Reconstruction and pathological localization method for intelligent fast magnetic resonance imaging

Abstract

The application discloses a reconstruction and pathological positioning method of intelligent fast magnetic resonance imaging, and specifically implements the following steps: step 1, constructing a pathological focusing nuclear magnetic resonance image reconstruction model; step 2, constructing a pathological focusing nuclear magnetic resonance image reconstruction network according to the pathological focusing nuclear magnetic resonance image reconstruction model constructed in step 1; step 3, training the pathological focusing nuclear magnetic resonance image reconstruction network; and step 4, applying the trained pathological focusing magnetic resonance image reconstruction network to nuclear magnetic resonance imaging and pathological positioning. The method can not only reduce the dependence on accurate pixel-level labels, but also can locate the pathological part of the image while enhancing the nuclear magnetic resonance image reconstruction effect, and has better reconstruction effect in the part, thereby effectively assisting a clinician in diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of medical magnetic resonance imaging technology, specifically relating to a method for reconstruction and pathological localization in intelligent rapid magnetic resonance imaging. Background Technology

[0002] Magnetic resonance imaging (MRI) is a widely used non-invasive, radiation-free imaging technique. MRI images provide high soft tissue contrast and rich anatomical and functional information for medical diagnosis. Despite its importance, MRI faces two main challenges in the clinical setting. First, the long time required to capture fully sampled images not only causes patient discomfort but also increases the risk of motion artifacts, thus reducing image quality. Second, there is a discrepancy between the scanning procedure and the specific requirements of subsequent clinical evaluation. Often, the general imaging procedure, which emphasizes improving the average image quality, does not guarantee accurate reconstruction of smaller lesions, which is crucial for accurate diagnosis and treatment planning.

[0003] To shorten scan time, improve imaging efficiency, and reduce patient discomfort, researchers and engineers have developed a series of accelerated imaging techniques. These can be broadly categorized into traditional methods and deep learning-based methods. In previous years, parallel imaging and sparse sampling methods based on compressed sensing theory were proposed and applied to undersampled k-spaces to accelerate imaging; however, the reconstruction quality, especially in terms of detail, often failed to meet practical clinical needs. In recent years, CNN-based deep learning networks have demonstrated strong dominance in computer vision and achieved commendable results in various interdisciplinary fields. Researchers have applied them to MRI image reconstruction with promising results, showcasing their immense potential. Generally speaking, deep learning methods for MRI image reconstruction can be divided into: direct inversion methods that directly map undersampled k-spaces to the reconstructed image; and model expansion methods that combine optimization algorithms to solve the MRI image reconstruction model.

[0004] However, these methods focus on improving the reconstruction quality of MRI images, neglecting downstream tasks related to MRI image reconstruction. Research on downstream tasks primarily revolves around the detection of pathological components in MRI images—specifically, how to quickly locate pathological areas to assist physicians in diagnosis. While some recent studies have begun evaluating downstream tasks related to MRI image reconstruction, aiming to focus on tasks such as segmentation, these studies also largely require accurate labeling for these tasks. For example, in segmentation, accurate pixel-level labels are crucial, and labeling these labels consumes significant human and material resources, posing a considerable challenge for clinical applications. Summary of the Invention

[0005] The purpose of this invention is to provide a method for intelligent and rapid magnetic resonance imaging (MRI) reconstruction and pathological localization, which can reduce the reliance on accurate pixel-level labels, enhance the reconstruction effect of MRI images, locate the pathological part of the image, and achieve better reconstruction results in that part.

[0006] The technical solution adopted in this invention is an intelligent rapid magnetic resonance imaging reconstruction and pathological localization method, which is implemented according to the following steps:

[0007] Step 1: Construct a pathologically focused MRI image reconstruction model;

[0008] Step 2: Construct a pathology-focused MRI image reconstruction network based on the pathology-focused MRI image reconstruction model built in Step 1.

[0009] Step 3: Train the pathology-focused MRI image reconstruction network;

[0010] Step 4: Apply the trained pathology-focused magnetic resonance image reconstruction network to perform magnetic resonance imaging and pathological localization.

