Identification and prediction system of breast cancer molecular subtypes based on ultrasound images

Abstract

The present invention relates to the field of breast cancer ultrasound image recognition and prediction, and specifically refers to a recognition and prediction system for breast cancer molecular subtypes based on ultrasound images, comprising a data acquisition module, an image preprocessing module, an image segmentation module, a feature extraction module, a recognition and prediction module and a user interface module; the scheme adopts an improved diffusion probability model for image segmentation, the model uses dynamic conditional coding to enhance the step-by-step regional attention of the diffusion probability model in medical image segmentation, and adopts a feature frequency resolver to eliminate the negative impact of high-frequency noise components in medical image segmentation; a visual transformer integrated with dendritic learning is used to recognize and predict breast cancer molecular subtypes, the architecture of the existing feature extraction network is optimized, the image features extracted are richer and more representative, the recognition accuracy of breast cancer molecular subtypes in ultrasound images is improved, the training time of the model is reduced and the interpretability is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of breast cancer ultrasound image recognition and prediction, and in particular to a breast cancer molecular subtype recognition and prediction system based on ultrasound images. Background Art

[0002] Breast cancer is a disease with a high incidence rate and serious harm to women's physical and mental health. At present, the treatment of breast cancer at home and abroad is mainly based on molecular classification and comprehensive treatment. The treatment plans for patients with different molecular subtypes may be completely different. In clinical practice, the detection of molecular subtypes of breast cancer patients is mainly through obtaining histological specimens of lesions, and then performing immunohistochemistry or even molecular in situ hybridization on the specimens. These methods are invasive, time-consuming, expensive, and have poor repeatability. Therefore, there is an urgent need for a non-invasive, rapid, economical, and accurate method for detecting molecular subtypes of breast cancer in clinical practice. In addition, in my country, breast ultrasound is often recommended as the preferred or even the first choice for screening breast cancer patients because of its advantages of no radiation, economy, convenience, and good imaging effect on dense breast lesions. However, there are few reports on related research using breast ultrasound images. In addition, existing segmentation algorithms such as U-Net and its variants perform well in many cases, but their accuracy will decrease when facing lesions with low contrast, blurred boundaries, or high heterogeneity; existing deep learning models require a large amount of data, long training time, and limited interpretability during training, and have low accuracy in identifying molecular subtypes of breast cancer. Therefore, in view of the above shortcomings, it is crucial to invent a breast cancer molecular subtype identification and prediction system based on breast ultrasound images with good performance for the accurate diagnosis and treatment of breast cancer. Summary of the invention

[0003] In view of the above-mentioned deficiencies, the present invention provides a recognition and prediction system for molecular subtypes of breast cancer based on ultrasound images. In view of the problem that the accuracy of existing segmentation algorithms decreases when facing lesion areas with low contrast, blurred boundaries or high heterogeneity, an improved diffusion probability model is used for image segmentation. The model uses dynamic conditional coding to enhance the step-by-step regional attention of the diffusion probability model in medical image segmentation, and uses a feature frequency parser to eliminate the negative impact of high-frequency noise components in medical image segmentation. In view of the problem that existing deep learning models require a large amount of data, a long training time and limited interpretability during the training process, and have a low accuracy in the recognition of molecular subtypes of breast cancer, a visual transformer integrated with dendritic learning is used to recognize and predict molecular subtypes of breast cancer, and the architecture of the existing feature extraction network is optimized to make the extracted image features richer and more representative. The dendritic learning model and the visual transformer are used to establish an effective and biologically interpretable structure, thereby improving the recognition accuracy of molecular subtypes of breast cancer in ultrasound images, reducing the training time of the model and enhancing the interpretability.

[0004] The ultrasonic image-based breast cancer molecular subtype recognition and prediction system provided by the present invention comprises a data acquisition module, an image preprocessing module, an image segmentation module, a feature extraction module, a recognition and prediction module and a user interface module;

[0005] The data acquisition module is responsible for collecting ultrasound images containing breast cancer molecular subtypes and related clinical data;

[0006] The image preprocessing module preprocesses the collected ultrasound images containing breast cancer molecular subtypes to obtain processed ultrasound images;

[0007] The image segmentation module uses an improved diffusion probability model to separate the image of the region of interest from the processed ultrasound image to obtain a segmentation feature map;

[0008] The feature extraction module uses a visual transformer integrated with dendritic learning for image recognition and extracts features of breast cancer molecular subtypes from the segmentation feature map for classification;

[0009] The recognition prediction module receives a new ultrasound image as input, uses the image segmentation module and the feature extraction module to perform classification, and outputs the probability distribution and direct classification result of the breast cancer molecular subtype represented by the new ultrasound image;

[0010] The user interface module provides a user-friendly interactive interface, allowing doctors and researchers to upload images, view the probability distribution and direct classification results of breast cancer molecular subtypes, and adjust the parameter settings of the recognition prediction module.

