Dynamic contrast-enhanced magnetic resonance image classification method, system, device and medium

Abstract

The application provides a dynamic contrast-enhanced magnetic resonance image classification method, system, device and medium, the method comprising: acquiring breast dynamic contrast-enhanced magnetic resonance images of different disease ages and preprocessing to generate a breast dynamic enhancement magnetic resonance image dataset; according to the breast dynamic enhancement magnetic resonance image dataset, an image classification prediction model comprising a multi-task classification network and a perception discrimination network connected in turn is constructed; the acquired breast dynamic contrast-enhanced magnetic resonance image to be classified is input into the image classification prediction model for multi-task classification, and an image classification result comprising the position, spatial structure and spatial signal of the tumor is obtained. The method can fully and accurately mine the rich multi-dimensional, multi-scale heterogeneous space, time and semantic correlation representation possessed by the breast contrast-enhanced magnetic resonance image data, and can supplement the missing different spatial dimension information, ensuring the reliability and accuracy of image classification prediction.

Description

Technical Field

[0001] This invention relates to the field of medical image data processing technology, and in particular to a dynamic contrast-enhanced magnetic resonance imaging classification method, system, device, and medium. Background Technology

[0002] With the continuous development of medical technology, breast cancer has become one of the most curable cancers through personalized precision treatment. Dynamic contrast-enhanced breast MRI (DCE-MRI) has high sensitivity, and research on breast MRI genomics has become a powerful tool for precision medicine, a key task in establishing clinical treatment plans and prognosis, and can effectively improve the 5-year survival rate and reduce the mortality rate of patients.

[0003] However, breast tumors of different natures exhibit multidimensional heterogeneity in dynamic magnetic resonance imaging, making it difficult to accurately correspond spatial, temporal, and semantic relationships. Furthermore, due to tumor infiltration, semantic relationships and interactions also exist between adjacent sub-regions, between anomalous sub-regions that are far apart, and between sub-regions that are related by temporal signals and spatial structures, making it difficult to effectively mine and utilize these relationships. At the same time, the imbalance of data and the lack of samples or annotations between different dimensions of magnetic resonance imaging not only make it difficult to train the representation model but also affect the robustness of the model, thereby affecting the accuracy of image classification. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic contrast-enhanced magnetic resonance imaging (MRI) image classification method that addresses the application deficiencies of existing dynamic contrast-enhanced MRI image omics data processing techniques. This method can fully and accurately mine the rich multidimensional and multi-scale heterogeneous spatial, temporal, and semantic association representations in breast contrast-enhanced MRI image data, and can supplement missing information in different spatial dimensions to achieve accurate and effective characterization, thereby ensuring the reliability and accuracy of image classification prediction results.

[0005] To achieve the above objectives, it is necessary to provide a dynamic contrast-enhanced magnetic resonance imaging classification method, system, computer equipment, and storage medium to address the aforementioned technical problems.

[0006] In a first aspect, embodiments of the present invention provide a dynamic contrast-enhanced magnetic resonance imaging classification method, the method comprising the following steps:

[0007] Dynamic contrast-enhanced magnetic resonance images of the breast at different disease stages were acquired and preprocessed to generate a dynamic contrast-enhanced magnetic resonance image dataset of the breast; the dynamic contrast-enhanced magnetic resonance images of the breast include contrast-enhanced magnetic resonance images of the entire breast and bilateral axillae.

[0008] Based on the aforementioned dynamic enhanced magnetic resonance imaging dataset of breast tissue, an image classification prediction model is constructed; the image classification prediction model includes a multi-task classification network and a perceptual discriminant network connected in sequence.

[0009] The acquired dynamic contrast-enhanced magnetic resonance images of the breast to be classified are input into the image classification prediction model for multi-task classification to obtain the image classification results; the image classification results include the location, spatial structure and spatial signal of the tumor.

[0010] Furthermore, the multi-task classification network includes a tumor location extraction module, a spatial structure extraction module, a spatial signal extraction module, a tower-style multi-scale dense residual network, a feature distillation module, and a multi-task classifier connected in sequence.

[0011] Furthermore, the loss function of the multi-task classification network includes a spatial content loss function and a total variable loss function;

[0012] The spatial content loss function is expressed as:

[0013] Loss C =||G(u′)-u||1

[0014] Among them, Loss c Let u' represent the spatial content loss; u' represent the image to be enhanced; u represents the target image; and G(u') represent the output of the multi-task classification network based on the image to be enhanced.

[0015] The total variable loss function is expressed as follows:

[0016]

[0017] Among them, Loss Tv G represents the total variable loss; u represents the target image; G(·) represents the output of the multi-task classification network; and represents the gradient of the output of the multi-task classification network based on the changes in variables x and y, respectively; C, H, and W represent the spatial dimensions of the augmented image.

[0018] Furthermore, the tower-type multi-scale dense residual network includes several multi-scale dense residual blocks; each multi-scale dense residual block is composed of three different sizes of tower-type convolutional kernels; the output of each multi-scale dense residual block is represented as:

[0019]

[0020] In the formula,

[0021]

[0022]

[0023]

[0024]

[0025] Among them, L n and L n-1 These represent the outputs of the nth and (n-1th)th multi-scale dense residual blocks, respectively. This represents a 5×5 convolutional layer of the i-th layer; This represents a 3×3 convolutional layer of the i-th layer; This indicates a 1×1 convolutional layer of the i-th layer; [*,…,*] represents the concat connection operation.

