Multimodal dMRI image distortion correction system based on unsupervised deep learning

Abstract

This invention belongs to the field of medical image processing technology, specifically a multimodal dMRI image distortion correction system based on unsupervised deep learning. The correction system comprises a coarse-to-fine deep learning network—a multimodal distortion correction network—consisting of cascaded rigid transform networks and dense transform networks; it also includes a cascaded training strategy module. This invention integrates rich information from two modalities and inputs it into the network. First, the rigid transform network is used to learn the linear misalignment between two images with opposite phase encodings. Then, through the dense transform network, a dual-stream encoder and a multimodal attention module are used to learn the distortion field from the images of both modalities. Furthermore, a dynamic weighted averaging (DWA) method is employed to simultaneously optimize the multimodal images. Compared to current magnetic susceptibility distortion correction methods, this invention can correct magnetic susceptibility distortion in dMRI images more efficiently and accurately.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to an unsupervised deep learning image correction system for dMRI images. Background Art

[0002] Diffusion magnetic resonance imaging (dMRI) is a non-invasive imaging technique that captures the diffusion patterns of water molecules in biological tissues and is widely used to track white matter fiber bundles in the human brain. However, the echo-planar imaging (EPI) sequences commonly used to acquire dMRI are susceptible to geometric and intensity distortions caused by magnetic susceptibility in the phase encoding direction. These distortions typically occur at the boundaries of different tissues and cavities, such as the brainstem region. These distortions significantly affect the accuracy and reliability of subsequent analysis and interpretation of brain data.

[0003] Currently, the most commonly used magnetic susceptibility distortion correction method is "blip-up blip-down," which involves scanning two B0 images with opposite phase encoding directions during acquisition and then estimating their distortion fields. The topup method in the FSL toolbox uses discrete cosine basis functions to derive a smooth displacement field from the two acquired B0 images and combines them using the least squares method to obtain the distortion field. In recent years, with the development of deep learning technology, some methods have used unsupervised deep learning to estimate the distortion field from two B0 images with opposite phases, achieving correction results similar to topup with a shorter estimation time. However, due to insufficient contrast in certain regions of the B0 image, the above B0 image-based distortion correction algorithms do not perform well in certain brain regions, such as the brainstem and corpus callosum.

[0004] To address this issue, some methods have begun to use other feature images to correct magnetic susceptibility distortion in dMRI, such as fiber orientation distribution (FOD) images. Qiao et al. proposed FODReg, which aims to correct residual distortion in topup-preprocessed dMRI data by estimating a fine displacement field. Subsequently, they further developed DrC-Net, using a deep learning-based approach to further improve the performance of FOD image-based distortion correction. However, FODReg and DrC-Net simply concatenate FOD image-based methods after topup correction, failing to fully integrate useful information from both modalities. Furthermore, magnetic susceptibility distortion is prevalent in different dMRI derived modalities, and learning the distortion field using only a single modality may lead to decreased generalization ability in other modalities. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal diffusion magnetic resonance imaging (dMRI) image distortion correction system to automatically correct image distortion caused by magnetic susceptibility in dMRI images, thereby improving the quality and accuracy of dMRI images.

[0006] The multimodal dMRI image distortion correction system provided by this invention is based on unsupervised deep learning, specifically including: a coarse-to-fine deep learning network—a multimodal distortion correction network (denoted as MMDC-Net), which consists of cascaded rigid transform and dense transform networks; in addition, a cascaded training strategy module is designed; wherein:

[0007] (1) The rigid transformation network is used to correct linear misalignment between two phase-encoded direction images. Specifically, it is a convolutional neural network, including an encoder, a multilayer perceptron, and a spatial transformation network. The encoder is responsible for extracting feature information from the stitched B0 and FOD images, where the B0 and FOD images each include images with two opposite encoding directions. The encoder contains four convolutional layers, each of which also includes a batch normalization layer. The feature maps output by each convolutional layer have 16, 16, 32, and 32 channels, respectively, and the sizes of the feature maps are 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original image size, respectively. The subsequent multilayer perceptron includes two consecutive fully connected layers that map the feature maps learned by the encoder to a set of rigid transformation parameters μ:

[0008] μ=(θ x ,θ y ,θ z ,t x ,t y ,t z (1)

[0009] Where x, y, and z are the three coordinate directions, and θ is the coordinate direction. x ,θ y ,θ z These represent the rotation angles of the corresponding coordinate axes, t and t'. x ,t y ,t z These represent the translation distances for the corresponding coordinate axes. According to the topup settings, the translation parameter t is set to 0 in the phase encoding because it is impossible to distinguish between the subject's translation and the deformation offset caused by magnetic susceptibility distortion in this direction.

