Mammary gland dynamic enhancement MRI intelligent analysis system and method thereof

Abstract

The invention relates to the technical field of medical information systems, in particular to a breast dynamic enhancement MRI intelligent analysis system and method, and the system comprises a data collection module which is used for obtaining breast dynamic enhancement magnetic resonance imaging (DCE)-MRI data; the image preprocessing module is used for carrying out standardization processing on the DCE-MRI data; the segmentation module is used for segmenting the preprocessed image data into a format suitable for model input; the model reasoning module is used for performing deep learning analysis and focus diagnosis on the segmented image; the report generation module is used for generating a diagnosis report according to the model reasoning result; and the data storage module is used for storing data of each stage and a final diagnosis result, so that the analysis capability of the mammary gland DCE-MRI is comprehensively improved. The accuracy and the efficiency of diagnosis are remarkably improved, particularly, the performance is excellent in the aspect of tiny focus detection, and the clinical credibility is improved by enhancing the interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information systems, and more specifically, to a breast dynamic enhanced MRI intelligent analysis system and method thereof. Background Art

[0002] In recent years, the incidence of breast cancer has continued to rise and has become one of the major diseases threatening women's health. Early diagnosis and precise treatment are crucial to improving the survival rate of breast cancer patients. Among many diagnostic methods, dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has gradually become an important means of breast cancer diagnosis due to its high sensitivity and radiation-free advantages. However, the large amount of multidimensional data generated by DCE-MRI has brought great interpretation pressure to clinicians, and also provided opportunities and challenges for the development of computer-aided diagnosis (CAD) systems.

[0003] Traditional DCE-MRI analysis methods mainly rely on manually extracted features, such as morphological features and time-intensity curve (TIC) features. Although these methods have improved diagnostic efficiency to a certain extent, they still have many limitations. First, manually designed features often cannot fully capture the complex spatiotemporal information in DCE-MRI data, resulting in insufficient recognition of certain special types of lesions. Secondly, these methods are sensitive to image quality and acquisition parameters, and have limited generalization capabilities when faced with data acquired from different devices and different scanning schemes. In addition, traditional methods are inefficient in processing large-scale data and are difficult to meet the growing clinical needs.

[0004] In recent years, with the development of deep learning technology, some methods based on convolutional neural networks (CNNs) have been introduced into breast DCE-MRI analysis. These methods have improved the automation and accuracy of feature extraction to a certain extent. However, there are still some problems with existing deep learning methods: First, most methods process DCE-MRI sequences as independent 2D or 3D images, ignoring the temporal correlation between sequences and failing to fully utilize dynamic enhancement information. Second, existing methods usually use a fixed network structure and lack the ability to adaptively extract features of different scales, resulting in insufficient detection of tiny lesions. Furthermore, these methods are often black-box and lack interpretability, making it difficult to gain the trust of clinicians and be widely used.

[0005] In view of the shortcomings of existing technologies, there is an urgent need for an intelligent analysis system that can fully utilize DCE-MRI spatiotemporal information and is highly adaptable and interpretable. Summary of the invention

[0006] The present invention is proposed to meet this demand. The breast dynamic enhanced MRI intelligent analysis system and method proposed in the present invention are intended to solve the following key technical problems: how to effectively extract and utilize the spatiotemporal information in the DCE-MRI sequence; how to improve the system's ability to identify lesions of different sizes and morphologies; how to enhance the interpretability of the model to improve clinical credibility; and how to improve the diagnostic efficiency and accuracy of the system.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] Breast dynamic enhanced MRI intelligent analysis system, including:

[0009] Data acquisition module, used to obtain breast dynamic enhanced magnetic resonance imaging DCE-MRI data;

[0010] An image preprocessing module, used for performing standardization processing on the DCE-MRI data;

[0011] The segmentation module is used to segment the preprocessed image data into a format suitable for model input;

[0012] Model inference module, used for deep learning analysis and lesion diagnosis of segmented images;

[0013] A report generation module is used to generate a diagnostic report based on the model reasoning results;

[0014] And a data storage module is used to store data at each stage and the final diagnosis results.

[0015] Preferably, the model reasoning module includes a dynamic spatiotemporal spectral graph attention network DSTSGAN, which performs the following steps in sequence: dynamic spatiotemporal graph construction, spectral graph convolution feature extraction, topological attention mechanism, Lie group feature transformation and Bessel kernel dynamic enhancement analysis.