[0026] Furthermore, the perception-discrimination network includes an autoencoder module and a perception-discrimination module; the perception-discrimination module is an improved VGG network; the loss function of the perception-discrimination network includes an adversarial semantic loss function and an adversarial loss function; the adversarial semantic loss function is expressed as:

[0027]

[0028] Among them, Loss AS Indicates adversarial semantic loss; F j (·) represents the j-th convolutional layer of the perception and discrimination module; u' represents the image to be enhanced; u represents the target image; G(u') represents the output of the multi-task classification network based on the image to be enhanced; C', H', and W' represent the spatial dimensions of the image to be enhanced;

[0029] The adversarial loss function is expressed as:

[0030]

[0031] Among them, Loss A represents the adversarial loss; D(·) represents the output of the perceptual discriminant network.

[0032] Furthermore, the step of constructing an image classification and prediction model based on the breast dynamic contrast-enhanced magnetic resonance imaging dataset includes:

[0033] Each dynamic enhanced magnetic resonance image of the breast is stored in tensor form, and geometric objects of different dimensions are obtained through spatial geometric analysis of algebraic hypersurface neurons; the geometric objects include three-dimensional spatial vectors, two-dimensional spatial vectors, two-dimensional spatiotemporal vectors and one-dimensional time signal vectors;

[0034] The geometric objects of different dimensions are input into the tumor location extraction module for high-dimensional image scaling and edge pixel detection to obtain tumor location features;

[0035] The geometric objects of different dimensions are input into the spatial structure extraction module to perform weakly supervised clustering of spatial images of the same dimension, so as to obtain superpixel features and supervoxel features of multidimensional spatial images.

[0036] The geometric objects of different dimensions are input into the spatial signal extraction module to perform cross-dimensional spatial signal weakly supervised clustering to obtain superpixel features and supervoxel features of the image signal.

[0037] Based on the tumor location features, multidimensional spatial image superpixel features, multidimensional spatial image supervoxel features, image signal superpixel features, and image signal supervoxel features, the number of hidden layers and the number of hidden layer neurons in the tower-type multiscale dense residual network are determined, and multi-task association features are obtained by extracting association features through the tower-type multiscale dense residual network.

[0038] The multi-task associated features are input into the feature distillation module for feature fusion and redundancy removal to obtain tumor heteroproton region features.

[0039] The tumor heterogeneous region features and the preset target image are input into the autoencoder module for image feature encoding to obtain the tumor heterogeneous region encoded features;

[0040] The tumor heteroprotic region encoding features are input into the perception and discrimination module, and adversarial learning is performed based on the loss function of the perception and discrimination network to obtain enhanced image features;

[0041] The enhanced image features and the tumor heterogeneous region features are input into the multi-task classifier, and jointly trained according to the loss function of the multi-task classification network to obtain the image classification prediction model.

[0042] Furthermore, the step of determining the number of hidden layers and the number of hidden layer neurons in the tower-type multi-scale dense residual network includes:

[0043] Principal component analysis is performed on the tumor location features, multidimensional spatial image superpixel features, multidimensional spatial image supervoxel features, image signal superpixel features, and image signal supervoxel features to obtain principal component features. Based on the number of principal component features, the number of hidden layers of the tower-type multiscale dense residual network is determined.

[0044] The number of hidden layer neurons in the tower-type multi-scale dense residual network is determined by performing cluster analysis or principal component singular value calculation on the principal component features.

[0045] Secondly, embodiments of the present invention provide a dynamic contrast-enhanced magnetic resonance imaging classification system, the system comprising:

[0046] The dataset construction module is used to acquire and preprocess dynamic contrast-enhanced magnetic resonance images of the breast at different disease stages to generate a dynamic contrast-enhanced magnetic resonance image dataset of the breast; the dynamic contrast-enhanced magnetic resonance images of the breast include contrast-enhanced magnetic resonance images of the entire breast and bilateral axillae.

[0047] The model building module is used to construct an image classification prediction model based on the dynamic enhanced magnetic resonance imaging dataset of the breast; the image classification prediction model includes a multi-task classification network and a perceptual discriminant network connected in sequence;

[0048] The image classification module is used to input the acquired dynamic contrast-enhanced magnetic resonance images of the breast to be classified into the image classification prediction model for multi-task classification to obtain the image classification result; the image classification result includes the location, spatial structure and spatial signal of the tumor.

[0049] Thirdly, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0050] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0051] The present application provides a method, system, computer device, and storage medium for classifying dynamic contrast-enhanced magnetic resonance images. The method acquires and preprocesses dynamic contrast-enhanced magnetic resonance images of the breast at different disease stages to generate a dynamic contrast-enhanced magnetic resonance image dataset. Based on the dataset, an image classification prediction model is constructed, comprising a multi-task classification network and a perceptual discriminant network connected in sequence. The acquired dynamic contrast-enhanced magnetic resonance images of the breast to be classified are input into the image classification prediction model for multi-task classification, yielding image classification results including tumor location, spatial structure, and spatial signal. Compared with existing technologies, this dynamic contrast-enhanced magnetic resonance imaging (MRI) image classification method not only presents a vector node embedding method based on geometric algebra, realizing the semantic association and correspondence of images and signals of different dimensions on the input nodes in a multidimensional consistent representation form, thus obtaining more comprehensive and scalable feature representations, but also designs and establishes a multilayer perceptual neural network that can automatically determine the number of hidden layers and corresponding nodes, effectively improving the learning effect and generalization ability of the joint representation model. Furthermore, it establishes a novel tower-style multi-scale dense residual deep convolutional neural network model based on spatial domain image and spatiotemporal signal image analysis, providing a representation model with better generalization and applicability in the field of multidimensional medical imaging. This method fully and accurately mines the rich multidimensional and multi-scale heterogeneous spatial, temporal, and semantic association representations in breast contrast-enhanced MRI image data, while supplementing missing information in different spatial dimensions, effectively improving the reliability and accuracy of image classification prediction results. Ultimately, it can provide more scientific and valuable auxiliary diagnostic analysis results for doctors at different levels with varying levels of knowledge and experience. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the application scenario of the dynamic contrast-enhanced magnetic resonance imaging classification method in this embodiment of the invention;

