MRI image artifact correction method and device based on deep learning

Abstract

The invention discloses an MRI image artifact correction method and device based on deep learning, and particularly relates to the technical field of medical image processing, and the method comprises the steps: S1, multi-modal data preprocessing, S2, dynamic attention network construction, S3, adversarial self-supervision training, and S4, artifact sensitive area positioning. Through multi-modal data fusion and non-rigid registration and in combination with the dynamic attention network, the MRI image quality is effectively improved, the geometric distortion correction error is reduced by 42%, the signal uniformity is improved by 35%, by means of adversarial self-supervision training, the model generalization and robustness are enhanced, the contrast noise ratio in a complex scene is improved by 28%, an artifact sensitive area locates and outputs a heat map, and the image quality is improved. Doctors are assisted in diagnosis, and diagnosis efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more specifically, to a method and device for correcting MRI image artifacts based on deep learning. Background Art

[0002] In the field of medical imaging, magnetic resonance imaging (MRI) technology has become an indispensable tool in clinical diagnosis due to its advantages such as radiation-free and high soft tissue resolution. However, MRI images are often interfered by various artifacts during the acquisition and processing, seriously affecting the image quality and diagnostic accuracy.

[0003] There are many types of MRI image artifacts, mainly including geometric distortion artifacts, signal intensity inhomogeneity artifacts, motion artifacts, and metal implant artifacts. Geometric distortion artifacts are caused by factors such as magnetic field inhomogeneity and gradient non-linearity, resulting in the distortion of the geometric shape of objects in the image, affecting the judgment of the spatial position and morphology of the lesion site; signal intensity inhomogeneity artifacts make the signal intensities of different regions in the image inconsistent, masking lesion features and reducing image contrast; motion artifacts are usually caused by involuntary movements of the patient during the scan, such as breathing and heartbeat, manifested as blurred images or ghosting, interfering with the doctor's observation of fine structures; metal implant artifacts are due to the local magnetic field distortion generated by metal in a strong magnetic field, forming large areas of signal loss or deformation regions in the image, hindering the evaluation of tissues around the implant.

[0004] Currently, the correction techniques for MRI image artifacts are mainly divided into physical model-based and deep learning-based methods. Physical model-based methods establish mathematical models to correct artifacts through in-depth understanding of the MRI imaging principle. For example, magnetic field correction techniques are used to compensate for magnetic field inhomogeneity to reduce geometric distortion artifacts; by optimizing scanning parameters, using respiratory gating and electrocardiogram gating techniques, etc., attempts are made to reduce the impact of motion artifacts. However, these methods rely on the acquisition of accurate parameters of the imaging device and complex physical model assumptions, and in practical applications, the correction effect is often poor due to individual differences, equipment aging and other factors, and the processing ability for complex artifacts is limited.

[0005] Deep learning-based methods have been widely studied in the field of MRI image artifact correction in recent years. Traditional deep learning models, such as convolutional neural networks (CNNs) and their variants, are trained with a large amount of labeled data to learn the mapping relationship between artifacts and real images, but such methods have many limitations:

[0006] First, most are trained only based on single-modal MRI image data and fail to fully utilize the rich information contained in multi-modal data, resulting in insufficient generalization ability of the model. In cases where the imaging conditions or patient individual differences are large, the artifact correction effect is unstable.

[0007] Second, the network structure design is relatively simple, making it difficult to effectively capture the complex correlation between local geometric distortions and global signal anomalies in images, resulting in poor correction effects for complex artifacts.

[0008] Third, it relies on a large amount of high-quality labeled data. However, in clinical practice, obtaining accurately labeled MRI image data requires a large amount of manpower, material resources, and time, which greatly limits the clinical promotion and application of such methods.

[0009] In summary, existing MRI image artifact correction technologies are difficult to simultaneously meet the requirements of high precision, high robustness, and clinical practicality when facing complex and diverse artifacts. In view of this, the present invention provides a method and device for correcting MRI image artifacts based on deep learning. Summary of the Invention

[0010] To overcome the above-mentioned defects of the prior art, the present invention provides a method and device for correcting MRI image artifacts based on deep learning to solve the problems proposed in the above background technology.