[0010] Based on the learned rigid transformation parameter μ, the geometric transformation matrix R can be calculated. x R y R z :

[0011]

[0012] Finally, calculate the affine transformation matrix: By using a spatial transformation network, the learned affine transformation matrix is ​​applied to the negative phase B0 and FOD matrices to correct the linear misalignment between the two phase-encoded direction images.

[0013] The spatial transformation network is an existing network module (Jaderberg M, Simonyan K, Zisserman A. Spatial transformer networks[J]. Advances in neural information processing systems, 2015, 28.) used to apply rigid or non-rigid deformation fields to images for deformation processing.

[0014] (2) The dense transform network is used to correct non-rigid distortion of images in two opposite encoding directions. Specifically, a U-shaped neural network is constructed, which includes a two-stream encoder, a decoder, a skip connection layer connecting the encoder and decoder, and a multimodal attention module. The two-stream encoder processes B0 and FOD information respectively: the B0 branch stitches the positive and negative phase B0 images into a 2-channel image as input; the FOD branch stitches the positive and negative phase FOD images into a 12-channel image as input. Each branch of the two-stream encoder includes four convolutional blocks, each consisting of a 3×3×3 convolutional layer and a LeakyReLU activation layer. By using small convolutional kernels and a stride of 2, the encoder can progressively extract deep information of the image while maintaining spatial resolution.

[0015] After processing by the encoder, the decoder fuses feature information from different modalities through several multimodal attention modules at different scales, thereby achieving more accurate distortion correction. Each multimodal attention module receives FOD and B0 feature information from different scales of the encoder, as well as the output feature map of the decoder at the previous scale, aggregates the fused information at the corresponding scale, and upsamples it. In the final stage of the decoder, a 3D convolutional layer is used to estimate a three-dimensional non-rigid deformation field from the decoder features. This deformation field only retains the deformation in the phase encoding direction, thereby achieving magnetic susceptibility distortion correction.

[0016] Furthermore, the multimodal attention module, including a region attention module and a channel attention module, is used to aggregate feature information from the BO and FOD images in the dual-stream encoder during the decoding stage.

[0017] The region attention module is a core component of the decoder, responsible for dynamically estimating and assigning attention weights to different regions across different modalities during distortion correction. This module focuses on the feature contributions of the same region across different modalities, generating an attention map representing the contributions of regions from different modalities by capturing latent features from the encoder and decoder. First, the module concatenates encoder and decoder feature maps of the same scale and feeds them into a convolutional layer to extract region features. Then, the output dual-channel feature map is passed through a softmax layer to obtain an attention map, where the two channels correspond to the contributions of B0 and FOD in each region, respectively. This attention map can be represented as:

[0018]

[0019] in, This is the attention map learned by the region attention module, where i represents the i-th layer of the network, m represents different modalities, including B0 and FOD images, [·; ·] represents the concatenation operation, and e i and d i f represents the feature maps of the encoder and decoder, respectively. 3 ×3×3 This represents a convolutional layer with a kernel size of 3×3×3, and M is the set of different modalities, i.e., {B0,FOD}.

[0020] Subsequently, the learned feature map is multiplied onto the input encoder features on a volumetric basis to obtain the attention-weighted feature output.

[0021]

[0022] The channel attention module is responsible for estimating the overall contribution of different modal inputs, dynamically adjusting the weights of each channel. Unlike the region attention module, the channel attention module focuses more on the contribution of the overall features rather than the feature weights of individual regions. This module first concatenates the decoder features and the output of the region attention module, then generates two distinct feature vectors through average pooling and max pooling operations. These two vectors are then input into a multilayer perceptron with shared weights. The output of the multilayer perceptron is activated by a sigmoid function, mapping the resulting weight matrix to a range of 0 to 1, representing the degree of contribution of different modal inputs to the output features in the current layer.

[0023]

[0024] Where σ represents the sigmoid activation function, f MLP A multilayer perceptron representing shared weights, fr i and d iThe output features of the region attention module and the encoder at different scales are represented by AvgPool(·), which represents the average pooling layer and MaxPool(·), which represents the max pooling layer.