[0016] Preferably, the step of constructing the dynamic space-time graph comprises:

[0017] Converting MRI sequences into dynamic graph structures in: A graph representing time t, is the node set, ε t is the edge set, A t ∈R N×N is the adjacency matrix; calculate the adjacency matrix A t , and its calculation formula is

[0018]

[0019] Among them: A t(i,j) is the element of the adjacency matrix at position (i,j), x i and x j are the spatial coordinates of pixels i and j, I t (i) and I t (j) is the intensity value of pixel i and at time t, σ s is the scale parameter of spatial similarity, σ i is the scale parameter of intensity similarity.

[0020] Preferably, the spectrogram convolution feature extraction step comprises:

[0021] Perform convolution operation on the dynamic graph to extract spatiotemporal features. The calculation formula is:

[0022]

[0023] in, is the feature matrix of the lth layer, is the normalized graph Laplacian matrix, is the graph Laplacian matrix, D t is the degree matrix, I is the identity matrix, T k (x) is a Chebyshev polynomial, is a learnable parameter, σ is the activation function, and K is the highest order of the polynomial.

[0024] Preferably, the topological attention mechanism step includes:

[0025] Calculating topological features using persistent homology theory Among them: TF t is the topological feature at time t, PH represents continuous homology calculation, and d is the maximum homology dimension; the attention weight is calculated, and the calculation formula is:

[0026]

[0027] Where: α ij is the attention weight of node i to node j, h i is the feature of node i, W is the weight matrix, a is the attention vector, || represents the concatenation operation, is the neighborhood set of node i, and LeakyReLU is the leaky rectified linear unit activation function.

[0028] Preferably, the Lie group feature transformation step comprises:

[0029] The special orthogonal group SO(3) is used to represent the rotation transformation, and its calculation formula is:

[0030]

[0031] Where: R is the rotation matrix, ω∈R 3 is the rotation vector, θ=||ω|| is the rotation angle, ω ∧ is the antisymmetric matrix of ω, and I is the 3×3 identity matrix;

[0032] Perform feature transformation, and the calculation formula is:

[0033]

[0034] Among them: F t is the transformed feature, is the feature processed by the attention mechanism, t∈R 3 is the translation vector.

[0035] Preferably, the Bessel kernel dynamic enhancement analysis step comprises:

[0036] Define the Bessel kernel function:

[0037]

[0038] Where: K v (x) is the Bessel kernel function, v is the order, and x>0 is the independent variable;

[0039] Extract dynamic enhancement features, and the calculation formula is:

[0040] E t =K ν (||Ft-F0||)⊙(F t -F0)

[0041] Where: E t is the dynamic enhancement feature at time t, F t is the feature at time t, F0 is the baseline feature (before contrast agent injection), and ⊙ represents element-by-element multiplication; the final feature fusion is calculated as follows:

[0042]

[0043] Among them: Y is the final fusion feature, MLP is a multi-layer perceptron, Concat represents the concatenation operation in the time dimension, and T is the length of the time series.

[0044] The intelligent analysis method of breast dynamic enhanced MRI based on the system includes the following steps:

[0045] Acquire breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) data;

[0046] Performing standardization processing on the DCE-MRI data;

[0047] Split the preprocessed image data into a format suitable for model input;

[0048] Use the dynamic spatiotemporal spectral attention network DSTSGAN to perform deep learning analysis and lesion diagnosis on the segmented images;

[0049] Generate a diagnostic report based on the model reasoning results;

[0050] Store data from each stage and the final diagnosis results.

[0051] Preferably, the step of using the dynamic spatiotemporal spectrogram attention network DSTSGAN for deep learning analysis and lesion diagnosis further includes:

[0052] Perform dynamic spatiotemporal graph construction to convert MRI sequences into dynamic graph structures;

[0053] Then, the spectral convolution feature extraction is performed, and the dynamic graph is convolved to extract the spatiotemporal features;

[0054] Then, the topological attention mechanism is implemented, and the topological features are calculated and the attention weights are calculated using the continuous homology theory;

[0055] Then, the Lie group feature transformation is performed, using a special orthogonal group to represent the rotation transformation and perform the feature transformation;

[0056] Finally, Bessel kernel dynamic enhancement analysis is performed, and the dynamic enhancement features are extracted using the Bessel kernel function and the final feature fusion is performed.

[0057] As a preference, the method further comprises the following steps:

[0058] The DSTSGAN is trained, and the training process includes:

[0059] Prepare a training dataset containing labeled breast DCE-MRI image sequences;

[0060] Initialize the network parameters of DSTSGAN;

[0061] Perform forward propagation and calculate the loss between the model output and the true label;

[0062] Perform back propagation and update network parameters;

[0063] Repeat the above steps until the model converges or reaches the preset number of training rounds;

[0064] The model performance was evaluated on the validation set, and the best model was selected for subsequent lesion diagnosis.