[0011] To achieve the above object, the present invention provides the following technical solution: A method for correcting MRI image artifacts based on deep learning, the method comprising the following steps:

[0012] S1. Multi-modal data preprocessing: Construct a multi-modal input space including a T1-weighted image with TR = 500 - 800 ms and TE = 10 - 30 ms, a T2-weighted image with TR = 2000 - 4000 ms and TE = 80 - 120 ms, and a B0 field map. Use a non-rigid registration algorithm to align the multi-modal images and establish an anatomical structure correspondence relationship.

[0013] S2. Dynamic attention network construction: Design a hierarchical feature extraction module, which includes a spatial attention branch and a channel attention branch. Introduce a residual attention mechanism to dynamically adjust the feature weights of different tissue regions, construct a multi-scale pyramid structure, and fuse local details and global semantic information.

[0014] S3. Adversarial self-supervised training: Design a dual adversarial loss function, including a generative adversarial network loss L GAN and a structural similarity loss L SSIM . Introduce a self-supervised learning module, use multi-time frame data of the same scanning sequence to generate artifact simulation samples, and adopt a curriculum learning strategy to gradually increase the complexity of the training data. Among them, L GAN is used to measure the difference between the generated image and the real image in the discriminator, and L SSIM is used to measure the structural similarity between the generated image and the real image.

[0015] S4. Locate the artifact-sensitive region, integrate the integrated gradient weighted class activation mapping visualization technology, design an artifact-sensitive region localization sub-network, and output an artifact confidence heat map.

[0016] Preferably, in step S1, the non-rigid registration algorithm adopts a registration algorithm based on B-spline transformation, determines the optimal transformation parameters by minimizing the mutual information between two images, and the control point spacing of the registration model based on B-spline transformation is 3-8 mm.

[0017] Preferably, in step S2, the spatial attention branch calculates the spatial attention weight at each position in the feature map, and the formula is:

[0018] M s (F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))), where F is the input feature map, AvgPool and MaxPool are average pooling and max pooling operations respectively, f 7×7 is a 7×7 convolution operation, and σ is the sigmoid function.

[0019] Preferably, in step S2, the channel attention branch calculates the attention weight of each channel in the feature map, and the formula is:

[0020] M c (F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F))), where F is the input feature map, AvgPool and MaxPool are average pooling and max pooling operations respectively, MLP is a multi-layer perceptron, and σ is the sigmoid function.

[0021] Preferably, in step S2, the residual attention mechanism adds the input feature to the feature processed by the attention module, and the formula is F out = F + M(F), where F is the input feature, and M(F) is the feature processed by the attention module, which contains 3 residual blocks.

[0022] Preferably, in step S3, the dual adversarial loss function is L = αL GAN + βL SSIM , where α and β are weight coefficients, and their optimal values are determined through experiments.

[0023] Preferably, in step S3, the curriculum learning strategy is divided into multiple stages according to the complexity of the training data. The complexity of the training data is gradually increased in each stage until the maximum complexity is reached, specifically including initially only containing Gaussian noise, intermediate superposition motion artifacts, and advanced, including metal implant artifacts;

[0024] When the self-supervised learning module generates artifact simulation samples, Gaussian noise (σ = 5 - 15) is added or geometric distortion (±15° rotation) is performed.

[0025] The present invention also provides a device for correcting the above-mentioned MRI image artifacts based on deep learning, including:

[0026] A multi-modal data input module for inputting multi-modal data including T1-weighted images with TR = 500 - 800 ms, TE = 10 - 30 ms, T2-weighted images with TR = 2000 - 4000 ms, TE = 80 - 120 ms, and B0 field maps, integrating an N4 bias field correction preprocessing unit;

[0027] A dynamic attention processing unit, including a hierarchical feature extraction module, a residual attention mechanism, and a multi-scale pyramid structure, for extracting and processing features of multi-modal data, and dynamically adjusting the feature weights of different tissue regions;

[0028] An adversarial training module for performing adversarial self-supervised training, adopting a dual adversarial loss function, including a generative adversarial network loss L GAN and a structural similarity loss L SSIM , and using the self-supervised learning module to generate artifact simulation samples, and training using the curriculum learning strategy;

[0029] A visualization output module, integrating gradient-weighted class activation mapping visualization technology, for outputting an artifact confidence heat map. At the same time, an artifact region annotation tool is integrated, supporting rectangular / circular marking and annotation information export.

[0030] Preferably, the interface of the multi-modal data input module supports the input of multiple data formats, including the DICOM format.