[0011] The invention is further characterized in that,

[0012] The specific process of step 1 is as follows:

[0013] For dataset Where n is the number of datasets, The set of natural numbers, where For full-sample MRI images, The resolution size of the full-sample MRI image. Fully sampled MRI images The corresponding pathological labels, given undersampled k-space data ;

[0014] The pathological focus MRI image reconstruction model is as follows:

[0015] (1)

[0016] In equation (1), Indicates the forward operator, ,in, Representing the Sensitivity diagram of a signal receiving coil. Pre-estimation is performed using the Espirit algorithm. and These represent the one-dimensional Fourier transform matrices in the horizontal and vertical directions, respectively. and These represent the number of sampling lines in the horizontal and vertical directions, respectively. and It is a regularization parameter used to balance data items and multiple regularization terms;

[0017] ,

[0018] in, Represents a set of learnable position parameters;

[0019] In formula (1), (2)

[0020] In equation (2), This is a denoising network based on CNN. The parameters that can be learned in this denoising network are the parameters of the denoising network in the th... The Taylor expansion formula for the step is:

[0021] (3)

[0022] In the formula, for about Jacobian matrix, For the first The value of the nth iteration. This represents the change in each iteration. ;

[0023] Then we get information about Approximate estimate:

[0024] (4)

[0025] When disturbance When the value is sufficiently small, the second term on the right approaches 0. Based on this, we can obtain:

[0026] (5)

[0027] Will Recorded as ;

[0028] In formula (1), (6)

[0029] In equation (6), To introduce a classification network, the aim is to accurately locate the position of lesions; To refine the denoising network, further refinement of the denoising at the lesion site is performed; refinement of the denoising network Network structure and denoising network Completely identical, among which and These are the refinement and denoising networks. and classification networks Learnable parameters;

[0030] For the nth iteration, This can be processed using a univariate Taylor expansion, resulting in:

[0031] (7)

[0032] in, To generate The Jacobian matrix corresponding to the classification network, so if there is ,So It can be expressed as

[0033] (8)

[0034] When disturbance Enough hours Approaching 0, based on formula (8), the following approximate estimate can be obtained:

[0035] (9)

[0036] in, For classification networks The output class activation graph;

[0037] Then the iterative format of the pathologically focused MRI image reconstruction model can be directly written out, that is:

[0038] (10).

[0039] The specific process of step 2 is as follows:

[0040] Formula (10) is expanded into a multi-stage cascaded network as the pathology-focused MRI image reconstruction network, which includes a denoising network. The lesion guidance module LF and the data consistency module DC.

[0041] Noise reduction module It is a seven-layer CNN network. The first six layers are composed of convolutional layers, batch normalization (BN) layers, and ReLU activation function layers, respectively, and the seventh layer is composed of convolutional layers and batch normalization (BN) layers, respectively.

[0042] The denoising network is represented as:

[0043] (11).

[0044] A classification network is introduced in the lesion guidance module LF. The predicted CAM map provides accurate location information of the lesion, and then the predicted CAM map is compared with the input... The connection is used to achieve the CAM guidance, and the connected values ​​are then input into the refinement and denoising network. This achieves the purpose of enhancing the lesion;

[0045] The lesion guidance module LF is represented as:

[0046] (12)

[0047] Classification Network This is a pre-trained improved VGG19 network whose parameters are frozen during reconstruction and not updated. The improved VGG19 network consists of a feature extraction module and a classification module. The feature extraction module is divided into five levels. The first and second levels each consist of two convolutional layers, a ReLU activation function layer, and a max pooling layer, respectively. The third, fourth, and fifth levels each consist of four convolutional layers, a ReLU activation function layer, and a max pooling layer, respectively. The classification module consists of a global average pooling layer, two fully connected layers, and a softmax activation layer.

