A multi-parameter four-dimensional magnetic resonance imaging generation, timing optimization and automatic quality control integrated method and system

Abstract

The application discloses a kind of multi-parameter four-dimensional magnetic resonance imaging generation, timing optimization and automatic quality control integrated method and system, it is related to medical image and magnetic resonance imaging field, first to low-quality four-dimensional magnetic resonance and high-quality static three-dimensional magnetic resonance sequence do pre-processing and realize space-time geometry consistency, then with 3D-T1 as pivot complete cross modality hierarchical cascade registration, obtain the multi-modal high-quality static image of unified reference space;Through SEDER model extraction time continuous timing deformation field, obtain the deformation field sequence in line with physiological law by PCA low rank reconstruction optimization, it is sequentially acted on the static image after registration, generates including DWI in high-quality multi-parameter 4DMR;Finally, through optical flow field motion evaluation, multidimensional image quality evaluation and three-level quality control framework, realize automatic quality control and exception handling.This method does not need special equipment, the 4D-MRI generated by this method has good space-time continuity, can generate DWI and other multi-parameter functional images, adapts to the clinical needs of tumor radiotherapy, and has excellent robustness and reproducibility.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging technology, specifically to an integrated method and system for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control. Background Technology

[0002] In the field of respiratory motion management, four-dimensional computed tomography (4D-CT) is a commonly used technique, but it has limitations in abdominal applications, as it struggles to provide sufficient soft tissue contrast, making precise target localization difficult. Magnetic resonance imaging (MRI), on the other hand, offers superior contrast between the tumor and surrounding soft tissue. 4D MRI can provide a more accurate and reliable target motion range in radiotherapy than 4D-CT.

[0003] Existing 4D-MRI technologies are mainly divided into three categories: The first category is retrospective reordering 4D-MRI technology, which requires the acquisition of data from multiple respiratory cycles. The imaging time is long, it is highly dependent on the patient's respiratory stability, is prone to suture artifacts, and has limited spatiotemporal resolution, making it difficult to meet the needs of high-precision radiotherapy. The second category is prospective gated 4D-MRI technology, which relies on fast imaging sequences, has limited image contrast selection, is sensitive to magnetic field inhomogeneity, is prone to artifacts affecting image quality, and has an overall imaging time that is still relatively long, making it difficult to use frequently in clinical procedures.

[0004] To overcome the efficiency and quality issues of directly acquiring four-dimensional magnetic resonance imaging (MRI), some existing technologies have proposed four-dimensional image generation methods based on motion modeling. These methods use a set of high-quality three-dimensional MRI images as prior knowledge, obtain the deformation vector field (DVF) between the prior image and the target image through the deformation registration (DIR) algorithm, and then apply the motion information to the prior image to generate a 4D sequence. However, the traditional iterative DIR algorithm has high computational complexity and heavy computational burden, and requires individual modeling for each patient, resulting in poor versatility. Furthermore, due to the large differences in signal distribution between different MRI parameters, the cross-contrast registration effect is poor.

[0005] In summary, existing 4D-MRI generation technologies cannot simultaneously meet the clinical needs of high image quality, high processing efficiency, and high generalization ability, thus limiting the clinical application of 4D-MRI in tumor radiotherapy. Summary of the Invention

[0006] The purpose of this invention is to provide an integrated method and system for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization and automated quality control, to solve the technical problems of insufficient motion temporal continuity, single 4D image modality and lack of quality control system in existing 4D-MRI generation technology.

[0007] To achieve the above objectives, the present invention employs the following technical solutions.

[0008] A method integrating multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control includes the following steps: S1. Multi-sequence magnetic resonance imaging preprocessing and spatiotemporal geometric consistency: acquire low-quality dynamic four-dimensional magnetic resonance imaging sequences (Original 4D) acquired routinely in clinical settings and multiple sets of static three-dimensional magnetic resonance imaging sequences to be registered. Based on DICOM geometric metadata, complete the spatial coordinate mapping of all sequences, global common field of view cropping and unified resampling to eliminate systematic errors caused by geometric and resolution differences. S2. Reference Frame Selection and Cross-Modal Multi-Sequence Adaptive Hierarchical Cascaded Registration Based on Gradient Mutual Information Anchoring and Viterbi Graph Optimization: Using a static 3D sequence as a reference, the time frame with the highest similarity in Original 4D is selected as the reference reference frame based on gradient normalized mutual information. A multi-stage state graph is constructed based on the respiratory amplitude signal, with the target respiratory phase as the stage and candidate time frames as nodes. The Viterbi dynamic programming algorithm is used to search for the globally optimal path to obtain a temporally continuous phase-sequential four-dimensional magnetic resonance image sequence. Then, with 3D-T1 as the core hub, a hierarchical cascaded registration path is constructed. The cross-modal registration of the multimodal static 3D magnetic resonance sequence and the reference reference frame is completed through the registration solver, resulting in a high-quality multimodal static image under a unified reference space. S3. Reference frame-guided temporal cyclic deformation registration and respiratory motion modeling: Construct a cyclic deformation registration model SEDER with temporal memory mechanism. The model uses the reference frame image, the current target phase image, the difference image between the two, and the deformation prior image obtained by recursion from the previous moment as joint input. It processes each respiratory phase sequentially along the time axis. Through a combination of preheating training and temporal recursion training, combined with online composite deformation field supervision and second-order time smoothing constraints, the model outputs a temporal deformation field sequence that covers the entire respiratory cycle and is time-continuous. S4. Global Spatiotemporal Motion Smoothing and Optimization: Based on principal component analysis (PCA), the temporal deformation field sequence is decomposed into a high-dimensional matrix, and the principal components representing the main respiratory motion patterns are screened for low-rank reconstruction to remove high-frequency jitter and obtain a spatiotemporally continuous optimized deformation field sequence that conforms to the respiratory physiological law. S5. Generation of high-quality four-dimensional magnetic resonance imaging with multiple parameters: The optimized deformation field sequence is applied frame by frame to a high-quality static image in a unified reference space to generate a high-4D high-quality four-dimensional magnetic resonance sequence containing anatomical and functional information. S6. Full-process automated quality control assessment and abnormal closed-loop processing: The generated high-4D sequence is subjected to respiratory motion consistency assessment based on optical flow field and multi-dimensional image quality assessment. At the same time, the input, registration, physiological motion consistency and clinical accuracy are monitored through a three-level quality control framework, with an abnormal circuit breaker and stratified rollback mechanism, and automatic output of case-level reports and alarms.