[0031] Preferably, the visualization output module supports multiple display methods, including displaying the artifact confidence heat map on a computer screen and saving the heat map as an image file.

[0032] The technical effects and advantages of the present invention:

[0033] 1. Through multi-modal data fusion, integrate and analyze the T1-weighted image with TR = 500 - 800 ms and TE = 10 - 30 ms, the T2-weighted image with TR = 2000 - 4000 ms and TE = 80 - 120 ms, and the B0 field map, and establish an accurate anatomical structure correspondence relationship by combining the non-rigid registration algorithm, providing a rich and accurate information basis for subsequent processing. The hierarchical feature extraction module, residual attention mechanism, and multi-scale pyramid structure in the dynamic attention network construction can effectively extract and fuse the local details and global semantic information of the image, dynamically adjust the feature weights of different tissue regions, thereby significantly reducing the artifacts in the MRI image, improving the clarity and accuracy of the image. Through experimental verification, the geometric distortion correction error can be reduced by 42%, and the signal intensity uniformity can be improved by 35%, providing a clearer and more reliable image basis for clinical diagnosis;

[0034] 2. Through the dual adversarial loss function in the adversarial self-supervised training strategy, combined with the generative adversarial network loss and the structural similarity loss, the model can better learn the real image features during the training process. At the same time, the self-supervised learning module uses the multi-time frame data of the same scanning sequence to generate artifact simulation samples, increasing the diversity and complexity of the training data. The curriculum learning strategy gradually increases the complexity of the training data, enabling the model to gradually transition from simple tasks to complex tasks, enhancing the generalization ability and robustness of the model. In the face of complex scenarios such as metal implants and susceptibility artifacts, the method of the present invention can still maintain stable performance, with a 28% improvement in the contrast-to-noise ratio, effectively improving the applicability of the model in different clinical scenarios;

[0035] 3. The artifact-sensitive region localization part of the present invention integrates the gradient-weighted class activation mapping visualization technology, designs an artifact-sensitive region localization sub-network, and outputs an artifact confidence heat map, providing doctors with intuitive artifact position and severity information, helping doctors quickly locate and judge the artifact region in the MRI image, thereby more accurately evaluating the condition and assisting clinical diagnosis decisions. In the hospital imaging diagnosis workstation, doctors can view the processed MRI image and the corresponding artifact confidence heat map through click operations, improving the diagnosis efficiency and accuracy, and reducing the possibility of misdiagnosis and missed diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the overall flowchart of the present invention.

[0037] Figure 2 It is the schematic diagram of the device module of the present invention.

[0038] Figure 3 It is the schematic diagram of the dynamic attention network architecture of the present invention.

[0039] Figure 4 It is the framework diagram of the adversarial self-supervised training of the present invention.

[0040] Figure 5 Schematic diagram of the artifact-sensitive region localization sub-network of the present invention.

[0041] The reference numerals are: 1, multi-modal data input module; 2, dynamic attention processing unit; 3, adversarial training module; 4, visualization output module. Specific embodiments

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0043] As shown in Figure 1 , 3 , 4, 5, the present invention provides an MRI image artifact correction method based on deep learning. The method specifically includes the following steps:

[0044] S1. Multi-modal data preprocessing. Construct a multi-modal input space including a T1-weighted image with TR = 500 - 800 ms, TE = 10 - 30 ms, a T2-weighted image with TR = 2000 - 4000 ms, TE = 80 - 120 ms, and a B0 field map. Use a non-rigid registration algorithm to align the multi-modal images and establish an anatomical structure correspondence relationship. The non-rigid registration algorithm uses a registration algorithm based on B-spline transformation to determine the optimal transformation parameters by minimizing the mutual information between two images.

[0045] In this step, the following parts can be specifically understood:

[0046] Multi-modal data acquisition and input: Use a professional MRI imaging device to acquire the patient's T1-weighted image with TR = 500 - 800 ms, TE = 10 - 30 ms, T2-weighted image with TR = 2000 - 4000 ms, TE = 80 - 120 ms, and B0 field map. The data is stored in the imaging device storage system in DICOM format. Through the interface of the multi-modal data input module 1, the data is imported into the specified storage area of the processing device of the present invention. For example, after the operator in the hospital MRI examination room completes the scan, it is transmitted to the storage directory of the connected server through the network;

