Ultrasonic image classification method for breast 4A type nodules

Abstract

The invention provides an ultrasonic image classification method for breast 4A type nodules, and relates to the field of medical image processing. The method comprises the following steps: sending a target breast ultrasound image into a U-Mamba network to segment a breast tumor to obtain a breast focus segmentation result; classifying the edge, echo, form, aspect ratio, calcification, rear echo attenuation and catheter relation of each breast focus according to the breast focus segmentation result, and constructing a corresponding single-factor matrix according to the classification result; performing interaction processing according to the single-factor matrix, and constructing a multi-factor graph interaction matrix; and sending a one-dimensional array obtained by carrying out position coding and cascading processing by utilizing the single-factor matrix and the multi-factor graph interaction matrix into a Graph-Mama classification network so as to calculate and obtain a classification result of breast 4A type nodules corresponding to each segmentation result. According to the technical scheme, the classification accuracy and efficiency of the breast 4A type nodule ultrasonic image can be improved.

Description

Technical Field

[0001] The present application relates to the field of medical image processing, and in particular, to an ultrasound image classification method for breast 4A nodules. Background Art

[0002] In the field of medical imaging diagnosis, especially in breast health examinations, accurately determining whether a nodule is benign or malignant is crucial for developing subsequent treatment plans. As a widely used examination method, B-mode ultrasound provides clinicians with an important window for observing breast structure and detecting nodules. However, for BI-RADS (Breast Imaging Reporting and Data System) 4a nodules, that is, those with a certain risk of malignancy, clinicians often find it difficult to make a clear benign or malignant judgment based solely on ultrasound images.

[0003] At present, the research on the classification of benign and malignant BI-RADS 4a nodules mainly relies on clinical analysis, such as medical history inquiry, physical examination, laboratory tests, etc. Although these methods are helpful in judging the nature of nodules to a certain extent, they still have problems such as strong subjectivity and limited accuracy. With the rapid development of artificial intelligence technology, the research on using deep learning models to assist in the classification of benign and malignant BI-RADS 4a nodules has gradually emerged. However, this field is still in its infancy and faces many challenges. Existing artificial intelligence-based classification methods often only focus on a single nodule feature and ignore the interaction between different features, which limits the accuracy and generalization ability of the classification model.

[0004] Therefore, how to improve the classification accuracy and efficiency of breast 4A nodule ultrasound images is a technical problem that needs to be solved urgently. Summary of the invention

[0005] The purpose of this application is to provide an ultrasound image classification method for breast 4A nodules, which can improve the classification accuracy and efficiency of ultrasound images of breast 4A nodules.

[0006] This application is implemented as follows:

[0007] In a first aspect, the present application provides a method for classifying ultrasound images of breast category 4A nodules, including the following steps: sending a target breast ultrasound image into a U-Mamba network to segment breast tumors, obtaining a breast lesion segmentation result; classifying the edges, echoes, shapes, aspect ratios, calcifications, posterior echo attenuation, and duct relationships of each breast lesion according to the breast lesion segmentation result, and constructing a corresponding single-factor matrix according to the classification results; performing interactive processing according to the single-factor matrix to construct a multi-factor graph interaction matrix; sending a one-dimensional array obtained by performing position encoding and concatenation processing on the single-factor matrix and the multi-factor graph interaction matrix into a Graph-Mamba classification network to calculate the classification result of breast category 4A nodules corresponding to each segmentation result.

[0008] In some implementation manners, before the step of sending the target breast ultrasound image into the U-Mamba network to segment breast tumors, it further includes: calculating the centroid point of the area where the breast lesion is located; calculating the horizontal and vertical distances from each point in the area where the breast lesion is located to the centroid point, and the distances between these distances and the maximum and minimum values of their respective coordinate axes; cropping and resizing the target breast ultrasound image according to the calculated distance values.

[0009] In some implementation manners, the classification of the edges, echoes, shapes, aspect ratios, calcifications, posterior echo attenuation, and duct relationships of each breast lesion includes: edge classification, including classifying according to whether the edge of the breast lesion is smooth and the shape of the edge, where the shape of the edge includes three types: spiculation, angulated edge, and microlobulation; echo classification, including classifying according to the echo type within the breast lesion, where the echo type includes isoecho, hypoecho, cystic mixed echo, heterogeneous echo, and hyperecho; shape classification, including classifying according to whether the shape of the breast lesion belongs to ellipse, circle, and non-ellipse and non-circle; aspect ratio classification, including classifying according to the aspect ratio of the breast lesion, including being divided into parallel position, non-parallel position, or circular; calcification classification, including classifying according to the diameter of the breast lesion, including being divided into microcalcification and coarse calcification; posterior echo attenuation classification, including classifying according to whether there is echo attenuation behind the breast lesion; duct relationship classification, including classifying according to whether the ducts around the breast lesion are dilated and whether there are visible solid / cystic-solid masses in the ducts.

[0010] In some implementation manners, the constructed single-factor matrix includes four 2-classification matrices for the edges of breast lesions, a 5-classification matrix for echoes, a 3-classification matrix for shapes, a 3-classification matrix for aspect ratios, a 2-classification matrix for calcifications, a 2-classification matrix for posterior echo attenuation, and a 2-classification matrix for duct relationships.

[0011] In some implementations, the U-Mamba network includes: using the U-Mamba encoder as the backbone network to perform multi-classification task processing on edge blurring, edge shape, echo, morphology, and posterior echo attenuation; at the end of the U-Mamba decoder, two 3×3 fully convolutional layers with 3 output channels are used in parallel to output two segmentation tasks, one of which is used to output the segmentation results of background, breast lesions, and ducts to obtain the first segmentation result, and the other is used to output the segmentation results of background, calcifications, and solid / cystic masses in the ducts to obtain the second segmentation result; according to the first segmentation result and the second segmentation result, calculate the aspect ratio of breast lesions, calcification diameter, duct width, and information on whether there are visible masses in the ducts.

[0012] In some implementations, the interactive processing according to the single-factor matrix to construct the multi-factor graph interaction matrix includes: converting the single-factor matrix into a two-dimensional matrix, and then performing multi-factor graph interaction calculations in a specified order to obtain the multi-factor graph interaction matrix of different factor combinations.

