Intelligent rapid magnetic resonance imaging reconstruction and pathological localization method

Abstract

The invention discloses an intelligent rapid magnetic resonance imaging reconstruction and pathological positioning method. The method is specifically implemented according to the following steps: step 1, constructing a pathological focused nuclear magnetic resonance image reconstruction model; 2, constructing a pathological focused nuclear magnetic resonance image reconstruction network according to the pathological focused nuclear magnetic resonance image reconstruction model constructed in the step 1; step 3, training a pathological focused nuclear magnetic resonance image reconstruction network; and step 4, using the trained pathologically focused magnetic resonance image reconstruction network to carry out magnetic resonance imaging and pathology positioning. According to the method, dependence on an accurate pixel-level label can be reduced, the pathological part of the image can be positioned while the reconstruction effect of the nuclear magnetic resonance image is enhanced, a better reconstruction effect is achieved on the part, and a clinician is effectively assisted in diagnosis.

Description

Technical Field

[0001] The invention belongs to the technical field of medical nuclear magnetic resonance imaging, and in particular relates to a method for reconstruction and pathological localization of intelligent rapid magnetic resonance imaging. Background Art

[0002] Magnetic resonance imaging is a widely used non-invasive, radiation-free imaging technique. MRI images provide high soft tissue contrast as well as rich anatomical and functional information for medical diagnosis. Despite its significance, MRI still faces two major 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, thereby reducing image quality. Second, there is a deviation between the scanning process and the specific requirements of subsequent clinical evaluation. Typically, the general imaging process that emphasizes improving the average quality of the image does not guarantee the precise reconstruction of smaller lesions, which is crucial for accurate diagnosis and treatment planning.

[0003] In order to shorten the scanning time, improve the imaging efficiency, and reduce the discomfort of patients, researchers and engineers have developed a series of accelerated imaging technologies. Generally speaking, they can be divided into traditional methods and methods based on deep learning. In the past few years, parallel imaging and sparse sampling methods based on compressed sensing theory were proposed, and they were applied to undersampled k-space to accelerate imaging, but the reconstruction quality, especially its details, are difficult to meet the actual clinical needs. In recent years, CNN-based deep learning networks have shown a strong dominance in the field of computer vision, and their applications in various cross-fields have achieved recognized results. Researchers have applied it to the field of nuclear magnetic resonance image reconstruction and achieved good results, showing its great potential. Generally speaking, the use of deep learning methods for nuclear magnetic resonance image reconstruction can be divided into: the direct inversion method of using undersampled k-space to directly map to the reconstructed image; the method of model expansion that combines optimization algorithms to solve the nuclear magnetic resonance image reconstruction model.

[0004] However, these methods are all focused on how to improve the reconstruction quality of MRI images, and do not pay attention to the downstream tasks of MRI reconstructed images. The exploration of downstream tasks of MRI reconstructed images mainly revolves around the detection of pathological parts, that is, how to quickly locate the pathological part of MRI reconstructed images to assist doctors in diagnosis. Of course, some studies have recently begun to evaluate the downstream tasks of MRI reconstructed images. The core idea of ​​these studies is to focus on downstream tasks of MRI reconstructed images, such as segmentation tasks. But again, most of them require accurate labels for downstream tasks, such as accurate pixel-level labels in segmentation tasks. The annotation of these labels will consume a lot of manpower and material resources, which will cause great trouble to actual clinical applications. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent and fast magnetic resonance imaging reconstruction and pathology localization method, which can reduce the dependence on accurate pixel-level labels, while enhancing the reconstruction effect of nuclear magnetic resonance images, locating the pathological part of the image and achieving a better reconstruction effect in this part.