[0053] Figure 2 This is a flowchart illustrating the dynamic contrast-enhanced magnetic resonance imaging classification method in an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the structure of the multi-task classification network in an embodiment of the present invention;

[0055] Figure 4 yes Figure 3 A schematic diagram of the structure of a mid-tower type multi-scale dense residual network (enhanced U-Net network);

[0056] Figure 5 This is a schematic diagram of the structure of the perception and discrimination network in an embodiment of the present invention;

[0057] Figure 6This is a schematic diagram of the principal component analysis process for determining the number of hidden layers in a tower-type multi-scale dense residual network in an embodiment of the present invention.

[0058] Figure 7 This is a schematic diagram of the structure of the dynamic contrast-enhanced magnetic resonance imaging classification system in an embodiment of the present invention;

[0059] Figure 8 This is an internal structural diagram of the computer device in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and beneficial effects of this application clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the embodiments described below are only part of the embodiments of the present invention and are used to illustrate the present invention, but are not intended to limit the scope of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0061] The dynamic contrast-enhanced magnetic resonance imaging classification method provided by this invention can be applied to... Figure 1 The terminal and server shown are described. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be a standalone server or a server cluster consisting of multiple servers. The server can employ the dynamic contrast-enhanced magnetic resonance imaging classification method of this invention to construct an image classification prediction model based on a multi-task adversarial training framework using acquired dynamic contrast-enhanced magnetic resonance images of breast tissue at different disease stages. This model is then used to perform multi-task classification of the dynamic contrast-enhanced magnetic resonance images of breast tissue to be classified. The obtained image classification results can be used for subsequent research on the server or sent to the terminal for users to view and analyze. The following embodiments will provide a detailed description of the dynamic contrast-enhanced magnetic resonance imaging classification method of this invention.

[0062] In one embodiment, such as Figure 2 As shown, a dynamic contrast-enhanced magnetic resonance imaging (MRI) image classification method is provided, including the following steps:

[0063] S11. Acquire and preprocess dynamic contrast-enhanced magnetic resonance images of the breast at different disease stages to generate a dynamic contrast-enhanced magnetic resonance image dataset of the breast. The dynamic contrast-enhanced magnetic resonance images of the breast include contrast-enhanced magnetic resonance images of the entire breast and bilateral axillae. These dynamic contrast-enhanced magnetic resonance images can be understood as axial and sagittal fast spiral echo (FSE) sequences of the breast, including the entire breast and bilateral axillae, obtained from relevant image databases through dynamic contrast-enhanced scanning of breast tumor patients at different disease stages in a prone position using an MRI machine. It should be noted that the preprocessing of each dynamic contrast-enhanced magnetic resonance image in this embodiment includes normalizing the MRI signal intensity using spatial interpolation and bias field correction methods, which can effectively solve the problem of inconsistent image intensity distribution due to different internal settings of different MRI scanning equipment and different preprocessing parameters. It also includes medical imaging experts using Mimics v20 medical image analysis software to annotate the lesion areas of the magnetic resonance images of breast cancer patients who have undergone surgery and have been pathologically confirmed, based on clinical diagnostic results, to facilitate subsequent prediction and classification through machine learning.

[0064] S12. Based on the aforementioned dynamic enhanced magnetic resonance imaging dataset of the breast, construct an image classification prediction model; wherein, the image classification prediction model can be understood as a multi-task classification prediction model obtained based on a multi-task adversarial training framework, including a multi-task classification network and a perceptual discriminant network connected in sequence.

[0065] like Figure 3As shown, the multi-task classification network includes a tumor location extraction module, a spatial structure extraction module, a spatial signal extraction module, a tower-style multi-scale dense residual network, a feature distillation module, and a multi-task classifier connected in sequence. The tumor location extraction module can be understood as using high-dimensional image low-dimensional reconstruction technology to effectively suppress the influence of false positives caused by image measurement noise, achieving the detection of cross-dimensional tumor-related regions. Furthermore, by detecting edge pixels, it enhances the recognition and fitting ability of tumor image boundary points, thereby filling in missing pixels. The spatial structure extraction module can be understood as using a multi-instance weakly supervised clustering method to perform clustering analysis on same-dimensional spatial image data in high-dimensional image data to mine potential multi-level semantic information in the image data. The spatial signal extraction module can be understood as using a multi-instance weakly supervised clustering method to cluster cross-dimensional spatial signal data in high-dimensional image data. Analysis is used to uncover potential time-series features in image data. The tower-style multi-scale dense residual network (an enhanced U-Net network) can be understood as an improved version of the U-Net network. It uses a residual neural network combined with U-Net to strengthen the connections between layers, resulting in a multi-dimensional, multi-scale dense residual neural network that can automatically extract rich multi-dimensional and multi-scale image features, enabling it to more stably and efficiently complete tumor image classification tasks. The feature distillation module can be understood as automatically fusing all image features and removing redundant features to minimize network model parameters while ensuring network model performance, thus improving computational efficiency. The multi-task classifier can be understood as simultaneously including classifiers for structural classification, signal classification, and tumor location classification, where the main and auxiliary tasks can be specifically set according to actual application needs; no specific limitations are made here.