[0011] Furthermore, the image segmentation module uses an improved diffusion probability model to separate the image of the region of interest from the processed ultrasound image, specifically comprising the following steps:

[0012] Step S1: original image segmentation, taking the processed ultrasound image as the original image, and using the segmentation label to segment the original image to obtain an original segmentation map;

[0013] Step S2: Model architecture, using the ResUNet model as the basic framework for training. ResUNet includes ResNet encoder and UNet decoder;

[0014] Step S3: In the forward process of the diffusion probability model, Gaussian noise is used to gradually destroy the original segmentation map until a segmentation map completely contaminated by noise is obtained after T steps. ;

[0015] Step S4: Backward diffusion, from the segmentation map First, we use the ResUNet model in the diffusion probability model to learn how to reverse the noise addition process through T steps and restore the original segmentation map;

[0016] Step S5: Conditioning step, using dynamic conditional encoding strategy for feature integration in the ResUNet model;

[0017] Step S6: Eigenfrequency resolver, using the eigenfrequency resolver to constrain the segmentation map The high-frequency noise in the features is obtained to obtain the feature map after frequency domain processing;

[0018] Step S7: Output the segmentation results, each input original image and segmentation map The encoding results are processed by independent ResNet encoders respectively, and the encoding results are added together for the UNet decoder to predict the final segmentation result and obtain the segmentation feature map;

[0019] Furthermore, in step S5, the dynamic conditional encoding strategy specifically includes the following steps:

[0020] Step S51: Feature embedding generation, using a ResNet encoder network to extract the original image at time step feature embedding, get the feature map of the original image, for the segmentation map of the current step , use another ResNet encoder network to extract the segmentation map The feature embedding of The feature map of

[0021] Step S52: Conditioning step estimation, using the prior information of the original image to condition the step estimation function, the formula used is as follows:

[0022] ;

[0023] in, is the step estimation function, is the current time step The segmentation diagram, is the original image, is the current time step index, express At time step The feature embedding of Indicates that at the current time step Next, segmentation diagram The feature embedding of represents the UNet decoder network;

[0024] Step S53: Calculate the affinity graph and convert the segmentation graph of the current step The original image is layer normalized and integrated using the fusion mechanism to obtain the affinity map. The formula used is as follows:

[0025] ;

[0026] in, Indicates the corresponding In the The feature map of the layer, Indicates the corresponding In the The feature map of the layer, is the layer normalization operation, represents element-wise multiplication, For affinity diagram;

[0027] Further, in step S6, the characteristic frequency analyzer specifically includes the following steps:

[0028] Step S61: Obtain a UNet decoder feature map, and use the UNet decoder to output the feature map;

[0029] Step S62: Perform a two-dimensional fast Fourier transform on the feature map along the spatial dimension to obtain a complex frequency spectrum. The formula used is as follows:

[0030] M = F [m]∈ C H * W * C ;

[0031] in, F [ ] represents the two-dimensional fast Fourier transform, represents the feature map, is the height, is the width, Indicates the number of channels, represents the complex domain, is the spectrum in complex form;

[0032] Step S63: parameterize the attention map, and use the parameterized attention map to adjust the complex spectrum to obtain an adjusted spectrum. The formula used is as follows:

[0033] ;

[0034] in, is the parameterized attention map, is the adjusted spectrum;

[0035] Step S64: Reverse back to the spatial domain, use inverse fast Fourier transform to convert the adjusted spectrum back to the spatial domain, and obtain the feature map after frequency domain processing. The formula used is as follows:

[0036] m ' = F -1 [ M ' ] ;

[0037] In the formula, F -1 [ ] represents the inverse fast Fourier transform, It is the feature map after frequency domain processing.