[0006] The technical solution adopted by the present invention is a method for intelligent rapid magnetic resonance imaging reconstruction and pathological localization, which is specifically implemented according to the following steps: Step 1, constructing a pathologically focused MRI image reconstruction model; Step 2, constructing a pathology-focused MRI image reconstruction network according to the pathology-focused MRI image reconstruction model constructed in step 1; Step 3, training a pathology-focused MRI image reconstruction network; Step 4: Use the trained pathology-focused magnetic resonance image reconstruction network to perform magnetic resonance imaging and pathology localization. The present invention is also characterized in that: The specific process of step 1 is: For the dataset , where n is the number of data sets, is a set of natural numbers, among which is a fully sampled MRI image, is the resolution of the fully sampled MRI image, Fully sampled MRI images Corresponding pathological labels, given undersampled k-space data ; Then the pathological focused MRI image reconstruction model is: (1) In formula (1), represents the forward operator, ,in, Representative Sensitivity diagram of a signal receiving coil, Estimated in advance by the Espirit algorithm, and Represent the one-dimensional Fourier transform matrix in the horizontal and vertical directions respectively, and are the number of sampling lines in the horizontal and vertical directions, respectively. and is a regularization parameter used to balance the data term and multiple regularization terms; ,

[0007] in, represents a set of learnable position parameters; In formula (1), (2) In formula (2), is a denoising network based on CNN. is the learnable parameter in the denoising network. The Taylor expansion formula of the step is: (3) In the formula, for about Jacobian matrix, For the The iteration value, is the change in each iteration, ; Then get about An approximate estimate of: (4) When the disturbance When it is small enough, the second term of the right-hand side approaches 0. Based on this, we can get: (5) Will Recorded as ; In formula (1), (6) In formula (6), A classification network is introduced to achieve the purpose of accurately locating the location information of the lesion; To refine the denoising network, further refine the lesion site for denoising; refine the denoising network Network structure and denoising network Exactly the same, and Refinement denoising network and classification network Learnable parameters; For the nth iteration, It can be processed using the univariate Taylor expansion, and then: (7) in, To generate The Jacobian matrix corresponding to the classification network, so if there is ,So It can be expressed as (8) When the disturbance Enough hours, Approaching 0, based on formula (8), we can get the following approximate estimate: (9) in, For classification network Output class activation map; Then the iterative format of the pathological focused MRI image reconstruction model can be directly written out, namely: (10).

[0008] The specific process of step 2 is: Formula (10) is expanded into a multi-stage cascade network as the pathology-focused MRI image reconstruction network. The pathology-focused MRI image reconstruction network includes a denoising network , lesion guidance module LF and data consistency module DC.

[0009] Denoising module It is a seven-layer CNN network. The first six layers are composed of convolutional layers, BN layers, and ReLU activation function layers in sequence, and the seventh layer is composed of convolutional layers and BN layers in sequence; The denoising network is expressed as: (11).

[0010] A classification network is introduced into the lesion guidance module LF The predicted CAM map is used to provide accurate location information of the lesion, and then the predicted CAM map is combined with the input Connect to achieve the role of CAM guidance, and input the connected value into the refinement denoising network , thereby achieving the purpose of lesion enhancement; The lesion guidance module LF is expressed as: (12) Classification Network It is a pre-trained improved VGG19 network, whose parameters are frozen during the reconstruction process and do not participate in the update. The improved VGG19 network consists of a feature extraction module and a classification module, where the feature extraction module is divided into five levels. The first and second levels are respectively composed of two convolutional layers, a ReLU activation function layer and a maximum pooling layer, and the third, fourth and fifth levels are respectively composed of four convolutional layers, a ReLU activation function layer and a maximum pooling layer; the classification module is composed of a global average pooling layer, two fully connected layers and a softmax activation layer; Classification Network The processing flow is: denoising network The output of is used as the input of the classification network. The feature map extracted by the feature extraction module of the classification network passes through a global average pooling layer and two fully connected layers to obtain the classification score. The classification score passes through the softmax activation layer to obtain the classification result. The classification result and the classification label Do cross loss entropy training, then get a training parameter and freeze it in subsequent training, and calculate the gradient of the feature map by the classification score, there are:

[0011] In the formula, is the number of feature maps, h and w are the length and width of the feature maps respectively; is the value representing the abnormal component in the classification score, is the kth component of the feature map, is the gradient of the classification score to the value of the i-th row and j-th column of the feature map; for the above value, after the ReLU activation function layer and interpolation upsampling to , and finally get the attention map CAM.

[0012] The data consistency module DC is expressed as: (13) The data consistency module DC obtains the reconstructed image by minimizing the difference between the solution of the subproblem and the input , input is - spatial data , denoised image and lesion enhanced images , then the analytical solution to this problem can be derived as: (14) In the formula, is the forward operator, is the identity matrix, since It is not necessarily reversible, so the conjugate gradient method is used to iteratively solve this subproblem.