[0009] Furthermore, step S1 specifically includes: S11. Read the DICOM geometric metadata of the Original 4D sequence and multiple sets of static 3D magnetic resonance sequences, including voxel spacing, image direction cosine, and physical space origin coordinates. Map the voxel indices of all sequences to the same physical world coordinate system. Calculate the 3D coverage of the Original 4D reference sequence and each set of static 3D magnetic resonance sequences in physical space. Process the slicing order of the static 3D magnetic resonance sequences to maintain their continuity and orientation consistency in physical space. S12. For each set of static three-dimensional magnetic resonance sequences, calculate the pairwise intersection range between them and the Original 4D sequence in physical space, remove invalid data that do not overlap with the Original 4D sequence in space, and then further calculate the intersection of all valid pairwise intersection ranges to obtain the unique global minimum common field of view region. Map the clipping boundary in reverse to the original voxel coordinate system of each sequence and perform uniform clipping. S13. Based on the physical size of the global common field of view, a unified target spatial size and voxel spacing are set. Trilinear interpolation is used to uniformly resample the clipped Original 4D sequence and each static three-dimensional magnetic resonance sequence to the same spatial size and resolution, thus completing the spatiotemporal geometric consistency processing. S14. For special scenarios where the static sequence field of view cannot fully cover the target organ, a reference frame selection, cropping and resampling process based on normalized cross-correlation (NCC) is independently executed for each static 3D magnetic resonance sequence, and the corresponding 4D output is saved to retain complete anatomical information.

[0010] Furthermore, step S2 specifically includes: S21. Normalize and extract gradient magnitude features for each time frame of the static three-dimensional magnetic resonance sequence and Original 4D. Calculate the similarity between each time frame and the static three-dimensional sequence based on the gradient normalized mutual information. Determine the time frame with the highest similarity as the reference frame. For low signal-to-noise ratio functional sequences, calculate the gradient normalized mutual information only within a pre-defined local region of interest. S22. Set a region of interest in the diaphragm region or lung boundary region in Original 4D, and construct a respiratory amplitude signal based on the area change or bright pixel statistics in the region; divide the respiratory amplitude into multiple target respiratory phases according to the preset number of phases, and select several time frames with the closest respiratory amplitude as candidate nodes for each target respiratory phase. S23. Construct a global cumulative cost function, where the node cost represents the amplitude matching degree between the candidate time frame and the target breathing phase, and the transition cost represents the image structure continuity of the candidate time frames in adjacent stages; use the reference frame as the forced traversal node, and use the Viterbi dynamic programming algorithm to search for the global optimal path to obtain a temporally continuous phase-sequential four-dimensional magnetic resonance image sequence. S24. Construct a three-level hierarchical cascaded registration path: First level, register the T1 weighted 3D sequence to the reference frame to establish a unified reference space; Second level, register the T2 weighted 3D sequence to the registered 3D-T1 and map it to the unified reference space through composite transformation; Third level, register the diffusion-weighted imaging DWI to the completed 3D-T2 and map it to the unified reference space through cascaded transformation; During the registration process, the registration parameters are adaptively adjusted based on the image sharpness evaluation results.

[0011] Furthermore, in step S3, the cyclic deformation registration model SEDER adopts a three-dimensional encoder-decoder deep learning architecture. For any time t, the model's input includes a reference frame image. The current Original 4D image The difference image between the two and the deformation prior image recursively generated at the previous time t-1. The model extracts image features through 3D convolutional layers and attention modules, fuses prior motion information from the previous time step, and predicts the 3D deformation field from the reference frame to the current time step. .

[0012] Furthermore, the training of the SEDER model includes a warm-up training phase and a time-series recursive training phase: During the warm-up training phase, the deformation prior image is set to zero, and direct supervised training is performed using long-distance image pairs with an interval of no less than the preset number of frames to obtain basic deformation estimation capabilities. During the time-series recursive training phase, the continuous target phases are recursively predicted along the time axis, and the deformation prior image at the current moment is generated by gradient blocking of the previous prediction results. A composite loss function is used for optimization, which includes end-to-endpoint error (EPE) loss, normalized cross-correlation (NCC) loss, deformation field spatial smoothing regularization term, and time smoothing constraint term based on the second-order difference of the deformation field at adjacent time points. For the missing long-distance reference deformation field, an approximate reference deformation field is obtained by online composite of adjacent deformation fields, and it participates in training with a supervision weight lower than that of the real reference deformation field.

[0013] Furthermore, step S4 specifically includes: S41. The time-series deformation field sequence output in step S3 is matrixed. The three-dimensional deformation field of each breathing phase is flattened into a high-dimensional vector in the spatial dimension to construct a two-dimensional motion matrix with time points as samples and voxel displacement as features. S42. Perform singular value decomposition (SVD) on the motion matrix or eigenvalue decomposition on the covariance matrix to obtain principal components and corresponding eigenvectors sorted by energy. Select the top K principal components, where the cumulative energy of the top K principal components accounts for no less than 90%. S43. Based on the selected principal components and their corresponding time phases, the original deformation field sequence is reconstructed in low rank. The motion state of each time phase is recovered by linear combination of the first K principal components. The reconstructed matrix is ​​restored to the original four-dimensional deformation field format, and the optimized spatiotemporal continuous deformation field sequence is obtained.

[0014] Furthermore, in step S5, the high-4D multi-parameter high-quality four-dimensional magnetic resonance sequence includes T1-weighted and T2-weighted anatomical four-dimensional sequences and diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) combined functional four-dimensional sequences. During the generation process, each phase deformation field of the optimized deformation field sequence is applied to the corresponding high-quality static image in a unified reference space to generate the corresponding four-dimensional sequence, thereby realizing the joint generation of anatomical and functional multi-parameter four-dimensional magnetic resonance imaging.