[0065] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0066] The present invention has achieved significant technological breakthroughs in many aspects through innovative system architecture design and algorithm optimization, bringing a series of beneficial effects. From a macro perspective, the overall system architecture proposed by the present invention fully considers the characteristics of DCE-MRI data and the needs of clinical diagnosis. Through the organic combination of modules such as data acquisition, image preprocessing, segmentation, model reasoning, report generation and data storage, a complete intelligent diagnosis process is constructed. This comprehensive system design not only improves the accuracy and efficiency of diagnosis, but also greatly enhances the availability and scalability of the system.

[0067] At the algorithm level, the core innovation of the present invention is the proposal of a dynamic spatiotemporal spectrogram attention network (DSTSGAN). This network achieves deep mining of DCE-MRI data through innovative steps such as dynamic spatiotemporal graph construction, spectral convolution feature extraction, topological attention mechanism, Lie group feature transformation, and Bessel kernel dynamic enhancement analysis. This design cleverly solves the problems of insufficient utilization of spatiotemporal information and incomplete feature extraction in the prior art. In particular, DSTSGAN can adaptively capture features of different scales, significantly improving the ability to detect tiny lesions.

[0068] From the perspective of module collaboration, the various modules of the present invention form a close connection and complementary relationship. For example, the image preprocessing module provides high-quality input data for subsequent segmentation and model reasoning, and the results of model reasoning are directly used for report generation, forming a closed-loop workflow. This kind of collaboration between modules not only improves the overall efficiency of the system, but also enhances the reliability of the results.

[0069] In terms of resolving technical contradictions, the present invention cleverly balances model complexity and computational efficiency. By adopting innovative technologies such as Lie group feature transformation, the system greatly reduces computational complexity while maintaining high accuracy, achieving rapid diagnosis. This solves the contradiction between accuracy and efficiency in traditional methods.

[0070] Another significant advantage of the present invention is its interpretability. By introducing the topological attention mechanism and visualization technology, the system can intuitively display the basis for diagnostic decisions, which greatly enhances the clinician's trust in the system and is conducive to the wide application of the system.

[0071] From the perspective of complementary and superimposed effects, the various innovative points of the present invention have formed a good synergistic effect. For example, the combination of spectral convolution feature extraction and topological attention mechanism not only improves the accuracy of feature extraction, but also enhances the model's ability to identify complex lesions. Bessel kernel dynamic enhancement analysis further optimizes the extraction of time series features, forming a complementary effect with the previous steps.

[0072] In summary, the breast dynamic enhanced MRI intelligent analysis system and method of the present invention comprehensively improve the analysis capability of breast DCE-MRI through innovative design and multiple technological breakthroughs. It not only significantly improves the accuracy and efficiency of diagnosis, especially in the detection of small lesions, but also improves clinical credibility by enhancing interpretability. These advantages make this system likely to play an important role in the early diagnosis and precision treatment of breast cancer, and make important contributions to improving patient survival rate and quality of life. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flowchart of the overall system of the present invention.

[0074] Figure 2 It is a logic block diagram of the data acquisition module of the present invention.

[0075] Figure 3 It is a logic block diagram of the image preprocessing module of the present invention.

[0076] Figure 4 It is a logic block diagram of the segmentation module of the present invention.

[0077] Figure 5 It is a logical block diagram of the model reasoning module of the present invention. DETAILED DESCRIPTION

[0078] The scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. The embodiments in the present invention and all other embodiments obtained by ordinary persons in the art without making creative work are within the scope of protection of the present invention.

[0079] like Figure 1-5 As shown, the present invention provides a breast dynamic enhanced MRI intelligent analysis system, which includes a data acquisition module 1, an image preprocessing module 2, a segmentation module 3, a model reasoning module 4, a report generation module 5 and a data storage module 6. These modules work together to realize the full process automation from the acquisition of raw MRI data to the generation of the final diagnosis report.

[0080] Specifically, the data acquisition module 1 is responsible for acquiring breast dynamic enhanced magnetic resonance imaging (DCE-MRI) data. In practical applications, the module can be directly connected to the MRI device to receive scan data in real time. Preferably, in one embodiment of the present invention, the data acquisition module 1 can also include a data quality control unit 1a, which is used to monitor the quality of the collected data in real time to ensure the accuracy of subsequent analysis.