[0013] In some implementations, the performing multi-factor graph interaction calculations in a specified order to obtain the multi-factor graph interaction matrix of different factor combinations includes: generating two-factor, three-factor, four-factor, five-factor, six-factor, and seven-factor multi-factor graph interaction matrices in the order of edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship through matrix dot product and flatten operations.

[0014] In some implementations, the Graph-Mamba classification network includes at least one encoding module and at least one fully connected layer, and the encoding module includes a convolutional layer for capturing local information of short sequences and a Mamba module for capturing global information of long sequences.

[0015] In some implementations, the Graph-Mamba classification network includes 5 encoding modules and 3 fully connected layers. Each encoding module includes a 7×1 convolutional layer and a Mamba module. Among them, the output channels of each 7×1 convolutional layer are half of the input channels, and the feature channels of the 3 fully connected layers are 2048, 512, and 2 respectively.

[0016] In some implementations, it also includes calculating the Softmax cross-entropy loss for the two-dimensional features output by the Graph-Mamba classification network to optimize the performance of the classification network.

[0017] In a second aspect, the present application provides an ultrasonic image classification system for breast nodules of category 4A, which includes: an image segmentation module for sending a target breast ultrasonic image into a U-Mamba network to segment breast tumors and obtain a breast lesion segmentation result; a first construction module for classifying the edges, echoes, shapes, aspect ratios, calcifications, posterior echo attenuation, and duct relationships of each breast lesion according to the breast lesion segmentation result, and constructing a corresponding single-factor matrix according to the classification results; a second construction module for performing interaction processing based on the single-factor matrix to construct a multi-factor graph interaction matrix; and an image classification module for sending a one-dimensional array obtained by performing position encoding and concatenation processing on the single-factor matrix and the multi-factor graph interaction matrix into a Graph-Mamba classification network to calculate the classification result of the breast nodules of category 4A corresponding to each segmentation result.

[0018] In a third aspect, the present application provides an electronic device, which includes a memory for storing one or more programs; a processor; when the above one or more programs are executed by the above processor, the method described in any one of the first aspects above is implemented.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method described in any one of the first aspects above is implemented.

[0020] Compared with the prior art, the present application has at least the following advantages or beneficial effects:

[0021] The present application proposes an ultrasonic image classification method for breast nodules of category 4A. First, it accurately segments breast ultrasonic images through a U-Mamba network, which can accurately identify and extract breast lesion areas. This step is the basis for subsequent feature extraction and classification, ensuring the pertinence and accuracy of subsequent processing. Then, after obtaining the breast lesion segmentation result, it further classifies key features such as the edges, echoes, shapes, aspect ratios, calcifications, posterior echo attenuation, and duct relationships of each lesion, and constructs a single-factor matrix. These features cover multiple important aspects in the diagnosis of breast nodules, providing rich information for comprehensively evaluating the nature of nodules. Then, considering the shortcoming of ignoring the interaction between features in the prior art, the present application constructs a multi-factor graph interaction matrix to effectively capture the complex relationships between different features. This step enhances the model's ability to identify the nature of nodules and improves the classification accuracy. Finally, the single-factor matrix and the multi-factor graph interaction matrix are subjected to position encoding and concatenation processing, and then sent into a Graph-Mamba classification network for classification. The Graph-Mamba classification network can make full use of the characteristics of graph-structured data to further explore the deep relationships between features, thereby achieving more accurate classification. Description of the Drawings

[0022] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the accompanying drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0023] Figure 1 It is a flowchart of an embodiment of a method for classifying ultrasonic images of breast category 4A nodules in the present application;

[0024] Figure 2 It is a schematic framework diagram of an embodiment of a method for classifying ultrasonic images of breast category 4A nodules in the present application;

[0025] Figure 3 It is a structural block diagram of an embodiment of a system for classifying ultrasonic images of breast category 4A nodules in the present application;

[0026] Figure 4 It is a structural block diagram of an electronic device provided by an embodiment of the present application.

[0027] Icons: 201, processor; 202, memory; 203, communication interface. Detailed implementation manners

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Usually, the components of the embodiments of the present application described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations.

[0029] The following will describe in detail some implementation manners of the present application in conjunction with the accompanying drawings. Without conflict, the following various embodiments and the various features in the embodiments can be combined with each other.

[0030] Embodiment 1

[0031] The embodiment of the present application provides a method for classifying ultrasonic images of breast category 4A nodules, which can improve the classification accuracy and efficiency of ultrasonic images of breast category 4A nodules.

[0032] Please refer to Figure 1-2 , the method for classifying ultrasonic images of breast category 4A nodules includes the following steps:

[0033] Step S101: Send the target breast ultrasonic image into the U-Mamba network to segment the breast tumor and obtain the segmentation result of the breast lesion;

[0034] It should be noted that the U-Mamba network is a deep learning model suitable for image segmentation tasks. By analyzing the input target breast ultrasound image, it can accurately identify and segment the breast tumor area, and then output the breast lesion segmentation result, that is, the probability map in which each pixel point in the target breast ultrasound image is divided into tumor or non-tumor. That is, through the U-Mamba network, a high-quality breast lesion segmentation map can be obtained, providing reliable input data for subsequent steps.

[0035] Step S102: Classify the edges, echoes, morphologies, aspect ratios, calcifications, posterior echo attenuations, and ductal relationships of each breast lesion according to the breast lesion segmentation result, and construct a corresponding single-factor matrix based on the classification result;

[0036] In the above step, after obtaining the breast lesion segmentation result, further classify the edges, echoes, morphologies, aspect ratios, calcifications, posterior echo attenuations, and ductal relationships of each breast lesion. The classification result is represented in the form of a single-factor matrix, where each row represents a lesion and each column represents a feature category. The construction of this single-factor matrix enables the intuitive visualization of the performance of each lesion on different features, providing basic data for subsequent feature interaction processing. At the same time, this step also realizes the quantitative description of lesion features, improving the accuracy of classification.

[0037] Step S103: Perform interaction processing based on the single-factor matrix to construct a multi-factor graph interaction matrix;

[0038] In the above step, based on the single-factor matrix, further consider the interaction between different features, and use a method similar to a graph neural network to construct a multi-factor graph interaction matrix, where nodes represent features and edges represent the interaction relationships between features. The construction of the multi-factor graph interaction matrix enables the capture of complex relationships between different features, which are difficult to discover in single-feature analysis. By considering the interaction between features, the nature of breast nodules can be evaluated more comprehensively, thereby improving the accuracy of classification.