[0015] Furthermore, in step S6, the respiratory motion consistency assessment based on the optical flow field specifically includes: S611. Construct an adaptive computational domain based on the target region ROI mask, and expand it outward with a set margin to form a clipping window; preprocess the original Original 4D sequence and the generated high-4D sequence respectively, perform contrast-limited adaptive histogram equalization and strong Gaussian smoothing on the Original 4D, and perform weak smoothing on the high-4D. S612. Using the first frame as a reference, estimate the dense displacement field frame by frame using the Demons optical flow method; based on the three-dimensional gradient magnitude of the reference frame, filter out high-texture effective points within the ROI and remove noise interference from flat areas. S613. Extract the motion trajectories of the effective point group in the front-back, left-right, and head-to-feet directions. Calculate the mean absolute error, maximum error, and error confidence interval of the motion trajectories of the two sequences. Output the motion statistics and visualization results.

[0016] Further, in step S6, the multi-dimensional image quality assessment includes calculating the tumor contrast-to-noise ratio (CNR), organ boundary full width at half maximum (FWHM), and cumulative probability of blur detection (CPBD) between the generated high-4D sequence and the original Original 4D sequence; the three-level quality control framework specifically includes: First-level input geometric consistency detection: Before registration, calculate the field overlap ratio and centroid offset of the input sequence. If the preset threshold is not met, the process will be terminated directly and an error message will be displayed. Physical rationality monitoring of the second-layer registration process: During the registration and deformation modeling process, the Jacobian determinant of the output deformation field is continuously monitored to detect tissue folding that is not physically deformed. If an anomaly is triggered, smoothing enhancement and parameter adaptive adjustment are automatically performed. The third layer of physiological motion consistency and clinical accuracy verification: After generating the high-4D sequence, the results are divided into three levels: pass, warning, and failure, based on motion consistency and image quality indicators. For failure results, a multi-path rollback process is automatically triggered: first, the deformation model is rerun through parameter adaptive adjustment; if the requirements are still not met, the traditional high robustness registration algorithm is switched to. In extreme cases, manual intervention prompts are triggered.

[0017] A multi-parameter four-dimensional magnetic resonance imaging (MRI) generation, temporal optimization, and automated quality control integrated system includes: The preprocessing and spatiotemporal geometry unification module is used to complete the spatial coordinate mapping, common field-of-view clipping and unified resampling of the Original 4D sequence and multiple sets of static 3D magnetic resonance sequences, and eliminate systematic geometric errors. The reference frame selection and hierarchical cascaded registration module is used to select reference reference frames based on gradient normalized mutual information, complete global phase reassembly through Viterbi dynamic programming, and construct a three-level registration path with the T1 weighted sequence as the hub to complete cross-modal registration between multimodal static sequences and reference reference frames. The temporal cyclic deformation registration module is used to run the SEDER model with temporal memory mechanism. It takes the reference frame image, the target temporal image, the difference image and the recursively generated deformation prior image as joint input to extract the temporal deformation field of respiratory motion that is continuous in time. The global spatiotemporal motion smoothing module is used to perform low-rank reconstruction and noise reduction optimization of temporal deformation fields through PCA principal component analysis. The multi-parameter four-dimensional imaging generation module is used to apply the optimized deformation field to multimodal high-quality static images to generate multi-parameter high-quality four-dimensional magnetic resonance sequences that combine anatomy and function. The automated quality control and anomaly handling module is used to perform pixel-level motion consistency assessment and multi-dimensional image quality assessment. It completes full-process quality monitoring and anomaly rollback processing through a three-level quality control framework, and outputs qualified imaging results and quality control reports.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) Multimodal cross-sequence spatiotemporal consistency, adapting to routine clinical multi-parameter MRI data. Through global common field of view cropping and unified resampling based on DICOM geometric data, the spatial coordinates of low-quality dynamic 4D sequences and multiple static 3D sequences are unified. It supports the joint processing of multimodal sequences such as T1, T2, DWI, and ADC. No special acquisition equipment is required. It directly adapts to routine clinical MRI data and has strong clinical applicability.

[0019] (2) Viterbi global optimization phase separation significantly improves temporal continuity. Gradient normalized mutual information is used to anchor the reference frame. A multi-stage state diagram is constructed by combining the respiratory amplitude signal. Viterbi dynamic programming searches for the global optimal phase separation path, which effectively avoids frame skipping, mismatch and artifact interference caused by traditional amplitude binning methods. The 4D sequence after phase separation has better temporal continuity and structural consistency.

[0020] (3) Hierarchical cascaded cross-modal registration with high precision in unified multimodal space. A three-level hierarchical cascaded registration path is constructed with 3D-T1 as the hub to achieve accurate mapping of different contrast sequences such as T2 and DWI to a unified reference space. For low signal-to-noise ratio sequences such as DWI, local ROI similarity calculation and adaptive parameter adjustment are adopted, resulting in strong robustness of cross-modal registration.

[0021] (4) SEDER time-series cyclic deformation modeling, motion estimation spatiotemporal continuity with time-series memory mechanism cyclic deformation registration model SEDER integrates the deformation prior information of the previous moment, combines the two-stage strategy of warm-up training and time-series recursive training, and the second-order time smoothing constraint, the output time-series deformation field has good time continuity and can accurately capture the organ motion law of the complete respiratory cycle.

[0022] (5) PCA low-rank reconstruction denoising, motion conforms to physiological laws. The temporal deformation field is reconstructed by low-rank through principal component analysis, retaining the first K principal components, removing high-frequency jitter and noise interference. The reconstructed deformation field is spatiotemporally smooth and conforms to the characteristics of respiratory physiological motion, avoiding non-physiological tissue deformation.

[0023] (6) Multi-parameter 4D-MRI combined generation, combining functional and anatomical information. The optimized deformation field is applied frame by frame to the registered multimodal high-quality static image, which can simultaneously generate T1, T2 anatomical four-dimensional sequences as well as DWI, ADC functional four-dimensional sequences, breaking through the limitation of traditional 4D-MRI single modality, and providing multi-dimensional image support for tumor radiotherapy target delineation and functional assessment.

[0024] (7) Three-level automated quality control closed loop ensures safe and reliable clinical application. A three-level quality control framework is established, which includes input geometric consistency, registration physical rationality, physiological motion and clinical accuracy. It is equipped with an abnormal circuit breaker and stratified rollback mechanism to realize full-process automated quality monitoring and abnormal handling. Case-level quality control reports and alarms can be output without manual intervention, ensuring the accuracy and safety of clinical application.