[0038] Furthermore, the feature extraction module uses a visual transformer integrated with dendrite learning to perform image recognition, specifically comprising the following steps:

[0039] Step P1: Linear projection and position encoding, use linear projection to convert the segmentation feature map to the feature space, and add position encoding to obtain the encoded feature vector;

[0040] Step P2: Feature extraction, use stacked Transformer blocks to extract features from the encoded feature vector to obtain global features;

[0041] Step P3: Local self-attention mechanism, using the self-attention mechanism in each Transformer block to enhance local attention capabilities. The formula used is as follows:

[0042] ;

[0043] ;

[0044] in, is the self-masking matrix, is a learnable parameter, , and represents the different eigenvectors after linear projection, is the activation function, is the transpose operation, is element-wise multiplication, is the local eigenvector, represents a matrix of all 1s, represents the identity matrix, is the sequence length, Indicates infinity;

[0045] Step P4: feature normalization, performing a normalization operation on the input segmentation feature map to obtain a normalized feature;

[0046] Step P5: Synaptic connection, using learnable weights and biases to map the normalized features to M dendritic branches, and apply an activation function to each dendritic branch to calculate the output probability of each category. The formula used is as follows:

[0047] ;

[0048] in, is the feature normalization operation, is the input segmentation feature map, Represents the category index, is the dendrite branch index, is the feature dimension index, represents the number of dendritic branches, is the number of feature dimensions, represents the learnable weights, is the bias value, Predict the probability for the classification.

[0049] Step P6: Final output, integrating the category probabilities output by all dendritic branches to obtain the final classification result.

[0050] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0051] (1) To address the problem that the accuracy of existing segmentation algorithms decreases when faced with lesions with low contrast, blurred boundaries, or high heterogeneity, an improved diffusion probability model is used for image segmentation. The model uses dynamic conditional coding to enhance the step-by-step regional attention of the diffusion probability model in medical image segmentation, and uses a characteristic frequency resolver to eliminate the negative impact of high-frequency noise components in medical image segmentation, thereby improving the effectiveness of existing models in the segmentation of breast cancer molecular subtype ultrasound images, enhancing image deblurring capabilities, and improving the accuracy of segmentation of lesions with high heterogeneity.

[0052] (2) In order to address the problems that existing deep learning models require a large amount of data, a long training time and limited interpretability during the training process, and have low accuracy in identifying molecular subtypes of breast cancer, we use a dendritic learning-integrated visual transformer to identify and predict molecular subtypes of breast cancer, optimize the architecture of the existing feature extraction network to make the extracted image features richer and more representative, and use the dendritic learning model and visual transformer to establish an effective and biologically interpretable structure, thereby improving the accuracy of identifying molecular subtypes of breast cancer in ultrasound images, reducing the training time of the model and enhancing interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of a system for identifying and predicting breast cancer molecular subtypes based on ultrasound images provided by the present invention;

[0054] Figure 2 It is a flowchart of the improved diffusion probability model;

[0055] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0057] Example 1, see Figure 1 Breast ultrasound is often recommended as the preferred or even the first choice for breast cancer screening because of its advantages of no radiation, economical, convenient, and good imaging effect on dense breast lesions. However, there are few results of related research using ultrasound images of breast cancer patients, and there are no good methods and models available to identify and predict molecular subtypes of breast cancer.

[0058] In order to overcome the above problems, the present invention provides a breast cancer molecular subtype recognition and prediction system based on ultrasound images, which includes a data acquisition module, an image preprocessing module, an image segmentation module, a feature extraction module, a recognition and prediction module and a user interface module;

[0059] The data acquisition module is responsible for collecting ultrasound images containing breast cancer molecular subtypes and related clinical data;

[0060] The image preprocessing module preprocesses the collected ultrasound images containing breast cancer molecular subtypes to obtain processed ultrasound images;

[0061] The image segmentation module uses an improved diffusion probability model to separate the image of the region of interest from the processed ultrasound image to obtain a segmentation feature map;

[0062] The feature extraction module uses a visual transformer integrated with dendritic learning for image recognition and extracts features of breast cancer molecular subtypes from the segmentation feature map for classification;

[0063] The recognition prediction module receives a new ultrasound image as input, uses the image segmentation module and the feature extraction module to perform classification, and outputs the probability distribution and direct classification result of the breast cancer molecular subtype represented by the new ultrasound image;

[0064] The user interface module provides a user-friendly interactive interface, allowing doctors and researchers to upload images, view the probability distribution and direct classification results of breast cancer molecular subtypes, and adjust the parameter settings of the recognition prediction module.