[0048] Classification Network The processing flow is as follows: denoising network The output of the algorithm is used as the input to the classification network. The feature map extracted by the feature extraction module of the classification network is passed through a global average pooling layer and two fully connected layers to obtain a classification score. The classification score is then passed through a softmax activation layer to obtain the classification result. The classification result is compared with the classification label. We perform training using cross-loss entropy, then obtain a set of training parameters which are frozen in subsequent training. The gradient is calculated using the classification score against the feature map, resulting in:

[0049]

[0050] In the formula, is the number of feature maps, and h and w are the length and width of the feature maps, respectively; The values ​​in the classification scores represent outlier components. For the k-th component of the feature map, The gradient of the classification score with respect to the value in the i-th row and j-th column of the feature map; for the above value, after passing through a ReLU activation function layer and interpolation upsampling to Finally, the attention map (CAM) is obtained.

[0051] The data consistency module DC is represented as:

[0052] (13)

[0053] The Data Consistency Module (DC) obtains the reconstructed image by minimizing the difference between the solution to the subproblem and the input. Input is spatial data Denoising images and enhanced images of lesions Then, the analytical solution to this problem can be derived as follows:

[0054] (14)

[0055] In the formula, Forward operator, It is the identity matrix, because If it is not necessarily invertible, then the conjugate gradient method is used to iteratively solve this subproblem.

[0056] The specific process of step 3 is as follows:

[0057] Step 3.1, Construct the training dataset

[0058] The training dataset consists of multiple data sets, each consisting of undersampled k-space data, fully sampled data reconstructed MRI images, and image-level labels for MRI images showing lesions.

[0059] Step 3.2: Train the network parameters for pathological focused magnetic resonance images using the training dataset from Step 3.1;

[0060] The specific process is as follows:

[0061] The MRI images reconstructed using full-sample data and the image-level labels of lesions in the MRI images were used as inputs to a classification network. The classification network was trained, and then the parameters of the trained classification network were frozen and integrated into the pathology-focused MRI image reconstruction network as a classification module to provide lesion location information.

[0062] Undersampled k-space data is used as the initial input to the pathology-focused magnetic resonance image reconstruction network. Then, backpropagation is used to calculate the gradient of the loss function with respect to the parameters of the pathology-focused magnetic resonance image reconstruction network. Finally, the gradient is used as the input of the Adam algorithm to update the parameters of the pathology-focused magnetic resonance image reconstruction network to obtain the optimal parameters of the pathology-focused magnetic resonance image reconstruction network.

[0063] During training, the loss function of the pathology-focused magnetic resonance image reconstruction network is:

[0064] (15)

[0065] In the formula, For the sample size, For full-sample MRI images, Magnetic resonance imaging reconstruction networks focused on pathology The final reconstruction result of the phase, To assist in overall image denoising, For classification networks The CAM diagram output at stage T. Represents cross-entropy loss, A label indicating the presence of a lesion. A label indicating the absence of disease. This represents the classification probability map output by the classification network at stage T. , , , All of these are hyperparameters.

[0066] The beneficial effects of this invention are as follows: The intelligent and rapid magnetic resonance imaging reconstruction and pathological localization method of this invention only requires weakly supervised training using image-level annotations, combined with attention mechanisms, to study downstream tasks of MRI reconstructed images. This reduces the dependence on accurate pixel-level labels and is more in line with actual conditions. The method of this invention can locate the pathological part of the image while enhancing the reconstruction effect of MRI images, and has a better reconstruction effect in this part, effectively assisting clinicians in diagnosis. Attached Figure Description

[0067] Figure 1 This is a flowchart of the intelligent rapid magnetic resonance imaging reconstruction and pathological localization method of the present invention;

[0068] Figure 2 This is a diagram of the network structure for pathological-focused MRI image reconstruction in the method of this invention.

[0069] Figure 3 This is a structural diagram of the pathological focusing and denoising network module in the method of the present invention;

[0070] Figure 4 This is a flowchart of the classification network and its generation of attention maps in the method of the present invention;

[0071] Figure 5 This is a visualization of reconstructed images and pathological localization results under high sampling rates in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0073] Example 1

[0074] The present invention provides a method for intelligent rapid magnetic resonance imaging reconstruction and pathological localization, such as... Figure 1 As shown, please follow these steps:

[0075] Step 1: Construct a pathologically focused MRI image reconstruction model;

[0076] The specific process is as follows:

[0077] For dataset Where n is the number of datasets, The set of natural numbers, where This is a fully sampled MRI image (reconstructed image). The resolution size of a fully sampled MRI image. Fully sampled MRI images The corresponding pathological labels, given undersampled k-space data The ultimate goal is to b was restored to a high-fidelity, fully sampled MRI image. ;

[0078] The pathological focus MRI image reconstruction model is as follows:

[0079] (1)

[0080] In equation (1), Indicates the forward operator, ,in, Representing the Sensitivity diagram of a signal receiving coil. Pre-estimation is performed using the Espirit algorithm. and These represent the one-dimensional Fourier transform matrices in the horizontal and vertical directions, respectively. and These represent the number of sampling lines in the horizontal and vertical directions, respectively. and It is a regularization parameter used to balance data items and multiple regularization terms;

[0081] ,

[0082] in, Represents a set of learnable position parameters;

[0083] This invention only considers a one-dimensional imaging scene, i.e., a fixed scene. ,make Only the position parameters of the phase encoding direction, i.e., the vertical direction, need to be learned. At this time, the acceleration factor is This can be further understood as the number of k-space lines sampled in the phase encoding direction in undersampling and full-sampling scenarios; acceleration factor The degree to which data acquisition is subsampled is usually expressed as the compression ratio of the acquired data relative to the complete data, where the sampling mask... The sampled k-space data must satisfy the acceleration factor constraint, i.e. ;

[0084] In formula (1), (2)

[0085] In equation (2), This is a denoising network based on CNN. The parameters that can be learned in this denoising network are the parameters of the denoising network in the th... The Taylor expansion formula for the step is:

[0086] (3)

[0087] In the formula, for about Jacobian matrix, For the first The value of the nth iteration. This represents the change in each iteration. ;

[0088] Then we get information about Approximate estimate:

[0089] (4)

[0090] When disturbance When the value is sufficiently small, the second term on the right approaches 0. Based on this, we can obtain:

[0091] (5)

[0092] Will Recorded as ;

[0093] In formula (1), (6)

[0094] In equation (6), To introduce a classification network, the aim is to accurately locate the position of lesions; To refine the denoising network, further refinement of the denoising at the lesion site is performed; refinement of the denoising network Network structure and denoising network Completely identical, among which and These are the refinement and denoising networks. and classification networks Learnable parameters;

[0095] For the nth iteration, This can be processed using a univariate Taylor expansion, resulting in:

[0096] (7)

[0097] in, To generate The Jacobian matrix corresponding to the classification network, so if there is ,So This can be expressed as:

[0098] (8)

[0099] When disturbance Enough hours Approaching 0, based on formula (8), the following approximate estimate can be obtained:

[0100] (9)

[0101] in, For classification networks The output class activation graph;

[0102] Then the iterative format of the pathologically focused MRI image reconstruction model can be directly written out, that is:

[0103] (10)

[0104] Step 2: Construct a pathology-focused MRI image reconstruction network based on the pathology-focused MRI image reconstruction model built in Step 1.

[0105] The specific process is as follows:

[0106] Formula (10) is expanded into a multi-stage cascaded network as the pathology-focused MRI image reconstruction network, which includes a denoising network. Lesion guidance module LF and data consistency module DC;

[0107] (1) Noise reduction module For a seven-layer CNN network, such as Figure 3 As shown, the first six layers consist of convolutional layers, batch normalization (BN) layers, and ReLU activation function layers, respectively, while the seventh layer consists of convolutional layers and BN layers. Each convolutional layer has a 3×3 kernel size, a stride of 1, and 1 padding. The number of kernels from the first to the seventh layer is 64, 64, 128, 128, 64, 64, and 2, respectively. The denoising network reconstructs the MRI image from the previous stage. As input, the output is an image with noise and aliasing artifacts removed;

[0108] The denoising network is represented as:

[0109] (11)

[0110] (2) A classification network is introduced into the lesion guidance module LF. This is combined with XGradCAM technology to provide accurate lesion location information through predicted CAM maps, and then the predicted CAM maps are compared with the input... The connection is used to achieve the CAM guidance, and the connected values ​​are then input into the refinement and denoising network. During the reconstruction process, greater attention is paid to the characteristics of the lesion area to achieve lesion enhancement, provide higher quality reconstructed images, and better assist clinical diagnosis. This includes refining the denoising network. Structure and denoising network completely consistent;