[0039] Step S104: Feed the one-dimensional array obtained by performing position encoding and concatenation processing on the single-factor matrix and the multi-factor graph interaction matrix into the Graph-Mamba classification network to calculate the classification result of breast category 4A nodules corresponding to each segmentation result.

[0040] It should be noted that after obtaining the single-factor matrix and the multi-factor graph interaction matrix, they are first position-encoded to ensure the integrity and sequentiality of the information. Then, these two matrices are concatenated into a one-dimensional array as the input of the Graph-Mamba classification network. The Graph-Mamba classification network is a deep learning model specifically designed for processing graph-structured data. It can make full use of the node and edge information in the graph structure to achieve accurate classification. Thus, through the Graph-Mamba classification network, the information in the single-factor matrix and the multi-factor graph interaction matrix can be fully utilized to achieve accurate classification of breast nodules of category 4A. This step not only improves the accuracy of classification but also enhances the generalization ability of the model, enabling it to be applicable to breast nodule classification tasks under different circumstances.

[0041] In summary, the present application first accurately segments breast ultrasound images through the U-Mamba network, which can accurately identify and extract breast lesion regions. This step is the basis for subsequent feature extraction and classification, ensuring the pertinence and accuracy of subsequent processing. Then, after obtaining the breast lesion segmentation results, key features such as the edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship of each lesion are further classified, and a single-factor matrix is constructed. These features cover multiple important aspects in breast nodule diagnosis, providing rich information for comprehensively evaluating the nature of nodules. Next, considering the shortcoming in the prior art of neglecting the interaction between features, the present application effectively captures the complex relationships between different features by constructing a multi-factor graph interaction matrix. This step enhances the model's ability to identify the nature of nodules and improves the accuracy of classification. Finally, the single-factor matrix and the multi-factor graph interaction matrix are position-encoded and concatenated, and then sent to the Graph-Mamba classification network for classification. The Graph-Mamba classification network can make full use of the characteristics of graph-structured data to further explore the deep relationships between features, thereby achieving more accurate classification.

[0042] That is, through accurate segmentation of breast lesions, multi-dimensional feature extraction and classification, and feature interaction processing, the present application can more comprehensively evaluate the nature of breast nodules, thereby significantly improving the accuracy of classification. Moreover, by constructing a multi-factor graph interaction matrix and introducing the Graph-Mamba classification network, it is possible to better adapt to nodule classification tasks under different circumstances and enhance the generalization ability of the model.

[0043] Based on the foregoing solution, in some implementation manners of the present application, before the step of inputting the target breast ultrasound image into the U-Mamba network for segmenting breast tumors, the following steps are further included: calculating the centroid point of the area where the breast lesion is located; calculating the horizontal and vertical distances from each point in the area where the breast lesion is located to the centroid point, and the distances between these distances and the maximum and minimum values of their respective coordinate axes; and cropping and resizing the target breast ultrasound image according to the calculated distance values.

[0044] In the above implementation manner, a series of preprocessing steps are performed before the step of inputting the target breast ultrasound image into the U-Mamba network for segmenting breast tumors, so as to further improve the accuracy of segmentation and classification. The preprocessing steps include: First, calculating the centroid point of the area where the breast lesion is located. The centroid point is the average value of the positions (coordinates) of all pixel points in the breast lesion area, and it represents the center of the breast lesion. By calculating the centroid point, a reference point can be provided for subsequent steps, which is used for cropping the breast lesion area and adjusting the image size. After determining the centroid point, calculate the horizontal and vertical distances from each pixel point in the breast lesion area to the centroid point. These distance values reflect the relative position relationship between each point in the breast lesion area and the center, which helps to understand the shape and distribution of the breast lesion. Then, by calculating the horizontal and vertical distances from each point in the area where the breast lesion is located to the centroid point, and the distances between these distances and the maximum and minimum values of their respective coordinate axes, the size and position of the breast lesion in the image can be understood, providing a basis for subsequent cropping and resizing the image. Finally, after obtaining the centroid point and the extended range of the breast lesion area, the target breast ultrasound image can be cropped to only retain the area containing the lesion. At the same time, in order to maintain the consistency of the image and facilitate subsequent processing, the size of the cropped image also needs to be adjusted to meet the input requirements of the U-Mamba network.

[0045] By preprocessing the image, breast lesion images of different sizes and positions can have similar sizes and layouts before being input into the U-Mamba network. This helps the network better learn the features of breast lesions and enhances its generalization ability on different datasets. At the same time, cropping and resizing the image can also reduce the consumption of computing resources in subsequent processing steps. Since only the breast lesion area needs to be processed instead of the entire image, the computational complexity and processing time can be significantly reduced.

[0046] Based on the foregoing solutions, in some implementation manners of the present application, the classification of the edges, echoes, shapes, aspect ratios, calcifications, posterior echo attenuations, and duct relationships of each breast lesion includes: Edge classification, including classifying according to whether the edge of the breast lesion is smooth and the shape of the edge, where the shape of the edge includes three types: spiculation, angulated margin, and microlobulation; Echo classification, including classifying according to the echo type within the breast lesion, where the echo type includes isoecho, hypoecho, cystic mixed echo, heterogeneous echo, and hyperecho; Shape classification, including classifying according to whether the shape of the breast lesion belongs to ellipse, circle, and non-ellipse and non-circle; Aspect ratio classification, including classifying according to the aspect ratio of the breast lesion, including being divided into parallel position, non-parallel position, or circular; Calcification classification, including classifying according to the diameter of the breast lesion, including being divided into microcalcification and coarse calcification; Posterior echo attenuation classification, including classifying according to whether there is echo attenuation behind the breast lesion; Duct relationship classification, including classifying according to whether the ducts around the breast lesion are dilated and whether there are visible solid / cystic-solid masses in the ducts.

[0047] In the above implementation manners, by carefully classifying the edges, echoes, shapes, aspect ratios, calcifications, posterior echo attenuations, and duct relationships of each breast lesion, the characteristics and properties of the breast lesion can be evaluated more comprehensively, thereby improving the accuracy of classification.