[0066] It should be noted that the tower-style multi-scale dense residual network used in this embodiment is based on the consideration that dynamic magnetic resonance images are composed of multi-dimensional geometric objects with different structures, which can increase the width of the representation model. If multi-dimensional consistency deep learning is to be obtained, the convolutional layers need to be reduced in dimensionality at multiple scales according to different data structures through multiple kernels to increase the depth of the representation model and improve its representation ability. The proposed novel tower-style multi-scale dense residual network, similar to U-Net, combines the residual network idea with dense convolutional networks as the main network structure, and integrates tower-style convolutional kernels into multi-scale dense residual blocks. The depth of the convolutional kernels gradually decreases as the kernel size increases. Each set of inputs is a feature input at all scales, and each set of outputs features at different scales. Its special structural design can automatically extract features of the image at various scales and effectively reduce dimensionality using 1*1 convolutions, thereby better meeting the needs of analyzing multi-dimensional geometric images of breast tumors with different scales. Specifically, the tower-style multi-scale dense residual network is as follows: Figure 4As shown, it includes several multi-scale dense residual blocks; each multi-scale dense residual block consists of three different sizes of tower-type convolutional kernels; the output of each multi-scale dense residual block is represented as:

[0067]

[0068] In the formula,

[0069]

[0070]

[0071]

[0072]

[0073] Among them, L n and L n-1 These represent the outputs of the nth and (n-1th)th multi-scale dense residual blocks, respectively. This represents a 5×5 convolutional layer of the i-th layer; This represents a 3×3 convolutional layer of the i-th layer; This indicates a 1×1 convolutional layer of the i-th layer; [*,…,*] represents the concat connection operation.

[0074] To ensure accurate and effective classification prediction, and considering features such as content, texture, and color (intensity) of magnetic resonance images, this embodiment preferably designs a series of new loss functions for the aforementioned multi-task classification network to better complete the learning and training of network parameters, including the spatial content loss function (hereinafter referred to as Loss). C ) and the total variation loss function (denoted as Loss) Tv ), and respectively represented as:

[0075] The spatial content loss function is expressed as:

[0076] Loss C =||G(u′)-u||1

[0077] Among them, Loss c Let u' represent the spatial content loss; u' represent the image to be enhanced; u represents the target image; and G(u') represent the output of the multi-task classification network based on the image to be enhanced.

[0078] The total variable loss function is expressed as:

[0079]

[0080] Among them, Loss TvG represents the total variable loss; u represents the target image; G(·) represents the output of the multi-task classification network; and represent the gradients of the output of the multi-task classification network based on the changes in variables x and y, respectively; C, H, and W represent the spatial dimensions of the augmented image;

[0081] Considering the imbalance and lack of annotations in different dimensions of magnetic resonance imaging, not only is it difficult to train the representation model, but it also affects the robustness of the model. This embodiment preferably introduces a generative adversarial network to supplement the missing information in different spatial dimensions, thereby improving the established spatiotemporal representation model (spatiotemporal feature extraction model) and helping the model to be trained effectively under conditions such as missing labels.

[0082] Perceptual discriminative networks can be understood as adversarial generative networks, such as... Figure 5 As shown, it includes an autoencoder module and a perceptual discrimination module. The autoencoder module is mainly used to encode the features of the input image so that it retains only the most important feature components, reducing computational complexity. It can be implemented using existing autoencoders, and no specific restrictions are imposed here. Correspondingly, the perceptual discrimination module is mainly used to extract the high-level semantic features of the image and further judge the authenticity of the generated image from the high-level semantic feature level. It is implemented using an improved VGG network, such as the existing improved VGG16 and VGG19 networks, and no specific restrictions are imposed here.

[0083] To ensure the reliability and effectiveness of the enhanced images generated by the perceptual discriminant network, this embodiment preferably designs a series of new loss functions to further discriminate the enhanced images. Specifically, the loss functions of the perceptual discriminant network include an adversarial semantic loss function and an adversarial loss function; the adversarial semantic loss function is expressed as follows:

[0084]

[0085] Among them, Loss AS Indicates adversarial semantic loss; F j (·) represents the j-th convolutional layer of the perception and discrimination module; u' represents the image to be enhanced; u represents the target image; G(u') represents the output of the multi-task classification network based on the image to be enhanced; C', H' and W' represent the spatial dimensions of the image to be enhanced.

[0086] The adversarial loss function is expressed as:

[0087]

[0088] Among them, Loss A represents the adversarial loss; D(·) represents the output of the perceptual discriminant network.

[0089] The construction process of the image classification prediction model can be understood as an iterative training process of jointly training a multi-task classification network and a perceptual discriminant network that meet the above requirements, resulting in a parameter-stable model that meets the classification requirements: First, considering the dimensionality loss that may be caused when instantiating features from the original image data, geometric objects are extracted from the dynamic enhanced magnetic resonance imaging of the breast based on geometric algebra theory. Related sub-regions of interest in the multidimensional image are embedded into node representations to form vector node representations. Then, superpixel and supervoxel clustering is performed on the spatial and spatiotemporal images, and the spatial and spatiotemporal images are clustered into several superpixel and supervoxel neurons based on texture, etc., thereby obtaining tumor heterogeneous sub-region features. Finally, knowledge transfer is performed through a generative adversarial network.