[0111] like Figure 4 As shown, classification network This is a pre-trained improved VGG19 network whose parameters are frozen during reconstruction and not updated. The improved VGG19 network consists of a feature extraction module and a classification module. The feature extraction module is divided into five levels. The first and second levels each consist of two convolutional layers, a ReLU activation function layer, and a max pooling layer, respectively. The third, fourth, and fifth levels each consist of four convolutional layers, a ReLU activation function layer, and a max pooling layer, respectively. The kernel size of each convolutional layer is 3×3, the stride is 1, and the number of convolutional kernels from the first to the fifth level are 64, 128, 256, 512, and 512, respectively, with 1 unit of padding. The max pooling size of each max pooling layer is 2×2. The classification module consists of a global average pooling layer, two fully connected layers, and a softmax activation layer. The number of neurons in the two fully connected layers are 128 and 2, respectively.

[0112] Classification network The processing flow is as follows: denoising network The output of the algorithm is used as the input to the classification network. The feature map extracted by the feature extraction module of the classification network is passed through a global average pooling layer and two fully connected layers to obtain a classification score. The classification score is then passed through a softmax activation layer to obtain the classification result. The classification result is compared with the classification label. We perform training using cross-loss entropy, then obtain a set of training parameters which are frozen in subsequent training. The gradient is calculated using the classification score against the feature map, resulting in:

[0113]

[0114] In the formula, is the number of feature maps, and h and w are the length and width of the feature maps, respectively; The values ​​in the classification scores represent outlier components. For the k-th component of the feature map, The gradient of the classification score with respect to the value in the i-th row and j-th column of the feature map; for the above value, after passing through a ReLU activation function layer and interpolation upsampling to Finally, the attention map CAM is obtained;

[0115] The lesion guidance module LF is represented as:

[0116] (12)

[0117] Compared with existing joint reconstruction and downstream segmentation or classification methods, our classification network can provide more accurate lesion location information and does not require pixel-level labels, which greatly saves the resources wasted on manual annotation.

[0118] (3) The data consistency module DC is represented as:

[0119] (13)

[0120] The data consistency module DC works by minimizing the solution to the subproblem and the input (i.e., k-space data). Denoising images and enhanced images of lesions The difference between the two is used to obtain the reconstructed image. Then, the analytical solution to this problem can be derived as follows:

[0121] (14)

[0122] In the formula, Forward operator, It is the identity matrix, because If it is not necessarily invertible, then the conjugate gradient method is used to iteratively solve this subproblem, specifically:

[0123]

[0124]

[0125] in, The number of algorithm iterations. The initial solution vector, For the initialized residual vector, The direction vector is initialized and initially defined as follows: , The step size parameter is used to determine the direction vector. The step size for moving upwards is then used to update the solution vector. residual vector and update the direction vector The iterative value is obtained after multiple iterations. That is the value we are looking for. ;

[0126] Step 3: Train the pathology-focused MRI image reconstruction network;

[0127] The specific process is as follows:

[0128] Step 3.1: Construct the training and testing datasets.

[0129] The training dataset consists of multiple data sets, each consisting of undersampled k-space data, fully sampled data reconstructed MRI images, and image-level labels for MRI images showing lesions.

[0130] The test dataset contains undersampled k-space data and pixel-level labels of specific bounding boxes of lesions in MRI images, as well as fully sampled data reconstructed MRI images to evaluate the effectiveness of our network in reconstructing and locating pathological sites.

[0131] Step 3.2: Train the network parameters for pathological focused magnetic resonance images using the training dataset from Step 3.1;

[0132] The specific process is as follows:

[0133] Using fully sampled data to reconstruct MRI images and image-level labels indicating whether lesions are present in the MRI images as input, a cross-loss entropy function is used to classify the network. Fine-tuning is performed until the CAM image and classification results show good performance. Then, the parameters of the trained classification network are frozen and integrated into the pathology-focused magnetic resonance image reconstruction network as the classification module of the pathology-focused magnetic resonance image reconstruction network to provide lesion location information. This ensures that the reconstruction performance and CAM image have good performance without training failure.

