Dual phase-cycled cardiac magnetic resonance imaging artifact removal method guided by magnetic susceptibility field map

Abstract

The application relates to a dual-phase cycle cardiac magnetic resonance imaging artifact removal method based on magnetic susceptibility field map guidance, which comprises the following steps: acquiring two cardiac movie images under different phase increments as image domain complex images; pre-processing the two image domain complex images and then inputting the pre-trained dual-branch deep learning model to remove artifacts to obtain artifact-free cardiac movie images, wherein the dual-branch deep learning model comprises a magnetic susceptibility field map estimation network and a conditional image reconstruction network, the artifact removal step comprises estimating a pixel-level magnetic susceptibility field map based on the pre-processing result by using the magnetic susceptibility field map estimation network; and based on the pre-processing result and the pixel-level magnetic susceptibility field map, the conditional image reconstruction network removes artifacts through a conditional strategy to output the artifact-free cardiac movie images. Compared with the prior art, the application has the advantages of improving the model physical interpretability and the magnetic susceptibility artifact suppression robustness.

Description

Technical Field

[0001] This invention relates to the field of medical magnetic resonance imaging, and in particular to a method for removing artifacts in dual-phase circulatory cardiac magnetic resonance imaging guided by magnetic field maps. Background Technology

[0002] Cardiac magnetic resonance cine imaging enables dynamic assessment of cardiac anatomy and function by recording the contraction and relaxation of the heart throughout the entire cardiac cycle. Balanced Steady-State Free Precession (bSSFP) sequences have become the mainstream sequence for clinical cardiac cine imaging due to their high signal-to-noise ratio, high temporal resolution, and excellent blood pool-myocardial contrast.

[0003] However, bSSFP sequences are extremely sensitive to local magnetic field homogeneity. When substances with significant differences in magnetic susceptibility exist within the imaging region (such as air-tissue interfaces, blood flow, or metal generators in implanted cardiac electronic devices), local resonant frequencies shift, leading to decreased signal intensity or even complete phase loss, which manifests as low-signal black band artifacts or signal voids on the image. As clinical MRI field strength increases to 3T and above, the problem of magnetic susceptibility artifacts becomes increasingly severe, and conventional bSSFP cardiac cine imaging often fails to meet the clinical needs for quantitative analysis of cardiac structure and function.

[0004] Phase cycling technology shifts the spatial position of black band artifacts in images by changing the radio frequency phase increment of the bSSFP sequence, as shown in patent application CN119090991A. Theoretically, acquiring multiple images with different phase increments and performing joint calculations can suppress black band artifacts to a certain extent. However, traditional phase cycling schemes have the following inherent defects: (1) Scanning time is multiplied: In order to obtain sufficient artifact complementary information, it is usually necessary to acquire 4 to 6 sets of images with different phase increments, which leads to a significant increase in breath-holding time, decreased patient tolerance, and increased risk of motion artifacts; (2) Introduction of blood flow artifacts: Phase cycling changes the radio frequency phase increment, which can easily introduce new blood flow artifacts in the joint images, thus reducing image quality; (3) Insufficient suppression of strong magnetic sensitive artifacts: For areas with non-uniform magnetic fields (such as the lung-heart interface) or extreme magnetic field distortions (such as near implantable electronic devices in the heart, where the frequency shift can reach hundreds of Hz), simply relying on the pixel-level average or maximum value projection of multiple images cannot recover the signal cavity area that is completely out of phase, and will blur anatomical details.

[0005] In recent years, deep learning-based post-processing techniques have been introduced into the field of cardiac MRI artifact suppression. Current techniques primarily employ end-to-end convolutional neural networks, using images with artifacts as input and images without artifacts as output, and undergoing supervised training with a large amount of paired data. However, existing deep learning solutions suffer from the following fundamental limitations: (1) Purely data-driven, lacking physical interpretability: Existing networks treat artifact suppression as a general image domain mapping problem, using only the appearance statistical features of the image for black-box mapping, without explicitly modeling the physical root of magnetic susceptibility artifacts—that is, local magnetic field inhomogeneity. This leads to the network being prone to anatomical illusions or unpredictable reconstruction biases when faced with extreme artifacts outside the training data distribution, such as strong signal voids around metal or severe black band artifacts, resulting in insufficient clinical robustness.

[0006] (2) Failure to utilize the physical complementarity of biphase loops: Although some existing technologies use biphase loop images as input, they only treat them as two images with complementary artifact appearances and perform abstract fusion at the feature level, such as simple channel stitching or cross-image attention mechanisms, without deeply exploring the effects of different phase increments on the physical complementarity of the two images. The key physical correlation lies in the differentiated signal response. Therefore, the network cannot distinguish between signal loss caused by magnetic field distortion and normal low-signal tissue, resulting in limited reconstruction quality in areas with severe artifacts.