[0090] Specifically, the step of constructing an image classification and prediction model based on the dynamic enhanced magnetic resonance imaging dataset of the breast includes:

[0091] Each breast dynamic enhanced magnetic resonance imaging (MRI) image is stored in tensor form, and geometric objects of different dimensions are obtained through spatial geometric analysis using algebraic hypersurface neurons. These geometric objects can be understood as four different levels of the multi-dimensional image: point, line, surface, and volume, as well as vectors in different directions on the sagittal, coronal, and horizontal planes, including three-dimensional spatial vectors, two-dimensional spatial vectors, two-dimensional spatiotemporal vectors, and one-dimensional time signal vectors. Specifically, the steps for obtaining geometric objects of different dimensions through spatial geometric analysis using algebraic hypersurface neurons include:

[0092] Clustering is performed on dynamic enhanced magnetic resonance imaging of the breast stored in tensor form to generate a three-dimensional spatial vector with multiple block-shaped regional blocks associated with each other. Each regional block in the three-dimensional spatial image data can be regarded as a series of two-dimensional spatial image data stacks arranged along three different directions: sagittal, coronal and horizontal planes.

[0093] Two-dimensional spatial image data of each block region in each direction are extracted to obtain a two-dimensional spatial vector;

[0094] Add time frame sequences to each two-dimensional spatial vector to obtain a three-dimensional spatiotemporal vector;

[0095] Adding a time frame sequence to each row or column of a two-dimensional spatial vector in a three-dimensional spatial vector yields a two-dimensional spatiotemporal vector.

[0096] Adding time frame sequences to each pixel in two-dimensional spatial image data yields a series of corresponding time dimension signal vectors.

[0097] This embodiment decomposes high-dimensional images into geometric objects and, based on geometric algebra theory, embeds related sub-regions of interest in multi-dimensional images into node representations, forming a vector node representation. It presents a vector node embedding method based on geometric algebra, which not only realizes the semantic association and correspondence between images and their signals of different dimensions at the input nodes, thus avoiding potential dimensionality loss during the instantiation of original data features, but also, by fully exploring the multi-scale semantic features of tumor morphology and signal intensity changes in different tissues in cross-dimensional images, it can form a more comprehensive cluster of input nodes with different representation methods and capabilities. This provides a reliable basis for constructing deeper and broader classification models suitable for high-dimensional medical images.

[0098] The geometric objects of different dimensions are input into the tumor location extraction module for high-dimensional image scaling and edge pixel detection to obtain tumor location features. The extraction of tumor location features can be understood as the process of analyzing and reconstructing multidimensional consistent temporal, spatial, and semantic related regions through tensor analysis, and mining the potential correlation between multidimensional medical image spatial, temporal, and semantic features.

[0099] Tumor growth often affects surrounding normal tissues. Within the same dimension of an image, the tumor region often shows adhesion to its surrounding neighborhood. That is, the associated regions within a dimension are mostly considered to be adjacent regions and related regions at the image edge. However, for cross-dimensional medical dynamic enhanced images, it is necessary to search for semantically associated regions in each dimension, both spatially and temporally, and to mine the correlation between abnormal regions between dimensions to provide a basis for the establishment of a representation model. Considering the characteristics of medical images, in order to better search for semantically associated regions between dimensions, this embodiment preferably adopts a high-dimensional image low-dimensional reconstruction method to effectively suppress the influence of false positives caused by image measurement noise and realize the detection of cross-dimensional tumor associated regions. In addition, in order to accurately reconstruct tumor images, various tensor decomposition methods, such as HOSVD, HOOI, 2DPCA, and 2DSVD, can be compared and analyzed to determine the tensor decomposition method with a higher degree of preservation of image data structure information as the method used for tensor analysis in this embodiment.

[0100] According to the definition of tensor decomposition, we have in, To reconstruct the high-dimensional image data decomposition formula for the previous τ time step, C τ The core tensor is used; to discover the image region with maximum enhancement, a principal component analysis strategy is adopted for the dynamic two-dimensional image with low-dimensional embedding in each direction. By calculating the covariance matrix Δ and its corresponding eigenvector E, a new low-dimensional matrix is ​​generated: The final tensor-reconstructed silhouette image can be redefined based on the generated low-dimensional matrix as follows: τ is the average core tensor, and γ represents multiple channels corresponding to time frame τ. While reducing the differences between the temporal and spatial dimensions through high-dimensional image scaling techniques, the detection of edge pixels enhances the ability to identify and fit tumor image boundary points, thereby filling in missing pixels and obtaining effective tumor location features.

[0101] The geometric objects of different dimensions are input into the spatial structure extraction module for weakly supervised clustering of spatial images in the same dimension, to obtain superpixel features and supervoxel features of multidimensional spatial images. The weakly supervised clustering of spatial images in the same dimension can be understood as performing multi-instance weakly supervised clustering of two-dimensional spatial vectors and three-dimensional spatial vectors based on the semantic annotation and orientation of the image data. The specific clustering method can be obtained by combining the fuzzy C-means clustering method with the superpixel segmentation method, and then integrating machine learning algorithms such as vector machines as needed. It will not be elaborated here.

[0102] The geometric objects of different dimensions are input into the spatial signal extraction module for cross-dimensional spatial signal weakly supervised clustering to obtain superpixel features and supervoxel features of the image signal. Cross-dimensional spatial signal weakly supervised clustering can be understood as performing multi-instance weakly supervised clustering on two-dimensional spatiotemporal vectors, three-dimensional spatiotemporal vectors and one-dimensional time signal vectors. The specific clustering method can also be obtained by combining the fuzzy C-means clustering method with the superpixel segmentation method, and then integrating machine learning algorithms such as vector machines as needed. It will not be elaborated here.