[0048] For ease of understanding, the imaging characteristics of edges, echoes, shapes, aspect ratios, calcifications, posterior echo attenuations, and duct relationships on B-mode ultrasound will be further described below.

[0049] Edges: The edge classification of breast lesions mainly involves two classifications. One is to judge whether the edge is smooth or blurred, and the other is to judge the shape of the edge. If all the edge boundaries are clear, it belongs to smooth edge. When there is no clear and distinct boundary between all or part of the breast lesion and the surrounding tissues, it belongs to the case of blurred edge. The edge shape is further divided into three cases: spiculation, angulated margin, and microlobulation: 1) Spiculation: Sharp needle-like linear structures protruding from the edge of the mass; 2) Angulated margin: Part or all of the edge of the lesion forms a sharp angle (<90°), similar to a crab's foot; 3) Microlobulation: The edge of the mass is not smooth, forming a gear-like / minute undulation (more than three waveforms, and the depression is an acute angle).

[0050] Echogenicity: The echogenicity information within breast lesions includes isoechoic, hypoechoic, cystic mixed echo, heterogeneous echo, and hyperechoic. Isoechoic indicates that the echo within the breast lesion is equal to that of the subcutaneous adipose tissue; hypoechoic or very hypoechoic indicates that the echo within the breast lesion is lower than that of the subcutaneous adipose tissue; cystic-solid mixed echo indicates that the breast lesion contains anechoic (cystic) and echogenic (solid) components; heterogeneous echo indicates that the internal echo of the breast lesion is a single or a mixture of unevenly distributed echoes, which is a solid mass rather than a cystic-solid composite echo mass; hyperechoic indicates that the echo within the breast lesion is higher than that of the subcutaneous adipose tissue or equivalent to the echo of the fibrous tissue of the breast gland.

[0051] Shape: The shape of breast lesions can be divided into two cases: regular and irregular. Regular includes oval and round, and irregular refers to other cases except round and oval. Oval includes two or three undulations, depressions with obtuse angles, that is, gently locally elevated lobulations "shallow lobulations" or large lobulations, and round means that the anteroposterior diameter and transverse diameter of the breast lesion are the same.

[0052] Aspect ratio: The aspect ratio of breast lesions includes three cases: parallel, non-parallel, and round. The aspect ratio of the parallel position < 1, the aspect ratio of the non-parallel position > 1, and the aspect ratio of the round = 1.

[0053] Calcification: Calcification is divided into microcalcification and coarse calcification. Microcalcification is calcification with a diameter < 0.5 mm, and coarse calcification has a diameter > 0.5 mm.

[0054] Posterior echo attenuation: Posterior echo attenuation is an area where the echo decreases behind the breast lesion.

[0055] Duct relationship: It includes two cases: 1) The ducts around the breast lesion are dilated, and the dilated ducts > 2 mm. 2) Mass within the duct: A solid / cystic-solid mass can be seen within the dilated duct.

[0056] Based on the foregoing solution, in some implementation manners of the present application, the constructed single-factor matrix includes four 2-classification matrices for the edge of the breast lesion, a 5-classification matrix for echogenicity, a 3-classification matrix for shape, a 3-classification matrix for aspect ratio, a 2-classification matrix for calcification, a 2-classification matrix for posterior echo attenuation, and a 2-classification matrix for duct relationship. It transforms the complex features of breast lesions into a series of single-factor matrices that are easy to process and analyze. Each matrix classifies a specific feature of the lesion in detail, thereby improving the accuracy and efficiency of feature extraction. Through this classification method, it is easier to identify and analyze the key features of the lesion.

[0057] Based on the aforementioned scheme, in some implementations of the present application, the U-Mamba network includes: using the U-Mamba encoder as the backbone network to perform multi-classification task processing on edge blur, edge shape, echo, morphology, and rear echo attenuation; at the end of the U-Mamba decoder, two 3×3 full convolution layers with 3 output channels are used in parallel to output two segmentation tasks, one of which is used to output the segmentation results of the background, breast lesions, and ducts to obtain a first segmentation result, and the other is used to output the segmentation results of the background, calcification, and solid / cystic-solid tumors in the duct to obtain a second segmentation result; based on the first segmentation result and the second segmentation result, the aspect ratio of the breast lesion, the calcification diameter, the duct width, and information on whether a mass is visible in the duct are calculated.

[0058] It should be noted that the U-Mamba encoder, as the backbone of the network, is responsible for processing the input target breast ultrasound image and extracting key features. These features include edge blur, edge shape, echo, morphology, and rear echo attenuation. Its encoder uses a multi-classification task processing method to carefully classify each feature to improve the accuracy of subsequent analysis. The decoder is located downstream of the encoder and is responsible for converting the features extracted by the encoder into specific segmentation results. At the end of the decoder, two 3×3 full convolutional layers with 3 output channels are used in parallel to perform two different segmentation tasks. The first segmentation task outputs the segmentation results of the background, breast lesions, and ducts, while the second segmentation task outputs the segmentation results of the background, calcification, and solid / cystic solid masses in the ducts. This design allows the network to process multiple segmentation tasks simultaneously, improving the efficiency and accuracy of information extraction. Then, based on the two segmentation results output by the decoder, the aspect ratio of the breast lesion, the diameter of the calcification, the width of the duct, and whether the mass is visible in the duct can be further calculated.

[0059] In summary, through the design of the U-Mamba network in the above implementation method, it is possible to achieve high-precision classification and segmentation of multiple features in breast images through its unique network structure and processing flow.

[0060] Based on the aforementioned scheme, in some implementations of the present application, the interaction processing is performed according to the single-factor matrix to construct a multi-factor graph interaction matrix, including: converting the single-factor matrix into a two-dimensional matrix, and then performing multi-factor graph interaction calculations in a specified order to obtain a multi-factor graph interaction matrix of different factor combinations.

[0061] Through the specific processing method in the above implementation, the single factor matrix can be converted into a multi-factor graph interaction matrix. This process aims to mine the interaction relationship between multiple factors from a single-dimensional data (i.e., a single factor matrix), and then construct a multi-factor graph interaction matrix that can reflect these complex interaction relationships.

