A method and system for cross-contrast dynamic alignment of magnetic resonance image reconstruction

Abstract

The application discloses a cross-contrast dynamic alignment magnetic resonance image reconstruction method and system, and relates to the technical field of magnetic resonance image processing. The method acquires target contrast undersampling k-space data and auxiliary contrast images, constructs two types of contrast features; determines the corresponding relationship and offset relationship of the target contrast spatial region and the auxiliary contrast spatial region, and performs spatial correction on the auxiliary features; based on the target features, the reliability of the aligned auxiliary features is evaluated and adaptively adjusted; in multi-scale reconstruction, the dynamic alignment information is updated according to the spatial offset relationship between the current scale reconstruction features and the reliable auxiliary features, the reliable auxiliary features are re-corrected and fused, and the decoding result is used for subsequent reconstruction scales to obtain a target contrast reconstruction image. Thus, the error caused by spatial misalignment and unreliable auxiliary information can be reduced, and the structural consistency and stability of the reconstruction image are improved.

Description

Technical Field

[0001] This application relates to the field of magnetic resonance image processing technology, and in particular to a magnetic resonance image reconstruction method and system with cross-contrast dynamic alignment. Background Technology

[0002] Magnetic resonance imaging (MRI) offers multi-parameter, multi-sequence imaging capabilities. By employing different scanning sequences, it can obtain MRI images with varying tissue contrast characteristics, such as T1-weighted images, T2-weighted images, proton density images, and fluid attenuation inversion recovery images. MRI images of different contrast ratios can reflect tissue structure and pathological information from different perspectives; therefore, multi-contrast MRI is widely used in medical image analysis and assisted diagnosis. However, different contrast sequences typically require separate data acquisition, resulting in a long overall scan time. To improve MRI efficiency, undersampling of the target contrast MRI data can be performed, and auxiliary information can be provided using other previously acquired contrast MRI images of the same subject to reconstruct the target contrast MRI image.

[0003] Existing multi-contrast magnetic resonance imaging (MRI) image reconstruction methods typically utilize shared tissue structure information between images of different contrast ratios. This is achieved through methods such as feature stitching, feature weighting, attention mechanisms, or joint reconstruction networks, incorporating information from auxiliary contrast images into the reconstruction process of the target contrast image. However, due to potential differences in acquisition time, scanning parameters, and the state of the examined object among MRI images of different contrast ratios, spatial shifts, local deformations, or inconsistencies in tissue boundaries can easily occur between these images. This can prevent the structural information in the auxiliary contrast images from accurately corresponding to the target contrast image. Furthermore, the degree of cross-contrast correlation varies across different spatial regions. Existing methods often lack effective assessment of the reliability of auxiliary information from different regions when fusing auxiliary information, easily introducing misaligned or inconsistent auxiliary information into the reconstruction process. This can lead to local structural mismatches, blurred boundaries, or distortion of image details, affecting the reconstruction quality and stability of MRI images under undersampling conditions.

[0004] Therefore, in the process of multi-contrast magnetic resonance image reconstruction, the unstable spatial correspondence between images of different contrasts, the difficulty in effectively judging the reliability of auxiliary contrast information, and the susceptibility of reconstruction quality to the mismatch of auxiliary information under undersampling conditions have become urgent problems to be solved. Summary of the Invention

[0005] This application provides a magnetic resonance image reconstruction method and system with cross-contrast dynamic alignment, aiming to solve the problems in the existing technology of unstable spatial correspondence between images with different contrasts, difficulty in effectively judging the reliability of auxiliary contrast information, and easy influence of auxiliary information mismatch on reconstruction quality under undersampling conditions during the reconstruction of multi-contrast magnetic resonance images.