[0007] (3) Reconstruction without physical prior guidance is unreliable: Existing technologies lack a mechanism to inject physical field map information as a prior condition into the artifact suppression process, which makes it difficult for the network to simultaneously achieve artifact suppression, resolution restoration and dissection fidelity under the dual degradation of k-space undersampling and magnetic susceptibility artifacts.

[0008] Therefore, there is an urgent need for a new artifact removal technique to address the problems of existing deep learning reconstructions lacking physical interpretability, insufficient robustness in suppressing magnetically sensitive artifacts, and failure to fully utilize the physical complementary information of two-phase cycles. Summary of the Invention

[0009] The purpose of this invention is to provide a method for removing artifacts in dual-phase cyclic cardiac magnetic resonance imaging based on magnetic field map guidance, which improves the quality and reliability of cardiac cine imaging.

[0010] The objective of this invention can be achieved through the following technical solutions: A method for artifact removal in dual-phase circulatory cardiac magnetic resonance imaging guided by magnetic field maps includes the following steps: Two cardiac cine images with different phase increments were acquired and used as complex images in the image domain. Two complex images in the image domain are preprocessed and then input into a pre-trained dual-branch deep learning model for artifact removal, resulting in an artifact-free cardiac cinematic image. The dual-branch deep learning model includes a magnetic susceptibility field estimation network and a conditional image reconstruction network. The artifact removal steps include... Based on the preprocessing results, a magnetic susceptibility field estimation network is used to estimate the pixel-level magnetic susceptibility field map. Based on the preprocessing results and pixel-level magnetic susceptibility field maps, the conditional image reconstruction network removes artifacts through a conditional strategy, outputting an artifact-free cardiac cinema image.

[0011] Furthermore, the acquisition steps for the two cardiac cine images with different phase increments include: Set the scanning parameters, wherein the scanning parameters include the phase increment; Under the scanning parameters, a double-phase cyclic sequence is executed within one breath-hold using multiple receiving coils to collect partial k-space data under different phase increments; Based on partial k-space data under different phase increments, a parallel imaging reconstruction algorithm is used to perform multi-coil joint reconstruction to obtain two cardiac cine images under different phase increments.

[0012] Furthermore, a k-space acceleration strategy is adopted to collect partial k-space data under different phase increments. The k-space acceleration strategy is as follows: 1D undersampling is implemented in the phase encoding direction, a set area in the center of k-space is retained as an automatic calibration signal, full sampling is performed, uniform undersampling is adopted in the outer area, and an equivalent acceleration factor is set.

[0013] Furthermore, the phase increment difference between the two cardiac film images is 180°.

[0014] Furthermore, the two complex images in the image domain include dual channels for real and imaginary parts, or dual channels for amplitude and phase, and the preprocessing steps include: One of the complex images in the image domain is used as the reference image, and the other complex image in the image domain is rigidly registered. Based on rigid registration, the image amplitude of the two complex images in the image domain is normalized to obtain the two normalized complex images in the image domain. For the two normalized complex images in the image domain, their real and imaginary parts or amplitude and phase are preserved, and they are concatenated into an input tensor containing multiple channels, represented as follows: Complete the preprocessing process, in which, To take the real part of the complex number, To take the imaginary part of a complex number, , These are two complex images in the image domain.

[0015] Furthermore, the magnetic susceptibility field estimation network utilizes local bias frequencies from two complex images in the image domain. and phase increment physical association , Inversely derived pixel-level magnetic susceptibility field map ,in For pixels, The repetition time.

[0016] Furthermore, the magnetic field estimation network adopts an encoder-decoder architecture based on a multi-layer 3D U-Net. The encoder architecture extracts multi-scale features from the preprocessed results through multi-layer convolution and downsampling, while the decoder architecture recovers spatial resolution through upsampling and skip connections. Finally, after scaling and spatial smoothing constraints at the output layer, a pixel-level magnetic field map is output. ,in In pixels.

[0017] Furthermore, the conditional image reconstruction network employs an encoder-decoder architecture.

