A method, system, and medium for multi-focus view morphological analysis of embryo images

Abstract

The application discloses a kind of multi-focus view embryo image morphological analysis method, system and medium, it is related to image processing technical field, including: using the multi-focus view of isomeric fusion segmentation network to the input multiple focal plane embryo image is carried out multi-scale semantic feature extraction and fusion;The isomeric fusion segmentation network fuses Swin Transformer and U-Net structure, utilizes multi-head cross attention module to realize the joint modeling of global and local semantic features, and outputs embryo segmentation result;Wherein, the isomeric fusion segmentation network is by cross-image contrast learning strategy to carry out intra-class compactness constraint and inter-class dispersion constraint optimization, and curvature regularization loss function is introduced to carry out geometric prior constraint training and obtain;The method effectively enhances the analytical ability to weak contrast structure and the recognition accuracy of complex morphological characteristics.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, system, and medium for morphological analysis of embryonic images with multifocal views. Background Technology

[0002] Embryonic morphodynamic analysis relies on high-precision morphological analysis of specific developmental stages, and accurate segmentation is a key prerequisite for achieving quantitative morphological assessment. Morphological assessment indicators for cleavage-stage embryos mainly include the degree of fragmentation, the number of blastomeres, homogeneity, and symmetry, while for blastocysts, the focus is on the size of the blastocyst cavity, the number of inner cell mass and trophoblast cells. Figure 5 Key morphological features of the cleavage and blastocyst stages and their corresponding segmentation examples are shown. However, the embryo exhibits a complex topological structure at the biological level, such as... Figure 5 The cell overlap, adhesion, and deformation phenomena shown in (a1) and (a2), and Figure 5 The low-contrast intracellular tissue boundaries are shown in (b1) and (b2). Existing segmentation methods exhibit poor robustness under complex morphological variations and low contrast, making it difficult to meet the requirements for accurate embryo morphological analysis. Summary of the Invention

[0003] Based on the technical problems existing in the background technology, this invention proposes a method, system and medium for morphological analysis of embryonic images with multifocal views, which effectively enhances the ability to resolve weak contrast structures and the recognition accuracy of complex morphological features.

[0004] This invention proposes a method for morphological analysis of embryonic images using multifocal views, comprising: A heterogeneous fusion segmentation network with multiple focal planes is used to extract and fuse multi-scale semantic features from the input embryo images. The heterogeneous fusion segmentation network integrates the Swing Transformer and U-Net structures, uses a multi-head cross-attention module to achieve joint modeling of global and local semantic features, and outputs embryo segmentation results. The heterogeneous fusion segmentation network is obtained by optimizing intra-class compactness constraints and inter-class dispersion constraints through a cross-image contrastive learning strategy, and by introducing a curvature regularization loss function for geometric prior constraint training.

[0005] Furthermore, the heterogeneous fusion segmentation network adopts a U-Net structure as the overall framework, including an encoder, a decoder, and skip connections; The encoder uses a cascaded structure of ResNet50 and Swin Transformer. ResNet50 is used to extract multi-level local feature maps, and the deepest local feature map is input into Swin Transformer to establish global dependencies. The skip connection is used to fuse features from corresponding levels of the encoder and decoder; The decoder is equipped with a multi-head cross-attention module, which is used to fuse the local feature map extracted by the encoder through skip connections with the deep features upsampled by the decoder, so as to achieve joint modeling of global and local semantic features.

[0006] Furthermore, the multi-head cross-attention module includes a gating mechanism, which utilizes the deep features output by the encoder at each stage to generate a gating weight tensor through an activation function; Element-wise weighted summation is performed using the fused features from the gating weight tensor and the multi-head cross-attention module output, and then passed to the local feature map of the encoder via skip connections. The features are concatenated to obtain the output feature map of the current decoder layer.

[0007] Furthermore, the output feature map of the current decoder layer is specifically as follows: ; ; ; in, This is the output feature map of the current decoder layer. For the gated weight tensor, The features are those fused by the multi-head cross-attention module. These are the deep features after upsampling by the decoder. As a gated intermediate variable, For dimensional reshaping operations, It is the Sigmoid activation function. These are learnable parameters. This represents the Hadamard product. This is for splicing operations.