[0103] Based on the tumor location features, multidimensional spatial image superpixel features, multidimensional spatial image supervoxel features, image signal superpixel features, and image signal supervoxel features, the number of hidden layers and the number of hidden layer neurons in the tower-like multiscale dense residual network are determined. Then, association features are extracted through the tower-like multiscale dense residual network to obtain multi-task association features. The tumor location features, multidimensional spatial image superpixel features, multidimensional spatial image supervoxel features, image signal superpixel features, and image signal supervoxel features can be understood as cross-scale multidimensional image features based on geometric algebra, achieving a preliminary representation of same-dimensional and cross-dimensional association relationships. Considering the rich and close connections between various local regions within breast images, and when using a multilayer perceptual convolutional network to complete the classification task, the network... Mining the connections between nodes and multiple hidden layers plays a crucial role in feature extraction and subsequent classification prediction. The selection of an appropriate number of layers and hidden layer nodes significantly affects the accuracy of neural network classification. In this embodiment, principal component analysis is preferably used to transform the original correlated variables into some new uncorrelated variables of smaller size. The number of hidden layers in the neural network is determined based on the cumulative variance, and then the number of hidden layer neurons is determined by clustering. The core of using a tower-style multi-scale dense residual network for aggregation lies in the aggregation and mining of neighbor node features based on the adjacency matrix. Dynamically sampling the adjacency matrix can obtain neurons of different sizes, representing data blocks of different scales, providing more potential information for classification prediction and providing a reliable guarantee for accurate prediction of multi-task classification.

[0104] Specifically, the steps of determining the number of hidden layers and the number of hidden layer neurons in the tower-type multi-scale dense residual network include:

[0105] Principal component analysis is performed on the tumor location features, multidimensional spatial image superpixel features, multidimensional spatial image supervoxel features, image signal superpixel features, and image signal supervoxel features to obtain principal component features. Based on the number of principal component features, the number of hidden layers of the tower-type multiscale dense residual network is determined.

[0106] The number of hidden layer neurons in the tower-type multi-scale dense residual network is determined by performing cluster analysis or principal component singular value calculation on the principal component features.

[0107] Specifically, in principal component analysis, the principal components (PCi) are linear combinations of the original feature variables, where the value of i ranges from 1 to p. That is, if there are p random variables, there are p principal components. The number of principal components retained in the analysis depends on the percentage of the cumulative variance of the retained components in the total variance (i.e., the cumulative contribution rate). This indicates how much feature information the first few principal components summarize. Figure 6As shown, the area above the horizontal line represents the variance of the composite variable, which is largely comprised of the two retained principal components. By transforming the input into several principal components, whose values ​​remain within a certain range (maximum value - minimum value), and by using methods such as clustering input nodes, such as the K-Means method, we can aggregate and mine node information and the relationships between nodes in the convolutional neural network.

[0108] like Figure 6 The multilayer perceptron feedforward neural network shown has each node in the upper layer connected to every node in the lower layer; if each layer k of the neural network contains p1, ..., p n A neuron's output can be represented as a vector (column) h1, ..., h2. k Each column vector consists of p1...pn (n=5 in the figure). The hidden layers h1,...,h can be obtained through clustering or by calculating the singular values ​​of principal components (as described above the horizontal line in the figure). k The number of nodes; deeper hidden layers imply more complex, deeper abstract features. The complexity of hidden layer features corresponds to the principal components with higher variance. That is, the number of hidden layers in a neural network is determined by the number of important principal components formed through PCA transformation, satisfying the formula: ∑Variance(PCA) i )=∑Complexity(h i ).

[0109] This embodiment takes into account that dimensional data increases the number of hidden layers, thereby reducing the performance of neural networks and the accuracy of tumor classification. Combining doctors' cognitive experience of tumor image morphology during diagnosis, a multilayer perceptual neural network that can automatically determine the number of hidden layers and corresponding nodes was designed and established. This results in a multidimensional consistent spatiotemporal representation model of potential semantic associations in multidimensional images in image genomics. Furthermore, with the support of artificial intelligence data analysis strategies, the learning effect and generalization ability of the joint representation model are effectively improved, providing more scientific and valuable auxiliary diagnostic analysis results for doctors at all levels with different knowledge and experience.

[0110] The multi-task associated features are input into the feature distillation module for feature fusion and redundancy removal to obtain tumor heteroproton region features.

[0111] The tumor heterogeneous region features and the preset target image are input into the autoencoder module for image feature encoding to obtain the tumor heterogeneous region encoded features;

[0112] The tumor heteroprotic region encoding features are input into the perception and discrimination module, and adversarial learning is performed based on the loss function of the perception and discrimination network to obtain enhanced image features;

[0113] The enhanced image features and the tumor heterogeneous region features are input into the multi-task classifier, and jointly trained according to the loss function of the multi-task classification network to obtain the image classification prediction model.

[0114] It should be noted that the process of constructing the image classification prediction model can be understood as a process of jointly iteratively training a multi-task classification network and a perceptual discriminant network based on a dynamic enhanced magnetic resonance imaging dataset of breast tissue. In practical applications, existing multi-task learning and adversarial training methods are used, combined with the loss function of the multi-task classification network and the loss function of the perceptual discriminant network designed in this invention, and trained according to preset iterative convergence conditions to obtain an image classification prediction model that meets the requirements of multi-task classification. This embodiment effectively alleviates the imbalance problems of dimensionality and annotation in multi-dimensional breast tissue image data by adopting a transfer mechanism based on anti-generative networks.

[0115] S13. Input the acquired dynamic contrast-enhanced magnetic resonance imaging of the breast to be classified into the image classification prediction model for multi-task classification to obtain the image classification result; the image classification result includes the location, spatial structure and spatial signal of the tumor; wherein, the process of obtaining the image classification result by classifying and predicting the dynamic contrast-enhanced magnetic resonance imaging of the breast to be classified by the image classification prediction model can participate in the processing of each module given in step S12 above, which will not be repeated here.