[0081] Next, the image preprocessing module 2 performs standardization on the collected DCE-MRI data. This step is crucial to improving the accuracy of subsequent analysis. In a preferred embodiment of the present invention, the image preprocessing module 2 includes an image denoising unit 2a, an intensity standardization unit 2b and a spatial registration unit 2c. Among them, the image denoising unit 2a can use a wavelet transform or a deep learning based denoising algorithm to effectively remove noise from the MRI image. The intensity standardization unit 2b can use methods such as histogram matching to normalize the image intensity collected by different patients and different devices to the same range. The spatial registration unit 2c ensures that the images collected by the same patient at different time points are spatially aligned, which is particularly important for subsequent dynamic enhancement analysis.

[0082] After preprocessing, the segmentation module 3 divides the image data into a format suitable for model input. Preferably, in one embodiment of the present invention, the segmentation module 3 adopts a three-dimensional U-Net network structure, which can effectively segment the breast tissue area and exclude interference from background and other non-related tissues. It is worth noting that the output of the segmentation module 3 directly affects the accuracy of subsequent analysis. Therefore, in practical applications, the present invention recommends setting the segmentation threshold to 0.5. This threshold shows the best balance in multiple experiments, which can effectively segment the area of ​​interest without excessively including non-related tissues.

[0083] The model inference module 4 is the core of the system, which is used to perform deep learning analysis and lesion diagnosis on the segmented images. The report generation module 5 is used to generate a diagnosis report based on the model inference results, including MRI images, model diagnosis results, and doctor's judgment results; and the data storage module 6 is used to store the data of each stage and the final diagnosis results.

[0084] In one embodiment of the present invention, the model reasoning module 4 includes an innovative dynamic spatiotemporal spectrogram attention network (DSTSGAN). DSTSGAN achieves efficient analysis of DCE-MRI data through five key steps: dynamic spatiotemporal graph construction, spectrogram convolution feature extraction, topological attention mechanism, Lie group feature transformation and Bessel kernel dynamic enhancement analysis. This design fully considers the spatiotemporal characteristics of DCE-MRI data and can effectively capture the dynamic enhancement pattern of lesions.

[0085] In one embodiment of the present invention, the dynamic spatiotemporal graph construction step converts the MRI sequence into a dynamic graph structure. here, A graph representing time t, is a node set representing each pixel in the image, ε t is an edge set, indicating the connection relationship between pixels, A t ∈R N×Nis the adjacency matrix, describing the similarity between nodes. t The calculation of is done using the Gaussian kernel function:

[0086]

[0087] Among them, A t (i,j) is the element of the adjacency matrix at position (i,j), x i and x j are the spatial coordinates of pixels i and j, I t (i) and I t (j) is the intensity value of pixels i and j at time t, σ s is the scale parameter of spatial similarity, σ i is the scale parameter of intensity similarity. In practical applications, the present invention finds that σ s =5 and σ i =0.1 This set of parameter values ​​can achieve good results in most cases because this set of parameters takes into account both spatial proximity and the similarity of pixel intensities.

[0088] Then, the spectral convolution feature extraction step performs convolution operation on the dynamic graph to extract spatiotemporal features. Its core formula is:

[0089]

[0090] here, is the feature matrix of the lth layer, is the normalized graph Laplacian matrix, is the graph Laplacian matrix, D t is the degree matrix, I is the identity matrix, T k (x) is a Chebyshev polynomial, is a learnable parameter, σ is the activation function, and K is the highest order of the polynomial.

[0091] In a preferred embodiment of the present invention, the present invention selects K=3, because lower-order Chebyshev polynomials can effectively capture local graph structure information while maintaining computational efficiency.

[0092] In one embodiment of the present invention, the topological attention mechanism step uses the continuous homology theory to calculate the topological features. Here, TF t is the topological feature at time t, PH represents continuous homology calculation, and d is the maximum homology dimension. In the application scenario of breast DCE-MRI, the present invention usually chooses d=2, because this is enough to capture the main topological features (such as connected components and holes) in the two-dimensional image. Then, the attention weight is calculated:

[0093]

[0094] Among them, α ij is the attention weight of node i to node j, h i is the feature of node i, W is the weight matrix, a is the attention vector, || represents the concatenation operation, is the neighborhood set of node i, and LeakyReLU is the activation function of the rectified linear unit with leakage. This design can effectively capture the local and global spatiotemporal dependencies in DCE-MRI sequences, and is particularly suitable for analyzing complex patterns in the dynamic enhancement process.

[0095] Through this innovative dynamic spatiotemporal spectrogram attention network, the present invention can more accurately analyze breast DCE-MRI data and improve the accuracy of lesion detection and classification. In particular, the system shows significant advantages when processing complex dynamic enhancement patterns.