[0025] Finally, the channel attention module applies these weights to the feature map output by the region attention module, thereby adjusting the influence of different modal features on the correction effect:

[0026]

[0027] (3) The cascaded training strategy module is used to improve the accuracy of the deep learning network MMDC-Net in estimating non-rigid deformation fields, including using the dynamic weight averaging (DWA) method to simultaneously optimize multimodal images.

[0028] For the dense transform network, the input is a pair of positive and negative phase FOD and B0 images to be corrected, and the output is a three-dimensional deformation field. A spatial transform module is used to apply the learned deformation field to the image pair to be corrected: applying the positive deformation field to the positive phase B0 and FOD images yields the corrected positive phase image; keeping the magnitude of the deformation field unchanged but reversing its direction, the reverse deformation field is applied to the negative phase B0 and FOD images to obtain the corrected negative phase image. Subsequently, this coarsely corrected positive and negative phase B0 and FOD image pair can be input again into the dense transform network to learn the residual error between them, resulting in a finer deformation field. After a series of cascaded operations, the difference between the two phase images tends to converge. Adding the deformation fields output by this series of networks yields the final deformation field of the original input image to be corrected.

[0029]

[0030] It is the deformed field output by each subnetwork, where n is the number of cascaded subnetworks.

[0031] The training method of the MMDC-Net of this invention is as follows:

[0032] Data preparation: acquire positive and negative phase dMRI data, preprocess the dMRI data using eddy current correction algorithm; extract B0 images from dMRI data and normalize them, while simultaneously calculating FOD images from dMRI data.

[0033] During training, the B0 and FOD images with positive and negative phases are concatenated and input into the neural network. A stochastic gradient descent strategy with momentum is used, and the learning rate is set to 0.0001. To ensure neural network convergence, a total of 500 epochs are trained, with each epoch consisting of 100 steps and a batch size of 1. Unsupervised learning is employed, with all subnetworks cascaded and trained end-to-end. The optimization objective of the training is the similarity between the B0 and FOD deformed images with positive and negative phases, thus requiring two similarity loss functions:

[0034] in, It is a rigid body deformation field. It is a dense deformation field. and This is a distorted FOD image. and It is the deformed B0 image, λ sim This is the weight hyperparameter used to balance the similarity loss functions of FOD and B0. Here, the similarity loss function for FOD images uses mean squared error; the similarity loss function for B0 images uses cross-correlation. For multimodal optimization tasks, the weight hyperparameter λ... sim Achieving a balance between the random iterations during network training and the loss function is difficult. Therefore, a dynamic weighted average (DWA) algorithm is introduced during optimization to dynamically adjust the weights based on the iteration rate of different loss functions during training. Thus, for k tasks, the weights λ... k It can be defined as:

[0035]

[0036] Where t is the iteration coefficient, and T is the temperature parameter used to control the task weights. Additionally, to ensure the smoothness of the generated deformation field, a regularization term needs to be added:

[0037]

[0038] Therefore, the total loss function is:

[0039]

[0040] λ1 and λ2 are the order of magnitude hyperparameters used to balance the order of magnitude difference between the two similarity loss functions, and are selected based on debugging / multiple experiments.

[0041] In this invention, for data that requires magnetic susceptibility distortion correction, the distortion field of dMRI data can be obtained through a single forward propagation, which greatly improves the efficiency of dMRI preprocessing.

[0042] This invention proposes a novel unsupervised learning-based distortion correction system for magnetic susceptibility. It employs a coarse-to-fine deep learning network—a multimodal distortion correction network, denoted as MMDC-Net—composed of a rigid transform network and a dense transform network. Unlike existing methods that use only B0 or ​​fiber orientation distribution (FOD) images, this invention integrates rich information from both modalities and inputs it into the network. Specifically, a rigid transform network is first used to learn the linear misalignment between two images encoded with opposite phases. Then, a cascaded dense transform network is designed to learn the distortion field from the images of both modalities through a dual-stream encoder and a multimodal attention module. Furthermore, a dynamic weighted averaging (DWA) method is used to simultaneously optimize the multimodal images. During training, the input images are B0 and FOD images with positive and negative phases, and the output is the corresponding deformation field used to correct the distortion. During inference, the input images are the B0 and FOD images with positive and negative phases from the dMRI image to be corrected, and the output is the deformation field. Experimental results demonstrate that the present invention can effectively correct the magnetic susceptibility distortion of dMRI, significantly outperforming other comparative methods, effectively improving the efficiency of dMRI preprocessing, and enhancing the accuracy of subsequent analyses such as fiber tract tracking and quantitative analysis. Attached Figure Description

[0043] Figure 1 Magnetic susceptibility distortion correction process.