[0008] Furthermore, the intra-class compactness constraint This is achieved by minimizing the squared Euclidean distance between feature vectors of the same type of pixels, specifically: ; in, For the set of positive sample pairs across images, For two different pixels, For pixels and The feature embedding vectors are extracted by the encoding network of the heterogeneous fusion segmentation network encoder. This is a discrimination function that takes the value 1 when the condition is true and 0 otherwise. Each pixel and The true category label, The distance between two feature vectors is the Euclidean distance.

[0009] Furthermore, the inter-class dispersion constraint This is achieved by maximizing the distance between the feature vectors of different pixel categories to a preset threshold, specifically: ; in, For the set of negative sample pairs across images, For two different pixels, For pixels and The feature embedding vector is output by the feature extraction part of the heterogeneous fusion segmentation network. Let be the Euclidean distance between two eigenvectors. Each pixel and The true category label, This is the interval threshold. This is a discrimination function that takes the value 1 when the condition is true and 0 otherwise.

[0010] Furthermore, the curvature regularization loss function is calculated based on curvature, specifically as follows: For three adjacent pixels on the embryo boundary contour, the curvature is calculated using the cross product and dot product of the edge vectors. A curvature regularization loss function is then constructed to constrain the geometric continuity of the boundary. The middle pixel of these three pixels... Located at the intersection of the boundary contours of the two embryos, the other two pixels They are located on the boundary outlines of the two embryos respectively.

[0011] Furthermore, the curvature regularization loss function is formulated as follows: ; ; ; in, Let curvature regularization loss function be used. For including A set of contour pixels For the index of the outline pixels, For the first Curvature loss function of a point For the first The curvature of a point For the first Point to number The edge vector of a point, For the first Point to number The edge vector of a point.

[0012] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0013] A computer-readable storage medium storing a plurality of computer programs for being invoked by a processor and executing the method described above.

[0014] The advantages of the multi-focal view embryo image morphological analysis method, system, and medium provided by this invention are as follows: A heterogeneous fusion segmentation network is introduced, which employs a global-local joint representation learning mechanism based on contrastive learning, effectively enhancing the analytical capability for weak-contrast structures and the recognition accuracy of complex morphological features. A curvature regularization loss function is also set to achieve a unified modeling expression of biophysical characteristics, thereby improving the accuracy and stability of the network in the embryonic development feature modeling process. Furthermore, by introducing a geometric prior constraint mechanism, the intrinsic correlation attributes between structures in the topological hierarchy are enhanced, thus achieving joint analysis of weak-contrast structures and complex morphologies within a unified heterogeneous fusion segmentation network framework, improving the overall system's representational robustness and structural recognition performance. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the structural process of the present invention; Figure 2 This is a schematic diagram of the structure and flow of the multi-head cross-attention module; Figure 3 This is a schematic diagram of cross-image contrastive learning. Figure 4 This is a diagram illustrating the curvature; Figure 5 The images show the important morphological features and corresponding segmentation diagrams of the cleavage and blastocyst stages. (a1) is a microscopic image of the embryo at the cleavage stage, (a2) is a segmentation diagram of an embryo instance corresponding to (a1), (b1) is a microscopic image of the embryo at the blastocyst stage, and (b2) is a segmentation diagram of an embryo instance corresponding to (b1). Detailed Implementation

[0016] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0017] like Figure 1 and Figure 4 As shown, the present invention proposes a method for morphological analysis of embryonic images using multifocal views, comprising: A heterogeneous fusion segmentation network with multiple focal planes is used to extract and fuse multi-scale semantic features from the input embryo images. The heterogeneous fusion segmentation network integrates the Swing Transformer and U-Net structures, uses a multi-head cross-attention module to achieve joint modeling of global and local semantic features, and outputs embryo segmentation results. The heterogeneous fusion segmentation network is obtained by optimizing intra-class compactness constraints and inter-class dispersion constraints through a cross-image contrastive learning strategy, and by introducing a curvature regularization loss function for geometric prior constraint training.