[0025] (8) The whole process is integrated, with high processing efficiency, good reproducibility, high degree of algorithm automation, short processing time per case, easy to integrate into the clinical radiotherapy workflow, and has good clinical translation and promotion value. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the overall method architecture of the present invention; Figure 3 This is a schematic diagram of the amplitude phase-separated 4D image sequence obtained by the Viterbi algorithm of this invention; Figure 4 This is a schematic diagram of the SEDER model of the present invention; Figure 5 This is a visualization result of the motion trajectory in the three directions of AP, ML, and SI output by this invention; Figure 6 The images show the high-quality, multi-parameter four-dimensional magnetic resonance image generated by this invention and the original low-quality four-dimensional magnetic resonance image. Detailed Implementation The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0027] Example 1 like Figure 1 As shown, this embodiment provides an integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, timing optimization, and automated quality control, specifically for radiotherapy scenarios in patients with abdominal liver cancer. The overall technical approach is as follows: Figure 2 As shown, the specific steps include:

[0028] S1. Preprocessing and spatiotemporal geometrical consistency of multi-sequence magnetic resonance imaging The input data in this embodiment is routinely acquired abdominal MRI data, including: 1) low-quality dynamic four-dimensional magnetic resonance sequence Original 4D, acquired using the commercial TWIST-VIBE sequence, covering the entire respiratory cycle, with high temporal resolution, but low signal-to-noise ratio, blurry images, and artifacts; 2) high-quality static three-dimensional magnetic resonance sequence, including 3D-T1 weighted sequence, 3D-T2 weighted sequence, DWI sequence and ADC map, with good spatial resolution and tissue contrast, but it is a static image and does not contain respiratory motion information.

[0029] S11. Read the DICOM geometric metadata of all sequences, including voxel spacing, image orientation cosine, and physical space origin coordinates. Map the discrete voxel indices of all sequences (Original 4D sequences and 3D static sequences) to the same physical world coordinate system, and calculate the 3D coverage of the 4D reference sequence and each 3D sequence in physical space. For 3D sequences, the system automatically processes the slicing order to ensure their continuity and orientation consistency in physical space.

[0030] S12. Calculate the 3D coverage of the Original 4D sequence and each 3D static sequence in physical space. For each 3D sequence, calculate the pairwise intersection range with the 4D reference sequence in physical space, and automatically remove invalid data that do not overlap with the 4D sequence in space. Then, further calculate the intersection of all valid pairwise intersection ranges to obtain a unique global minimum common field of view region, which represents the minimum anatomical range commonly covered by all input sequences in physical space.

[0031] S13. Reverse map the global clipping boundary to the original voxel coordinate system of each sequence, perform uniform clipping on all sequences, set a uniform target space size based on the physical size of the global common field of view, calculate the corresponding target voxel spacing, and use trilinear interpolation to uniformly resample all sequences to this spatial resolution to ensure that all sequences have a completely consistent spatial definition before entering subsequent processing.

[0032] S14. In view of the special case of a small field of view of the DWI sequence in this embodiment, if global cropping is forced, key anatomical regions may be lost. Therefore, the system independently performs reference frame selection for the DWI sequence: calculates the similarity between each frame of Original 4D and the DWIb0 image based on normalized cross-correlation NCC, selects the most similar frame as the exclusive reference frame of the DWI sequence, and independently performs cropping and resampling to preserve the anatomical information of the tumor region to the greatest extent.

[0033] S2. Reference frame selection and cross-modal multi-sequence adaptive hierarchical concatenation registration based on gradient mutual information anchoring and Viterbi graph optimization. S21. Normalize and extract gradient magnitude features for each time frame of the static 3D magnetic resonance imaging (MRI) sequence and Original 4D. Calculate the similarity between each time frame and the 3D-T1 weighted sequence based on gradient-normalized mutual information, and determine the time frame with the highest similarity (end-expiratory phase) as the global registration reference frame. For low signal-to-noise ratio functional sequences such as DWI, gradient-normalized mutual information is calculated only within pre-defined regions of interest in the tumor and liver to avoid background noise interference.

[0034] S22. On a slice at 55% depth of the coronal plane of the Original 4D image, draw a region of interest in the diaphragm region. Use a fixed signal intensity threshold (50) to binarize the pixels in the ROI and count the area of ​​bright pixels to construct the initial respiratory amplitude signal. Use the 5%-95% quantile range to eliminate irregular deep breathing or extreme artifacts of the patient. Divide the respiratory amplitude into 10 target respiratory phases according to the preset number of phases K=10. For each target respiratory phase, select the 5 time frames with the closest respiratory amplitude from all time frames as candidate nodes.

[0035] S23. Construct the global cumulative cost function: Node cost: Characterizing candidate time frames The target amplitude of the breathing amplitude and the target breathing phase k The smaller the difference in the degree of matching between them, the lower the cost. Transfer cost: Characterizing adjacent stages and Between candidate frames Transfer to The higher the structural continuity and structural similarity (SSIM) of the image, the lower the transfer cost.

[0036] The global path cost is then: , in =1.2, , This represents the candidate node path from phase 1 to phase K. If the path includes the node corresponding to the reference frame, then... , ,in This is a preset large penalty constant.

[0037] like Figure 3 As shown, by using the reference frame as the forced traversal node, the Viterbi dynamic programming algorithm is used to search for the globally optimal path to obtain a temporally continuous phase-sequential four-dimensional magnetic resonance image sequence, which effectively avoids frame skipping, mismatch and extreme artifact interference caused by the traditional amplitude binning method.

[0038] S24. Construct a three-level hierarchical cascaded registration path and use a self-developed adjustable parameter general robust registration solver to complete cross-modal registration: First level: Register high-resolution 3D-T1 images as floating images to the reference frame to establish a unified reference space; The second level: The 3D-T2 weighted sequence is used as a floating image, and the registered T1 weighted image is used as a fixed image. After registration, the T2 sequence is mapped to the 4D reference space through the composite transformation from T1 to the reference frame. The third level: DWI (especially low b-value or b0 images) is used as a floating image, and the registered T2-weighted image is used as a fixed image. Registration is performed and the geometric distortion of the EPI sequence is corrected. Then, through the cascaded transformation from T2 to T1 and from T1 to 4D, the DWI sequence is accurately mapped to the reference space.