[0065] Embodiment 2: This embodiment is based on the above embodiment. The data acquisition module collects relevant clinical data including the ultrasound image itself, the patient's basic information and the pathology report. The patient's basic information includes personal identity information, medical history, current health status, physiological indicators, treatment history and allergy history. The pathology report includes pathology reports related to tissue biopsy and cytology examination, which contains key information such as the specific type and grade of the tumor.

[0066] Embodiment 3: This embodiment is based on the above embodiment. The preprocessing operation in the image preprocessing module includes denoising the image and using median filtering to reduce speckle noise; applying contrast adjustment to enhance image details so that tumors and cancer lesions are more clearly identifiable; for color Doppler images, it is necessary to separate blood flow signals from tissue background; and performing standardization to improve the consistency between images from different sources by unifying parameters such as image size and grayscale range, thereby ensuring data consistency when training the model.

[0067] Example 4, see Figure 1 and Figure 2 This embodiment is based on the above embodiment, and the image segmentation module uses an improved diffusion probability model to separate the image of the region of interest from the processed ultrasound image, and specifically includes the following steps:

[0068] Step S1: original image segmentation, taking the processed ultrasound image as the original image, and using the segmentation label to segment the original image to obtain an original segmentation map;

[0069] Step S2: Model architecture, using the ResUNet model as the basic framework for training. ResUNet includes a ResNet encoder and a UNet decoder. The residual block of the ResNet encoder contains convolution, group normalization, and SiLU activation.

[0070] Step S3: In the forward process of the diffusion probability model, Gaussian noise is used to gradually destroy the original segmentation map until a segmentation map completely contaminated by noise is obtained after T steps. ;

[0071] Step S4: Backward diffusion, from the segmentation map First, we use the ResUNet model in the diffusion probability model to learn how to reverse the noise addition process through T steps and restore the original segmentation map;

[0072] Step S5: Conditioning step, using dynamic conditional encoding strategy for feature integration in the ResUNet model;

[0073] Step S6: Eigenfrequency resolver, using the eigenfrequency resolver to constrain the segmentation map The high-frequency noise in the features is obtained to obtain the feature map after frequency domain processing;

[0074] Step S7: Output the segmentation results, each input original image and segmentation map The encoding results are processed by independent ResNet encoders respectively, and the encoding results are added together for the UNet decoder to predict the final segmentation result and obtain the segmentation feature map.

[0075] Embodiment 5. This embodiment is based on the above embodiment. In step S1, the segmentation label is an initial segmentation label obtained by manual annotation by professional doctors and radiologists. According to the expression of four molecular markers, namely estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2) and Ki-67, the segmentation label divides breast cancer into four types: luminal A type, luminal B type, HER2 positive type and triple negative type. By marking and segmenting different types of breast cancer, the improved diffusion probability model is more accurate in the segmentation and recognition prediction in new breast cancer ultrasound images.

[0076] Embodiment 6, based on the above embodiment, in step S5, the dynamic conditional encoding strategy specifically includes the following steps:

[0077] Step S51: Feature embedding generation, using a ResNet encoder network to extract the original image at time step feature embedding, get the feature map of the original image, for the segmentation map of the current step , use another ResNet encoder network to extract the segmentation map The feature embedding of The feature map of

[0078] Step S52: Conditioning step estimation, using the prior information of the original image to condition the step estimation function, the formula used is as follows:

[0079] ;

[0080] in, is the step estimation function, is the current time step The segmentation diagram, is the original image, is the current time step index, express At time step The feature embedding of Indicates that at the current time step Next, segmentation diagram The feature embedding of represents the UNet decoder network;

[0081] Step S53: Calculate the affinity graph and convert the segmentation graph of the current step The original image is layer normalized and integrated using the fusion mechanism to obtain the affinity map. The formula used is as follows:

[0082] ;

[0083] in, Indicates the corresponding In the The feature map of the layer, Indicates the corresponding In the The feature map of the layer, is the layer normalization operation, represents element-wise multiplication, For affinity diagram.