[0006] In a first aspect, a magnetic resonance image reconstruction method with dynamic cross-contrast alignment is provided, the method comprising: Acquire target contrast undersampled k-space data and at least one auxiliary contrast image of the same subject, and construct target contrast features and auxiliary contrast features respectively; Based on the target contrast feature and the auxiliary contrast feature, the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region is determined. Based on the spatial correspondence, the offset relationship between different spatial regions is determined to obtain cross-contrast dynamic alignment information. Based on the cross-contrast dynamic alignment information, the auxiliary contrast feature is spatially corrected to obtain alignment auxiliary features. Based on the target contrast features, the alignment auxiliary features are reliably evaluated and adaptively adjusted to obtain reliable auxiliary features; The target contrast features and the reliable auxiliary features are fused together. During the multi-scale reconstruction process, the cross-contrast dynamic alignment information is updated according to the spatial offset relationship between the reconstruction features at the current reconstruction scale and the reliable auxiliary features. The reliable auxiliary features are spatially corrected according to the updated cross-contrast dynamic alignment information. The spatially corrected reliable auxiliary features are then fused with the reconstruction features at the current reconstruction scale. The fused decoding result is used for subsequent reconstruction scales. The update, spatial correction, fusion, and decoding processes are repeated to obtain the target contrast reconstructed image.

[0007] Optionally, in the above scheme, constructing the target contrast feature and auxiliary contrast feature respectively includes: The target contrast undersampled k-space data is subjected to data format conversion, multi-coil correction and normalization to obtain preprocessed target k-space data; Perform an inverse Fourier transform on the preprocessed target k-space data to obtain the initial image of the target contrast, and retain the sampling mask corresponding to the undersampled k-space data of the target contrast. Feature extraction is performed on the initial image of the target contrast to obtain the target contrast features; Feature extraction is performed on the auxiliary contrast image to obtain the auxiliary contrast features.

[0008] Optionally, in the above scheme, determining the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region, and determining the offset relationship between different spatial regions based on the spatial correspondence, includes: Perform cross-contrast structure correlation analysis on the target contrast feature and the auxiliary contrast feature to obtain the cross-contrast structure correlation results; Based on the cross-contrast structure association results, the spatial correspondence between the target contrast space region and the auxiliary contrast space region is determined, and the cross-contrast space correspondence results are obtained. Based on the cross-contrast space correspondence results, the offset relationship between different spatial regions is determined, and the cross-contrast dynamic alignment information is obtained.

[0009] Optionally, in the above scheme, determining the offset relationship between different spatial regions based on the cross-contrast space correspondence result to obtain the cross-contrast dynamic alignment information includes: Based on the cross-contrast space correspondence results, motion estimation or deformation field prediction is performed on the target contrast features and the auxiliary contrast features to obtain pixel-level or region-level spatial offset information. A cross-contrast dynamic displacement field is constructed based on the spatial offset information, and the cross-contrast dynamic displacement field is used as the cross-contrast dynamic alignment information. Based on the cross-contrast dynamic displacement field, the auxiliary contrast feature is spatially mapped to obtain the alignment auxiliary feature corresponding to the target contrast feature.

[0010] Optionally, in the above scheme, the step of performing reliability assessment and adaptive adjustment of the alignment auxiliary features based on the target contrast features includes: The target contrast feature and the alignment auxiliary feature are divided into multiple corresponding spatial regions to obtain multiple corresponding region feature pairs. The correlation between the target region features and the auxiliary region features in each corresponding region feature pair is determined to obtain region correlation information; Based on the regional correlation information, the reliability of the auxiliary information corresponding to each spatial region is determined to obtain the reliability information of the auxiliary information. The reliable auxiliary features are obtained by enhancing the auxiliary region features that are consistent with the target contrast feature structure based on the reliability information of the auxiliary information, and suppressing the auxiliary region features that are misaligned or structurally inconsistent with the target contrast feature.

[0011] Optionally, in the above scheme, the step of performing reliability assessment and adaptive adjustment of the alignment auxiliary feature based on the target contrast feature, and fusing the target contrast feature and the reliable auxiliary feature, includes: The reliability of the target contrast feature and the alignment auxiliary feature is evaluated to obtain the reliability evaluation result. Based on the reliability assessment results, the degree of auxiliary feature fusion corresponding to different spatial regions is determined, and the region fusion weight is obtained. The alignment auxiliary features are weighted and modulated according to the region fusion weights to obtain the reliable auxiliary features; The reliable auxiliary features are combined with the target contrast features through feature splicing, weighted modulation, or attention-guided fusion to obtain cross-contrast fusion features.