[0044] Figure 2 Structural details of the distortion correction network.

[0045] Figure 3 The correction effects of different distortion correction methods in the brainstem region. Detailed Implementation

[0046] The embodiments of the present invention will be described in detail below, but the scope of protection of the present invention is not limited to the embodiments described.

[0047] This invention employs, as follows Figure 1 The illustrated magnetic susceptibility distortion correction process is as follows:

[0048] During acquisition, dMRI images of the same subject in both positive and negative phases are scanned, denoted as I. + and I - Subsequently, eddy current correction was performed on the original dMRI images, and FOD images with positive and negative phases were estimated to obtain the eddy current-corrected B0 and FOD images as inputs to the subsequent correction network.

[0049] During training, the B0 and FOD images are first normalized to minimax. Then, as... Figure 2As shown, the B0 and FOD images are concatenated along the channel dimension to form inputs of [N,2,H,W,T] and [N,12,H,W,T], respectively, where N is the batch size and H, W, and T are the three dimensions of the input image. For the FOD image, we use the first 6 channels to enhance sufficient feature information. The correction algorithm is trained using the ADAM optimizer with a learning rate set to 2e-4.

[0050] like Figure 2 As shown, the training process includes three loss functions: B0 similarity loss function L... simB0 FOD similarity loss function L simFOD and deformation field regularization L reg In addition, to balance the convergence speed of the two similarity losses, a Dynamically Weighted Average (DWAM) algorithm was added, using temperature T to balance the convergence speed of the two loss functions. The weight parameter of the similarity loss was set to 1e-4, and the regularization parameter was set to 5e-3. The temperature T of the DWAM method was set to 2. During training, 400 epochs were used, with each epoch containing 100 steps. Finally, a trained neural network was obtained to correct the magnetic susceptibility distortion of dMRI.

[0051] During testing, B0 and FOD images for both positive and negative phases were obtained using the same preprocessing method as during training. These images were then input into a neural network with frozen parameters to generate a 3D deformation field. Subsequently, this deformation field was used to deform the positive-phase dMRI image to obtain a positive-phase corrected image; the deformation field was then reversed to deform the negative-phase deformation field to obtain a negative-phase corrected image. Finally, the corrected dMRI images for both positive and negative phases were averaged to obtain a merged, distortion-free image.

[0052] Table 1. Correction effects of different correction methods on HCP data

[0053]

[0054] Table 1 shows the correction performance of different correction methods on the HCP dataset. To quantify the distortion correction performance of the method of this invention, two metrics were used to evaluate the differences between dMRI images of two phases after correction by different methods. The mean squared deviation (MSD) of the fractional anisotropy (FA) generated by the tensor model of the dMRI data was measured. The minimum angular difference (MAD) of the major fiber directions of the FOD between two opposite PEs was also measured. The angular difference was obtained by calculating the minimum angle between the major fiber of the positive PE FOD (with the maximum amplitude) and the top three fibers of the negative PE FOD (sorted by amplitude). To better suit the needs of practical applications, special attention was paid to local brain regions that still exhibit significant residual distortion after processing by other correction methods, including the brainstem and corpus callosum. The brainstem region is crucial for studying the pathological changes of brain diseases such as Alzheimer's disease, and this region is often severely distorted in dMRI images. Therefore, studying the correction effect in this region has important clinical significance, and the Pons region of the brainstem was selected as the region of interest (ROI) in the experiment. The corpus callosum (cc) is the largest white matter pathway and neural tract in the brain, and it is often significantly distorted in dMRI images. Correction of the corpus callosum is crucial for white matter tract fiber tracking. Therefore, the corpus callosum region was designated as a Region of Interest (ROI). The brainstem and corpus callosum regions were segmented from T1 images using Freesurfer 6.0, and then linearly mapped from T1 and dMRI images to the dMRI image space using linear registration.

[0055] like Figure 3 The figure illustrates the correction effects of different distortion correction methods—topup, DrC-Net, and the distortion correction method of this invention—on the brainstem region. As can be seen from the figure, even with topup or DrC-Net correction, the FOD fiber tract trajectories generated from the two opposite phase encoding directions remain irregular and contain significant noise, which may lead to erroneous analysis in subsequent analyses. In contrast, the method of this invention can completely correct the distortion of both phase images, resulting in clear and regular fiber tract trajectories in the brainstem region.