[0039] During the registration process, the image sharpness and the overlap of the registered structures are evaluated in real time, and the number of registration iterations and the weight of the regularization term are adaptively adjusted to ensure the registration accuracy of different modal sequences. The Dice coefficient of the liver region after registration is greater than 0.95.

[0040] S3. Reference frame-guided temporal cyclic deformation registration and respiratory motion modeling This step constructs and runs the sequential cyclic deformation registration model SEDER, extracting the time-continuous respiratory motion deformation field from the phase-separated Original 4D sequence. The model architecture is as follows: Figure 4 As shown. The SEDER model adopts a three-dimensional encoder-decoder deep learning architecture. In this embodiment, a three-dimensional U-Net with shared weights is used. The network front end receives four-channel input and performs feature extraction, and the network end outputs a three-channel three-dimensional displacement field, corresponding to the displacement components in the left-right, front-back, and head-to-toe directions, respectively.

[0041] S31. During model inference, the selected reference frame is used. For a globally fixed reference, 10 respiratory phases are processed sequentially along the time axis. For any time t, the model input includes: a reference frame image. The current Original4D image The difference image of the two and the deformation prior image generated recursively at the previous time step t-1 Among them, deformation prior images It is recursively generated from the deformation field predicted at the previous moment, used to characterize the local breathing motion trend and constrain the temporal continuity of the deformation prediction at the current moment; in the initial recursive step, the deformation prior image is set to zero.

[0042] S32. The model extracts image features through a 3D convolutional layer and an attention module, focusing on key anatomical regions that undergo significant deformation during respiration, such as the diaphragm region or the liver boundary. It also integrates prior motion information from the previous moment to predict the 3D deformation field from the reference frame to the current moment. The reference frame is deformed to the current target time phase through a spatial transformation layer to obtain a deformed image. ,in This indicates a spatial resampling operation based on trilinear interpolation.

[0043] The training of the S33 and SEDER models includes a warm-up training phase and a time-series recursive training phase: Preheating training phase: The deformation prior image input is set to zero. In this embodiment, a total of 48 deformation field samples are generated for a single patient. Direct supervised training is performed using 30 pairs of long-distance images with frame number intervals greater than or equal to 5 and 18 pairs of adjacent temporal image pairs, enabling the model to acquire the basic deformation estimation capability from the reference frame to the target temporal phase. The loss function for this phase is: , in, , , , .

[0044] Temporal recursive training phase: Recursion is performed along both decreasing and increasing frame number directions around multiple candidate reference time phases. Previous prediction results are only used to generate the deformation prior image required for the current time step and do not participate in backpropagation. A temporal smoothing loss based on the second-order difference of the deformation field is introduced. This constrains the deformation evolution trend between consecutive target phases. The composite loss function for this stage is: ,in , , The parameters and preheating stage are the same. .

[0045] S34. For missing long-distance reference deformation fields, online composite of adjacent two-step or multi-step deformation fields is used to obtain an approximate reference deformation field, which is then assigned a low supervision weight of 0.4 to participate in model training. Without needing to pre-calculate the real reference deformation field for all long-distance registration pairs, the time-series recursive training samples are expanded and the model's ability to learn long-range motion evolution is enhanced.

[0046] The model ultimately outputs a temporal deformation field sequence covering 10 respiratory phases, including voxel displacement values ​​in three directions, which fully characterizes the organ movement and deformation throughout the respiratory cycle.

[0047] S4, Global Spatiotemporal Motion Smoothing and Optimization This step uses PCA to globally optimize the temporal deformation field sequence and remove non-physiological jitter.

[0048] S41. Temporal Deformation Field Matrixing: The input is a 3D deformation field with dimensions D×H×W and 10 time phases. Each time phase's deformation field contains displacement components in three directions. For each time phase, its corresponding 3D DVF is flattened in space, and the three-directional displacements of all voxels are concatenated into a high-dimensional vector, which serves as a sample for that time phase. By performing this operation on all T time phases, a 2D data matrix is ​​constructed, where each row corresponds to the complete 3D motion state at a given time point, and the column dimensions correspond to the displacement features of all voxels in the three directions.

[0049] S42. Principal Component Decomposition and Screening: Singular Value Decomposition (SVD) is performed on the motion matrix to obtain principal components sorted by energy. In this embodiment, the cumulative energy of the first three principal components accounts for more than 95%, which can cover most of the respiratory physiological motion. Therefore, the first three principal components are screened for subsequent reconstruction, and higher-order noise components are discarded.

[0050] S43, Low-Rank Reconstruction: Using the overall average motion field as a basis, the motion vector of each time phase is reconstructed through a linear combination of the first three principal components. The reconstructed vectors are then restored to a four-dimensional deformation field format of D×H×W×3, resulting in an optimized spatiotemporal continuous deformation field sequence. The reconstructed deformation field eliminates local high-frequency jitter while fully preserving the amplitude and phase characteristics of the overall respiratory motion.

[0051] S5, high-quality four-dimensional magnetic resonance imaging generation with multiple parameters This step applies the optimized deformation field sequence frame by frame to the registered multimodal high-quality static images to generate a multi-parameter high-4D sequence. For the T1-weighted high-4D sequence, the optimized deformation field for each phase is applied to the registered 3D-T1-weighted image, and spatial transformation is achieved through trilinear interpolation to generate the corresponding T1-weighted four-dimensional image. The 10 phases are stitched together to obtain the complete T1-weighted high-4D sequence. Using the same method, T2-weighted, DWI, and ADC four-dimensional sequences are generated respectively, ultimately yielding a multi-parameter high-quality four-dimensional magnetic resonance imaging sequence containing anatomical and functional information.

[0052] The generated high-quality four-dimensional magnetic resonance sequence with multiple parameters looks like this. Figure 6As shown, from top to bottom, the generated high-4D sequences (labeled as UQ-DWI, UQ-T1, and UQ-T2) of three modalities—DWI, T1-weighted, and T2-weighted—are compared with the original low-quality Original 4D (labeled as lowQ). The red arrows indicate the tumor region. It can be seen that the generated sequences maintain clear anatomical structures and tumor contrast at each respiratory phase, and their spatiotemporal continuity is significantly better than that of the original low-quality sequences.