[0018] Furthermore, the conditionalization strategy includes at least one of the following: (1) The preprocessing result and the pixel-level magnetic field map are spliced ​​along the channel dimension in the input layer of the conditional image reconstruction network, and the splicing result is used as the joint input of the conditional image reconstruction network; (2) In the encoder-decoder architecture, a cross-attention mechanism is inserted at the same-scale connection between each encoder and decoder layer, except at the connection between the first encoder layer and its same-scale decoder layer. The processing steps of the encoder-decoder architecture include: The preprocessed results are sequentially passed through a multi-layer encoder for downsampling and feature extraction to obtain image features at each corresponding scale of each layer. , For the first Image features output by the layer encoder; Based on the image features at the minimum scale, the features are sequentially upsampled and transformed through multiple layers of decoders to obtain the decoded features at each corresponding scale. Meanwhile, in each layer, the decoded features are... Projecting these into key vectors and value vectors respectively, the pixel-level magnetic susceptibility map output by the magnetic susceptibility map estimation network at the corresponding scale is obtained. The query vector is obtained by convolutional linear projection. A spatial cross-attention weight map is calculated through a cross-attention mechanism, where the smallest-scale image feature is the image feature output by the last encoder layer. The cross-attention mechanism is as follows: , In the formula, These are the query vector, key vector, and value vector, respectively. For spatial cross-attention weights, For feature dimensions.

[0019] Furthermore, the dual-branch deep learning model employs an end-to-end joint training strategy or a phased training strategy during training, wherein the training steps of the phased training strategy include: Get training labels: Multiple sets of fully sampled complex images in the image domain were acquired, and a coarse pixel-level magnetic susceptibility field map was obtained by fitting a bSSFP signal model through pixel-by-pixel nonlinear least squares fitting. Reference image , serving as a weak supervision label for the magnetic susceptibility field estimation network; Multiple sets of fully sampled complex images in the image domain under phase increment are acquired and averaged to obtain artifact-free labeled images for the conditional image reconstruction network. ; Phased training strategy and loss function: The first stage of training involves fixing the parameters of the image reconstruction network and training the magnetic susceptibility map estimation network separately. This enables the magnetic susceptibility map estimation network to learn how to invert pixel-level magnetic susceptibility maps from the magnetic susceptibility map estimation network. The loss function used during training is: , , , In the formula, For pixel-level magnetic susceptibility map physical consistency loss, Pixel-level magnetic susceptibility map output by the network for estimating magnetic susceptibility map Compared with the reference figure L1 supervision loss, The total variational smoothing loss is used to constrain the continuity of the field diagram space. These are the weighting coefficients. For image horizontal First-order finite-difference gradient operator in the direction, For vertical along the image First-order finite difference gradient operator in direction, subscript , For pixel space coordinate index, where Indicates row index, Indicates column index; The second stage of training involves fixing the parameters of the magnetic susceptibility map estimation network and training the conditional image reconstruction network separately, enabling the conditional image reconstruction network to estimate the magnetic susceptibility map at the pixel level. Image reconstruction is performed conditionally, where the loss function used during training is: , In the formula, To rebuild the losses, Artifact-free heart movie images and artifact-free label images output by the conditional image reconstruction network. L1 norm difference Artifact-free heart movie images and artifact-free labeled images Structural similarity loss, Artifact-free heart movie images and artifact-free labeled images Perceptual loss, , These are the weighting coefficients; The third stage of training involves jointly optimizing the parameters of the unfrozen conditional image reconstruction network and the magnetic susceptibility field map estimation network at a set learning rate. The joint loss function during this optimization process is as follows: , In the formula, For joint losses; Training terminates when the set termination condition is met, resulting in a well-trained dual-branch deep learning model. , These are the weighting coefficients.

[0020] Compared with the prior art, the present invention has the following beneficial effects: (1) Improve imaging quality and reliability: This invention utilizes the physical complementarity of dual-phase cyclic images to explicitly estimate pixel-level magnetic susceptibility field maps, and uses these physical field maps as prior conditions to guide image reconstruction, thereby suppressing magnetic susceptibility artifacts in cardiac magnetic resonance cine imaging, achieving deep synergy between physical and data-driven approaches, and improving the quality and reliability of cardiac cine imaging.

[0021] (2) Breakthrough in physical interpretability: Compared with the existing pure data-driven black box mapping, this invention explicitly estimates the pixel-level magnetic susceptibility field map, links deep learning reconstruction with the MRI physical signal formation mechanism, so that the network's suppression of magnetic susceptibility artifacts has a clear physical basis, and improves the model's generalization ability and clinical credibility in conventional black band artifacts and extreme metal artifact scenarios.

[0022] (3) Precise suppression of the root cause of magnetic susceptibility artifacts: The Field-Net of this invention directly encodes the frequency shift caused by the local magnetic field shift. Recon-Net uses this physical field map as a priori to perform conditional reconstruction, so that the network's repair of signal holes and structural distortions conforms to the magnetic field distribution law, which is better than the traditional scheme that only relies on the similarity of image appearance.