[0096] In practical applications, the system of the present invention can effectively assist radiologists in early diagnosis of breast cancer. For example, by analyzing the shape and speed of the dynamic enhancement curve, the system can distinguish between benign and malignant lesions. Typically, malignant lesions often show an enhancement pattern of rapid rise followed by rapid outflow, while benign lesions usually show a pattern of continuous enhancement or slow enhancement. The system can automatically identify these patterns and provide doctors with objective quantitative analysis results.

[0097] In addition, another significant advantage of this system is its interpretability. By visualizing the attention weights and topological features, doctors can intuitively understand the basis on which the system makes diagnostic decisions, which greatly improves the credibility and clinical acceptance of the system.

[0098] In summary, the breast dynamic enhanced MRI intelligent analysis system and method provided by the present invention provide a powerful tool for early diagnosis and precise treatment of breast cancer through innovative algorithm design and system architecture. It can not only improve the accuracy and efficiency of diagnosis, but also provide more objective basis for clinical decision-making, and is expected to play an important role in future breast cancer screening and diagnosis.

[0099] In one embodiment of the present invention, the Lie group feature transformation step of the present invention is an important innovation in the model reasoning module 4. This step uses the special orthogonal group SO(3) to represent the rotation transformation, and its calculation formula is:

[0100]

[0101] In this formula, R represents the rotation matrix, ω∈R 3 is the rotation vector, θ=||ω|| represents the rotation angle, ω ∧is the antisymmetric matrix of ω, and I is the 3×3 unit matrix. The advantage of this representation method is that it can effectively capture the spatial features in DCE-MRI images while maintaining rotation invariance. In practical applications, the present invention finds that limiting the rotation angle θ to the range of [0,π] usually achieves the best results. This is because too large a rotation angle may cause excessive deformation of the feature, while a smaller rotation angle can enhance the adaptability of the model to different viewing angles while maintaining the basic form of the feature.

[0102] In one embodiment of the present invention, the present invention performs feature transformation, and its calculation formula is:

[0103]

[0104] here, is the transformed feature, is the feature processed by the attention mechanism, t∈R 3 is the translation vector. This transformation can not only enhance the expressiveness of features, but also improve the robustness of the model to changes in lesion location and morphology.

[0105] It is worth noting that in the analysis of breast DCE-MRI, the present invention usually limits the range of the translation vector t to a smaller interval, such as [-5,5] pixels. This is because excessive translation may cause the shift of the region of interest, while moderate translation helps the model learn richer local features.

[0106] The Bessel kernel dynamic enhancement analysis step of the present invention is another innovation of the whole system. This step first defines the Bessel kernel function:

[0107]

[0108] Among them, K v (x) is the Bessel kernel function, v represents the order, and x>0 is the independent variable. In the analysis of breast DCE-MRI, the present invention usually selects v=1 or v=2. This is because the first-order and second-order Bessel functions can well simulate the enhancement curve shape commonly seen in DCE-MRI.

[0109] Then, the present invention extracts the dynamic enhancement feature, and its calculation formula is:

[0110] E t =K v (||F t -F0||)⊙(F t -F0)

[0111] Here, E t represents the dynamic enhancement feature at time t, F tis the feature at time t, F0 is the baseline (before contrast agent injection) feature, and ⊙ represents element-by-element multiplication. This design can effectively capture the dynamic enhancement process after contrast agent injection, especially in distinguishing benign and malignant lesions.

[0112] Finally, the present invention performs final feature fusion, and its calculation formula is:

[0113]

[0114] in, is the final fusion feature, MLP stands for multi-layer perceptron, Concat represents the concatenation operation in the time dimension, and T is the length of the time series. In practical applications, the present invention usually selects T=5 or T=6, which corresponds to the DCE-MRI acquisition scheme commonly used in clinical practice.

[0115] In one embodiment of the present invention, the present invention also provides a method for intelligent analysis of breast dynamic enhanced MRI based on the above system. This method includes the steps of data acquisition, preprocessing, segmentation, deep learning analysis, report generation and data storage. Each step corresponds closely to the corresponding system module to form a complete workflow. First, obtain breast dynamic enhanced magnetic resonance imaging DCE-MRI data; standardize the DCE-MRI data; divide the preprocessed image data into a format suitable for model input; use the dynamic spatiotemporal spectral attention network DSTSGAN to perform deep learning analysis and lesion diagnosis on the segmented image; generate a diagnosis report based on the model reasoning results; store the data of each stage and the final diagnosis results.