[0056] Besides achieving better correction results, as a learning-based distortion correction method, this invention corrects a pair of dMRI data on a GPU in approximately 3 seconds, while traditional iterative methods like topup and FODReg require over an hour. Furthermore, DrC-Net first requires coarse correction via topup, which consumes additional time.

Claims

1. A multimodal dMRI image distortion correction system based on unsupervised deep learning, characterized in that, Specifically include: The constructed deep learning network, a multimodal distortion correction network (MMDC-Net), is a coarse-to-fine model. This network consists of cascaded rigid transformation networks and dense transformation networks; it also includes a designed cascaded training strategy module. Among these components: (1) The rigid transformation network is used to correct linear misalignment between two phase-encoded direction images; the rigid transformation network is a convolutional neural network, including an encoder, a multilayer perceptron, and a spatial transformation network; the encoder is responsible for extracting feature information from the stitched B0 and FOD images, wherein the B0 and FOD images respectively include images with two opposite encoding directions; the subsequent multilayer perceptron maps the feature map learned by the encoder to a set of rigid transformation parameters μ: μ=(θ x ,i y ,i z ,t x ,t y ,t z ), (1) Where x, y, and z are the three coordinate directions, and θ is the coordinate direction. x ,θ y ,θ z These represent the rotation angles of the corresponding coordinate axes, t and t'. x ,t y ,t z These represent the translation distances of the corresponding coordinate axes. According to the topup settings, the translation parameter t is set to 0 in the phase encoding because it is impossible to distinguish the subject's translation from the deformation offset caused by magnetic susceptibility distortion in this direction. The geometric transformation matrix R is calculated based on the learned rigid transformation parameter μ. x R y R z : Finally, calculate the affine transformation matrix: A = R x R y R z T; By using a spatial transformation network, the learned affine transformation matrix is ​​applied to the negative phase B0 and FOD matrices to correct the linear misalignment between the two phase-encoded direction images; (2) The dense transform network is used to correct the non-rigid distortion of images in two opposite encoding directions. Specifically, a U-shaped neural network is constructed, which includes a dual-stream encoder, a decoder, a skip connection layer connecting the encoder and the decoder, and a multimodal attention module. The dual-stream encoder processes B0 and FOD information respectively: the B0 branch stitches the positive and negative phase B0 images into a 2-channel image as input; the FOD branch stitches the positive and negative phase FOD images into a 12-channel image as input; the dual-stream encoder extracts the deep information of the image step by step while maintaining the spatial resolution. After processing by the encoder, the decoder fuses feature information from different modalities through several multimodal attention modules of different scales, thereby achieving more accurate distortion correction. Each multimodal attention module receives FOD and B0 feature information from different scales of the encoder and the output feature map of the decoder at the previous scale, aggregates the fused information at the corresponding scale, and performs upsampling. In the final stage of the decoder, a 3D convolutional layer is used to estimate a three-dimensional non-rigid deformation field from the decoder features. This deformation field only retains the deformation in the phase encoding direction, thereby achieving magnetic susceptibility distortion correction. (3) The cascaded training strategy module is used to improve the accuracy of the deep learning network MMDC-Net in estimating non-rigid deformation fields. For the dense transformation network, its input is the positive and negative phase FOD and B0 image pairs to be corrected, and the output is a three-dimensional deformation field. The spatial transformation module is used to apply the learned deformation field to the image pairs to be corrected respectively: the positive deformation field is applied to the positive phase B0 and FOD images to obtain the corrected positive phase image; the magnitude of the deformation field is kept unchanged, but the direction is reversed to obtain the reverse deformation field, which is applied to the negative phase B0 and FOD images to obtain the corrected negative phase image; then, the corrected positive and negative phase B0 and FOD image pairs can be input into the dense transformation network again to learn the residual error between them to obtain a more refined deformation field; after a series of cascaded operations, the difference between the two phase images tends to converge, and the deformation fields output by this series of networks are added together to obtain the final deformation field of the original input image to be corrected: It is the deformed field output by each subnetwork, where n is the number of cascaded subnetworks.