[0062] Exemplarily, the specific operation can be as follows: First, preprocess the original single-factor matrix to convert it into the form of a two-dimensional matrix. This step may involve operations such as data sorting and normalization to ensure the accuracy and effectiveness of subsequent calculations. Then, after obtaining the two-dimensional matrix, perform multi-factor graph interaction calculations on the elements in the matrix according to a predetermined order (such as according to the priority, correlation, etc. of the factors). Such calculations may involve operations such as matrix multiplication, addition, and comparison between elements, aiming to reveal the mutual influence and correlation between different factors. Through these calculations, a series of results reflecting the interaction relationships between different factor combinations can be obtained. These results can be organized into a new matrix, namely the multi-factor graph interaction matrix. Each element in this matrix represents the interaction strength or correlation degree between the corresponding factor combinations.

[0063] Based on the foregoing solution, in some implementation manners of the present application, performing multi-factor graph interaction calculations in the specified order to obtain a multi-factor graph interaction matrix of different factor combinations includes: generating two-factor, three-factor, four-factor, five-factor, six-factor, and seven-factor multi-factor graph interaction matrices through matrix dot product and flatten operations in the order of edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship.

[0064] In the above implementation manner, according to the professional knowledge and practical experience of medical image analysis, the priority and order of the factors of edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship are first determined. This order aims to ensure that the mutual influence and correlation between different factors can be gradually revealed in the subsequent calculation process. After determining the order, perform dot product operations on the feature vectors or matrices corresponding to each factor. This operation aims to calculate the similarity or correlation between different factors, thereby generating intermediate results reflecting the interaction relationship between the factors. In order to integrate the results obtained from the dot product operation into a unified matrix, a flatten operation is required. This operation converts a multi-dimensional matrix or tensor into a one-dimensional vector or a two-dimensional matrix for subsequent calculation and analysis. Through the above steps, two-factor, three-factor, four-factor, five-factor, six-factor, and seven-factor multi-factor graph interaction matrices can be gradually generated. Each element in these matrices represents the interaction strength or correlation degree between the corresponding factor combinations, thereby revealing the complex relationships between different features in medical images.

[0065] Based on the foregoing solution, in some implementation manners of the present application, the Graph-Mamba classification network includes at least one encoding module and at least one fully connected layer, and the encoding module includes a convolutional layer for capturing local information of short sequences and a Mamba module for capturing global information of long sequences.

[0066] In the above implementation, in the Graph-Mamba classification network, convolutional layers are used to process short sequence data, and local information in the sequence is extracted by means of a sliding window. This local information is crucial for understanding the detailed features of the data. The Mamba module is designed specifically to capture the global information of long sequences. Compared with traditional recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), the Mamba module can adopt a more efficient mechanism to process long sequence data, reducing the computational complexity while maintaining the ability to capture global information. This design enables the Graph-Mamba classification network to be more efficient and accurate in processing long sequence data. The fully connected layer is located after the encoding module and is used to receive the features extracted by the encoding module and perform further classification or regression tasks. The fully connected layer performs a linear transformation on the input features through a weight matrix and a bias term, and introduces non-linearity through an activation function, thereby enhancing the classification ability of the network.

[0067] Its working process includes: the input data is first processed by the encoding module, where the convolutional layer extracts the local information of the short sequence and the Mamba module captures the global information of the long sequence. The feature vector output by the encoding module is then fed into the fully connected layer for further classification or regression calculation. Finally, the network outputs the classification result or regression value, completing the entire processing process.

[0068] In summary, the above implementation optimizes the Graph-Mamba classification network, effectively captures the local information of short sequences and the global information of long sequences, improves the accuracy and computational efficiency of the classification task, and enhances the generalization ability of the model.

[0069] Based on the foregoing solution, in some implementations of the present application, the Graph-Mamba classification network includes 5 encoding modules and 3 fully connected layers. Each encoding module includes a 7×1 convolutional layer and a Mamba module. Among them, the output channels of each 7×1 convolutional layer are half of the input channels, and the feature channels of the 3 fully connected layers are 2048, 512, and 2 respectively.

[0070] In the above implementation, the Graph-Mamba classification network consists of 5 encoding modules, each equipped with a 7×1 convolutional layer and a Mamba module. This design aims to gradually extract the deep features of the input data. The 7×1 convolutional layer uses a 7×1 convolutional kernel to capture the local features of the input data. The number of output channels is set to half of the number of input channels, which helps reduce the computational complexity while maintaining the effective extraction of key features. Following the convolutional layer, the Mamba module is responsible for capturing the global information of the long sequence. This module adopts an efficient mechanism to process long sequence data to reduce the computational complexity and maintain the ability to capture global information. The Graph-Mamba classification network contains 3 fully connected layers, which are located after the encoding modules, used to receive the features extracted by the encoding modules and perform further classification tasks. The number of feature channels of the 3 fully connected layers is 2048, 512, and 2 respectively. This design helps gradually reduce the feature dimension while maintaining the effective representation of key classification information. Finally, the network outputs 2 classification results corresponding to the binary classification task.

[0071] Based on the foregoing solution, in some implementations of the present application, it further includes calculating the Softmax cross-entropy loss for the two-dimensional features output by the Graph-Mamba classification network to optimize the performance of the classification network.

[0072] Under the guidance of the Softmax cross-entropy loss function, the Graph-Mamba classification network can learn more accurate classification boundaries, thereby improving the classification accuracy. The Softmax cross-entropy loss function can effectively measure the difference between the probability distribution predicted by the model and the true distribution, and guide the training process of the model, enabling the model to better adapt to different data sets and scenarios, and enhancing the generalization ability of the model. Moreover, the Graph-Mamba classification network itself has high computational performance. Combining the optimization of the Softmax cross-entropy loss function can further reduce the computational complexity and improve the computational efficiency.

[0073] To enable those skilled in the art to more intuitively understand the present application, a specific example will be used to illustrate it here.

[0074] Taking the i-th target breast ultrasound image x i and the i-th breast lesion as an example, the detailed steps of the above implementation can be as follows:

[0075] (1) Calculate the centroid point of, and calculate the average value of the abscissa and ordinate of all points within the breast lesion area on to obtain the centroid point

[0076] (2) Calculate The abscissa c x and The distance between the maximum and minimum values of the abscissa, and calculate The ordinate cy and The distance between the maximum and minimum values of the ordinate, a total of 4 values, denoted as d x1 , d x2 , d y1 and d y2 .

[0077] (3) Calculate d x1 , d x2 , d y1 and d y2 The maximum value of, denoted as d max .