[0116] In one embodiment of the present invention, the step of using the dynamic spatiotemporal spectral attention network DSTSGAN for deep learning analysis and lesion diagnosis further includes: performing dynamic spatiotemporal graph construction to convert the MRI sequence into a dynamic graph structure; then performing spectral graph convolution feature extraction, performing a convolution operation on the dynamic graph, and extracting spatiotemporal features; then implementing a topological attention mechanism, using the continuous homology theory to calculate the topological features and calculate the attention weights; then performing Lie group feature transformation, using a special orthogonal group to represent the rotation transformation and perform feature transformation; finally, performing Bessel kernel dynamic enhancement analysis, using the Bessel kernel function to extract dynamic enhancement features and perform final feature fusion.

[0117] In one embodiment of the present invention, the following steps are also included: training the DSTSGAN, the training process including: preparing a training data set, the data set containing a labeled breast DCE-MRI image sequence; initializing the network parameters of the DSTSGAN; performing forward propagation to calculate the loss between the model output and the true label; performing back propagation to update the network parameters; repeating the above steps until the model converges or reaches a preset number of training rounds; evaluating the model performance on the validation set, and selecting the best model for subsequent lesion diagnosis.

[0118] In particular, the deep learning analysis step of this method uses a dynamic spatiotemporal spectral graph attention network (DSTSGAN). This network is processed in the order of dynamic spatiotemporal graph construction, spectral graph convolution feature extraction, topological attention mechanism, Lie group feature transformation and Bessel kernel dynamic enhancement analysis. This design can not only effectively extract spatiotemporal features in DCE-MRI data, but also capture complex dynamic enhancement patterns.

[0119] In practical applications, the present invention has found that this method is particularly suitable for cases with complex enhancement patterns. For example, for some early invasive ductal carcinomas, traditional methods may be difficult to distinguish from benign lesions, while this method can more accurately identify such lesions by analyzing fine spatiotemporal features and dynamic enhancement patterns.

[0120] In one embodiment of the present invention, the present invention also includes a training process for DSTSGAN. This process first prepares a training data set, which usually contains hundreds or even thousands of labeled breast DCE-MRI image sequences. In practice, the present invention finds that using a data set of at least 1,000 cases can achieve good training results.

[0121] During the training process, the present invention adopts a dynamic learning rate adjustment strategy. The initial learning rate is set to 0.001, and then reduced by 10% every 50 epochs. This strategy can converge quickly in the early stage of training, while fine-tuning the model parameters in the later stage. In addition, the present invention also introduces an early stopping mechanism, which stops training when the performance on the validation set does not improve for 10 consecutive epochs. This helps prevent the model from overfitting.

[0122] It is worth noting that in the model evaluation stage, the present invention not only focuses on the overall classification accuracy, but also pays special attention to the performance of the model in dealing with complex cases. For example, the present invention calculates the sensitivity and specificity of the model in identifying small lesions (diameter <1cm). These indicators can more comprehensively reflect the clinical application value of the model.

[0123] In general, the breast dynamic enhanced MRI intelligent analysis system and method provided by the present invention provides a powerful tool for early diagnosis of breast cancer through innovative algorithm design and system architecture. It can not only improve the accuracy and efficiency of diagnosis, but also provide more objective basis for clinical decision-making. Especially when dealing with complex cases, this system shows significant advantages and is expected to play an important role in future breast cancer screening and accurate diagnosis.

[0124] In order to fully verify the superiority of the breast dynamic enhanced MRI intelligent analysis system of the present invention, the present invention has carried out a series of rigorous experiments. Here, the present invention will introduce a representative embodiment in detail and compare it with the current clinical commonly used computer-aided diagnosis (CAD) system.

[0125] Example 1: In this example, the present invention uses a large-scale breast DCE-MRI dataset from multiple medical centers. The dataset contains 5,000 cases, including 2,000 benign lesions and 3,000 malignant lesions. Each case contains a complete DCE-MRI sequence, usually consisting of 3D images at 6 time points: 1 baseline image before contrast agent injection and 5 dynamic enhancement images after injection.

[0126] The present invention randomly divides the data set into a training set, a validation set, and a test set in a ratio of 7:1:2. In the model training stage, the present invention adopts the dynamic learning rate adjustment strategy and early stopping mechanism described above. It is particularly worth mentioning that the present invention introduces data enhancement technology in the training process, including random rotation, flipping, and brightness adjustment, which greatly improves the generalization ability of the model.

[0127] Comparative Example 1: As a control, the present invention selected a commercial CAD system currently widely used in clinical practice. This system is based on traditional machine learning methods and mainly relies on manually designed features, such as morphological features and dynamic enhancement curve features. This system represents a common method in current clinical practice.