[0012] Optionally, in the above scheme, updating the cross-contrast dynamic alignment information during multi-scale reconstruction includes: Multi-scale feature recovery is performed on the cross-contrast fusion feature formed by fusing the target contrast feature and the reliable auxiliary feature to obtain scale reconstruction features corresponding to multiple reconstruction scales; For each of the reconstruction scales, the cross-contrast dynamic alignment information is updated based on the spatial offset relationship between the scale reconstruction features and the reliable auxiliary features at the corresponding scale to obtain scale alignment information; For each reconstruction scale, the reliable auxiliary features are spatially adjusted according to the corresponding scale alignment information, and the spatially adjusted reliable auxiliary features are fused with the scale reconstruction features corresponding to the reconstruction scale to obtain scale-up reconstruction features; The scale-up reconstruction features are progressively decoded, and the scale alignment information is continuously updated during the decoding process to obtain the target contrast reconstruction image.

[0013] Optionally, in the above scheme, the progressive decoding of the scale-updated reconstructed features includes: Determine the current reconstruction scale, and update the scale alignment information according to the spatial offset relationship between the scale reconstruction features corresponding to the current reconstruction scale and the reliable auxiliary features to obtain the current scale alignment information; Based on the current scale alignment information, the reliable auxiliary features are spatially corrected to obtain the current scale alignment auxiliary features; The current scale alignment auxiliary feature is fused with the scale reconstruction feature corresponding to the current reconstruction scale to obtain the current scale updated reconstruction feature; The current scale-up reconstruction features are decoded, and the decoding results are used for feature recovery at subsequent reconstruction scales. The scale alignment information update, spatial correction, fusion, and decoding processes are repeated until the progressive decoding of each reconstruction scale is completed, resulting in a target contrast prediction image.

[0014] Optionally, in the above scheme, obtaining the target contrast reconstructed image includes: The contrast image of the target to be corrected is obtained through the multi-scale reconstruction process. Perform a Fourier transform on the contrast image of the target to be corrected to obtain the predicted k-space data; Based on the sampling mask corresponding to the target contrast undersampled k-space data, the data at the sampled positions in the predicted k-space data are replaced with the actual measurement data in the target contrast undersampled k-space data, while the predicted data at the unsampled positions are retained, to obtain the corrected k-space data. The target contrast reconstruction image is obtained by performing an inverse Fourier transform on the corrected k-space data.

[0015] Secondly, a magnetic resonance image reconstruction system with dynamic cross-contrast alignment, the system comprising: The feature construction module is used to acquire the target contrast undersampled k-space data and at least one auxiliary contrast image of the same inspected object, and to construct the target contrast features and auxiliary contrast features respectively. The dynamic alignment module is used to determine the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region based on the target contrast feature and the auxiliary contrast feature, determine the offset relationship between different spatial regions based on the spatial correspondence, obtain cross-contrast dynamic alignment information, and perform spatial correction on the auxiliary contrast feature based on the cross-contrast dynamic alignment information to obtain alignment auxiliary features. A reliability adjustment module is used to perform reliability assessment and adaptive adjustment of the alignment auxiliary feature based on the target contrast feature to obtain a reliable auxiliary feature. The fusion reconstruction module is used to fuse the target contrast features and the reliable auxiliary features. During the multi-scale reconstruction process, it updates the cross-contrast dynamic alignment information according to the spatial offset relationship between the reconstruction features at the current reconstruction scale and the reliable auxiliary features. It then performs spatial correction on the reliable auxiliary features based on the updated cross-contrast dynamic alignment information and fuses the spatially corrected reliable auxiliary features with the reconstruction features at the current reconstruction scale. The fused decoding result is used for subsequent reconstruction scales. The update, spatial correction, fusion, and decoding processes are repeated to obtain the target contrast reconstructed image.

[0016] Compared with the prior art, this application has at least the following beneficial effects: This application determines the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region, and further determines the offset relationship between different spatial regions, so that the cross-contrast dynamic alignment information has a clear spatial correction basis, thereby enabling the adjustment of auxiliary features according to the actual misalignment state of different regions. After the reliability assessment and adaptive adjustment of the spatially corrected auxiliary features, the auxiliary information consistent with the target structure is preserved and enhanced, while the auxiliary information that still has misalignment or structural inconsistency is suppressed. Furthermore, in the multi-scale reconstruction process, the cross-contrast dynamic alignment information is updated according to the spatial offset relationship between the reconstruction features and reliable auxiliary features at the current reconstruction scale, and the updated dynamic alignment information is used to re-correct and fuse reliable auxiliary features, so that the target structure recovered at the current scale can have a feedback effect on the subsequent spatial alignment and image restoration process, thereby reducing the continuous propagation of initial alignment error and unreliable auxiliary information in multiple reconstruction scales, and improving the structural consistency, detail fidelity and reconstruction stability of the target contrast reconstructed image. Attached Figure Description