[0078] (4) Let c xl = c x - 2d max , c xr = c x + 2d max , c yl = c y - 2d max , c yr = c y + 2d max . Let W and H represent the width and height of xi. When 2d max < W and 2d max < H, continue to perform the following operations.

[0079] (5) When c xl ≥ 0, c yl ≥ 0, c xr < W and c yr < H, the values of c xl and c yl remain unchanged. When c xl < 0, let c xl = 0. When c yl < 0, let c yl = 0. When c xr ≥ W, let c xl = c xl -(c xr - W). When c yr ≥ H, let c yl = c yl -(c yr - H).

[0080] (6) Taking (c xl , c yl ) as the starting point in the upper left corner, 2dmax are the length and width, respectively, and crop the corresponding regions at xi and and resize (operation of changing the image size) to the size of 256*256, denoted as and At the same time, retain the scaling factor corresponding to each image, denoted as and retain the distance d corresponding to each pixel value in the original image i .

[0081] Next, calculate the multi-classification tasks of edge blur, edge shape, echo, morphology, and posterior echo attenuation information. Stack and as the input, use U-Mamba as the backbone network, concatenate two fully connected layers with 2048 and 512 channels at the end of the encoder, output features of 512 dimensions, and then respectively use 5 fully connected layers with 2, 2, 2, 2, 5, 3, 2 output channels in parallel. The first 4 fully connected layers with 2 output channels respectively perform 2-classification tasks of whether the edge of the mass is blurred, whether it has a spiculated shape, whether the edge of the mass is angulated, and whether the shape of the mass presents differential lobulation. The fourth fully connected layer with 5 output channels performs a 5-classification task of isoechoic, hypoechoic, cystic-solid mixed echo, heterogeneous echo, and hyperechoic of the mass. The fifth fully connected layer with 3 output channels performs a 3-classification task of whether the mass is oval, round, and non-oval and round. The sixth fully connected layer with 2 output channels performs a 2-classification task of whether there is attenuation behind the mass. All classification tasks use the Softmax cross-entropy loss function.

[0082] The edge shape of the mass includes whether there are spicules, whether the edge of the mass is angulated, and whether the shape of the mass presents differential lobulation. The classification losses are respectively denoted as L c1 , L c2 and L c3 . The classification loss L c4 of the mass echo, and the classification loss of the mass morphology is denoted as L c5 , and the classification loss of the posterior echo attenuation is denoted as L c6 .

[0083] Then calculate the segmentation of calcification and ducts: At the end of the U-Mamba decoder network, use 2 3×3 fully convolutional layers with 3 output channels in parallel to output two segmentation tasks. The first segmentation task outputs the segmentation results of the background, breast lesions, and ducts The second segmentation task outputs the segmentation results of the background, calcification, and solid / cystic-solid masses in the ducts For the two segmentation tasks, first, the Softmax cross-entropy loss function is used to calculate pixel-level optimization. Second, the prediction results of the second segmentation task are in an inclusion relationship with those of the first task. That is, the calcified area is within the breast mass lesion, and the solid / cystic mass within the duct is within the duct. Therefore, for the second segmentation task, the features belonging to the calcification prediction results and the features belonging to the solid / cystic mass within the duct are extracted respectively, and then the Softmax cross-entropy loss function is calculated with the breast lesion prediction features and duct prediction features at the corresponding positions of the first segmentation task. The loss of the first segmentation task is defined as L s1 , and the loss of the second segmentation task is defined as L s2 .

[0084] Finally, calculate the mass aspect ratio, calcification diameter, duct width, and whether there is a mass visible within the duct: After obtaining the prediction results of the two segmentation tasks and , divide and by the scaling factor r i to restore to the original resolution, then calculate the number of pixels occupied by the transverse and longitudinal diameters of the longest distances of the breast mass, calcification, and solid / cystic mass within the duct, and multiply by the distance d i corresponding to each pixel value to obtain the transverse and longitudinal diameters of the mass, calcification, and duct. Then calculate the aspect ratio of the mass. Considering that it is difficult for the transverse and longitudinal diameters of the breast mass to be exactly the same. Therefore, a range can be set. For example, when the aspect ratio of the mass is between [0.9, 1.1], the mass belongs to a round shape; when the aspect ratio of the mass < 0.9, it belongs to the parallel position; when the aspect ratio of the mass > 0.9, it belongs to the non-parallel position. Then, discriminate the calcification diameter, take the larger value of the transverse and longitudinal diameters of the calcification as the diameter, and then judge whether its diameter is greater than 5 mm. Microcalcifications are calcifications with a diameter < 0.5 mm, and coarse calcifications have a diameter > 0.5 mm. Discriminate the duct width, take the smaller value of the transverse and longitudinal diameters of the duct as the duct width, and judge whether the duct width is greater than 2 mm. A width greater than 2 mm indicates duct dilation.

[0085] Multi-task single-factor matrix: Through the above calculations, a multi-task single-factor matrix related to the benign and malignant classification of breast BI-RADS 4A can be obtained. The mass margin includes 4 two-class single-factor matrices, which are whether the mass margin is blurred, whether the mass has a spiculated shape, whether the mass margin is angulated, and whether the mass shape shows micro-lobulation, denoted as [a 11 , a 12 , [a 21 , a 22 , [a 31 , a 32 and [a 41 , a 42. The mass echo is a single-factor matrix with 5 classifications, including isoechoic mass, hypoechoic, cystic-solid mixed echo, heterogeneous echo, and hyperechoic, denoted as [a 51 , a 52 , a 53 , a 54 , a 55 . The mass shape is a single-factor matrix with 3 classifications, including the mass being oval, round, and non-oval and non-round cases, denoted as [a 61 , a 62 , a 63 . The mass aspect ratio is a single-factor matrix with 3 classifications, including parallel position, round, and non-parallel position, denoted as [a 71 , a 72 , a 73 . The mass calcification is a single-factor matrix with 2 classifications, including microcalcification and coarse calcification, denoted as [a 81 , a 82 . Whether the posterior echo is attenuated is a single-factor matrix with 2 classifications for mass calcification, denoted as [a 91 , a 92 . Whether the duct is dilated is a single-factor matrix with 2 classifications, denoted as [a 101 , a 102 . Whether a mass can be seen in the duct is a matrix with 2 classifications, denoted as [a 111 , a 112 .