[0018] This embodiment addresses the challenge of uniformly modeling the spatiotemporal dynamic evolutionary characteristics and individual phenotypic heterogeneity during embryonic development by proposing a heterogeneous fusion segmentation network based on multi-focal plane views. This network employs a global-local joint representation learning mechanism based on contrastive learning, effectively enhancing its ability to resolve weak-contrast structures and improve the recognition accuracy of complex morphological features. Furthermore, a curvature regularization loss function is implemented to achieve a unified modeling representation of biophysical characteristics, thereby improving the network's accuracy and stability in modeling embryonic developmental features. Building upon this, a geometric prior constraint mechanism is introduced to enhance the intrinsic correlation between structures within the topological hierarchy, thus achieving joint resolution of weak-contrast structures and complex morphologies within a unified heterogeneous fusion segmentation network framework, improving the overall system's representational robustness and structural recognition performance.

[0019] I. Global and local semantic mining based on Heterogeneous Fusion Segmentation Network (HFSNet); To address the issue of ambiguous feature representation in low-contrast regions, this embodiment employs a heterogeneous fusion segmentation network architecture. Through multi-focal plane feature fusion, the heterogeneous fusion segmentation network is guided to supplement attention to key embryonic features from multiple perspectives. Its core lies in constructing a global-local dependency collaborative modeling mechanism. This mechanism captures long-range global context through the SwinTransformer and achieves refined fusion of multi-level features through a designed multi-head cross-attention module, thereby realizing joint optimization of low-contrast cell morphology and complex topology within a unified framework. The SwinTransformer is a hierarchical visual Transformer architecture.

[0020] 1. Global context modeling and feature extraction; Given multiple focal planes of a single embryo. , Indicates the height of the input embryo image. This indicates the width of the input embryo image. To indicate the number of focal planes, a ResNet50-based encoder is first used to extract multi-level local feature maps. This process can be described as follows: ; in, For the first Local feature map of the hierarchy, For the first Local feature map of the hierarchy, For ResNet50 The forward propagation process consists of several phases. In the first phase of ResNet50... Corresponding to .

[0021] As an example of a multi-level local feature map: ; ; ; ; in, Corresponding to local feature maps The number of channels, It consists of four levels of local feature maps, from low to high, which are rich in information at different levels, ranging from spatial details to high-level semantics. It retains rich edge and texture details, while This encodes the most abstract category and contextual information. For resolution.

[0022] Subsequently, the deepest local feature map ( Input the Swin Transformer module to establish global dependencies. Swin Transformer effectively expands the receptive field of heterogeneous fusion segmentation networks through a moving window-based self-attention mechanism. Specifically, the moving window-based self-attention mechanism works by first restricting each attention operation to an independent small window to reduce computation, and then periodically moving the window to allow information from different windows to be fused, thus balancing local efficiency with global modeling.

[0023] The core operations of Swin Transformer can be simplified as follows: ; ; ; in, For layer normalization, right The feature map obtained after layer normalization; For a self-attention mechanism based on a moving window, This is an intermediate feature map calculated by window attention and superimposed with residual connections. The residual connections preserve the original input information and alleviate gradient vanishing. A multilayer perceptron typically consists of two fully connected layers and a nonlinear activation function (such as GELU) to perform nonlinear transformations on features and fuse channel information. This is an enhanced feature map output by the Swing Transformer module, incorporating global contextual information. This feature map is used by subsequent decoders to guide segmentation.

[0024] This embodiment uses a fixed-window self-attention mechanism, which integrates feature maps... The system divides the data into non-overlapping local windows, and self-attention is computed within each local window. The core objective is to reduce complexity and address the computational explosion problem by restricting attention computation to each local window. However, completely independent windows cannot exchange information, limiting the heterogeneous fusion segmentation network's ability to understand the global picture.

[0025] Therefore, this embodiment sets up a self-attention mechanism based on moving windows. Specifically, the window grid is usually moved to the lower right by a set number of pixels. The new local window after the movement will naturally cross the boundary of the original old local window, thereby connecting the originally isolated window areas and realizing cross-window information interaction.