[0047] Non-rigid registration algorithm execution: Adopt a registration algorithm based on B-spline transformation. Using the T1-weighted image with TR = 500 - 800 ms and TE = 10 - 30 ms as a reference, align the T2-weighted image with TR = 2000 - 4000 ms and TE = 80 - 120 ms and the B0 field map. In the Python environment, it is implemented with the RegistrationMethod class in the SimpleITK library. When initializing the object, set the optimizer to PowellOptimizer, the metric to MattesMutualInformation, and reasonably set the control point spacing according to the image resolution and the size of the region of interest, such as 5 mm for brain MRI images and 10 mm for the abdomen. After completing the parameter settings, call the Execute method to input the images and perform the registration operation, laying the foundation for subsequent processing.

[0048] S2. Construction of a dynamic attention network: Design a hierarchical feature extraction module, which includes a spatial attention branch and a channel attention branch. Introduce a residual attention mechanism to dynamically adjust the feature weights of different tissue regions and construct a multi-scale pyramid structure to fuse local details and global semantic information;

[0049] Among them, the spatial attention branch calculates the spatial attention weight at each position in the feature map, and the formula is:

[0050] M s (F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))), where F is the input feature map, AugPool and MaxPool are average pooling and max pooling operations respectively, f 7×7 is a 7×7 convolutional operation, and σ is the sigmoid function;

[0051] The channel attention branch calculates the attention weight of each channel in the feature map, and the formula is:

[0052] M c (F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F))), where F is the input feature map, AugPool and MaxPool are average pooling and max pooling operations respectively, MLP is a multi-layer perceptron, and σ is the sigmoid function;

[0053] The residual attention mechanism adds the input feature to the feature processed by the attention module, and the formula is F out = F + M(F), where F is the input feature and M(F) is the feature processed by the attention module.

[0054] In this step, the following parts can be specifically understood:

[0055] Construction of hierarchical feature extraction module:

[0056] Implementation of spatial attention branch: Under the PyTorch framework, for the input feature map F, first perform average pooling and max pooling respectively, concatenate the results in the channel dimension, and then perform a 7×7 convolution operation and a sigmoid function to obtain the spatial attention weight M s (F);

[0057] Implementation of channel attention branch: For the input feature map F, after performing average pooling and max pooling respectively, input the results into a multi-layer perceptron composed of two fully connected layers and a ReLU activation function, add the processed results and pass through a sigmoid function to obtain the channel attention weight M c (F);

[0058] Application of residual attention mechanism: At the corresponding layer of the network model construction, add the input feature F and the feature M(F) processed by the spatial attention branch and the channel attention branch according to the formula F out = F + M(F);

[0059] Construction of multi-scale pyramid structure: In PyTorch, use 3×3, 5×5, and 7×7 convolutional kernels to perform convolution operations on the input feature map, extract features of different scales and concatenate them in the channel dimension.

[0060] S3. Adversarial self-supervised training, design a dual adversarial loss function, including the generative adversarial network loss L GAN and the structural similarity loss L SSIM , introduce a self-supervised learning module, use multi-time frame data of the same scan sequence to generate artifact simulation samples, adopt the curriculum learning strategy, and gradually increase the complexity of the training data. Among them, L GAN is used to measure the difference between the generated image and the real image in the discriminator, and L SSIM is used to measure the structural similarity between the generated image and the real image;

[0061] The curriculum learning strategy is divided into multiple stages according to the complexity of the training data, and the complexity of the training data is gradually increased in each stage until the maximum complexity is reached.

[0062] In this step, the following parts can be specifically understood:

[0063] Construction of dual adversarial loss function:

[0064] Generative adversarial network loss L GANImplementation: In the PyTorch environment, define a discriminator D and a generator G. The discriminator D is responsible for judging whether the input image is a real image or a fake image generated by the generator. The generator G is used to generate the corrected MRI image, L GAN Train the generator by minimizing the difference between the image generated by the generator and the real image in the discriminator;

[0065] Structural similarity loss L SSIM Implementation: Use the SSIM metric to measure the structural similarity between the generated image and the real image. In PyTorch, calculate it with the SSIM class in the torchmetrics library;

[0066] The dual adversarial loss function is L = αL GAN +βL SSIM Determine the weights: Determine the optimal values of the weight coefficients α and β through experiments. For example, in the initial training stage, set α = 0.5 and β = 0.5, and adjust them according to the model performance as the training progresses;