[0053] S6. Full-process automated quality control assessment and abnormal closed-loop handling S61. The input for the respiratory motion consistency assessment based on optical flow field consists of three parts: the original 4D, the generated high-quality 4D sequence, and the ROI mask used to limit the assessment scope.

[0054] S611. Based on the ROI mask drawn by the doctor, an adaptive computation domain is constructed: the 3D bounding box of the ROI is calculated, and it is expanded outward by 30mm to form a computation clipping window. Boundary truncation is performed to ensure that all operations are performed within the clipping window, reducing computational load and background noise interference. Extreme value normalization is performed on each sequence within the computation domain; for the original low-quality Original 4D sequence, contrast-limited adaptive histogram equalization (CLAHE, with the contrast limit parameter set to 0.02) and strong Gaussian smoothing (smoothing coefficient σ=1.0) are performed; for the generated high-4D high-quality sequence, CLAHE is disabled and weak smoothing (σ=0.5) is used to preserve high-frequency texture details to the greatest extent.

[0055] S612. For the original Original 4D sequence and the generated T1-weighted high-4D sequence, using the first frame as the fixed reference frame and the k-th frame as the floating frame, the dense displacement field is estimated frame-by-frame using the multi-resolution Demons optical flow method. The optical flow parameters are set as follows: a 3-layer image pyramid with iterations of 500, 400, and 200 for each layer from coarse to fine, and a cumulative displacement field smoothing coefficient of 1.5. The 3D gradient magnitude is calculated on the reference frame, and the 60th percentile of the gradient magnitude within the ROI is selected as the threshold to filter high-texture effective points and remove noise interference from flat areas.

[0056] S613. Extract the motion trajectories of the effective point group in the three directions AP (front / back), ML (left / right), and SI (head / foot). Calculate the median displacement of the effective point group in each time phase as the representative value of the overall displacement. Extract the 25% and 75% quantiles to construct confidence intervals for local motion uncertainty. Further calculate the mean absolute error, maximum error, and corresponding time phases between the corresponding trajectories of the two sequences. Output an Excel report and a trajectory comparison chart with unified coordinates, such as... Figure 5 As shown.

[0057] S62. Multidimensional image quality assessment quantitatively compares and analyzes the generated high-quality 4D sequence with the original Original 4D at the image level, calculates the tumor contrast-to-noise ratio (CNR), full width at half maximum (FWHM) of the lung-liver boundary, and cumulative probability of blur detection (CPBD), and outputs an image quality assessment statistical report to verify the improvement effect of the generated sequence on visual quality and anatomical details.

[0058] S63. Three-level quality control and abnormal closed-loop handling First-layer input geometric consistency check: Before registration, calculate the field-of-view overlap ratio and centroid offset of the input sequence. When the field-of-view overlap ratio... And centroid offset When, it is judged as passed; when or When this occurs, it is judged as a warning, and the parameter adaptive adjustment is initiated; when or If the process fails, the process will be terminated immediately and a data error message will be displayed.

[0059] The second-level registration process involves physical rationality monitoring: during registration and deformation modeling, the Jacobian determinant of the deformation field is calculated in real time to detect tissue folding caused by non-physical deformation. When the Jacobian determinant... When the percentage of voxels is no more than 0.1%, it is considered a pass; when the percentage is between 0.1% and 0.5%, it is considered a warning, the smoothing regularization weight is automatically increased and the prediction is re-predicted; when the percentage is greater than 0.5%, it is considered a failure, and a high smoothing weight recalculation is triggered.

[0060] The third layer of physiological motion consistency and clinical accuracy verification: After generating the high-4D sequence, motion consistency and image quality indicators are combined for grading. When the mean absolute error (MAE) is... Maximum error MaxE If the generated high-4D model improves upon the original Original 4D model in at least two of the following metrics—CNR, FWHM, and CPBD—it is considered passed; when... MAE or MaxE A warning is issued when only one image quality metric reaches the improvement threshold; when MAE Or MaxE If the image quality index fails to improve, or if all image quality metrics do not improve, the image is considered to have failed.

[0061] For failed results, a multi-path rollback process is automatically triggered: first, the weights of the regularization terms in the SEDER model are adjusted, and deformation modeling is rerun; if the requirements are still not met, the traditional pTVreg registration algorithm is switched to recalculate the deformation field; in extreme cases, a manual intervention prompt is triggered to ensure clinical safety.

[0062] Ultimately, the system outputs a high-quality, multi-parameter 4D-MRI sequence that has passed quality control, including 10-phase four-dimensional images in four modalities: T1, T2, DWI, and ADC, along with a complete motion assessment quality control report.

[0063] Example 2 This embodiment provides an integrated system for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control to implement the above method, including: The preprocessing and spatiotemporal geometry unification module is used to complete the spatial coordinate mapping, common field-of-view clipping and unified resampling of the Original 4D sequence and multiple sets of static 3D magnetic resonance sequences, and eliminate systematic geometric errors. The reference frame selection and hierarchical cascaded registration module is used to select reference reference frames based on gradient normalized mutual information, complete global phase reassembly through Viterbi dynamic programming, and construct a three-level registration path with the T1 weighted sequence as the hub to complete cross-modal registration between multimodal static sequences and reference reference frames. The temporal cyclic deformation registration module is used to run the SEDER model with temporal memory mechanism. It takes the reference frame image, the target temporal image, the difference image and the recursively generated deformation prior image as joint input to extract the temporal deformation field of respiratory motion that is continuous in time. The global spatiotemporal motion smoothing module is used to perform low-rank reconstruction and noise reduction optimization of temporal deformation fields through PCA principal component analysis. The multi-parameter four-dimensional imaging generation module is used to apply the optimized deformation field to multimodal high-quality static images to generate multi-parameter high-quality four-dimensional magnetic resonance sequences that combine anatomy and function. The automated quality control and anomaly handling module is used to perform pixel-level motion consistency assessment and multi-dimensional image quality assessment. It completes full-process quality monitoring and anomaly rollback processing through a three-level quality control framework, and outputs qualified imaging results and quality control reports.