[0023] (4) Compatible with clinical parallel imaging workflow: This invention directly receives complex images of the image domain after preliminary reconstruction by standard parallel imaging algorithms such as SENSE / GRAPPA as input, and seamlessly connects with existing clinical reconstruction pipelines.

[0024] (5) Field map assists clinical decision-making: The magnetic field map output by this invention not only serves the reconstruction process, but can also be directly used for clinical display, helping physicians to quantitatively assess the range of magnetic field distortion and provide objective reference for functional analysis of high-risk reconstruction areas. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a diagram of the dual-branch deep learning model architecture of the present invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0027] This embodiment provides a method for artifact removal in dual-phase cyclic cardiac magnetic resonance imaging (MRI) guided by magnetic field maps. This method is based on magnetic field map-guided bi-branch collaborative reconstruction. By explicitly estimating the magnetic field map and using this physical field map as a priori condition to guide image restoration, it achieves deep synergy between physical and data-driven approaches, improving the quality and reliability of cardiac cine imaging. Specifically, as... Figure 1 As shown, the method includes the following steps: Step S1: Acquisition and preliminary reconstruction of dual-phase cyclic under-sampling data.

[0028] Using multiple receiving coils, a double-phase cyclic bSSFP sequence is executed within a single breath-hold to acquire partial k-space data of two cardiac cine images with different phase increments. Subsequently, a parallel imaging reconstruction algorithm (such as SENSE, GRAPPA, or k-tGRAPPA) is used to perform multi-coil joint reconstruction on the undersampled k-space data to obtain complex images in the image domain corresponding to the two phase-cyclic images, represented as dual channels of real and imaginary parts (or equivalent amplitude and phase).

[0029] Preferably, the phase increment difference between the two images is 180° (e.g., =0°, =180°), to maximize the complementarity of magnetic susceptibility artifact distribution; or, adaptively select the phase increment combination according to the imaging scene to maximize the signal difference between the two images in areas with severe magnetic susceptibility artifacts.

[0030] k-space acquisition methods include, but are not limited to: acquiring two images using the same trajectory portion, such as retaining 25% full sampling of the center in the phase encoding direction and undersampling of the periphery; or using staggered k-space trajectories to combine the two images to achieve accelerated scanning.

[0031] In this embodiment, a 3T clinical MRI scanner was used on 28 healthy volunteers. Cardiac short-axis cine imaging was performed using a modified dual-phase cyclic bSSFP sequence, with the following specific scanning parameters: Phase increment: =90°, =270°; repetition time TR=3.0ms, echo time TE=1.5ms, flip angle =60°; Matrix size = 192×192, Field of view (FOV) = 300mm×300mm, Slice thickness 8mm; Temporal resolution: 25 frames per cardiac cycle; k-space acceleration strategy: 1D undersampling in the phase encoding direction, retaining 25% of the fully sampled data (autocalibrated signal, ACS) at the center of k-space, with uniform undersampling at the periphery, equivalent acceleration factor R = 4; Breath-holding time: approximately 8-10 seconds, completing the acquisition of two cine images. After acquisition, the k-tGRAPPA algorithm is used to perform multi-coil joint reconstruction of the two undersampled k-space complex data to obtain the cardiac cine image, i.e., the aforementioned complex image in the image domain. and Each image domain complex image contains two channels, real and imaginary, with a spatial dimension of 192×192 and a temporal dimension of 25 frames.

[0032] Step S2: Image preprocessing.

[0033] The following preprocessing steps are performed on the image domain complex image obtained from the initial reconstruction: Image registration: using one of the image domains as a complex image (e.g.: Using a reference image, another complex image in the image domain (e.g.: Rigid registration is performed to compensate for minute displacements during the breath-holding process; Normalization: Based on rigid registration, the amplitudes of the two complex images in the image domain are normalized to [0,1]. Complex channel construction: retain the real and imaginary dual channels, or convert them into amplitude and phase maps to construct the network input tensor.

[0034] Step S3: Construct a dual-branch deep learning model.

[0035] like Figure 2 The parallel structure and feature interaction method of the magnetic field estimation branch (Field-Net) and the conditional image reconstruction branch (Recon-Net) in the dual-branch deep learning model are shown.

[0036] Branch A: Field-Net network for estimating magnetic susceptibility maps Physical basis: In a bSSFP sequence, the steady-state signal strength affects the local bias frequency. (and Proportional and radio frequency phase increment It has a clear resolution dependency relationship , For pixels, For the same anatomical location, when... When they exist, two different The image will present predictable and distinguishable signal differences. Field-Net utilizes this physical correlation to invert pixel-level magnetic susceptibility maps from complex images in the image domain based on biphase cycles. .