[0013] The specific process of step 3 is: Step 3.1: Build a training dataset The training data set consists of multiple data groups, each of which consists of undersampled k-space data, MRI images reconstructed from fully sampled data, and image-level labels of whether the MRI images have lesions or not; Step 3.2, using the training data set of step 3.1 to train the pathology-focused magnetic resonance image reconstruction network parameters; The specific process is: The fully sampled data are used to reconstruct the MRI images and the image-level labels of whether the MRI images have lesions are used as the input of the classification network, and the classification network is trained. Then, the parameters of the trained classification network are frozen and integrated into the pathology-focused MRI image reconstruction network, and the network is used as a classification module of the pathology-focused MRI image reconstruction network to provide lesion location information. The undersampled k-space data is used as the initial input of the pathology-focused magnetic resonance image reconstruction network, and the gradient of the loss function with respect to the parameters of the pathology-focused magnetic resonance image reconstruction network is calculated by back propagation. 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, so as to obtain the optimal parameters of the pathology-focused magnetic resonance image reconstruction network. During the training process, the loss function of the pathology-focused MRI reconstruction network is: (15) In the formula, is the number of samples, is a fully sampled MRI image, Pathology-focused magnetic resonance image reconstruction network The final reconstruction result of the stage, To assist the overall denoising image, For classification network The CAM diagram output at the T stage, represents the cross entropy loss, Labels representing the presence of lesions, The label represents the absence of lesions, Represents the classification probability map output by the classification network at stage T, , , , are all hyperparameters.

[0014] The beneficial effects of the present invention are as follows: the intelligent fast magnetic resonance imaging reconstruction and pathology localization method of the present invention only needs to use image-level annotations for weak supervision training, and combines the attention mechanism to study the downstream tasks of nuclear magnetic resonance reconstruction images, which can reduce the dependence on accurate pixel-level labels and is more in line with actual conditions; the method of the present invention can locate the pathological part of the image while enhancing the nuclear magnetic resonance image reconstruction effect, and has a better reconstruction effect in this part, effectively assisting clinicians in diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the intelligent rapid magnetic resonance imaging reconstruction and pathology localization method of the present invention; Figure 2 A network structure diagram for reconstructing a pathologically focused nuclear magnetic resonance image in the method of the present invention; Figure 3 It is a structural diagram of the pathological focusing and denoising network module in the method of the present invention; Figure 4 A flowchart of a classification network and an attention map generated by the method of the present invention; Figure 5 This is the visualization of the reconstructed image and the pathological localization result at a high sampling rate in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Example 1 The intelligent rapid magnetic resonance imaging reconstruction and pathology localization method of the present invention is as follows: Figure 1 As shown, the specific implementation steps are as follows: Step 1, constructing a pathologically focused MRI image reconstruction model; The specific process is: For the dataset , where n is the number of data sets, is a set of natural numbers, among which is the fully sampled MRI image (reconstructed image), is the resolution of the fully sampled MRI image, Fully sampled MRI images Corresponding pathological labels, given undersampled k-space data , the ultimate goal is to and b restored to a high-fidelity fully sampled MRI image ; Then the pathological focused MRI image reconstruction model is: (1) In formula (1), represents the forward operator, ,in, Representative Sensitivity diagram of a signal receiving coil, Estimated in advance by the Espirit algorithm, and Represent the one-dimensional Fourier transform matrix in the horizontal and vertical directions respectively, and are the number of sampling lines in the horizontal and vertical directions, respectively. and is a regularization parameter used to balance the data term and multiple regularization terms; ,