[0017] Figure 1 A schematic flowchart of a magnetic resonance image reconstruction method with cross-contrast dynamic alignment provided in one embodiment of this application; Figure 2 This is a schematic diagram of a cross-contrast dynamic alignment magnetic resonance image reconstruction framework provided in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In one embodiment, such as Figure 1 and Figure 2 As shown, a magnetic resonance image reconstruction method with cross-contrast dynamic alignment is provided. The image information corresponding to the undersampled data of the target contrast and the auxiliary contrast image information are respectively entered into their respective feature processing branches. The spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region is determined based on the target contrast features and auxiliary contrast features. The offset relationship between different spatial regions is also determined based on the spatial correspondence, forming cross-contrast dynamic alignment information. Based on this, spatial correction, reliability assessment, and adaptive adjustment are performed on the auxiliary contrast features. During multi-scale reconstruction, the cross-contrast dynamic alignment information is updated according to the spatial offset relationship between the reconstruction features and reliable auxiliary features at the current reconstruction scale. The reliable auxiliary features are then re-spatialized and fused. The decoding result of the current scale is used for subsequent reconstruction scales, and the above process is repeated to obtain the target contrast reconstructed image.

[0020] In one implementation, target contrast undersampled k-space data and at least one auxiliary contrast image are first acquired for the same subject. The target contrast undersampled k-space data is frequency domain data actually acquired by the magnetic resonance imaging (MRI) scanner according to the corresponding undersampling method, used to provide true observation information about the target contrast itself; the auxiliary contrast image is used to provide information about anatomical structures or tissue complementarities shared with the target contrast. For example, a T2-weighted image can be used as the target contrast to be recovered, and an already acquired T1-weighted image can be used as the auxiliary contrast; other contrast combinations can also be used depending on the clinical scanning task.

[0021] After acquiring the undersampled k-space data of the target contrast, preprocessing and initial image restoration can be performed to obtain an image representation that reflects the spatial structure of the target contrast itself. During the target contrast data processing, auxiliary contrast images are not used to replace the observed target contrast data; instead, the undersampled target contrast data is used to establish the foundation of target contrast information. Subsequently, feature extraction is performed on both the target contrast and auxiliary contrast data to obtain target contrast features and auxiliary contrast features. Target contrast features primarily characterize the currently recoverable tissue structure, spatial distribution, and image content of the target contrast, while auxiliary contrast features characterize the information in the auxiliary contrast image that can be used to supplement the target contrast reconstruction.

[0022] After obtaining the target contrast features and auxiliary contrast features, instead of directly assuming a strict spatial correspondence between them, a cross-contrast structural correlation analysis is used to determine the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region. Then, based on this spatial correspondence, the offset relationship between different spatial regions is determined, resulting in cross-contrast dynamic alignment information. Specifically, the cross-contrast relationship modeling process can be used to analyze the structural changes and spatial correspondence states between different contrasts, establishing a dynamic correspondence between the tissue regions in the target space and the corresponding tissue regions in the auxiliary space. This cross-contrast dynamic alignment information reflects the spatial offset or spatial change state of the auxiliary contrast features relative to the target contrast features.

[0023] Spatial correction is performed on auxiliary contrast features based on cross-contrast dynamic alignment information, adjusting the corresponding structures in the auxiliary contrast features to their spatial positions corresponding to the target contrast features, thus obtaining aligned auxiliary features. This spatial correction does not require a fixed registration before the image enters the reconstruction network, but rather allows for dynamic adjustment of the spatial position of the auxiliary information based on the current feature states of the target and auxiliary contrasts, thereby adapting to cross-contrast spatial misalignment caused by patient movement, differences in acquisition time between different sequences, differences in scanning parameters, and local tissue deformation.

[0024] After obtaining the alignment auxiliary features, a reliability assessment is performed on them using the target contrast features as a reference. The reliability assessment determines whether the spatially corrected auxiliary information maintains sufficient consistency with the target region. For auxiliary information that corresponds closely to the target contrast structure, its participation in subsequent reconstruction is increased; for auxiliary information that still exhibits misalignment, local deformation, or structural inconsistency, its participation is reduced. Through this reliability assessment and adaptive adjustment process, reliable auxiliary features are obtained.