[0037] Network Structure: An encoder-decoder architecture, such as the 3DU-Net structure, is adopted. The input is a multi-channel input tensor (dimension [B,C,H,W,T], where C≥4, composed of the real and imaginary parts of the two images, or represented by amplitude and phase) obtained by concatenating two complex images in the image domain based on biphase cyclic imaging. The encoder extracts multi-scale features through multi-layer convolution and downsampling; the decoder restores spatial resolution through upsampling and skip connections.

[0038] Output layer: Single-channel output, appropriately scaled to represent pixel-level magnetic susceptibility field maps. The unit can be Hz. Spatial smoothing constraints can be optionally applied after the output layer to suppress high-frequency noise.

[0039] In this embodiment, the encoder has [32, 64, 128, 256] channels, and each layer contains two sets of 3×3×3 convolutions, GroupNorm and ReLU activations. Downsampling uses max pooling with a stride of 2×2×1. The decoder performs symmetrical upsampling and fuses multi-scale features through skip connections. The output layer is followed by a 3×3×1 average pooling layer for spatial smoothing to suppress high-frequency noise. The pixel-level magnetic susceptibility field map output range is mapped to ±500Hz (the typical magnetic susceptibility offset range of conventional 3T cardiac imaging) through a learnable scaling factor.

[0040] Branch A utilizes two bSSFP sequences with different phase increments to estimate pixel-level magnetic susceptibility maps explicitly via a convolutional neural network, based on the differentiated signal responses to local magnetic field shifts. Physically, the intensity and spatial distribution of magnetic field distortion caused by magnetic susceptibility artifacts are quantified.

[0041] Branch B: Conditional Image Reconstruction Network (Recon-Net) Network structure: The encoder-decoder architecture is adopted using backbone networks such as 3DU-Net, V-Net or Transformer, with both input and output being image domain data.

[0042] Inputs: (a) a preliminary reconstructed image domain complex image (real / imaginary or amplitude / phase, multi-channel) based on a two-phase loop; (b) a pixel-level magnetic susceptibility map output from branch A. .

[0043] Output: A heart movie image with no magnetic sensitivity artifacts and no undersampling artifacts after normalization by 1×1×1 convolution and Sigmoid activation.

[0044] In branch B, a conditional injection mechanism is used to inject the physics map. To effectively embed the image reconstruction process, at least one of the following conditionalization strategies should be employed: (1) Channel dimension stitching: stitching pixel-level magnetic susceptibility field maps The original image domain complex image is concatenated with the original image domain complex image along the channel dimension at the input layer and used as the joint input of Recon-Net. Recon-Net removes artifacts according to its processing logic. (2) Cross-Attention Conditioning: Pixel-level magnetic susceptibility map The field image is encoded as a query vector, and Recon-Net features are used as key-value pairs. A cross-attention mechanism focuses the reconstructed features on the physical root cause regions of artifacts. Specifically, this embodiment uses a 3DU-Net backbone with a 4-layer encoder and a 4-layer decoder structure. Each encoder layer contains residual blocks with [32, 64, 128, 256 channels]. Cross-attention conditionalization units are inserted at the scale connections of the encoder-decoder layers 2, 3, and 4 for cross-attention conditionalization. In this case, the processing steps of the encoder-decoder architecture include: The preprocessed input tensor is sequentially passed through four encoder layers for downsampling and feature extraction, yielding image features at each corresponding scale. ( ), For the first The image features output by the layer encoder are based on the minimum scale image features. The features are sequentially upsampled and transformed through four layers of decoders to obtain the decoded features at each corresponding scale. ( Meanwhile, in each layer, the decoded features are... Projecting these vectors as key and value vectors respectively, the pixel-level magnetic susceptibility field map at the corresponding scale in branch A will be generated. The vector is linearly projected through convolution to form a query vector, and attention weights are calculated using a spatial cross-attention mechanism. , in, Derived from pixel-level magnetic field map features, , Derived from signal reconstruction features, As a feature dimension. This mechanism enables the reconstruction network to... Intense region adaptive enhancement feature response, in Maintaining standard reconstruction behavior in flat areas to achieve physical field mapping Pixel-level adaptive guidance for image restoration.

[0045] Branch B takes the initially reconstructed biphase cyclic image domain complex image as input and the Field-Net output as input. The field map serves as a conditional prior, and physical information is embedded into the image reconstruction process through a feature modulation mechanism to recover an artifact-free image.

[0046] Step S4: Construct the joint loss function.