[0026] In this embodiment, the Swin Transformer alternately stacks a self-attention mechanism based on a fixed window and a self-attention mechanism based on a moving window, enabling the heterogeneous fusion segmentation network to enjoy the efficiency of local window computation while allowing information to flow freely throughout the entire image, ultimately achieving a near-global receptive field.

[0027] Through the above The formula, heterogeneous fusion segmentation network overcomes the limitation of traditional convolutional neural networks in local receptive fields, laying the foundation for understanding the macroscopic structure of the entire embryo.

[0028] 2. Feature fusion based on multi-head cross-attention; In the decoder section, in order to accurately recover cell boundaries and maintain topological consistency, a multi-head cross-attention module is introduced to fuse the high-resolution features of the encoder with the deep semantic features of the decoder.

[0029] In the ResNet50 architecture configured above, a multi-head cross-attention module is set in stages 1 to M-1. The multi-head cross-attention module is set between each upsampling layer of the decoder and the corresponding skip connection. It is used to fuse the local feature map extracted by the encoder and the deep features upsampled by the decoder passed by the skip connection, so as to achieve joint modeling of global and local semantic features. M is the total number of stages in ResNet50.

[0030] Suppose that at a certain layer of the decoder, the upsampled deep features are... (As a query), the local feature map of the encoder passed by the skip connection is: (as key and value); The spatial size of the feature map of the current decoder layer. For the number of channels. The Multi-Head Cross-Attention Module (MHCA module) first uses linear projection to... and Mapping to different representation spaces, and then computing the output of each attention head: ; ; in, For the index of attention head, For the first Query projection weight matrix for each attention head For the first Key projection weight matrix for each attention head, For the first Value projection weight matrix for each attention head, The projected weight matrix is ​​a learnable matrix. It is the dimension of Value. For the first A query matrix with attention heads It is the first The key matrix of each attention head. It is the first The value matrix of each attention head, For normalized exponential functions, For the first The output features of each attention head.

[0031] The output features of all attention heads are concatenated and projected again to form preliminary fused features. This is a feature resulting from the fusion of multiple cross-attention modules.

[0032] ; in, The projected weight matrix is ​​a learnable matrix. For splicing operations, For the first The output features of an attention head The total number of attention heads.

[0033] 3. Feature selection for gating mechanisms; The multi-head cross-attention module includes a gating mechanism that utilizes deep features upsampled by the encoder. A gated weight tensor is generated through an activation function, and this gated weight tensor is fused with the features output by the multi-head cross-attention module. Perform element-wise weighted summation and pass the local feature maps of the encoder to the skip connections. The components are then assembled into the final output.

[0034] To dynamically filter and enhance important features and suppress background noise, this embodiment introduces a gating mechanism in the multi-head cross-attention module. This gating mechanism uses a deep feature upsampled by the decoder. Controlled gating weight tensor To modulate the fusion result.

[0035] ; ; ; in, This is the output feature map of the current decoder layer. For the gated weight tensor, The features are those fused by the multi-head cross-attention module. These are the deep features after upsampling by the decoder. As a gated intermediate variable, For dimensional reshaping operations, It is the Sigmoid activation function. These are learnable parameters. This represents the Hadamard product (element-by-element multiplication). This is for the splicing operation. Each element in the gating tensor Z has a value between 0 and 1, acting as a feature valve.

[0036] In particular, the multi-head cross-attention module introduces a gating mechanism in skip connections and Activation functions dynamically adjust feature weights, effectively suppressing noise interference in non-embryonic regions while preserving crucial local details. For example... Figure 2 As shown, specifically, the multi-head cross-attention module applies the upsampled deep features... Apply a gating operation to generate a gating weight tensor. Among them, the gate weight tensor Low-amplitude elements with median values ​​close to 0 correspond to background or irrelevant regions that need to be suppressed, and are processed through the Hadamard product. Achieve feature-selective enhancement and background suppression.

[0037] This gating mechanism not only ensures the effective preservation of high-resolution structural information, but also makes full use of the rich semantic features of deep networks, providing solid technical support for modeling the topological consistency of embryonic cell populations, thereby improving segmentation accuracy and enhancing the model's generalization ability.