[0084] Embodiment 7, based on the above embodiment, in step S6, the characteristic frequency analyzer specifically includes the following steps:

[0085] Step S61: Obtain a UNet decoder feature map, and use the UNet decoder to output the feature map;

[0086] Step S62: Perform a two-dimensional fast Fourier transform on the feature map along the spatial dimension to obtain a complex frequency spectrum. The formula used is as follows:

[0087] M = F [m]∈ C H * W * C ;

[0088] in, F [ ] represents the two-dimensional fast Fourier transform, represents the feature map, is the height, is the width, Indicates the number of channels, represents the complex domain, is the spectrum in complex form;

[0089] Step S63: parameterize the attention map, and use the parameterized attention map to adjust the complex spectrum to obtain an adjusted spectrum. The formula used is as follows:

[0090] ;

[0091] in, is the parameterized attention map, is the adjusted spectrum;

[0092] Step S64: Reverse back to the spatial domain, use inverse fast Fourier transform to convert the adjusted spectrum back to the spatial domain, and obtain the feature map after frequency domain processing. The formula used is as follows:

[0093] m ' = F -1 [ M ' ] ;

[0094] In the formula, F -1 [ ] represents the inverse fast Fourier transform, It is the feature map after frequency domain processing.

[0095] By performing the operations, an improved diffusion probability model is used for image segmentation to address the problem that the accuracy of existing segmentation algorithms decreases when faced with lesion areas with low contrast, blurred boundaries or high heterogeneity. The model uses dynamic conditional coding to enhance the step-by-step regional attention of the diffusion probability model in medical image segmentation, and uses a characteristic frequency resolver to eliminate the negative impact of high-frequency noise components in medical image segmentation, thereby improving the effectiveness of existing models in segmenting breast cancer molecular subtype ultrasound images, enhancing image deblurring capabilities, and improving the accuracy of segmentation of lesions with high heterogeneity.

[0096] Embodiment 8, based on the above embodiment, the feature extraction module uses a visual transformer integrated with dendrite learning to perform image recognition, specifically comprising the following steps:

[0097] Step P1: Linear projection and position encoding, use linear projection to convert the segmentation feature map to the feature space, and add position encoding to obtain the encoded feature vector;

[0098] Step P2: Feature extraction, use stacked Transformer blocks to extract features from the encoded feature vector to obtain global features;

[0099] Step P3: Local self-attention mechanism, using the self-attention mechanism in each Transformer block to enhance local attention capabilities. The formula used is as follows:

[0100] ;

[0101] ;

[0102] in, is the self-masking matrix, is a learnable parameter, , and represents the different eigenvectors after linear projection, is the activation function, is the transpose operation, is element-wise multiplication, is the local eigenvector, represents a matrix of all 1s, represents the identity matrix, is the sequence length, Indicates infinity;

[0103] Step P4: feature normalization, performing a normalization operation on the input segmentation feature map to obtain a normalized feature;

[0104] Step P5: Synaptic connection, using learnable weights and biases to map the normalized features to M dendritic branches, and apply an activation function to each dendritic branch to calculate the output probability of each category. The formula used is as follows:

[0105] ;

[0106] in, is the feature normalization operation, is the input segmentation feature map, Represents the category index, is the dendrite branch index, is the feature dimension index, represents the number of dendritic branches, is the number of feature dimensions, represents the learnable weights, is the bias value, Predict the probability for the classification.

[0107] Step P6: Final output, integrating the category probabilities output by all dendritic branches to obtain the final classification result.

[0108] By performing the operations, in order to address the problems that existing deep learning models require a large amount of data, a long training time and limited interpretability during the training process, and have a low accuracy in identifying molecular subtypes of breast cancer, a visual transformer integrated with dendritic learning is used to identify and predict molecular subtypes of breast cancer, and the architecture of the existing feature extraction network is optimized to make the extracted image features richer and more representative. The dendritic learning model and the visual transformer are used to establish an effective and biologically interpretable structure, thereby improving the recognition accuracy of molecular subtypes of breast cancer in ultrasound images, reducing the training time of the model and enhancing the interpretability.