[0134] Undersampled k-space data is used as the initial input to the pathology-focused magnetic resonance image reconstruction network. Then, backpropagation is used to calculate the gradient of the loss function with respect to the parameters of the pathology-focused magnetic resonance image reconstruction network. Finally, the gradient is used as the input of the Adam algorithm to update the parameters of the pathology-focused magnetic resonance image reconstruction network to obtain the optimal parameters of the pathology-focused magnetic resonance image reconstruction network. The Adam algorithm is an optimization algorithm for network training.

[0135] During training, the loss function of the pathology-focused magnetic resonance image reconstruction network is:

[0136] (15)

[0137] In the formula, For the sample size, For full-sample MRI images, Magnetic resonance imaging reconstruction networks focused on pathology The final reconstruction result of the phase, To assist in overall image denoising, For classification networks The CAM diagram output at stage T. Represents cross-entropy loss, A label indicating the presence of a lesion. A label indicating the absence of disease. This represents the classification probability map output by the classification network at stage T. , , , All are hyperparameters;

[0138] Step 4: Apply the trained pathology-focused magnetic resonance image reconstruction network to perform MRI and pathology localization. Input the test set into the trained pathology-focused magnetic resonance image reconstruction network to perform MRI and pathology localization.

[0139] Through the above training process, this invention can determine the optimal parameters of the pathology-focused magnetic resonance image reconstruction network. Based on the trained network, multiple k-space undersampled data are input, and the network outputs the reconstructed MRI image and its pathological localization attention map. Because the training process of the pathology-focused MRI image reconstruction network parameters makes the output image as close as possible to the standard MRI image, the trained network can still obtain high-quality reconstructed images under different or even low data sampling rates. It is worth noting that we process the original complex data into two channels, a real part and a complex part, using a denoising network. It accepts two channels of data; for the denoising network Since the input is connected to the CAM image, it accepts three-channel data.

[0140] Example 2

[0141] In numerical experiments, this invention uses two contrast datasets from the public MRI dataset fastMRI: PDFS (fat suppression) and PD (non-fat suppression), as well as the SKM-TEA dataset. The k-space sampling mode is selected from fixed 1D Cartesian sampling and learnable 1D Cartesian sampling, with a sampling rate of 1 / 8. During training and testing, this invention uses datasets of size... (fastMRI dataset) and (SKM_TEA dataset) A pathology-focused magnetic resonance imaging (MRI) image reconstruction network was trained and its reconstruction accuracy was tested using 60% of the middle 2D slices with anatomical details selected from the 3D volume. To objectively evaluate different methods, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) were used to measure the average reconstruction accuracy on the test set. In addition to evaluating the overall image reconstruction quality, this project also specifically evaluated the reconstruction quality of specific pathological sites. Specifically, the PSNR and SSIM values ​​of the pathological sites within the bounding boxes were measured, denoted as LF-PSNR and LF-SSIM, respectively.

[0142] Example 3

[0143] The difference from Example 2 is that the k-space sampling mode selects a fixed 1-dimensional Cartesian sampling with a sampling rate of 1 / 4.

[0144] Table 1: Comparison results of different methods on the test data set under 1D Cartesian sampling at a 1 / 4 sampling rate.

[0145]

[0146] Table 2: Comparison results of different methods on the test data set under 1D Cartesian sampling at a sampling rate of 1 / 8.

[0147]

[0148] Table 3: Evaluation results of learnable sampling on the test data set under 1D Cartesian sampling at a sampling rate of 1 / 8.

[0149]

[0150] As shown in Tables 1, 2, and 3, the pathology-focused magnetic resonance imaging reconstruction network (LF_MRI) of this invention is compared with traditional reconstruction methods, purely data-driven reconstruction methods, and model-driven reconstruction methods at different sampling rates. Traditional reconstruction methods include GRAPPA; purely data-driven reconstruction methods include UNet and SwinMR; and model-driven reconstruction methods include ISTA-Net and MoDL. The pathology-focused magnetic resonance imaging reconstruction network designed in this invention achieves the best reconstruction accuracy under different sampling modes and sampling rates (where LS-LF-MRI is the result of adding learnable operators to LF-MRI). Figure 5 The visualization results of the reconstructed image under high sampling rate show that the reconstructed MRI image of the present invention has clear structural details and no obvious artifacts; and can be well located in the pathological part (i.e. within the marked box).