[0086] Multi-factor graph interaction matrix: First, combine the 4 single-factor matrices of the mass edge into a group of single-factor matrices of the mass edge, denoted as [a 11 , a 12 , a 21 , a 22 , a 31 , a 32 , a 41 , a 42 . Combine the 2 single-factor matrices of whether the duct is dilated and whether a mass can be seen in the duct into a single-factor matrix of the duct relationship, denoted as [a 101 , a 102 , a 111 , a 112. Thus, 7 graph nodes are generated, including edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship, with matrix sizes of 8, 5, 3, 3, 2, 2, and 4 respectively. Then, the one-dimensional matrices are transformed into two-dimensional matrices with sizes of [1,8], [1,5], [1,3], [1,3], [1,2], [1,2], and [1,4] respectively. Then, the multi-factor graph interaction matrix is calculated. Taking the three-factor graph interaction matrix of edge, echo, and morphology as an example. First, the [1,8] matrix of the edge is transposed to obtain an edge matrix with a size of [8,1]. Then, the [8,1] edge matrix is multiplied element-wise with the [1,5] echo matrix to obtain an edge and echo two-factor graph interaction matrix with a size of [8,5]. Then, the [8,5] two-dimensional matrix is flattened into a one-dimensional two-factor graph interaction matrix with a size of 40, and then it is converted into a two-dimensional matrix with a size of [40,1]. Finally, the [40,1] two-factor graph interaction matrix is multiplied matrix-wise with the [1,3] morphology matrix to obtain a three-factor graph interaction matrix of edge, echo, and morphology with a size of [40,3]. Then, [40,3] is flattened and dimension-increased to obtain a three-factor graph interaction matrix with a size of [120,1].

[0087] Since the determination of the benign and malignant nature of breast BI-RADS 4A is analyzed by integrating multi-factor diagnosis and treatment information without considering the order of using which diagnosis and treatment information. To avoid invalid operations, for repeated interaction methods, according to the sorting order of edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship, directed multi-factor graph interaction calculation is performed. Then, the position information of edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship is corresponding to 1, 2, 3, 4, 5, 6, 7, and then the interactions of multiple graph nodes are position-encoded to determine whether to calculate the graph interaction matrix. For example, for the two connection methods of edge-echo-morphology and edge-morphology-echo of graph nodes, their corresponding position encodings are [1,2,3] and [1,3,2]. Among them, the position encoding of [1,2,3] conforms to the position value of the subsequent graph node being greater than that of the previous graph node, so only the graph interaction matrix of edge-echo-morphology is calculated. Thus, for 7 graph nodes, 21 two-factor graph interaction matrices, 15 three-factor graph interaction matrices, 10 four-factor graph interaction matrices, 6 five-factor graph interaction matrices, 3 six-factor graph interaction matrices, and 1 seven-factor graph interaction matrix can be calculated.

[0088] Benign and malignant classification network for breast BI-RADS 4A: Perform positional encoding on all single-factor matrices and multi-factor graph interaction matrices, then concatenate them into a one-dimensional array, and design a Graph-Mamba classification network for benign and malignant classification of breast BI-RADS 4A. Graph-Mamba includes 5 encoding modules and 3 fully connected layers. Each encoding module includes a 7×1 convolution and a Mamba module. The 7×1 convolution is used to capture local information of short sequences, and the Mamba module is used to capture global information of long sequences. At the same time, the output channels of the 7×1 convolution are half of the input channels, which is used to reduce the feature dimension. Finally, there are three fully connected layers with feature channels of 2048, 512, and 2 respectively. Finally, calculate the Softmax cross-entropy loss for the two-dimensional features output by the benign and malignant classification network of breast BI-RADS 4A.

[0089] Example 2

[0090] Please refer to Figure 3 , this embodiment of the present application provides an ultrasonic image classification system for breast 4A nodules, which includes:

[0091] An image segmentation module for sending the target breast ultrasound image into the U-Mamba network to segment the breast tumor and obtain the breast lesion segmentation result; a first construction module for classifying the edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship of each breast lesion according to the breast lesion segmentation result, and constructing a corresponding single-factor matrix according to the classification result; a second construction module for performing interaction processing according to the single-factor matrix to construct a multi-factor graph interaction matrix; an image classification module for sending the one-dimensional array obtained by performing positional encoding and concatenation processing on the single-factor matrix and the multi-factor graph interaction matrix into the Graph-Mamba classification network to calculate the classification result of the breast 4A nodules corresponding to each segmentation result.

[0092] For the specific implementation process of the above system, please refer to the ultrasonic image classification method for breast 4A nodules provided in Example 1, which will not be elaborated here.

[0093] Example 3

[0094] Please refer to Figure 4, an embodiment of the present application provides an electronic device, which includes at least one processor 201 and at least one memory 202; wherein, the processor 201 is directly connected to the memory 202, or communicates with each other through the communication interface 203, or is electrically connected through one or more communication buses or signal lines to realize data transmission or interaction; the memory 202 stores program instructions executable by the processor 201, and the processor 201 calls the program instructions to execute an ultrasonic image classification method for breast nodules of category 4A. For example, it realizes:

[0095] Send the target breast ultrasound image into the U-Mamba network to segment the breast tumor and obtain the breast lesion segmentation result; according to the breast lesion segmentation result, classify the edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation and duct relationship of each breast lesion, and construct a corresponding single-factor matrix according to the classification result; perform interactive processing according to the single-factor matrix to construct a multi-factor graph interaction matrix; send the one-dimensional array obtained by position encoding and concatenation processing using the single-factor matrix and the multi-factor graph interaction matrix into the Graph-Mamba classification network to calculate the classification result of breast nodules of category 4A corresponding to each segmentation result.

[0096] Among them, the memory 202 can be, but is not limited to, random access memory (Random Access Memory, RAM), read only memory (Read Only Memory, ROM), programmable read only memory (Programmable Read-Only Memory, PROM), erasable programmable read only memory (Erasable Programmable Read-Only Memory, EPROM), electrically erasable programmable read only memory (Electric Erasable Programmable Read-Only Memory, EEPROM), etc.