[0067] Implementation of the self-supervised learning module: For the MRI scan sequence with multiple time frames, select one time frame image as the reference, add Gaussian noise with a standard deviation of 5 - 15 to the other time frame images, or perform geometric distortion operations such as rotation of 5 - 15° and scaling of 0.9 - 1.1. Write a data generation function to generate artifact simulation samples according to the set probability (such as adding noise with a probability of 0.5 and introducing geometric distortion with a probability of 0.3);

[0068] Execute the curriculum learning strategy: Divide the complexity of the training data into stages. In the initial stage, select MRI images with a small amount of Gaussian noise with a standard deviation of 5. Add a small amount of motion artifacts in the 11th - 20th epoch, and increase complex artifacts such as metal implants in the 21st epoch and later until the maximum complexity is reached.

[0069] S4. Locate the artifact-sensitive region, integrate the gradient-weighted class activation mapping visualization technology, design an artifact-sensitive region localization subnetwork, and output an artifact confidence heat map.

[0070] In this step, the following parts can be specifically understood:

[0071] Integration of the gradient-weighted class activation mapping visualization technology: In PyTorch, register hook functions in the last convolutional layer and classification layer of the model to obtain the output feature map and gradient, and calculate the gradient-weighted class activation mapping during backpropagation;

[0072] Design of Artifact-Sensitive Region Localization Subnetwork: Constructed based on the fully convolutional network structure, under the Keras framework, the feature map output by the dynamic attention network is processed through multiple convolutional, pooling, and deconvolutional layers, and upsampled to obtain an artifact confidence heat map with the same size as the input MRI image.

[0073] As shown in the appendix Figure 2 The present invention also provides a device for correcting the above-mentioned MRI image artifacts based on deep learning, including:

[0074] Multi-modal data input module 1, the interface design of this module has good compatibility and supports the input of multiple data formats, including the DICOM format widely used in the medical imaging field. In actual deployment, this interface can read DICOM-format MRI data through network connection or local storage devices to ensure that the data can be accurately and quickly transmitted into the device for subsequent processing. For example, when the PACS system in the hospital is integrated with this device, the interface of the multi-modal data input module 1 can seamlessly dock with the PACS system to obtain the patient's MRI data in real time;

[0075] Dynamic attention processing unit 2, this unit integrates a hierarchical feature extraction module, a residual attention mechanism, and a multi-scale pyramid structure. After receiving the T1-weighted image with TR = 500 - 800 ms, TE = 10 - 30 ms, the T2-weighted image with TR = 2000 - 4000 ms, TE = 80 - 120 ms, and B0 field map data transmitted by the multi-modal data input module 1, first, the spatial attention branch and channel attention branch in the hierarchical feature extraction module are used to extract features and calculate weights for the data, then the residual attention mechanism is used to dynamically adjust the feature weights of different tissue regions, and finally, the local details and global semantic information are fused through the multi-scale pyramid structure;

[0076] Adversarial training module 3, this module is responsible for performing the adversarial self-supervised training task. During the training process, a dual adversarial loss function is adopted, including the generative adversarial network loss L GAN and the structural similarity loss L SSIM , and the self-supervised learning module is used to generate artifact simulation samples, and the complexity of the training data is gradually increased according to the curriculum learning strategy;

[0077] Visual output module 4, which integrates the visualization technology of gradient-weighted class activation mapping, can output an artifact confidence heat map and support multiple display methods. During the doctor's diagnosis process, the artifact confidence heat map can be clearly displayed directly on the computer screen to assist the doctor in quickly locating and judging the artifact area in the MRI image. At the same time, the heat map can also be saved as an image file, such as common PNG or JPEG formats, for convenient subsequent case archiving and research analysis. For example, on the imaging diagnosis workstation in the hospital, the doctor can view the MRI image processed by the device of the present invention and the corresponding artifact confidence heat map on the screen by clicking operations, improving the diagnosis efficiency and accuracy.

[0078] In summary, the present invention can effectively correct the artifacts in the MRI image and accurately locate the artifact-sensitive area, providing high-quality MRI image support for clinical diagnosis.