[0038] The heterogeneous fusion segmentation network set up in this embodiment integrates the long-range dependency modeling capability of SwinTransformer and the multi-scale feature extraction advantage of U-Net by constructing a global-local feature collaboration mechanism, and establishes a cross-scale feature interaction network for multi-focal views, which effectively improves the structural representation accuracy of low-contrast regions.

[0039] 4. Cross-image semantic extraction optimized based on contrastive learning strategy; To address the critical issue of limited cross-image feature alignment in low-contrast embryo image segmentation—namely, the variations in morphology, brightness, and texture across images caused by microscopic imaging conditions, embryonic developmental stages, and individual biological differences—existing segmentation networks struggle to effectively model pixel-level semantic relationships between images, thus significantly limiting the model's generalization ability. This embodiment proposes a cross-image contrastive learning method to overcome this bottleneck.

[0040] Specifically, such as Figure 3 As shown, this embodiment introduces a cross-image contrastive learning strategy to fully utilize the rich contextual information between samples and construct a more compact feature embedding space, thereby enhancing the robustness of the heterogeneous fusion segmentation network to cross-image feature variations. The challenge of cross-image feature alignment for embryonic images can be attributed to differences in data distribution and the lack of pixel-level semantic associations.

[0041] To model semantic associations across image pixels, this embodiment introduces intra-class compactness constraints and inter-class dispersion constraints.

[0042] (1) Intra-class compactness constraint: This constraint makes samples of the same class closer in the feature space, thereby improving intra-class consistency. For example Figure 3 As shown, the optimized features should be highly clustered in the embedding space.

[0043] Intra-class compactness constraints This is achieved by minimizing the squared Euclidean distance between feature vectors of the same type of pixels, specifically:

[0044] in, For the set of positive sample pairs across images, For two different pixels, For pixels and The feature embedding vectors are extracted by the encoding network of the heterogeneous fusion segmentation network. This is a discrimination function that takes the value 1 when the condition is true and 0 otherwise. Each pixel and The true category label, The distance between two feature vectors is the Euclidean distance.

[0045] For pixel embedding vectors, This is a feature extraction network (usually the encoding network in a heterogeneous fusion segmentation network), and its parameters are: It receives input pixels. (In practice, it is usually the image patch where the pixel is located or the feature vector at that location), outputting the corresponding feature embedding vector. This embedding vector is a compact representation in a high-dimensional space, used to measure the semantic similarity between pixels. In this embodiment, when Pick , correspond and , correspond and .

[0046] In low-contrast image processing, feature representations are often not significant enough and are susceptible to noise interference. Introducing intra-class compactness constraints can enhance the feature consistency of samples within the same class, improve the discriminative ability of heterogeneous fusion segmentation networks, and reduce the impact of noise interference on feature learning, thereby improving the robustness of heterogeneous fusion segmentation networks. This intra-class compactness constraint mechanism effectively improves the segmentation accuracy of low-contrast images by enhancing intra-class consistency, expanding inter-class discriminative power, and optimizing feature representation.

[0047] (2) Inter-class dispersion constraint: This constraint strengthens the class boundaries by increasing the distance between pixels of different classes in the feature space, thereby improving the generalization ability and inter-class discrimination of the heterogeneous fusion segmentation network.

[0048] Inter-class dispersion constraints This is achieved by maximizing the distance between the feature vectors of different pixel categories to a preset threshold, specifically: ; in, For the set of negative sample pairs across images, For two different pixels, For pixels and The feature embedding vectors are extracted by the encoding network of the heterogeneous fusion segmentation network. Let be the Euclidean distance between two eigenvectors. This is the interval threshold. This is a discrimination function that takes the value 1 when the condition is true and 0 otherwise.

[0049] In low-contrast image processing, inter-class feature similarity can easily lead to blurring of the feature space, affecting the discriminative ability of heterogeneous fusion segmentation networks. Introducing inter-class dispersion constraints can effectively widen the feature distribution of different categories, optimize the feature space structure, and thus improve the ability of heterogeneous fusion segmentation networks to distinguish category boundaries.