[0025] Subsequently, target contrast features and reliable auxiliary features are fused. Reliable auxiliary features compensate for the structural information loss caused by undersampling of the target contrast, while the target contrast features maintain the correspondence between the final reconstruction result and the target's own imaging characteristics. A multi-scale reconstruction approach is employed during image restoration. For the current reconstruction scale, cross-contrast dynamic alignment information is updated based on the spatial offset relationship between the reconstruction features and reliable auxiliary features at that scale. The reliable auxiliary features are then spatially corrected based on the updated cross-contrast dynamic alignment information, and fused with the reconstruction features at the current reconstruction scale. The fused decoding result is used for subsequent reconstruction scales, and the update, spatial correction, fusion, and decoding processes are repeated. Finally, the target contrast reconstructed image is obtained based on the multi-scale reconstruction results.

[0026] Through the above processing, the spatial relationship between the target contrast and the auxiliary contrast can be dynamically adjusted according to the actual feature state. Before using the auxiliary information, the reliability of the auxiliary information is further judged, so that the auxiliary information that corresponds to the spatial position and is consistent with the target structure participates more in the target image restoration, thereby reducing the interference caused by misaligned or inconsistent information. At the same time, the target structure restored at the current reconstruction scale participates in the feature restoration of subsequent reconstruction scales through the decoding results and is used to continue to update the dynamic alignment state. This makes cross-contrast alignment, information fusion and target image restoration form a feedback and collaborative process, reducing the continuous propagation of initial alignment errors and unreliable auxiliary information in multiple reconstruction scales, thereby improving the structural consistency, detail fidelity and reconstruction stability of the target contrast reconstructed image.

[0027] In this embodiment, the target contrast undersampled k-space data is preprocessed. The preprocessing may sequentially include data format conversion, multi-coil correction, and normalization. Data format conversion converts the target contrast undersampled data output by the magnetic resonance imaging (MRI) device into a data format suitable for subsequent reconstruction processing. Multi-coil correction processes data corresponding to different coils under multi-coil MRI acquisition conditions, ensuring the target contrast data forms a unified reconstruction input. Normalization reduces the impact of different data amplitude ranges on subsequent feature extraction and image reconstruction. After completing these processes, the preprocessed target k-space data is obtained.

[0028] An inverse Fourier transform is performed on the preprocessed target k-space data to convert the undersampled data in the frequency domain to the image space, resulting in an initial target contrast image. Because the target k-space data is undersampled, the initial target contrast image may contain aliasing artifacts or missing information corresponding to the undersampling method. However, this initial target contrast image still contains the basic organizational structure and image content corresponding to the true sampled information of the target contrast. Simultaneously with forming the initial target contrast image, a sampling mask corresponding to the undersampled k-space data of the target contrast is retained. This sampling mask is used to record the sampled and unsampled positions in k-space and can distinguish between real acquired data and predicted completion data during subsequent magnetic resonance physical consistency processing.

[0029] Subsequently, feature extraction is performed on the initial target contrast image to obtain target contrast features. The feature extraction process can be completed by the target contrast feature extraction branch in the magnetic resonance image reconstruction network, which gradually transforms the initial target contrast image into a feature representation for cross-contrast relationship modeling and image restoration.

[0030] For the auxiliary contrast image, feature extraction is performed through an auxiliary contrast feature extraction branch to obtain auxiliary contrast features. Unlike the processing method that requires strict spatial registration between the auxiliary contrast image and the target contrast image beforehand, the spatial differences that may exist between different contrasts are preserved in the auxiliary contrast feature construction stage, and adjustments are made according to the actual feature relationship between the target contrast and the auxiliary contrast during the subsequent cross-contrast dynamic alignment process.

[0031] Through the above processing, we can establish independent and further correlated feature representations from the actual undersampled observation data of the target contrast and the auxiliary contrast image, respectively. While preserving the observation constraints of the target contrast itself, we introduce the structural information of the auxiliary contrast and avoid assuming that different contrasts are strictly spatially corresponding in the feature construction stage. This provides a feature basis with actual spatial differences for subsequent modeling of cross-contrast dynamic spatial relationships.