[0109] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0110] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0111] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A system for identifying and predicting molecular subtypes of breast cancer based on ultrasound images, characterized by: It includes a data acquisition module, an image preprocessing module, an image segmentation module, a feature extraction module, a recognition prediction module and a user interface module; The data acquisition module is responsible for collecting ultrasound images containing breast cancer molecular subtypes and related clinical data; The image preprocessing module preprocesses the collected ultrasound images containing breast cancer molecular subtypes to obtain processed ultrasound images; The image segmentation module uses an improved diffusion probability model to separate the image of the region of interest from the processed ultrasound image to obtain a segmentation feature map; The feature extraction module uses a visual transformer integrated with dendritic learning for image recognition and extracts features of breast cancer molecular subtypes from the segmentation feature map for classification; The recognition prediction module receives a new ultrasound image as input, uses the image segmentation module and the feature extraction module to perform classification, and outputs the probability distribution and direct classification result of the breast cancer molecular subtype represented by the new ultrasound image; The user interface module provides a user-friendly interactive interface, allowing doctors and researchers to upload images, view the probability distribution and direct classification results of breast cancer molecular subtypes, and adjust the parameter settings of the recognition prediction module; The image segmentation module uses an improved diffusion probability model to separate the image of the region of interest from the processed ultrasound image, specifically comprising the following steps: Step S1: original image segmentation, taking the processed ultrasound image as the original image, and using the segmentation label to segment the original image to obtain the original segmentation map; Step S2: Model architecture, using the ResUNet model as the basic framework for training. The ResUNet model includes a ResNet encoder and a UNet decoder. Step S3: In the forward process of the diffusion probability model, Gaussian noise is used to gradually destroy the original segmentation map until a segmentation map completely contaminated by noise is obtained after T steps. ; Step S4: Backward diffusion, from the segmentation map First, we use the ResUNet model in the diffusion probability model to learn how to reverse the noise addition process through T steps and restore the original segmentation map; Step S5: Conditioning step, using dynamic conditional encoding strategy for feature integration in the ResUNet model; Step S6: Eigenfrequency resolver, using the eigenfrequency resolver to constrain the segmentation map The high-frequency noise in the features is obtained to obtain the feature map after frequency domain processing; Step S7: Output the segmentation results, each input original image and segmentation map The encoding results are processed by independent ResNet encoders respectively, and the encoding results are added together for the UNet decoder to predict the final segmentation result and obtain the segmentation feature map; In step S5, the dynamic conditional encoding strategy specifically includes the following steps: Step S51: Feature embedding generation, using a ResNet encoder network to extract the original image at time step feature embedding, get the feature map of the original image, for the segmentation map of the current step , use another ResNet encoder network to extract the segmentation map The feature embedding of The feature map of Step S52: Conditioning step estimation, using the prior information of the original image to condition the step estimation function, the formula used is as follows: ； in, is the step estimation function, is the current time step The segmentation diagram, is the original image, is the current time step index, express At time step The feature embedding of Indicates that at the current time step Next, segmentation diagram The feature embedding of represents the UNet decoder network; Step S53: Calculate the affinity graph and convert the segmentation graph of the current step Perform layer normalization with the original image and integrate it using a fusion mechanism to obtain an affinity map; In step S6, the characteristic frequency analyzer specifically includes the following steps: Step S61: Obtain a UNet decoder feature map, and use the UNet decoder to output the feature map; Step S62: Perform a two-dimensional fast Fourier transform on the feature map along the spatial dimension to obtain a complex frequency spectrum. The formula used is as follows: ； in, represents the two-dimensional fast Fourier transform, represents the feature map, is the height, is the width, Indicates the number of channels, represents the complex domain, is the spectrum in complex form; Step S63: parameterize the attention map, and use the parameterized attention map to adjust the complex spectrum to obtain an adjusted spectrum. The formula used is as follows: ； in, is the parameterized attention map, is the adjusted spectrum; Step S64: Reverse back to the spatial domain, use the inverse fast Fourier transform to convert the adjusted spectrum back to the spatial domain, and obtain the feature map after frequency domain processing.

2. The system for identifying and predicting breast cancer molecular subtypes based on ultrasound images according to claim 1, characterized in that: The feature extraction module uses a visual transformer integrated with dendrite learning to perform image recognition, specifically including the following steps: Step P1: Linear projection and position encoding, use linear projection to convert the segmentation feature map to the feature space, and add position encoding to obtain the encoded feature vector; Step P2: Feature extraction, use stacked Transformer blocks to extract features from the encoded feature vector to obtain global features; Step P3: Local self-attention mechanism, using the self-attention mechanism in each Transformer block to enhance local attention capabilities. The formula used is as follows: ； ； in, is the self-masking matrix, is a learnable parameter, , and represents the different eigenvectors after linear projection, is the activation function, is the transpose operation, is element-wise multiplication, is the local eigenvector, represents a matrix of all 1s, represents the identity matrix, is the sequence length, Indicates infinity; Step P4: feature normalization, performing a normalization operation on the input segmentation feature map to obtain a normalized feature; Step P5: Synaptic connection, using learnable weights and biases to map the normalized features onto M dendritic branches, and applying an activation function to each dendritic branch to calculate the output probability of each category; Step P6: Final output, integrating the category probabilities output by all dendritic branches to obtain the final classification result.

Images