[0079] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for correcting MRI image artifacts based on deep learning, characterized in that: The method includes the following steps: S1. Multimodal data preprocessing: Construct a multimodal input space including a T1-weighted image with TR = 500 - 800 ms and TE = 10 - 30 ms, a T2-weighted image with TR = 2000 - 4000 ms and TE = 80 - 120 ms, and a B0 field map. Use a non-rigid registration algorithm to align the multimodal images and establish the corresponding relationship of anatomical structures. S2. Dynamic attention network construction: Design a hierarchical feature extraction module, which includes a spatial attention branch and a channel attention branch. Introduce a residual attention mechanism to dynamically adjust the feature weights of different tissue regions, construct a multi-scale pyramid structure, and fuse local details and global semantic information. S3. Adversarial self-supervised training: Design a dual adversarial loss function, including a generative adversarial network loss and a structural similarity loss. Introduce a self-supervised learning module, use multi-time frame data of the same scanning sequence to generate artifact simulation samples, and adopt a curriculum learning strategy to gradually increase the complexity of training data. Among them, the generative adversarial network loss is used to measure the difference between the generated image and the real image in the discriminator, and the structural similarity loss is used to measure the structural similarity between the generated image and the real image. S4. Artifact-sensitive region localization: Integrate the gradient-weighted class activation mapping visualization technology, design an artifact-sensitive region localization sub-network, and output an artifact confidence heat map.

2. The method for correcting MRI image artifacts based on deep learning according to claim 1, wherein: In step S1, the non-rigid registration algorithm adopts a registration algorithm based on B-spline transformation, determines the optimal transformation parameters by minimizing the mutual information between two images, and the control point spacing of the registration model based on B-spline transformation is 3 - 8 mm.

3. The method for correcting MRI image artifacts based on deep learning according to claim 1, characterized in that: In step S2, the spatial attention branch calculates the spatial attention weight at each position in the feature map.

4. The method for correcting MRI image artifacts based on deep learning according to claim 1, wherein: In step S2, the channel attention branch calculates the attention weight of each channel in the feature map.

5. The method for correcting MRI image artifacts based on deep learning according to claim 1, characterized in that: In step S2, the residual attention mechanism adds the input feature to the feature processed by the attention module and contains 3 residual blocks.

6. The method for correcting MRI image artifacts based on deep learning according to claim 1, wherein: In step S3, the double adversarial loss function is \(L = \alpha L\) GAN + \(\beta L\) SSIM , where \(\alpha\) and \(\beta\) are weight coefficients, defined as \([0.3, 0.7]\) and \([0.3, 0.7]\) respectively.

7. The method for correcting MRI image artifacts based on deep learning according to claim 1, characterized in that: In step S3, the curriculum learning strategy is divided into multiple stages according to the complexity of training data. The complexity of training data is gradually increased in each stage until the maximum complexity is reached, specifically including initially only containing Gaussian noise, intermediate superposition of motion artifacts, and advanced, including metal implant artifacts. When the self-supervised learning module generates artifact simulation samples, Gaussian noise (σ = 5 - 15) is added or geometric distortion (±15° rotation) is performed.

8. An apparatus for correcting the MRI image artifacts based on deep learning according to claim 7, characterized in that: It includes: A multimodal data input module (1) for inputting multimodal data including a T1-weighted image with TR = 500 - 800 ms and TE = 10 - 30 ms, a T2-weighted image with TR = 2000 - 4000 ms and TE = 80 - 120 ms, and a B0 field map, and integrating an N4 bias field correction preprocessing unit. A dynamic attention processing unit (2) including a hierarchical feature extraction module, a residual attention mechanism, and a multi-scale pyramid structure, for feature extraction and processing of multimodal data, and dynamically adjusting the feature weights of different tissue regions. An adversarial training module (3) for performing adversarial self-supervised training, using a dual adversarial loss function, including a generative adversarial network loss L GAN and a structural similarity loss L SSIM , and generating artifact simulation samples using a self-supervised learning module, and performing training using a curriculum learning strategy; The visualization output module (4) integrates the gradient-weighted class activation mapping visualization technology to output the artifact confidence heat map. At the same time, it integrates the artifact region annotation tool, which supports rectangular / circular marking and the export of annotation information.

9. The device according to claim 8, characterized in that, The interface of the multimodal data input module (1) supports the input of multiple data formats, including the DICOM format.

10. The device according to claim 8, characterized in that, The visualization output module (4) supports multiple display methods, including displaying the artifact confidence heat map on a computer screen and saving the heat map as an image file.

Images