[0018] in, represents a set of learnable position parameters; The present invention only considers one-dimensional imaging scenes, that is, fixed ,make , we only need to learn the position parameters of the phase encoding direction, i.e. the vertical direction , the acceleration factor is , which can be further understood as the number of k-space lines sampled in the phase encoding direction in undersampling and full sampling scenarios; the acceleration factor The degree of subsampling of data acquisition 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 meet the constraints of the acceleration factor, that is, ; In formula (1), (2) In formula (2), is a denoising network based on CNN. is the learnable parameter in the denoising network. The Taylor expansion formula of the step is: (3) In the formula, for about Jacobian matrix, For the The iteration value, is the change in each iteration, ; Then get about An approximate estimate of: (4) When the disturbance When it is small enough, the second term of the right-hand side approaches 0. Based on this, we can get: (5) Will Recorded as ; In formula (1), (6) In formula (6), A classification network is introduced to achieve the purpose of accurately locating the location information of the lesion; To refine the denoising network, further refine the lesion site for denoising; refine the denoising network Network structure and denoising network Exactly the same, and Refinement denoising network and classification network Learnable parameters; For the nth iteration, It can be processed using the univariate Taylor expansion, and then: (7) in, To generate The Jacobian matrix corresponding to the classification network, so if there is ,So It can be expressed as: (8) When the disturbance Enough hours, Approaching 0, based on formula (8), we can get the following approximate estimate: (9) in, For classification network Output class activation map; Then the iterative format of the pathological focused MRI image reconstruction model can be directly written out, namely: (10) Step 2, constructing a pathology-focused MRI image reconstruction network according to the pathology-focused MRI image reconstruction model constructed in step 1; The specific process is: Formula (10) is expanded into a multi-stage cascade network as the pathology-focused MRI image reconstruction network. The pathology-focused MRI image reconstruction network includes a denoising network , lesion guidance module LF and data consistency module DC; (1) Denoising module is a seven-layer CNN network, such as Figure 3 As shown in the figure, the first six layers are composed of convolutional layers, BN layers (batch normalization layers), and ReLU activation function layers in sequence, and the seventh layer is composed of convolutional layers and BN layers in sequence; the convolution kernel size of each convolution layer is 3×3, the step size is 1, the surrounding padding is 1, and the number of convolution 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, it outputs an image with noise and aliasing artifacts removed; The denoising network is expressed as: (11) (2) Introducing a classification network into the lesion guidance module LF , and combined with XGradCAM technology to provide accurate location information of the lesion through the predicted CAM map, and then use the predicted CAM map and the input Connect to achieve the role of CAM guidance, and input the connected value into the refinement denoising network In the reconstruction process, more attention is paid to the characteristics of the lesion area, so as to achieve the purpose of lesion enhancement, provide higher quality reconstructed images, and better assist clinical diagnosis. Structure and denoising network completely consistent; like Figure 4 As shown, the classification network It is a pre-trained improved VGG19 network, whose parameters are frozen during the reconstruction process and do not participate in the update. The improved VGG19 network consists of a feature extraction module and a classification module, where the feature extraction module is divided into five levels. The first and second levels are respectively composed of two convolutional layers, a ReLU activation function layer and a maximum pooling layer, and the third, fourth and fifth levels are respectively composed of four convolutional layers, a ReLU activation function layer and a maximum pooling layer. The convolution kernel size of each convolutional layer is 3×3, the step size is 1, and the number of convolution kernels from the first to the fifth level is 64, 128, 256, 512, and 512, respectively. The surrounding padding is 1, and the maximum pooling size of each maximum 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 is 128 and 2, respectively. Classification Network The processing flow is: denoising network The output of is used as the input of the classification network. The feature map extracted by the feature extraction module of the classification network passes through a global average pooling layer and two fully connected layers to obtain the classification score. The classification score passes through the softmax activation layer to obtain the classification result. The classification result and the classification label Do cross loss entropy training, then get a training parameter and freeze it in subsequent training, and calculate the gradient of the feature map by the classification score, there are:

[0019] In the formula, is the number of feature maps, h and w are the length and width of the feature maps respectively; is the value representing the abnormal component in the classification score, is the kth component of the feature map, is the gradient of the classification score to the value of the i-th row and j-th column of the feature map; for the above value, after the ReLU activation function layer and interpolation upsampling to , and finally get the attention map CAM; The lesion guidance module LF is expressed as: (12) 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, greatly saving the waste of manual annotation resources; (3) The data consistency module DC is expressed as: (13) The data consistency module DC minimizes the solution of the subproblem and the input (i.e., k-space data , denoised image and lesion enhanced images ) to obtain the reconstructed image , then the analytical solution to this problem can be derived as: (14) In the formula, is the forward operator, is the identity matrix, since It is not necessarily reversible, so the conjugate gradient method is used to iteratively solve this subproblem, specifically:

[0020]

[0021] in, is the number of algorithm iterations, is the initialization solution vector, is the initial residual vector, is the initialization direction vector, initially defined as , is the step size parameter, which is used to determine the direction vector The step size of the upward movement, and then update the solution vector , residual vector And update the direction vector , after multiple iterations, the iterative value is obtained This is the value we require ; Step 3, training a pathology-focused MRI image reconstruction network; The specific process is: Step 3.1: Build training and test datasets The training data set consists of multiple data groups, each of which consists of undersampled k-space data, MRI images reconstructed from fully sampled data, and image-level labels of whether the MRI images have lesions or not; The test dataset contains undersampled k-space data and pixel-level labels of specific marked boxes of MRI image lesions, and contains fully sampled data to reconstruct MRI images for evaluating the effectiveness of our network in pathological reconstruction and localization; Step 3.2, using the training data set of step 3.1 to train the pathology-focused magnetic resonance image reconstruction network parameters; The specific process is: The fully sampled data is used to reconstruct the MRI image and the image-level label of whether the MRI image has lesions is used as input, and the cross-loss entropy function is used to classify the classification network. Fine-tune until the CAM image and classification results have good performance, then stop fine-tuning, then freeze the parameters of the trained classification network and integrate it into the pathology-focused magnetic resonance image reconstruction network, and use it as the classification module of the pathology-focused magnetic resonance image reconstruction network to provide lesion location information, so as to ensure that both the reconstruction performance and the CAM image have good performance without training crash; The undersampled k-space data is used as the initial input of the pathology-focused magnetic resonance image reconstruction network, and the gradient of the loss function with respect to the parameters of the pathology-focused magnetic resonance image reconstruction network is calculated by back propagation. 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, wherein the Adam algorithm is an optimization algorithm for network training; During the training process, the loss function of the pathology-focused MRI reconstruction network is: (15) In the formula, is the number of samples, is a fully sampled MRI image, Pathology-focused magnetic resonance image reconstruction network The final reconstruction result of the stage, To assist the overall denoising image, For classification network The CAM diagram output at the T stage, represents the cross entropy loss, Labels representing the presence of lesions, The label represents the absence of lesions, Represents the classification probability map output by the classification network at stage T, , , , All are hyperparameters; Step 4, applying the trained pathology-focused magnetic resonance image reconstruction network to perform nuclear magnetic resonance imaging and pathology localization, and inputting the test set into the trained pathology-focused magnetic resonance image reconstruction network to perform nuclear magnetic resonance imaging and pathology localization.

[0022] Through the above-mentioned training process, the present invention can determine the optimal parameters of the pathology-focused magnetic resonance image reconstruction network. Based on the trained pathology-focused magnetic resonance image reconstruction network, multiple k-space undersampled data are input, and the output of the pathology-focused magnetic resonance image reconstruction network is the reconstructed nuclear magnetic resonance image and its pathology localization attention map. Because the parameter training process of the pathology-focused magnetic resonance image reconstruction network makes the output image of the pathology-focused magnetic resonance image reconstruction network as close as possible to the standard nuclear magnetic resonance image, the trained pathology-focused magnetic resonance image reconstruction network can still obtain high-quality reconstructed images under different or even low data sampling rates. It is worth noting that we divide the original complex data into the real part and the complex part and transform them into two channels for processing, namely the denoising network Accepts two-channel data; for the denoising network ,Since the input is connected to the CAM diagram, it accepts three-channel data.

[0023] Example 2 In the numerical experiments, the present invention uses PDFS (fat suppression) and PD (non-fat suppression) contrasts in the public MRI dataset fastMRI and the SKM-TEA dataset. The k-space sampling mode selects fixed 1D Cartesian sampling and learnable 1D Cartesian sampling, with a sampling rate of 1 / 8. During training and testing, the present invention starts with a size of (fastMRI dataset) and The middle 60% of continuous 2D slices with anatomical details were selected from the 3D volume of the (SKM_TEA dataset) to train the pathology-focused magnetic resonance image reconstruction network and test the reconstruction accuracy. In order to objectively evaluate different methods, the 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 conducted a specific evaluation of the reconstruction quality of specific pathological sites. That is, the PSNR and SSIM values ​​of the pathological sites inside the marked box were measured, which were recorded as LF-PSNR and LF-SSIM, respectively.