[0128] The system of the present invention (Example 1) and the control CAD system (Comparative Example 1) were evaluated on the same test set. The evaluation indicators include accuracy, sensitivity, specificity, AUC (area under the curve), and the detection rate for small lesions (diameter <1 cm). In addition, the present invention also evaluates the average diagnosis time of the two systems to reflect their efficiency in actual clinical applications.

[0129] The testing standards and methods are as follows:

[0130] 1. Accuracy, sensitivity and specificity: calculated based on the confusion matrix.

[0131] 2. AUC: obtained by drawing the ROC curve and calculating the area under the curve.

[0132] 3. Detection rate of small lesions: The sensitivity is calculated separately for lesions with a diameter less than 1 cm.

[0133] 4. Average diagnosis time: the average time required from inputting raw DCE-MRI data to outputting the final diagnosis result.

[0134] The following are the detailed test results:

[0135] Evaluation Metrics Example 1 Comparative Example 1 Accuracy 92.5％ 85.3％ Sensitivity 94.2％ 87.1％ Specificity 90.1％ 82.8％ AUC 0.96 0.89 Detection rate of micro-lesions 88.7％ 72.4％ Average diagnosis time 45 seconds 120 seconds

[0136] It can be clearly seen from the above results that the breast dynamic enhanced MRI intelligent analysis system of the present invention is significantly superior to the traditional CAD system in all indicators. It is particularly noteworthy that:

[0137] 1. Overall diagnostic performance: The system of the present invention has significantly improved accuracy, sensitivity and specificity, with an AUC of 0.96, which indicates that the system has excellent classification ability and can effectively distinguish between benign and malignant lesions.

[0138] 2. Detection of tiny lesions: The system of the present invention performs particularly well in detecting tiny lesions with a diameter of less than 1 cm, with a detection rate 16.3 percentage points higher than that of the traditional CAD system. This is of great significance for the early diagnosis of breast cancer, because early tiny lesions are often the easiest to be missed.

[0139] 3. Diagnostic efficiency: The system of the present invention shortens the average diagnosis time from 120 seconds to 45 seconds, thus improving the diagnostic efficiency. In clinical practice, this means that radiologists can complete the diagnosis of more cases in a shorter time, thus improving work efficiency.

[0140] These excellent results are mainly due to the innovative algorithms used in the system of the present invention, especially the dynamic spatiotemporal spectral attention network (DSTSGAN). This network can effectively capture the spatiotemporal features and dynamic enhancement patterns in DCE-MRI sequences, thereby improving the accuracy of diagnosis. In addition, innovations such as Lie group feature transformation and Bessel kernel dynamic enhancement analysis also provide the system with stronger feature extraction and expression capabilities, enabling it to better handle complex cases.

[0141] It is worth mentioning that the excellent performance of the system of the present invention in detecting tiny lesions fully demonstrates its potential in the diagnosis of early breast cancer. This can not only improve the detection rate of early breast cancer, but also reduce the unnecessary biopsy rate, thereby reducing the physical and mental burden of patients.

[0142] In general, these experimental results strongly demonstrate the advancedness and practical value of the system of the present invention in the field of intelligent analysis of breast dynamic enhanced MRI. It not only significantly improves the diagnostic accuracy, but also improves the diagnostic efficiency. It is expected to play an important role in clinical practice and provide strong support for the early diagnosis and precise treatment of breast cancer.

[0143] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto; any person familiar with the art who, within the scope disclosed by the present invention, makes substitutions or changes based on the scheme and improved concepts of the present invention shall be covered within the protection scope of the present invention.

Claims

1. Breast dynamic enhanced MRI intelligent analysis system, characterized by: include: Data acquisition module, used to obtain breast dynamic enhanced magnetic resonance imaging DCE-MRI data; An image preprocessing module, used for performing standardization processing on the DCE-MRI data; The segmentation module is used to segment the preprocessed image data into a format suitable for model input; Model inference module, used for deep learning analysis and lesion diagnosis of segmented images; A report generation module is used to generate a diagnostic report based on the model reasoning results; And a data storage module is used to store data at each stage and the final diagnosis results.

2. The breast dynamic enhancement MRI intelligent analysis system according to claim 1, characterized in that: The model reasoning module includes a dynamic spatiotemporal spectral attention network DSTSGAN, which performs the following steps in sequence: dynamic spatiotemporal graph construction, spectral graph convolution feature extraction, topological attention mechanism, Lie group feature transformation and Bessel kernel dynamic enhancement analysis.