[0047] Define the total loss function as .in, For image reconstruction loss, L1 norm, L2 norm, or SSIM loss is used to measure the difference between the reconstructed image (i.e., the artifact-free heart movie image output by the conditional image reconstruction network during training) and the artifact-free label. To perceive the loss, the reconstructed image and the labeled image are input into the network model respectively, and the L1 difference of the multi-stage feature maps is calculated to preserve high-frequency texture and anatomical details. The pixel-level magnetic susceptibility map physical consistency loss includes: spatial smoothness loss. Such as TotalVariation loss or Laplacian regularization, constraint field plot Within the anatomically continuous region, physical gradual continuity is satisfied to avoid unreasonable jumps; optional physical deviation penalty: if a theoretical field map reference based on the bSSFP signal equation exists, such as a coarse solution obtained through multiple sets of phase cycle analytical methods. This increases the L2 bias loss between the estimated field map and the reference field map.

[0048] Step S5: Training the dual-branch deep learning model.

[0049] The training process mainly includes the following steps: Training label acquisition: Field-Net labels (reference image) Using data from the same batch of healthy volunteers, 12 sets of fully sampled image domain complex images based on biphase cyclic imaging were collected. A rough reference image was obtained through analytical solving or iterative fitting of multiple sets of images. , as a weakly supervised label for Field-Net; Recon-Net labels (artifact-free images): The average result of SENSE reconstruction of 5 fully sampled images with phase increments around 180° is used as the artifact-free label. For example, multiple sets of phase increments around 180° (such as 120°, 150°, 180°, 210°, and 240°) can be averaged to obtain artifact-free label images. .

[0050] Training strategy: End-to-end joint training, with Field-Net and Recon-Net propagating forward simultaneously, and the joint loss being backpropagated to update the parameters of both branches. This strategy enables the field map estimation to automatically adapt to reconstruction requirements.

[0051] Preferably, a phased alternating training approach can also be adopted: In the first phase, Recon-Net is fixed as a simplified version or with pre-trained weights, and Field-Net is trained separately until convergence; in the second phase, Field-Net is fixed, and Recon-Net is trained to learn reconstruction conditioned on the field map; in the third phase, both branches are unfrozen for end-to-end fine-tuning. This strategy improves training stability. The training strategy and loss function under phased alternating training are as follows: Phase 1: Field-Net Pre-training (50 epochs): In this phase, the Recon-Net parameters are fixed (randomly initialized but not updated), and Field-Net is trained separately to learn how to invert complex differences from the biphase cyclic image domain. The loss function is: , in, The L1 supervised loss represents the Field-Net output field map and the fitted reference labels, providing a direct physical regression target for Field-Net; The total variational smoothing loss is used to constrain the continuity of the field diagram space. Along the horizontal direction of the image ( The first-order finite difference gradient operator (direction), For the vertical direction of the image ( First-order finite difference gradient operator (direction), subscript , For pixel space coordinate index, where Indicates row index, Indicates column index.

[0052] Phase 2: Recon-Net Training (200 epochs): In this phase, the converged Field-Net parameters are fixed, and Recon-Net is trained to learn image reconstruction conditioned on the field map. The loss function is: , in, To reconstruct the L1 norm difference between the image and the label, To reconstruct the structural similarity loss between the image and the label, To reconstruct the perceptual loss of images and labels.

[0053] Phase 3: End-to-end joint fine-tuning (100 epochs): Unfreeze both branches and perform joint optimization with a small learning rate to adapt field map estimation and image reconstruction to each other. The joint loss function is: , The definitions of each item are the same as above. The Field-Net supervision term is retained at this stage to prevent the joint optimization of the field graph estimation from deviating from the physical reality, while reducing the TV weights to avoid excessive smoothing that suppresses details.

[0054] Optimizer and Hyperparameters: The Adam optimizer was used with an initial learning rate of 1×10⁻⁶. -4 Combined with cosine annealing (to 1×10⁻⁶) -6 Alternatively, a polynomial decay strategy can be used; the batch size can be adjusted according to the GPU memory; training can be stopped early when the combined loss on the validation set no longer decreases after 10 consecutive epochs, thus obtaining a well-trained dual-branch deep learning model.

[0055] Step S6: Reasoning and Clinical Output.

[0056] The trained bi-branch deep learning model was then generalized to patients with cardiac implantable electronic devices (CIEDs). The specific process is as follows: Taking 15 patients with implanted dual-chamber pacemakers or implantable cardioverter defibrillators (ICDs) as an example, the same scanning parameters as in step S1 were set, and undersampled biphase cyclic k-space data with R=4 were acquired. The data were then reconstructed into complex images in the image domain using the kt GRAPPA algorithm. Due to the more severe metal artifacts in CIEDs, the same phase increments (90° and 270°) as in step S1 were used.