Claims

1. A method for intelligent rapid magnetic resonance imaging reconstruction and pathological localization, characterized in that, The specific steps are as follows: Step 1: Construct a pathologically focused MRI image reconstruction model; The specific process of step 1 is as follows: For the dataset {(x,Y)}n, Where n is the number of datasets. Let C be the set of natural numbers, where x∈C N*N Let x be a fully sampled MRI image, N be the resolution of the fully sampled MRI image, and Y∈{0,1} be the pathological label corresponding to the fully sampled MRI image x. Given undersampled k-space data b∈C N*N ; The pathological focus MRI image reconstruction model is as follows: In equation (1), A θ Indicates the forward operator, Among them, s r The sensitivity diagram of the r-th signal receiving coil, s r Pre-estimation is performed using the Espirit algorithm. and N represents the one-dimensional Fourier transform matrix in the horizontal and vertical directions, respectively. q and N p These represent the number of sampling lines in the horizontal and vertical directions, respectively. μ1 and μ2 are regularization parameters used to balance data items and multiple regularization items. in, Represents a set of learnable position parameters; In formula (1), N(x)=xD w (x) (2) In equation (2), D w (x) is a CNN-based denoising network, w is the learnable parameter in the denoising network, and the Taylor expansion formula of the denoising network at step n is: In the formula, D w (x) about x (n-1) Jacobian matrix, x (n-1) Let x be the value of the (n-1)th iteration, and Δx be the change in value during each iteration. (n) =x (n-1) +Δx; Then we obtain an estimate of N(x): When the perturbation Δx is sufficiently small, the second term on the right-hand side approaches 0. Based on this, we obtain: D w (x (n-1) ) is denoted as z (n) ; In formula (1), In equation (6), Net c (·) represents a classification network introduced to achieve the goal of accurately locating lesion information; G u (Net c (·),·) are used to refine the denoising network, further refining the denoising at the lesion site; the refined denoising network G u Network structure and denoising network D w They are exactly the same, where u and c are the refinement denoising network G. u and classification network Net c Learnable parameters; For the nth iteration, Net c (x n Using a univariate Taylor expansion to process +Δx), we have: in, To generate the Jacobian matrix corresponding to the classification network of CAM, if x = x n +Δx, then G u (Net c (·),·) is expressed as When the disturbance Δx is small enough Approaching 0, based on formula (8), the following estimate is obtained: Among them, CAM (n) For classification network Net c (x (n-1) The output class activation graph; The iterative format of the pathologically focused MRI image reconstruction model can then be directly written out, i.e.: Step 2: Construct a pathology-focused MRI image reconstruction network based on the pathology-focused MRI image reconstruction model built in Step 1. Step 3: Train the pathology-focused MRI image reconstruction network; Step 4: Apply the trained pathology-focused magnetic resonance image reconstruction network to perform magnetic resonance imaging and pathological localization.

2. The intelligent rapid magnetic resonance imaging reconstruction and pathological localization method according to claim 1, characterized in that, The specific process of step 2 is as follows: Formula (10) is expanded into a multi-stage cascaded network as the pathology-focused MRI image reconstruction network. The pathology-focused MRI image reconstruction network includes a denoising network D. w The lesion guidance module LF and the data consistency module DC.