[0116] This application's embodiments utilize a preprocessing method to generate a dynamic contrast-enhanced MRI dataset of breast tissue at different disease stages. Based on this dataset, an image classification prediction model is constructed, comprising a multi-task classification network and a perceptual discriminant network connected sequentially. This model performs multi-task classification prediction on the dynamic contrast-enhanced MRI images of breast tissue to be classified, yielding image classification results including tumor location, spatial structure, and spatial signals. This approach not only provides a vector node embedding method based on geometric algebra, achieving semantic association and correspondence between images and signals of different dimensions at input nodes in a multi-dimensional consistent representation, thus obtaining more comprehensive and scalable feature representations, but also designs and establishes a... A multilayer perceptron capable of automatically determining the number of hidden layers and corresponding nodes effectively enhances the learning effect and generalization ability of the joint representation model. Furthermore, a novel tower-style multi-scale dense residual deep convolutional neural network model based on spatial domain image and spatiotemporal signal image analysis has been established. This provides a representation model with better generalization and applicability for the field of multidimensional medical imaging. Specifically, it fully and accurately mines the rich multidimensional, multi-scale heterogeneous spatial, temporal, and semantic associations in breast contrast-enhanced magnetic resonance imaging data, while supplementing missing information in different spatial dimensions. This effectively improves the reliability and accuracy of image classification and prediction results, thereby providing more scientific and valuable auxiliary diagnostic analysis results for doctors at different levels with varying levels of knowledge and experience.

[0117] In one embodiment, such as Figure 7 As shown, a dynamic contrast-enhanced magnetic resonance imaging classification system is provided, the system comprising:

[0118] Dataset construction module 1 is used to acquire and preprocess dynamic contrast-enhanced magnetic resonance images of the breast at different disease stages to generate a dynamic contrast-enhanced magnetic resonance image dataset of the breast; the dynamic contrast-enhanced magnetic resonance images of the breast include contrast-enhanced magnetic resonance images of the entire breast and bilateral axillae.

[0119] Model building module 2 is used to construct an image classification prediction model based on the dynamic enhanced magnetic resonance imaging dataset of the breast; the image classification prediction model includes a multi-task classification network and a perceptual discriminant network connected in sequence;

[0120] The image classification module 3 is used to input the acquired dynamic contrast-enhanced magnetic resonance images of the breast to be classified into the image classification prediction model for multi-task classification to obtain the image classification result; the image classification result includes the location, spatial structure and spatial signal of the tumor.

[0121] Specific limitations regarding the dynamic contrast-enhanced magnetic resonance imaging (MRI) classification system can be found in the limitations of the dynamic contrast-enhanced MRI classification method described above, and will not be repeated here. Each module in the aforementioned dynamic contrast-enhanced MRI classification system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0122] Figure 8 An internal structural diagram of a computer device is shown in one embodiment. This computer device may specifically be a terminal or a server. Figure 8 As shown, the computer device includes a processor, memory, network interface, display, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a dynamic contrast-enhanced magnetic resonance imaging classification method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0123] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0124] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0125] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0126] In summary, the present invention provides a dynamic contrast-enhanced magnetic resonance imaging (MRI) image classification method, system, computer device, and storage medium. The method acquires and preprocesses dynamic contrast-enhanced MRI images of the breast at different disease stages to generate a dynamic contrast-enhanced MRI image dataset. Based on this dataset, an image classification prediction model is constructed, comprising a multi-task classification network and a perceptual discriminant network connected sequentially. The acquired dynamic contrast-enhanced MRI images of the breast to be classified are input into the image classification prediction model for multi-task classification, yielding image classification results including tumor location, spatial structure, and spatial signals. This method not only provides a vector node embedding method based on geometric algebra, but also realizes the semantic relationships of images and their signals in different dimensions at the input nodes in a multi-dimensional consistent representation. By combining and corresponding methods, more comprehensive and scalable feature representations can be obtained. Furthermore, a multilayer perceptual neural network that can automatically determine the number of hidden layers and corresponding nodes has been designed and established, effectively improving the learning effect and generalization ability of the joint representation model. A novel tower-style multi-scale dense residual deep convolutional neural network model based on spatial domain image and spatiotemporal signal image analysis has also been established, providing a representation model with better generalization and applicability in the field of multidimensional medical imaging. This model fully and accurately mines the rich multidimensional and multi-scale heterogeneous spatial, temporal, and semantic association representations in breast contrast-enhanced magnetic resonance imaging data, while supplementing the missing information in different spatial dimensions. This effectively improves the reliability and accuracy of image classification prediction results, and thus provides more scientific and valuable auxiliary diagnostic analysis results for doctors at all levels with different knowledge and experience.

[0127] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0128] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method of dynamic contrast-enhanced magnetic resonance image classification, characterized in that, The method includes the following steps: Dynamic contrast-enhanced magnetic resonance images of the breast at different disease stages were acquired and preprocessed to generate a dynamic contrast-enhanced magnetic resonance image dataset of the breast; the dynamic contrast-enhanced magnetic resonance images of the breast include contrast-enhanced magnetic resonance images of the entire breast and bilateral axillae. Based on the aforementioned dynamic enhanced magnetic resonance imaging dataset of breast tissue, an image classification prediction model is constructed. The image classification prediction model includes a multi-task classification network and a perception discrimination network connected in sequence. The multi-task classification network includes a tumor location extraction module, a spatial structure extraction module, a spatial signal extraction module, a tower-style multi-scale dense residual network, a feature distillation module, and a multi-task classifier connected in sequence. The acquired dynamic contrast-enhanced magnetic resonance images of the breast to be classified are input into the image classification prediction model for multi-task classification to obtain the image classification results; the image classification results include the location, spatial structure and spatial signal of the tumor.