[0057] Subsequently, the trained dual-branch deep learning model and the fixed Field-Net and Recon-Net are used for inference. Optionally, mild domain adaptation fine-tuning is performed on the CIED patient data: the last two layers of Field-Net are unfrozen and fine-tuned for 10-20 epochs with physical consistency loss (bSSFP signal equation constraint) to adapt to extreme field map distributions.

[0058] Finally, the results are evaluated. Field-Net output local frequency shift distribution map. This pixel-level magnetic susceptibility field map can be displayed as a pseudo-color overlay on a clinical workstation, intuitively presenting the range and intensity of magnetic field distortion. Results show significantly higher distortion than in typical healthy volunteer scenarios, consistent with the expected strong magnetic susceptibility effect caused by metals. Recon-Net outputs images with artifact suppression, significantly restoring large signal voids around the leads and the original black band artifact areas. Myocardial-blood pool contrast is clear, and the endocardial boundary is identifiable, meeting the needs for quantitative analysis of left ventricular volume and ejection fraction. Clinical decision support: Physicians combine field maps... Determine the reliability of the reconstruction results in areas with heavy artifacts: If the field map shows a certain area A significantly elevated value suggests that residual artifacts may still exist in the area, requiring careful functional quantitative analysis and, if necessary, cross-validation using low-resolution fully sampled images.

[0059] In summary, the above method addresses the technical challenges of insufficient accuracy, poor robustness, and high undersampling dependence in cardiac magnetic resonance imaging (MRI) by introducing magnetic susceptibility maps as a physical prior and combining biphase loops with deep learning fusion. This method is applicable to conventional cardiac MRI magnetic susceptibility artifact suppression and can also be extended to patients with implantable cardiac electronic devices (CIEDs).

[0060] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0061] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0062] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0065] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0066] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for artifact removal in dual-phase circulatory cardiac magnetic resonance imaging guided by magnetic field maps, characterized in that, Includes the following steps: Two cardiac cine images with different phase increments were acquired and used as complex images in the image domain. Two complex images in the image domain are preprocessed and then input into a pre-trained dual-branch deep learning model for artifact removal, resulting in an artifact-free cardiac cinematic image. The dual-branch deep learning model includes a magnetic susceptibility field estimation network and a conditional image reconstruction network. The artifact removal steps include... Based on the preprocessing results, a magnetic susceptibility field estimation network is used to estimate the pixel-level magnetic susceptibility field map. Based on the preprocessing results and pixel-level magnetic susceptibility field maps, the conditional image reconstruction network removes artifacts through a conditional strategy, outputting an artifact-free cardiac cinema image.

2. The method for artifact removal in dual-phase cyclic cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 1, characterized in that, The acquisition steps for the two cardiac cine images with different phase increments include: Set the scanning parameters, wherein the scanning parameters include the phase increment; Under the scanning parameters, a double-phase cyclic sequence is executed within one breath-hold using multiple receiving coils to collect partial k-space data under different phase increments; Based on partial k-space data under different phase increments, a parallel imaging reconstruction algorithm is used to perform multi-coil joint reconstruction to obtain two cardiac cine images under different phase increments.

3. The method for artifact removal in dual-phase circulatory cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 2, characterized in that, A k-space acceleration strategy is adopted to collect partial k-space data under different phase increments. The k-space acceleration strategy is as follows: 1D undersampling is implemented in the phase encoding direction, a set area in the center of k-space is retained as an automatic calibration signal, full sampling is performed, uniform undersampling is adopted in the outer area, and an equivalent acceleration factor is set.

4. The method for artifact removal in dual-phase cyclic cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 1, characterized in that, The phase increment difference between the two cardiac film images is 180°.

5. The method for artifact removal in dual-phase circulatory cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 1, characterized in that, The two complex images in the image domain include dual channels of real and imaginary parts, or dual channels of amplitude and phase. The preprocessing steps include: One of the complex images in the image domain is used as the reference image, and the other complex image in the image domain is rigidly registered. Based on rigid registration, the image amplitude of the two complex images in the image domain is normalized to obtain the two normalized complex images in the image domain. For the two normalized complex images in the image domain, their real and imaginary parts or amplitude and phase are preserved, and they are concatenated into an input tensor containing multiple channels, represented as follows: Complete the preprocessing process, in which, To take the real part of the complex number, To take the imaginary part of a complex number, , These are two complex images in the image domain.

6. The method for artifact removal in dual-phase cyclic cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 1, characterized in that, The magnetic susceptibility field estimation network utilizes local bias frequencies from two complex images in the image domain. and phase increment physical association , Inversely derived pixel-level magnetic susceptibility field map ,in For pixels, The repetition time.