[0064] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method integrating multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control, characterized in that, Includes the following steps: S1. Multi-sequence magnetic resonance imaging preprocessing and spatiotemporal geometric consistency: acquire low-quality dynamic four-dimensional magnetic resonance sequences (Original 4D) acquired routinely in clinical practice and multiple sets of static three-dimensional magnetic resonance sequences to be registered. Based on DICOM geometric metadata, complete the spatial coordinate mapping of all sequences, global common field of view cropping and unified resampling to eliminate systematic errors caused by geometric and resolution differences. S2. Reference Frame Selection and Cross-Modal Multi-Sequence Adaptive Hierarchical Cascaded Registration Based on Gradient Mutual Information Anchoring and Viterbi Graph Optimization: Using a static 3D sequence as a reference, the time frame with the highest similarity in Original 4D is selected as the reference reference frame based on gradient normalized mutual information. A multi-stage state graph is constructed based on the respiratory amplitude signal, with the target respiratory phase as the stage and candidate time frames as nodes. The Viterbi dynamic programming algorithm is used to search for the globally optimal path to obtain a temporally continuous phase-sequential four-dimensional magnetic resonance image sequence. Then, with 3D-T1 as the core hub, a hierarchical cascaded registration path is constructed. The cross-modal registration of the multimodal static 3D magnetic resonance sequence and the reference reference frame is completed through the registration solver, resulting in a high-quality multimodal static image under a unified reference space. S3. Reference frame-guided temporal cyclic deformation registration and respiratory motion modeling: Construct a cyclic deformation registration model SEDER with temporal memory mechanism. The model uses the reference frame image, the current target phase image, the difference image between the two, and the deformation prior image obtained by recursion from the previous moment as joint input. It processes each respiratory phase sequentially along the time axis. Through a combination of preheating training and temporal recursion training, combined with online composite deformation field supervision and second-order time smoothing constraints, the model outputs a temporal deformation field sequence that covers the entire respiratory cycle and is time-continuous. S4. Global Spatiotemporal Motion Smoothing and Optimization: Based on principal component analysis (PCA), the temporal deformation field sequence is decomposed into a high-dimensional matrix, and the principal components representing the main respiratory motion patterns are screened for low-rank reconstruction to remove high-frequency jitter and obtain a spatiotemporally continuous optimized deformation field sequence that conforms to the respiratory physiological law. S5. Generation of high-quality four-dimensional magnetic resonance imaging with multiple parameters: The optimized deformation field sequence is applied frame by frame to a high-quality static image in a unified reference space to generate a high-4D high-quality four-dimensional magnetic resonance sequence containing anatomical and functional information. S6. Full-process automated quality control assessment and abnormal closed-loop processing: The generated high-4D sequence is subjected to respiratory motion consistency assessment based on optical flow field and multi-dimensional image quality assessment. At the same time, the input, registration, physiological motion consistency and clinical accuracy are monitored through a three-level quality control framework, with an abnormal circuit breaker and stratified rollback mechanism, and automatic output of case-level reports and alarms.

2. The integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control according to claim 1, characterized in that, Step S1 specifically includes: S11. Read the DICOM geometric metadata of the Original 4D sequence and multiple sets of static 3D magnetic resonance sequences, including voxel spacing, image direction cosine, and physical space origin coordinates. Map the voxel indices of all sequences to the same physical world coordinate system. Calculate the 3D coverage of the Original 4D reference sequence and each set of static 3D magnetic resonance sequences in physical space. Process the slicing order of the static 3D magnetic resonance sequences to maintain their continuity and orientation consistency in physical space. S12. For each set of static three-dimensional magnetic resonance sequences, calculate the pairwise intersection range between them and the Original 4D sequence in physical space, remove invalid data that do not overlap with the Original 4D sequence in space, and then further calculate the intersection of all valid pairwise intersection ranges to obtain the unique global minimum common field of view region. Map the clipping boundary in reverse to the original voxel coordinate system of each sequence and perform uniform clipping. S13. Based on the physical size of the global common field of view, a unified target spatial size and voxel spacing are set. Trilinear interpolation is used to uniformly resample the clipped Original 4D sequence and each static three-dimensional magnetic resonance sequence to the same spatial size and resolution, thus completing the spatiotemporal geometric consistency processing. S14. For special scenarios where the static sequence field of view cannot fully cover the target organ, a reference frame selection, cropping and resampling process based on normalized cross-correlation (NCC) is independently executed for each static 3D magnetic resonance sequence, and the corresponding 4D output is saved to retain complete anatomical information.

3. The integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control according to claim 1, characterized in that, Step S2 specifically includes: S21. Normalize and extract gradient magnitude features for each time frame of the static three-dimensional magnetic resonance sequence and Original 4D. Calculate the similarity between each time frame and the static three-dimensional sequence based on the gradient normalized mutual information. Determine the time frame with the highest similarity as the reference frame. For low signal-to-noise ratio functional sequences, calculate the gradient normalized mutual information only within a pre-defined local region of interest. S22. Set a region of interest in the diaphragm region or lung boundary region in Original 4D, and construct a respiratory amplitude signal based on the area change or bright pixel statistics in the region; divide the respiratory amplitude into multiple target respiratory phases according to the preset number of phases, and select several time frames with the closest respiratory amplitude as candidate nodes for each target respiratory phase. S23. Construct a global cumulative cost function, where the node cost represents the amplitude matching degree between the candidate time frame and the target breathing phase, and the transition cost represents the image structure continuity of the candidate time frames in adjacent stages; use the reference frame as the forced traversal node, and use the Viterbi dynamic programming algorithm to search for the global optimal path to obtain a temporally continuous phase-sequential four-dimensional magnetic resonance image sequence. S24. Construct a three-level hierarchical cascaded registration path: First level, register the T1 weighted 3D sequence to the reference frame to establish a unified reference space; Second level, register the T2 weighted 3D sequence to the registered 3D-T1 and map it to the unified reference space through composite transformation; Third level, register the diffusion-weighted imaging DWI to the completed 3D-T2 and map it to the unified reference space through cascaded transformation; During the registration process, the registration parameters are adaptively adjusted based on the image sharpness evaluation results.

