enhancement gradient to calculate the joint score of each contour point: S(p)=γ·G c (p)+δ·Δr(p) S(p) represents the joint score of the contour points, Δr(p) represents the change in the radius of adjacent points, γ represents the weight of the curvature enhancement gradient in the selection of segmentation points, and δ represents the weight of the radius change in the selection of segmentation points; (4.1.8) Use density clustering algorithm to cluster high joint score points, and each cluster center point represents a segmentation point group; (4.1.9) Based on the clustering results, a segmentation line from one segmentation point group to another segmentation point group is generated through a dynamic elastic deformation function model; (4.1.10) Perform morphological operations on the generated segmentation lines to remove pseudo-boundaries and ensure that the segmentation lines are closed. Combined with the biological characteristics of chromosome length, width, and aspect ratio, the segmentation results are verified and adjusted to eliminate abnormal segmentation results; the normal segmentation results are used as the first chromosome image sample set.

4. The chromosome image segmentation method based on artificial intelligence according to claim 3, characterized in that: The step (4.1.9) includes the following: E=E elastic +E boundary +E curvature E represents the optimization objective function, and the goal is to optimize the objective function to make the generated segmentation line smooth and accurately fit the boundary characteristics of the chromosome; E elastic It is the second-order derivative integral of the dividing line, which indicates the curvature or smoothness of the dividing line in space. represents the second-order derivative of the dividing line, indicating the degree of bending of the dividing line; E boundary Indicates the matching degree between the segmentation line and the target boundary point, P represents the contour point, and I(p) represents the image boundary point corresponding to the segmentation line at the contour point P; E curvature It indicates the matching degree between the segmentation line and the high curvature area, ensuring that the segmentation line passes through the key points of the chromosome overlapping area. K(p) is the curvature value of the high curvature point, which indicates the intensity of the local mutation or concave of the contour of the high curvature point.

5. The chromosome image segmentation method based on artificial intelligence according to claim 2, characterized in that: In step (4), constructing a chromosome image training sample set includes the following steps: (4.2.1) Convert the input chromosome contour into a reference shape that matches the reference shape, and use the Procrustes transform to perform a first deformation operation on the input shape so that the key points of the two are aligned; (4.2.2) Through the second Procrustes transformation, the initial deformed reference shape is guided back to the coordinate system of the input contour point set; (4.2.3) Generate segmentation lines based on the optimized shape point set to obtain the second chromosome image sample set.

6. The chromosome image segmentation method based on artificial intelligence according to claim 5, characterized in that: The specific formula of step (4.2.2) is as follows: Among them, pts is the shape point set of the input chromosome, representing the input target shape, c is the geometric center of the input shape, which is used to adjust the input shape point set to align with the geometric center of the reference shape to ensure that the transformed shape is centered on the reference shape, T -1 is the inverse of the rotation and reflection matrix, which reflects the angle and direction adjustment of the input shape to the reference shape to ensure the consistency of the direction between the input shape and the reference shape. b is the scaling component, which is used to adjust the size of the input shape to make it consistent with the size ratio of the reference shape. A correction parameter ∈ is added to prevent the denominator from being zero to prevent numerical overflow. global is the global weight, which controls the influence of the global transformation and is related to the global matching degree of the overall shape of the chromosome. Its value is dynamically adjusted according to the overall difference between the input shape and the reference shape. local is the local weight, used to enhance the local characteristics of the overlapping area, Δ local It is the local error feedback used to correct the local transformation deviation and is calculated as follows: Among them, pts i is the i-th point in the input point set, ref i is the position of the corresponding point in the reference shape.

7. The chromosome image segmentation method based on artificial intelligence according to claim 6, characterized in that: The step (4.2.3) comprises the following steps: (4.2.3.1) Use density clustering algorithm to cluster high joint score points, and each cluster center point represents a segmentation point group; (4.2.3.2) Based on the clustering results, a segmentation line from one segmentation point group to another segmentation point group is generated using a dynamic elastic deformation function model; (4.2.3.3) Perform morphological operations on the generated segmentation lines to remove pseudo-boundaries and ensure that the segmentation lines are closed. Combined with the biological characteristics of chromosome length, width, and aspect ratio, verify and adjust the segmentation results to eliminate abnormal segmentation results; the normal segmentation results are used as the first chromosome image sample set.

8. The chromosome image segmentation method based on artificial intelligence according to claim 4 or 7, characterized in that: In the step (4), is the encoding feature χ in The output of the i-th encoding block, where i = 1, 2, 3, is expressed as: represents the maximum pooling operation, l *×* Represents the result of batch normalization operation, represents depthwise separable convolution, (*×*) represents kernel size, represents the ReLU activation function, β N Represents batch normalization operation; There are three encoding blocks in the chromosome image segmentation model, and the spatial input dimension gradually becomes smaller. is W×H×C, for for After the encoding block, three densely connected depthwise separable convolutional blocks Used with spatial input dimensions The output of the first dense block is calculated by applying two consecutive operations l 3×3 and the ReLU activation on the last encoding block Then STB(S ViT ) and connect it with itself to get θ represents the concatenation operation; The output of the second dense block is obtained by applying two consecutive l 3×3 Operation and ReLU activation and connect and two consecutive operations l 3×3 Followed by ReLU activation Get The output of the third dense block is obtained by applying two consecutive l 3×3 operation and ReLU activation and connect and two consecutive operations l 3×3 Followed by ReLU activation Get Correspondingly, the output of the decoder block is calculated as follows: Among them, T c is the transposed convolution operation of the i-th encoding block, S ViT Represents the lightweight Swin Transformer Block, represents BConvLSTM; The final output of the chromosome image segmentation model is χ out It is obtained by the following calculation method: Output of the batch normalization step It is fed into the ConvLSTM layer. The structure and operation of ConvLSTM are as follows: and Represent Hadamard and convolutional operations respectively, and the input tensor and hidden tensor are expressed as and The 2D convolution masks of hidden and input states are represented by The bias terms of storage unit, output, input and forget gate are represented by and β f express.

9. The chromosome image segmentation method based on artificial intelligence according to claim 8, characterized in that: The output of BConvLSTM in step (4) is calculated as follows: and The hidden state tensors representing the forward and backward states respectively, β is the bias component, and the hyperbolic tangent function tanh is used to nonlinearly combine the outputs of the forward and backward states in BConvLSTM.

10. The chromosome image segmentation method based on artificial intelligence according to claim 9, characterized in that: The training of the chromosome image segmentation model in step (4) comprises the following steps: (4.3.1) Inputting the first chromosome image sample set, the second chromosome image sample set, and the contour data involved in generating the first chromosome image sample set and the second chromosome image sample set into the chromosome image segmentation model to obtain a model segmentation result; (4.3.2) Compare the model segmentation results with the actual segmentation results and calculate the loss difference between the two; (4.3.3) The chromosome image segmentation model is iteratively trained using the loss difference until the loss difference of the chromosome image segmentation model meets the preset threshold. The iterative update is stopped and the final model is used as the trained chromosome image segmentation model.

Images