7. The method for artifact removal in dual-phase cyclic cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 1, characterized in that, The magnetic susceptibility field estimation network adopts an encoder-decoder architecture based on a multi-layer 3D U-Net. The encoder architecture extracts multi-scale features from the preprocessed results through multi-layer convolution and downsampling, while the decoder architecture restores spatial resolution through upsampling and skip connections. Finally, after scaling and spatial smoothing constraints at the output layer, a pixel-level magnetic susceptibility field map is output. ,in In pixels.

8. The method for artifact removal in dual-phase cyclic cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 1, characterized in that, The conditional image reconstruction network employs an encoder-decoder architecture.

9. A method for artifact removal in dual-phase cyclic cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 8, characterized in that, The conditionalization strategy includes at least one of the following: (1) The preprocessing result and the pixel-level magnetic field map are spliced ​​along the channel dimension in the input layer of the conditional image reconstruction network, and the splicing result is used as the joint input of the conditional image reconstruction network; (2) In the encoder-decoder architecture, a cross-attention mechanism is inserted at the same-scale connection between each encoder and decoder layer, except at the connection between the first encoder layer and its same-scale decoder layer. The processing steps of the encoder-decoder architecture include: The preprocessed results are sequentially passed through a multi-layer encoder for downsampling and feature extraction to obtain image features at each corresponding scale of each layer. , For the first Image features output by the layer encoder; Based on the minimum scale image features, the data are sequentially upsampled and transformed through multiple layers of decoders to obtain the decoded features at each corresponding scale. Meanwhile, in each layer, the decoded features are... Projecting these into key vectors and value vectors respectively, the pixel-level magnetic susceptibility map output by the magnetic susceptibility map estimation network at the corresponding scale is obtained. The query vector is obtained by convolutional linear projection. A spatial cross-attention weight map is calculated through a cross-attention mechanism, where the smallest-scale image feature is the image feature output by the last encoder layer. The cross-attention mechanism is as follows: ， In the formula, These are the query vector, key vector, and value vector, respectively. For spatial cross-attention weights, For feature dimensions.

10. A method for artifact removal in dual-phase cyclic cardiac magnetic resonance imaging based on magnetic field map guidance according to claim 1, characterized in that, The dual-branch deep learning model employs either an end-to-end joint training strategy or a phased training strategy during training. The phased training strategy includes the following training steps: Get training labels: Multiple sets of fully sampled complex images in the image domain were acquired, and a coarse pixel-level magnetic susceptibility field map was obtained by fitting a bSSFP signal model through pixel-by-pixel nonlinear least squares fitting. Reference image , serving as a weak supervision label for the magnetic susceptibility field estimation network; Multiple sets of fully sampled complex images in the image domain under phase increment are acquired and averaged to obtain artifact-free labeled images for the conditional image reconstruction network. ; Phased training strategy and loss function: The first stage of training involves fixing the parameters of the image reconstruction network and training the magnetic susceptibility map estimation network separately. This enables the magnetic susceptibility map estimation network to learn how to invert pixel-level magnetic susceptibility maps from the magnetic susceptibility map estimation network. The loss function used during training is: ， ， ， In the formula, For pixel-level magnetic susceptibility map physical consistency loss, Pixel-level magnetic susceptibility map output by the network for estimating magnetic susceptibility map Compared with the reference figure L1 supervision loss, The total variational smoothing loss is used to constrain the continuity of the field diagram space. These are the weighting coefficients. For image horizontal First-order finite-difference gradient operator in the direction, For vertical along the image First-order finite difference gradient operator in direction, subscript , For pixel space coordinate index, where Indicates row index, Indicates column index; The second stage of training involves fixing the parameters of the magnetic susceptibility map estimation network and training the conditional image reconstruction network separately, enabling the conditional image reconstruction network to estimate the magnetic susceptibility map at the pixel level. Image reconstruction is performed conditionally, where the loss function used during training is: ， In the formula, To rebuild the losses, Artifact-free heart movie images and artifact-free label images output by the conditional image reconstruction network. L1 norm difference Artifact-free heart film images and artifact-free labeled images Structural similarity loss, Artifact-free heart film images and artifact-free labeled images Perceptual loss, , These are the weighting coefficients; The third stage of training involves jointly optimizing the parameters of the unfrozen conditional image reconstruction network and the magnetic susceptibility field map estimation network at a set learning rate. The joint loss function during this optimization process is as follows: ， In the formula, For joint losses; Training terminates when the set termination condition is met, resulting in a well-trained dual-branch deep learning model. , These are the weighting coefficients.

Images