2. The multimodal dMRI image distortion correction system according to claim 1, characterized in that, In the rigid transformation network: The encoder contains four convolutional layers, each of which also includes a batch normalization layer. The feature maps output by each convolutional layer have 16, 16, 32, and 32 channels, respectively, and the feature map sizes are 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original image size, respectively. The multilayer perceptron includes two consecutive fully connected layers; The spatial transformation network is used to apply rigid or non-rigid deformation fields to an image for deformation processing.

3. The multimodal dMRI image distortion correction system according to claim 1, characterized in that, In the dense transform network, each branch of the dual-stream encoder includes four convolutional blocks, each consisting of a 3×3×3 convolutional layer and a LeakyReLU activation layer; using small convolutional kernels and a convolutional stride of 2.

4. The multimodal dMRI image distortion correction system according to claim 1, characterized in that, The multimodal attention module in the dense transform network includes a region attention module and a channel attention module, which are used to aggregate feature information from the B0 and FOD images in the dual-stream encoder during the decoding stage; wherein: The region attention module is responsible for dynamically estimating and allocating attention weights for different regions of different modalities during distortion correction. It generates an attention map representing the contributions of different modal regions by capturing latent features from the encoder and decoder. In this module, encoder and decoder feature maps of the same scale are first concatenated and fed into a convolutional layer to extract region features. Then, the output dual-channel feature map is passed through a softmax layer to obtain an attention map, where the two channels correspond to the contributions of B0 and FOD in each region, respectively. This attention map is represented as follows: in, This is the attention map learned by the region attention module, where i represents the i-th layer of the network, m represents different modalities, including B0 and FOD images, [·; ·] represents the concatenation operation, and e i and d i f represents the feature maps of the encoder and decoder, respectively. 3×3×3 This represents a convolutional layer with a kernel size of 3×3×3; M is the set of different modalities, i.e., {B0,FOD}. Subsequently, the learned feature map is multiplied onto the input encoder features on a volumetric basis to obtain the attention-weighted feature output. The channel attention module is responsible for estimating the overall contribution of different modal inputs and dynamically adjusting the weights of each channel. This module focuses more on the contribution of the overall features. First, the decoder features and the output of the region attention module are concatenated. Then, average pooling and max pooling operations are used to generate two different feature vectors. These two vectors are then input into a multilayer perceptron with shared weights. The output of the multilayer perceptron is activated by a sigmoid function, mapping the resulting weight matrix to a value between 0 and 1, representing the degree of contribution of different modal inputs to the output features in the current layer. Where σ represents the sigmoid activation function, f MLP A multilayer perceptron representing shared weights, fr i and d i The output features of the region attention module and the encoder at different scales are represented by AvgPool(·), which represents the average pooling layer. Finally, the channel attention module applies these weights to the feature map output by the region attention module, thereby adjusting the influence of different modal features on the correction effect: 。 5. The multimodal dMRI image distortion correction system according to claim 1, characterized in that, In the cascaded training strategy module, the training method for MMDC-Net is as follows: Data preparation: Acquire positive and negative phase dMRI data, preprocess the dMRI data using the eddy current correction algorithm; extract B0 images from the dMRI data and normalize them, while simultaneously calculating FOD images from the dMRI data; During training, the B0 and FOD images with positive and negative phases are stitched together and input into the neural network. Using a momentum-driven stochastic gradient descent strategy and employing unsupervised learning, all subnetworks are cascaded and trained end-to-end. The optimization objective of the training is the similarity between the positive and negative phases B0 and the FOD deformed image. To achieve this, two similarity loss functions are used: in, It is a rigid body deformation field. It is a dense deformation field. and This is a distorted FOD image. and It is the deformed B0 image, λ sim This is the weight hyperparameter used to balance the similarity loss functions of FOD and B0; here, the similarity loss function for FOD images uses mean squared error, while the similarity loss function for B0 images uses cross-correlation.

6. The multimodal dMRI image distortion correction system according to claim 5, characterized in that, For multimodal optimization tasks, the weight hyperparameter λ sim Achieving a balance between the two is difficult due to the random iterations during network training; therefore, a Dynamic Weighted Average (DWA) algorithm is introduced during optimization to dynamically adjust the weights based on the iteration rate of different loss functions during training; for k tasks, the weights λ... k Defined as: Where t is the iteration coefficient, and T is the temperature parameter used to control the task weights; at the same time, in order to ensure the smoothness of the generated deformation field, a regularization term also needs to be added: Therefore, the total loss function is: Here, λ1 and λ2 are order of magnitude hyperparameters used to balance the order of magnitude differences between the two similarity loss functions.

Images