4. The integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control according to claim 1, characterized in that, In step S3, the cyclic deformation registration model SEDER adopts a three-dimensional encoder-decoder deep learning architecture. For any time t, the model's input includes a reference frame image. The current Original 4D image The difference image between the two and the deformation prior image recursively generated at the previous time t-1. ; The model extracts image features through 3D convolutional layers and an attention module, fuses prior motion information from the previous time step, and predicts the 3D deformation field from the reference frame to the current time step. .

5. The integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control according to claim 4, characterized in that, The training of the SEDER model includes a warm-up training phase and a time-series recursive training phase: During the warm-up training phase, the deformation prior image is set to zero, and direct supervised training is performed using long-distance image pairs with an interval of no less than a preset number of frames to obtain basic deformation estimation capabilities. During the time-series recursive training phase, the continuous target phases are recursively predicted along the time axis, and the deformation prior image at the current moment is generated by gradient blocking of the previous prediction results. A composite loss function is used for optimization, which includes end-to-endpoint error (EPE) loss, normalized cross-correlation (NCC) loss, deformation field spatial smoothing regularization term, and time smoothing constraint term based on the second-order difference of the deformation field at adjacent time points. For the missing long-distance reference deformation field, an approximate reference deformation field is obtained by online composite of adjacent deformation fields, and it participates in training with a supervision weight lower than that of the real reference deformation field.

6. The integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control according to claim 1, characterized in that, Step S4 specifically includes: S41. The time-series deformation field sequence output in step S3 is matrixed. The three-dimensional deformation field of each breathing phase is flattened into a high-dimensional vector in the spatial dimension to construct a two-dimensional motion matrix with time points as samples and voxel displacement as features. S42. Perform singular value decomposition (SVD) on the motion matrix or eigenvalue decomposition on the covariance matrix to obtain principal components and corresponding eigenvectors sorted by energy. Select the top K principal components, where the cumulative energy of the top K principal components accounts for no less than 90%. S43. Based on the selected principal components and their corresponding time phases, the original deformation field sequence is reconstructed in low rank. The motion state of each time phase is recovered by linear combination of the first K principal components. The reconstructed matrix is ​​restored to the original four-dimensional deformation field format, and the optimized spatiotemporal continuous deformation field sequence is obtained.

7. The integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control according to claim 1, characterized in that, In step S5, the high-4D multi-parameter high-quality four-dimensional magnetic resonance sequence includes T1-weighted and T2-weighted anatomical four-dimensional sequences and diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) combined functional four-dimensional sequences. During the generation process, each phase deformation field of the optimized deformation field sequence is applied to the corresponding high-quality static image of the corresponding modality under a unified reference space to generate the corresponding modality four-dimensional sequence, thereby realizing the joint generation of anatomical and functional multi-parameter four-dimensional magnetic resonance imaging.

8. The integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control according to claim 1, characterized in that, In step S6, the respiratory motion consistency assessment based on optical flow field specifically includes: S611. Construct an adaptive computational domain based on the target region ROI mask, and expand it outward with a set margin to form a clipping window; preprocess the original Original 4D sequence and the generated high-4D sequence respectively, perform contrast-limited adaptive histogram equalization and strong Gaussian smoothing on the Original 4D, and perform weak smoothing on the high-4D. S612. Using the first frame as a reference, estimate the dense displacement field frame by frame using the Demons optical flow method; based on the three-dimensional gradient magnitude of the reference frame, filter out high-texture effective points within the ROI and remove noise interference from flat areas. S613. Extract the motion trajectories of the effective point group in the front-back, left-right, and head-to-feet directions. Calculate the mean absolute error, maximum error, and error confidence interval of the motion trajectories of the two sequences. Output the motion statistics and visualization results.

9. The integrated method for multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control according to claim 1, characterized in that, In step S6, the multi-dimensional image quality assessment includes calculating the tumor contrast-to-noise ratio (CNR), organ boundary full width at half maximum (FWHM), and cumulative probability of blur detection (CPBD) between the generated high-4D sequence and the original Original 4D sequence; the three-level quality control framework specifically includes: First-level input geometric consistency detection: Before registration, calculate the field overlap ratio and centroid offset of the input sequence. If the preset threshold is not met, the process will be terminated directly and an error message will be displayed. Physical rationality monitoring of the second-layer registration process: During the registration and deformation modeling process, the Jacobian determinant of the output deformation field is continuously monitored to detect tissue folding that is not physically deformed. If an anomaly is triggered, smoothing enhancement and parameter adaptive adjustment are automatically performed. The third layer of physiological motion consistency and clinical accuracy verification: After generating the high-4D sequence, the results are divided into three levels: pass, warning, and failure, based on motion consistency and image quality indicators. For failure results, a multi-path rollback process is automatically triggered: first, the deformation model is rerun through parameter adaptive adjustment; if the requirements are still not met, the traditional high robustness registration algorithm is switched to. In extreme cases, manual intervention prompts are triggered.

10. A multi-parameter four-dimensional magnetic resonance imaging generation, temporal optimization, and automated quality control integrated system, characterized in that, include: The preprocessing and spatiotemporal geometry unification module is used to complete the spatial coordinate mapping, common field-of-view clipping and unified resampling of the Original 4D sequence and multiple sets of static 3D magnetic resonance sequences, and eliminate systematic geometric errors. The reference frame selection and hierarchical cascaded registration module is used to select reference reference frames based on gradient normalized mutual information, complete global phase reassembly through Viterbi dynamic programming, and construct a three-level registration path with the T1 weighted sequence as the hub to complete cross-modal registration between multimodal static sequences and reference reference frames. The temporal cyclic deformation registration module is used to run the SEDER model with temporal memory mechanism. It takes the reference frame image, the target temporal image, the difference image and the recursively generated deformation prior image as joint input to extract the temporal deformation field of respiratory motion that is continuous in time. The global spatiotemporal motion smoothing module is used to perform low-rank reconstruction and noise reduction optimization of temporal deformation fields through PCA principal component analysis. The multi-parameter four-dimensional imaging generation module is used to apply the optimized deformation field to multimodal high-quality static images to generate multi-parameter high-quality four-dimensional magnetic resonance sequences that combine anatomy and function. The automated quality control and anomaly handling module is used to perform pixel-level motion consistency assessment and multi-dimensional image quality assessment. It completes full-process quality monitoring and anomaly rollback processing through a three-level quality control framework, and outputs qualified imaging results and quality control reports.

Images