2. The dynamic contrast-enhanced magnetic resonance image classification method of claim 1, wherein, The loss function of the multi-task classification network includes a spatial content loss function and a total variable loss function; The spatial content loss function is expressed as: in, This indicates a loss of spatial content; Indicates the image to be enhanced; Represents the target image. This represents the output of a multi-task classification network based on the image to be enhanced. The total variable loss function is expressed as follows: in, Indicates the total variable loss; Represents the target image; This represents the output of the multi-task classification network; and represents the gradient of the output of the multi-task classification network based on the changes in variables x and y, respectively; C, H, and W represent the spatial dimensions of the augmented image.

3. The dynamic contrast-enhanced magnetic resonance imaging classification method as described in claim 1, characterized in that, The tower-type multi-scale dense residual network comprises several multi-scale dense residual blocks; each multi-scale dense residual block is composed of three different sizes of tower-type convolutional kernels; the output of each multi-scale dense residual block is represented as: In the formula, in, and These represent the outputs of the nth and (n-1th)th multi-scale dense residual blocks, respectively. This represents a 5×5 convolutional layer of the i-th layer; This represents a 3×3 convolutional layer of the i-th layer; This indicates a 1×1 convolutional layer of the i-th layer; [*,…,*] represents the concat connection operation.

4. The dynamic contrast-enhanced magnetic resonance imaging classification method as described in claim 1, characterized in that, The perception-discrimination network includes an autoencoder module and a perception-discrimination module; the perception-discrimination module is an improved VGG network; the loss function of the perception-discrimination network includes an adversarial semantic loss function and an adversarial loss function; the adversarial semantic loss function is expressed as: in, This indicates an attempt to counteract semantic loss; This represents the j-th convolutional layer of the perception and discrimination module; Indicates the image to be enhanced; Represents the target image. This represents the output of a multi-task classification network based on the image to be enhanced. , and Indicates the spatial dimension of the image to be enhanced; The adversarial loss function is expressed as: in, Indicates resistance to loss; This represents the output of the perception and discrimination network.

5. The dynamic contrast-enhanced magnetic resonance imaging classification method as described in claim 1, characterized in that, The step of constructing an image classification and prediction model based on the dynamic contrast-enhanced magnetic resonance imaging dataset of the breast includes: Each dynamic enhanced magnetic resonance image of the breast is stored in tensor form, and geometric objects of different dimensions are obtained through spatial geometric analysis of algebraic hypersurface neurons; the geometric objects include three-dimensional spatial vectors, two-dimensional spatial vectors, two-dimensional spatiotemporal vectors and one-dimensional time signal vectors; The geometric objects of different dimensions are input into the tumor location extraction module for high-dimensional image scaling and edge pixel detection to obtain tumor location features; The geometric objects of different dimensions are input into the spatial structure extraction module to perform weakly supervised clustering of spatial images of the same dimension, so as to obtain superpixel features and supervoxel features of multidimensional spatial images. The geometric objects of different dimensions are input into the spatial signal extraction module to perform cross-dimensional spatial signal weakly supervised clustering to obtain superpixel features and supervoxel features of the image signal. Based on the tumor location features, multidimensional spatial image superpixel features, multidimensional spatial image supervoxel features, image signal superpixel features, and image signal supervoxel features, the number of hidden layers and the number of hidden layer neurons in the tower-type multiscale dense residual network are determined, and multi-task association features are obtained by extracting association features through the tower-type multiscale dense residual network. The multi-task associated features are input into the feature distillation module for feature fusion and redundancy removal to obtain tumor heteroproton region features. The tumor heterogeneous region features and the preset target image are input into the autoencoder module for image feature encoding to obtain the tumor heterogeneous region encoded features; The tumor heteroprotic region encoding features are input into the perception and discrimination module, and adversarial learning is performed based on the loss function of the perception and discrimination network to obtain enhanced image features; The enhanced image features and the tumor heterogeneous region features are input into the multi-task classifier, and jointly trained according to the loss function of the multi-task classification network to obtain the image classification prediction model.

6. The dynamic contrast-enhanced magnetic resonance imaging classification method as described in claim 5, characterized in that, The steps for determining the number of hidden layers and the number of hidden layer neurons in the tower-type multi-scale dense residual network include: Principal component analysis is performed on the tumor location features, multidimensional spatial image superpixel features, multidimensional spatial image supervoxel features, image signal superpixel features, and image signal supervoxel features to obtain principal component features. Based on the number of principal component features, the number of hidden layers of the tower-type multiscale dense residual network is determined. The number of hidden layer neurons in the tower-type multi-scale dense residual network is determined by performing cluster analysis or principal component singular value calculation on the principal component features.

7. A dynamic contrast-enhanced magnetic resonance imaging classification system, characterized in that, The system includes: The dataset construction module is used to acquire and preprocess dynamic contrast-enhanced magnetic resonance images of the breast at different disease stages to generate a dynamic contrast-enhanced magnetic resonance image dataset of the breast; the dynamic contrast-enhanced magnetic resonance images of the breast include contrast-enhanced magnetic resonance images of the entire breast and bilateral axillae. The model building module is used to construct an image classification prediction model based on the dynamic enhanced magnetic resonance imaging dataset of the breast. The image classification prediction model includes a multi-task classification network and a perception discrimination network connected in sequence. The multi-task classification network includes a tumor location extraction module, a spatial structure extraction module, a spatial signal extraction module, a tower-style multi-scale dense residual network, a feature distillation module, and a multi-task classifier connected in sequence. The image classification module is used to input the acquired dynamic contrast-enhanced magnetic resonance images of the breast to be classified into the image classification prediction model for multi-task classification to obtain the image classification result; the image classification result includes the location, spatial structure and spatial signal of the tumor.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images