[0050] In this embodiment, intra-class compactness constraints and inter-class dispersion constraints jointly construct a discriminative feature embedding space, establishing a feature modeling paradigm based on spatial correlation between samples. Pixels of the same class are compactly clustered, while pixels of different classes are clearly separated. This feature representation can significantly improve the generalization ability and segmentation accuracy of heterogeneous fusion segmentation networks under complex microscopic imaging conditions (low contrast, cell overlap, individual differences).

[0051] 5. Topological constraints based on curvature loss: For three adjacent pixels on the embryo boundary contour, the curvature is calculated by the cross product and dot product of the edge vectors, and a curvature regularization loss function is constructed to constrain the geometric continuity of the boundary.

[0052] To address the topological complexity issues arising from cell overlap, adhesion, and morphological deformation in embryo image segmentation, this embodiment proposes a topological constraint mechanism based on differential geometric properties. This embodiment constructs a rotation-invariant curvature loss function to suppress contour oscillations and improve segmentation stability. Curvature, as a key geometric quantity measuring the degree of local bending, has significant biological importance in embryo image analysis: the curvature change of a single embryo contour is usually smooth, but exhibits abrupt features in areas of cell overlap or contact. Based on this, this embodiment uses curvature features as a discriminant index for detecting adhesion regions. Through a curvature-driven loss constraint mechanism, it enhances the adaptability of heterogeneous fusion segmentation networks to complex topological structures, providing a new paradigm for accurate embryo image segmentation.

[0053] The curvature loss function, based on discrete differential geometry theory, discretizes the definition of the curvature differential of a continuous curve to adapt to pixel-level boundary calculations, and achieves a local approximation of curvature through the cross product of adjacent edge vectors. This embodiment first constructs a pixel-level embryonic boundary representation method, modeling the embryonic boundary as a set of discrete contour points with a thickness of one pixel. For any three adjacent pixels on the embryonic boundary contour... , , Their coordinates are denoted as follows: , , Based on this, Point to number The edge vector of a point can be represented as ,like Figure 4 As indicated by the purple arrow in the diagram. Point to number The edge vector of a point can be represented as ,like Figure 4 As indicated by the red arrow in the middle. Curvature of a point It can be represented as: .

[0054] The curvature The formula eliminates scale sensitivity through normalization (introducing a side-length weighting term into the denominator), ensuring the comparability of curvature measurements across images of different resolutions. Its core idea stems from the discrete Gaussian curvature integral theorem, which states that the curvature at a vertex is proportional to the sine of the angle between adjacent sides, thus guaranteeing consistency in the physical meaning of local curvature features. Furthermore, to enhance the constraint on embryonic morphological features, this embodiment defines a biologically meaningful curvature loss function. Its discrete form is as follows: .

[0055] For including A collection of contour pixels Curvature regularization loss function of the overall image It can be represented as: .

[0056] This curvature regularization loss function is introduced into a heterogeneous fusion segmentation network to constrain the embryonic boundary contour, thereby enhancing the influence of cell contour characteristics on the segmentation results. This embodiment deeply embeds differential geometric prior knowledge into the network training process, constructing rotationally invariant topological constraints through discrete curvature calculation. This ensures the geometric continuity and morphological consistency of embryonic cell boundary segmentation, achieving sub-pixel-level recognition accuracy for weak edge regions. This provides a new regularization paradigm for solving the accurate segmentation of weak-boundary, highly deformed targets in biomedical images.

[0057] 6. The training process of the heterogeneous fusion segmentation network is as follows: Suppose the training set contains Image of embryo, number 1 Zhang Image The set of pixels is ( For image resolution, Indicates the height of the input embryo image. (representing the width of the input embryo image), its true segmentation label is , For pixel index, This represents the number of categories.

[0058] The heterogeneous fusion segmentation network is trained using the training set, and the trainable parameters are adjusted by the intra-class compactness constraint, inter-class dispersion constraint, and topological constraint based on curvature loss set above.