3. The intelligent rapid magnetic resonance imaging reconstruction and pathological localization method according to claim 2, characterized in that, Noise reduction module D w It is a seven-layer CNN network. The first six layers are composed of convolutional layers, batch normalization (BN) layers, and ReLU activation function layers, respectively, and the seventh layer is composed of convolutional layers and batch normalization (BN) layers, respectively. The denoising network is represented as: z (n) =D w (x (n-1) ) (11)。 4. The intelligent rapid magnetic resonance imaging reconstruction and pathological localization method according to claim 2, characterized in that, A classification network Net is introduced into the lesion guidance module LF. c (·) The predicted CAM map provides accurate location information of the lesion, and then the predicted CAM map is compared with the input x-axis. (n-1) The connection is used to achieve the CAM guidance function, and the connected values ​​are input into the thinning and denoising network G. u This achieves the purpose of enhancing the lesion; The lesion guidance module LF is represented as: r (n) =LF(x (n-1) ) (12) Classification Networks (Net) c (·) represents a pre-trained improved VGG19 network whose parameters are frozen during reconstruction and do not participate in updates. The improved VGG19 network consists of a feature extraction module and a classification module. The feature extraction module is divided into five levels. The first and second levels each consist of two convolutional layers, a ReLU activation function layer, and a max pooling layer, respectively. The third, fourth, and fifth levels each consist of four convolutional layers, a ReLU activation function layer, and a max pooling layer, respectively. The classification module consists of a global average pooling layer, two fully connected layers, and a softmax activation layer. Classification Networks (Net) c The processing flow of (·) is as follows: Denoising network D w The output of the algorithm is used as the input to the classification network. The feature map extracted by the feature extraction module of the classification network is passed through a global average pooling layer and two fully connected layers to obtain a classification score. The classification score is then passed through a softmax activation layer to obtain the classification result. The classification result and the classification label Y are used for cross-loss entropy training to obtain a set of training parameters, which are then frozen in subsequent training. The gradient of the feature map is calculated from the classification score: In the formula, s is the number of feature maps, and h and w are the length and width of the feature maps, respectively; C 1 A represents the value in the classification score that indicates an outlier component. k For the k-th component of the feature map, The gradient of the classification score with respect to the value in the i-th row and j-th column of the feature map; for the above value, after passing through the ReLU activation function layer and interpolation upsampling to N*N, the attention map CAM is finally obtained.

5. The intelligent rapid magnetic resonance imaging reconstruction and pathological localization method according to claim 2, characterized in that, The data consistency module DC is represented as: x (n) =DC(b,z (n) ,r (n) ) (13) The data consistency module DC obtains the reconstructed image x by minimizing the difference between the solution to the subproblem and the input. (n) The input is spatial data b and the denoised image z. (n) and enhanced images of lesions r (n) Then, the analytical solution to this problem can be derived as follows: In the formula, A θ Let I be the forward operator, and I be the identity matrix. If it is not necessarily invertible, then the conjugate gradient method is used to iteratively solve this subproblem.

6. The intelligent rapid magnetic resonance imaging reconstruction and pathological localization method according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3.1, Construct the training dataset The training dataset consists of multiple data sets, each consisting of undersampled k-space data, fully sampled data reconstructed MRI images, and image-level labels for MRI images showing lesions. Step 3.2: Train the network parameters for pathological focused magnetic resonance images using the training dataset from Step 3.1; The specific process is as follows: The MRI images reconstructed using full-sample data and the image-level labels of lesions in the MRI images were used as inputs to a classification network. The classification network was trained, and then the parameters of the trained classification network were frozen and integrated into the pathology-focused MRI image reconstruction network as a classification module to provide lesion location information. Undersampled k-space data is used as the initial input to the pathology-focused magnetic resonance image reconstruction network. Then, backpropagation is used to calculate the gradient of the loss function with respect to the parameters of the pathology-focused magnetic resonance image reconstruction network. Finally, the gradient is used as the input of the Adam algorithm to update the parameters of the pathology-focused magnetic resonance image reconstruction network to obtain the optimal parameters of the pathology-focused magnetic resonance image reconstruction network. During training, the loss function of the pathology-focused magnetic resonance image reconstruction network is: In the formula, K is the number of samples, x i For full-sample MRI images, The final reconstruction result of the pathology-focused magnetic resonance imaging reconstruction network in the T phase. To assist in overall image denoising, For classification network Net c The CAM diagram output at stage T, L cls Represents cross-entropy loss, A label indicating the presence of a lesion. A label indicating the absence of disease. The diagram represents the classification probability output by the classification network at stage T, where α, β, γ1, and γ2 are all hyperparameters.

Images