3. The breast dynamic enhancement MRI intelligent analysis system according to claim 2, characterized in that: The dynamic space-time graph construction step includes: Converting MRI sequences into dynamic graph structures in: A graph representing time t, is the node set, ε t is the edge set, A t ∈R N×N is the adjacency matrix; calculate the adjacency matrix A t , and its calculation formula is Among them: A t (i,j) is the element of the adjacency matrix at position (i,j), x i and x j are the spatial coordinates of pixels i and j, I t (i) and I t (j) is the intensity value of pixel i and at time t, σ s is the scale parameter of spatial similarity, σ i is the scale parameter of intensity similarity.

4. The breast dynamic enhanced MRI intelligent analysis system according to claim 3, characterized in that: The spectrogram convolution feature extraction step comprises: Perform convolution operation on the dynamic graph to extract spatiotemporal features. The calculation formula is: in, is the feature matrix of layer I, is the normalized graph Laplacian matrix, is the graph Laplacian matrix, D t is the degree matrix, I is the identity matrix, T k (x) is a Chebyshev polynomial, is a learnable parameter, σ is the activation function, and K is the highest order of the polynomial.

5. The breast dynamic enhancement MRI intelligent analysis system according to claim 4, characterized in that: The topological attention mechanism steps include: Calculating topological features using persistent homology theory Among them: TF t is the topological feature at time t, PH represents continuous homology calculation, and d is the maximum homology dimension; the attention weight is calculated, and the calculation formula is: Where: α ij is the attention weight of node i to node j, h i is the feature of node i, W is the weight matrix, a is the attention vector, || represents the concatenation operation, is the neighborhood set of node i, and LeakyReLU is the leaky rectified linear unit activation function.

6. The breast dynamic enhanced MRI intelligent analysis system according to claim 5, characterized in that: The Lie group feature transformation step comprises: The special orthogonal group SO(3) is used to represent the rotation transformation, and its calculation formula is: Where: R is the rotation matrix, ω∈R 3 is the rotation vector, θ=∥ω∥ is the rotation angle, ω ∧ is the antisymmetric matrix of ω, and I is the 3×3 identity matrix; Perform feature transformation, and the calculation formula is: Among them: F t is the transformed feature, is the feature processed by the attention mechanism, t∈R 3 is the translation vector.

7. The breast dynamic enhancement MRI intelligent analysis system according to claim 6, characterized in that: The Bessel kernel dynamic enhancement analysis step comprises: Define the Bessel kernel function: Where: K v (x) is the Bessel kernel function, v is the order, and x>0 is the independent variable; Extract dynamic enhancement features, and the calculation formula is: E t =K v (||F t -F0||)⊙(F t -F0) Where: E t is the dynamic enhancement feature at time t, F t is the feature at time t, F0 is the baseline feature (before contrast agent injection), and ⊙ represents element-by-element multiplication; the final feature fusion is calculated as follows: Among them: Y is the final fusion feature, MLP is a multi-layer perceptron, Concat represents the concatenation operation in the time dimension, and T is the length of the time series.

8. A method for intelligent analysis of breast dynamic enhanced MRI based on the system according to any one of claims 1 to 7, characterized in that: The following steps are involved: Acquire breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) data; Performing standardization processing on the DCE-MRI data; Split the preprocessed image data into a format suitable for model input; Use the dynamic spatiotemporal spectral attention network DSTSGAN to perform deep learning analysis and lesion diagnosis on the segmented images; Generate a diagnostic report based on the model reasoning results; Store data from each stage and the final diagnosis results.

9. The intelligent analysis method for breast dynamic enhanced MRI according to claim 8, characterized in that: The step of using the dynamic spatiotemporal spectrogram attention network DSTSGAN for deep learning analysis and lesion diagnosis further includes: Perform dynamic spatiotemporal graph construction to convert MRI sequences into dynamic graph structures; Then, the spectral convolution feature extraction is performed, and the dynamic graph is convolved to extract the spatiotemporal features; Then, the topological attention mechanism is implemented, and the topological features are calculated and the attention weights are calculated using the continuous homology theory; Then, the Lie group feature transformation is performed, using a special orthogonal group to represent the rotation transformation and perform the feature transformation; Finally, Bessel kernel dynamic enhancement analysis is performed, and the dynamic enhancement features are extracted using the Bessel kernel function and the final feature fusion is performed.

10. The intelligent analysis method for breast dynamic enhanced MRI according to claim 8, characterized in that: The following steps are also included: The DSTSGAN is trained, and the training process includes: Prepare a training dataset containing labeled breast DCE-MRI image sequences; Initialize the network parameters of DSTSGAN; Perform forward propagation and calculate the loss between the model output and the true label; Perform back propagation and update network parameters; Repeat the above steps until the model converges or reaches the preset number of training rounds; The model performance was evaluated on the validation set, and the best model was selected for subsequent lesion diagnosis.

Images