[0059] Based on the above description of the embodiments, those skilled in the art will understand that the multi-focal view embryo image morphological analysis method, system, and medium described in this embodiment can be implemented purely in software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for morphological analysis of embryonic images using multifocal views, comprising: A heterogeneous fusion segmentation network with multiple focal planes is used to extract and fuse multi-scale semantic features from the input embryo images with multiple focal planes. The heterogeneous fusion segmentation network integrates the Swing Transformer and U-Net structures, utilizes a multi-head cross-attention module to jointly model global and local semantic features, and outputs embryo segmentation results; The heterogeneous fusion segmentation network is obtained by optimizing intra-class compactness constraints and inter-class dispersion constraints through a cross-image contrastive learning strategy, and by introducing a curvature regularization loss function for geometric prior constraint training.

2. The method according to claim 1, characterized in that, The heterogeneous fusion segmentation network adopts a U-Net structure as the overall framework, including an encoder, a decoder, and skip connections; The encoder uses a cascaded structure of ResNet50 and Swin Transformer. ResNet50 is used to extract multi-level local feature maps, and the deepest local feature map is input into Swin Transformer to establish global dependencies. The skip connection is used to fuse features from corresponding levels of the encoder and decoder; The decoder is equipped with a multi-head cross-attention module, which is used to fuse the local feature map extracted by the encoder through skip connections with the deep features upsampled by the decoder, so as to achieve joint modeling of global and local semantic features.

3. The method according to claim 2, characterized in that, The multi-head cross-attention module includes a gating mechanism, which utilizes the deep features output by the encoder at each stage to generate a gating weight tensor through an activation function. Element-wise weighted summation is performed using the fused features from the gating weight tensor and the multi-head cross-attention module output, and then passed to the local feature map of the encoder via skip connections. The features are concatenated to obtain the output feature map of the current decoder layer.

4. The method according to claim 3, characterized in that, The output feature map of the current decoder layer is specifically as follows: in, This is the output feature map of the current decoder layer. For the gated weight tensor, The features are those fused by the multi-head cross-attention module. These are the deep features upsampled by the decoder. As a gated intermediate variable, For dimensional reshaping operations, It is the Sigmoid activation function. These are learnable parameters. This represents the Hadamard product. This is for splicing operations.

5. The method according to claim 1, characterized in that, The intra-class compactness constraint This is achieved by minimizing the squared Euclidean distance between feature vectors of the same type of pixels, specifically: in, For the set of positive sample pairs across images, For two different pixels, For pixels and The feature embedding vectors are extracted by the encoding network of the heterogeneous fusion segmentation network encoder. This is a discrimination function that takes the value 1 when the condition is true and 0 otherwise. Each pixel and The true category label, The distance between two feature vectors is the Euclidean distance.

6. The method according to claim 1, characterized in that, The inter-class dispersion constraint This is achieved by maximizing the distance between the feature vectors of different pixel categories to a preset threshold, specifically: in, For the set of negative sample pairs across images, For two different pixels, For pixels and The feature embedding vector is output by the feature extraction part of the heterogeneous fusion segmentation network. Let be the Euclidean distance between two eigenvectors. Each pixel and The true category label, This is the interval threshold. This is a discrimination function that takes the value 1 when the condition is true and 0 otherwise.

7. The method according to claim 1, characterized in that, The curvature regularization loss function is calculated based on curvature, specifically as follows: For three adjacent pixels on the embryo boundary contour, the curvature is calculated using the cross product and dot product of the edge vectors. A curvature regularization loss function is then constructed to constrain the geometric continuity of the boundary. The middle pixel of these three pixels... Located at the intersection of the boundary contours of the two embryos, the other two pixels They are located on the boundary outlines of the two embryos respectively.

8. The method according to claim 7, characterized in that, The formula for the curvature regularization loss function is as follows: in, Let curvature regularization loss function be used. For including A set of contour pixels For the index of the outline pixels, For the first Curvature loss function of a point For the first The curvature of a point For the first Point to number The edge vector of a point, For the first Point to number The edge vector of a point.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of computer programs, which are used to be invoked by a processor and to execute the method as described in any one of claims 1-8.

Images