[0024] Example 3 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.

[0025] Table 1: Comparison results of different methods on the data test set under 1D Cartesian sampling 1 / 4 sampling rate

[0026] Table 2: Comparison results of different methods on the data test set under 1D Cartesian sampling 1 / 8 sampling rate

[0027] Table 3: Evaluation results of learnable sampling on the data test set under 1D Cartesian sampling with a sampling rate of 1 / 8

[0028] As shown in Tables 1, 2 and 3, the pathology-focused magnetic resonance image reconstruction network (LF_MRI) of the present invention is compared with traditional reconstruction methods, pure data-driven reconstruction methods and model-driven reconstruction methods at different sampling rates. Traditional reconstruction methods include: GRAPPA; pure data-driven reconstruction methods include: UNet and SwinMR; model-driven reconstruction methods include: ISTA-Net and MoDL. The pathology-focused magnetic resonance image reconstruction network designed by the present invention achieves the best reconstruction accuracy under different sampling modes and sampling rates (where LS-LF-MRI is the result of adding a learnable operator to LF-MRI). Figure 5 This is the visualization result of the reconstructed image at a high sampling rate. It can be seen that the reconstructed MRI image of the present invention has clear structural details and no obvious artifacts; and can be well positioned to the pathological part (i.e., within the marked frame).

Claims

1. Intelligent fast magnetic resonance imaging reconstruction and pathology localization method, characterized in that: Follow the steps below to implement it: Step 1, constructing a pathologically focused MRI image reconstruction model; Step 2, constructing a pathology-focused MRI image reconstruction network according to the pathology-focused MRI image reconstruction model constructed in step 1; Step 3, training a pathology-focused MRI image reconstruction network; Step 4: Use the trained pathology-focused magnetic resonance image reconstruction network to perform magnetic resonance imaging and pathology localization.

2. The intelligent rapid magnetic resonance imaging reconstruction and pathology localization method according to claim 1, characterized in that: The specific process of step 1 is: For the dataset , where n is the number of data sets, is a set of natural numbers, among which is a fully sampled MRI image, is the resolution of the fully sampled MRI image, Fully sampled MRI images Corresponding pathological labels, given undersampled k-space data ; Then the pathological focused MRI image reconstruction model is: （1） In formula (1), represents the forward operator, ,in, Representative Sensitivity diagram of a signal receiving coil, Estimated in advance by the Espirit algorithm, and Represent the one-dimensional Fourier transform matrix in the horizontal and vertical directions respectively, and are the number of sampling lines in the horizontal and vertical directions, and is a regularization parameter used to balance the data term and multiple regularization terms; , in, represents a set of learnable position parameters; In formula (1), (2) In formula (2), is a denoising network based on CNN. is the learnable parameter in the denoising network. The Taylor expansion formula of the step is: （3） In the formula, for about Jacobian matrix, For the The iteration value, is the change in each iteration, ; Then get about An approximate estimate of: （4） When the disturbance When it is small enough, the second term of the right-hand side approaches 0. Based on this, we can get: （5） Will Recorded as ; In formula (1), (6) In formula (6), A classification network is introduced to achieve the purpose of accurately locating the location information of the lesion; To refine the denoising network, further refine the lesion site for denoising; refine the denoising network Network structure and denoising network Exactly the same, and Refinement denoising network and classification network Learnable parameters; For the nth iteration, It can be processed using the univariate Taylor expansion, and then: （7） in, To generate The Jacobian matrix corresponding to the classification network, so if there is ,So It can be expressed as （8） When the disturbance Enough hours, Approaching 0, based on formula (8), we can get the following approximate estimate: （9） in, For classification network Output class activation map; Then the iterative format of the pathological focused MRI image reconstruction model can be directly written out, namely: （10）。 3. The intelligent rapid magnetic resonance imaging reconstruction and pathology localization method according to claim 2, characterized in that: The specific process of step 2 is: Formula (10) is expanded into a multi-stage cascade network as the pathology-focused MRI image reconstruction network. The pathology-focused MRI image reconstruction network includes a denoising network , lesion guidance module LF and data consistency module DC.