[0097] The processor 201 can be an integrated circuit chip with signal processing capabilities. The processor 201 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0098] It can be understood that Figure 4 The structure shown is only schematic, and the electronic device may also include more or fewer components than those shown Figure 4 in the figure, or have a different configuration from that shown Figure 4 in the figure. Figure 4 Each component shown in the figure can be implemented by hardware, software, or a combination thereof.

[0099] Embodiment 4

[0100] This application provides a computer-readable storage medium with a computer program stored thereon. When the computer program is executed by the processor 201, it implements a method for classifying ultrasonic images of breast nodules of category 4A. For example, it implements:

[0101] Sending the target breast ultrasound image into the U-Mamba network to segment the breast tumor, obtaining the segmentation result of the breast lesion; according to the segmentation result of the breast lesion, classifying the edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation, and duct relationship of each breast lesion, and constructing a corresponding single-factor matrix according to the classification results; performing interactive processing according to the single-factor matrix to construct a multi-factor graph interaction matrix; sending the one-dimensional array obtained by performing position encoding and concatenation processing on the single-factor matrix and the multi-factor graph interaction matrix into the Graph-Mamba classification network to calculate the classification result of the breast nodule of category 4A corresponding to each segmentation result.

[0102] If the above functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0103] For those skilled in the art, it is obvious that this application is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of this application, this application can be implemented in other specific forms. Therefore, from any perspective, the embodiments should be regarded as exemplary and non-restrictive. The scope of this application is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in this application. Any reference signs in the claims should not be regarded as limiting the claimed rights.

Claims

1. A method for classifying breast 4A nodules using ultrasound images, characterized in that: The following steps are involved: The target breast ultrasound image is sent to the U-Mamba network to segment the breast tumor and obtain the breast lesion segmentation result; According to the breast lesion segmentation results, the edge, echo, morphology, aspect ratio, calcification, posterior echo attenuation and ductal relationship of each breast lesion were classified, and the corresponding single factor matrix was constructed according to the classification results; Based on the single-factor matrix, the interaction processing was performed to construct a multi-factor graph interaction matrix; The one-dimensional array obtained by position encoding and cascading the single-factor matrix and the multi-factor graph interaction matrix is ​​sent to the Graph-Mamba classification network to calculate the classification result of breast 4A nodules corresponding to each segmentation result.

2. The method according to claim 1, characterized in that Before the step of sending the target breast ultrasound image to the U-Mamba network to segment the breast tumor, the method further includes: Calculate the centroid point of the area where the breast lesion is located; Calculate the horizontal and vertical distances from each point in the area where the breast lesion is located to the centroid point, as well as the distances between these distances and the maximum and minimum values ​​of the respective coordinate axes; According to the calculated distance value, the target breast ultrasound image is cropped and the image size is adjusted.

3. The method according to claim 1, characterized in that The margins, echogenicity, morphology, aspect ratio, calcifications, posterior echo attenuation, and ductal relationship of each breast lesion are classified, including: Edge classification, including whether the edge of the breast lesion is smooth and the shape of the edge. The edge shape includes three types: burr, edge angle and microlobule. Echo classification, including classification according to the echo type in the breast lesion, wherein the echo type includes isoechoic, hypoechoic, cystic mixed echoic, inhomogeneous echoic and hyperechoic; Morphological classification, including classification based on whether the shape of the breast lesion is oval, round, or neither oval nor round; Aspect ratio classification, including classification according to the aspect ratio of breast lesions, including parallel, non-parallel, or round; Calcification classification, including classification according to the diameter of breast lesions, including microcalcification and gross calcification; Posterior echo attenuation classification, including classification based on whether the posterior part of the breast lesion has echo attenuation; Ductal relationship classification includes whether the ducts around the breast lesions are dilated and whether there are visible solid / cystic-solid masses in the ducts.

4. The method according to claim 1 or 3, characterized in that: The constructed univariate matrices included four 2-category matrices of breast lesion margins, a 5-category matrix of echogenicity, a 3-category matrix of morphology, a 3-category matrix of aspect ratio, a 2-category matrix of calcification, a 2-category matrix of posterior echo attenuation, and a 2-category matrix of ductal relationship.

5. The method according to claim 1 or 3, characterized in that: The U-Mamba network includes: Using U-Mamba encoder as the backbone network, it performs multi-classification tasks on edge blur, edge shape, echo, morphology, and rear echo attenuation; At the end of the U-Mamba decoder, two 3×3 full convolutional layers with 3 output channels are used in parallel to output two segmentation tasks, one of which is used to output the segmentation results of the background, breast lesions, and ducts to obtain the first segmentation result, and the other is used to output the segmentation results of the background, calcifications, and solid / cystic-solid tumors in the ducts to obtain the second segmentation result; According to the first segmentation result and the second segmentation result, the aspect ratio, calcification diameter, duct width and information on whether a mass is visible in the duct of the breast lesion are calculated.

6. The method according to claim 1, characterized in that The interactive processing is performed according to the single factor matrix to construct a multi-factor graph interaction matrix, including: The single-factor matrix is ​​converted into a two-dimensional matrix, and then the multi-factor graph interaction calculation is performed in the specified order to obtain the multi-factor graph interaction matrix of different factor combinations.

7. The method according to claim 6, characterized in that The multi-factor graph interaction calculation is performed in a specified order to obtain a multi-factor graph interaction matrix with different factor combinations, including: In the order of edge, echogenicity, morphology, aspect ratio, calcification, posterior echo attenuation, and catheter relationship, multifactorial graph interaction matrices with two, three, four, five, six, and seven factors were generated by matrix dot product and flatten operations.

8. The method according to claim 1, characterized in that The Graph-Mamba classification network includes at least one encoding module and at least one fully connected layer, wherein the encoding module includes a convolutional layer for capturing local information of short sequences and a Mamba module for capturing global information of long sequences.

9. The method according to claim 8, characterized in that The Graph-Mamba classification network includes 5 encoding modules and 3 fully connected layers, each encoding module includes a 7×1 convolutional layer and a Mamba module, wherein the output channel of each 7×1 convolutional layer is half of the input channel, and the feature channels of the three fully connected layers are 2048, 512 and 2 respectively.

10. The method according to claim 1, characterized in that It also includes calculating the Softmax cross entropy loss on the two-dimensional features output by the Graph-Mamba classification network to optimize the performance of the classification network.

Images