4. The intelligent rapid magnetic resonance imaging reconstruction and pathology localization method according to claim 3, characterized in that: Denoising module It is a seven-layer CNN network. The first six layers are composed of convolutional layers, BN layers, and ReLU activation function layers in sequence, and the seventh layer is composed of convolutional layers and BN layers in sequence; The denoising network is expressed as: （11）。 5. The intelligent rapid magnetic resonance imaging reconstruction and pathology localization method according to claim 3, characterized in that: A classification network is introduced into the lesion guidance module LF The predicted CAM map is used to provide accurate location information of the lesion, and then the predicted CAM map is combined with the input Connect to achieve the role of CAM guidance, and input the connected value into the refinement denoising network , thereby achieving the purpose of lesion enhancement; The lesion guidance module LF is expressed as: （12） Classification Network It is a pre-trained improved VGG19 network, whose parameters are frozen during the reconstruction process and do not participate in the update. 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 are composed of two convolutional layers, ReLU activation function layers and maximum pooling layers in sequence. The third, fourth and fifth levels are composed of four convolutional layers, ReLU activation function layers and maximum pooling layers in sequence. The classification module consists of a global average pooling layer, two fully connected layers, and a softmax activation layer; Classification Network The processing flow is: denoising network The output of is used as the input of the classification network. The feature map extracted by the feature extraction module of the classification network passes through a global average pooling layer and two fully connected layers to obtain the classification score. The classification score passes through the softmax activation layer to obtain the classification result. The classification result and the classification label Do cross loss entropy training, then get a training parameter and freeze it in subsequent training, and calculate the gradient of the feature map by the classification score, there are: In the formula, is the number of feature maps, h and w are the length and width of the feature maps respectively; is the value representing the abnormal component in the classification score, is the kth component of the feature map, is the gradient of the classification score to the value of the i-th row and j-th column of the feature map; for the above value, after the ReLU activation function layer and interpolation upsampling to , and finally get the attention map CAM.

6. The intelligent rapid magnetic resonance imaging reconstruction and pathology localization method according to claim 3, characterized in that: The data consistency module DC is expressed as: （13） The data consistency module DC obtains the reconstructed image by minimizing the difference between the solution of the subproblem and the input , input is - spatial data , denoised image and lesion enhanced images , then the analytical solution to this problem can be derived as: （14） In the formula, is the forward operator, is the identity matrix, since It is not necessarily reversible, so the conjugate gradient method is used to iteratively solve this subproblem.

7. The intelligent rapid magnetic resonance imaging reconstruction and pathology localization method according to claim 1, characterized in that: The specific process of step 3 is: Step 3.1: Build a training dataset The training data set consists of multiple data groups, each of which consists of undersampled k-space data, MRI images reconstructed from fully sampled data, and image-level labels of whether the MRI images have lesions or not; Step 3.2, training the pathology-focused magnetic resonance image reconstruction network parameters using the training data set of step 3.1; The specific process is: The fully sampled data are used to reconstruct the MRI images and the image-level labels of whether the MRI images have lesions are used as the input of the classification network, and the classification network is trained. Then, the parameters of the trained classification network are frozen and integrated into the pathology-focused MRI image reconstruction network, and the network is used as a classification module of the pathology-focused MRI image reconstruction network to provide lesion location information. The undersampled k-space data is used as the initial input of the pathology-focused magnetic resonance image reconstruction network, and the gradient of the loss function with respect to the parameters of the pathology-focused magnetic resonance image reconstruction network is calculated by back propagation. 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, so as to obtain the optimal parameters of the pathology-focused magnetic resonance image reconstruction network. During the training process, the loss function of the pathology-focused MRI reconstruction network is: （15） In the formula, is the number of samples, is a fully sampled MRI image, Pathology-focused magnetic resonance image reconstruction network The final reconstruction result of the stage, To assist the overall denoising image, For classification network The CAM diagram output at the T stage, represents the cross entropy loss, Labels representing the presence of lesions, The label represents the absence of lesions, Represents the classification probability map output by the classification network at stage T, , , , are all hyperparameters.

Images