[0032] In this embodiment, the target contrast features and auxiliary contrast features are input into the cross-contrast relationship modeling process, and cross-contrast structural correlation analysis is performed on them. The purpose of structural correlation analysis is to identify correlation information that reflects the same or corresponding tissue structure from the feature representations of different magnetic resonance contrasts, rather than directly identifying them as belonging to the same tissue region based on the same coordinate positions in two images. After completing the cross-contrast structural correlation analysis, the cross-contrast structural correlation results are obtained.

[0033] Based on the cross-contrast structural correlation results, the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region is further determined. Specifically, based on the correlation between the target contrast features and the auxiliary contrast features in terms of spatial distribution and structural response, it can be determined that the corresponding region in the auxiliary space should correspond to the target region in the target space, thus obtaining the cross-contrast spatial correspondence result. This cross-contrast spatial correspondence result can vary with different inspected objects, different contrast combinations, and different scanning states, rather than using a fixed spatial correspondence template.

[0034] After obtaining the cross-contrast spatial correspondence results, the offset relationship between different spatial regions is determined based on the positional change state between each corresponding spatial region, and the cross-contrast dynamic alignment information is formed from the offset relationship. The cross-contrast dynamic alignment information can be used to characterize the spatial changes between the target contrast and the auxiliary contrast caused by motion, scanning differences, or local deformation, and serves as the basis for subsequent auxiliary feature space correction.

[0035] Through the above processing, the fixed or predetermined strict coordinate correspondence between images with different contrast is no longer relied upon. Instead, the spatial correspondence and offset state are gradually determined from the structural relationship between the target contrast features and the auxiliary contrast features themselves. This enables regional-level dynamic modeling of spatial offsets generated at different positions during the actual scanning process, providing a dynamic spatial basis for the accurate transmission of auxiliary contrast information to the target space, and reducing the impact of fixed registration errors on the subsequent reconstruction process.

[0036] Based on the aforementioned method for determining cross-contrast spatial relationships, motion estimation or deformation field prediction is performed on the target contrast features and auxiliary contrast features according to the cross-contrast spatial correspondence results to further quantify the offset state between different spatial regions. Motion estimation can be used to characterize positional changes between different contrasts caused by factors such as the movement of the inspected object, while deformation field prediction can be used to characterize non-rigid spatial changes occurring in local tissue regions. Depending on the granularity of different spatial changes, pixel-level spatial offset information or region-level spatial offset information can be obtained.

[0037] A cross-contrast dynamic displacement field is constructed based on the obtained pixel-level or region-level spatial offset information. This field represents the spatial adjustment state of a corresponding position in the auxiliary contrast feature relative to the corresponding position in the target contrast feature, and serves as a specific form of cross-contrast dynamic alignment information. The dynamic displacement field is not statically maintained throughout the reconstruction process after its initial generation, but rather can be further updated in conjunction with subsequent multi-scale recovery processes.

[0038] Spatial mapping of auxiliary contrast features is performed based on a cross-contrast dynamic displacement field. Specifically, according to the spatial offset direction and offset state indicated by the dynamic displacement field, the corresponding region in the auxiliary contrast feature is adjusted to the spatial position corresponding to the target contrast feature. This allows the auxiliary structure, which was originally spatially misaligned due to motion, local deformation, or differences in scanning conditions, to establish a more accurate spatial relationship with the corresponding structure in the target space, thus obtaining aligned auxiliary features.

[0039] In practical implementation, motion estimation can employ optical flow estimation or deformation field prediction that outputs spatial deformation information; as long as the dynamic displacement information for auxiliary feature space correction can be obtained based on the current feature state between the target contrast and the auxiliary contrast, it is acceptable. Thus, the spatial alignment process can be embedded into the magnetic resonance image reconstruction workflow without requiring a fixed external registration result as a constant input for subsequent reconstruction processes.

[0040] Through the above processing, the abstract cross-contrast spatial correspondence is further transformed into dynamic displacement information that can directly guide the adjustment of auxiliary feature positions. Based on the dynamic displacement field, the auxiliary features are adaptively corrected to the target space. Therefore, it can adapt to local offsets and non-rigid deformations under complex clinical scanning conditions, improve the spatial matching accuracy between the auxiliary contrast structure and the target contrast structure, and reduce structural mismatch and boundary distortion caused by misaligned auxiliary information directly entering the reconstruction network.

[0041] After obtaining the alignment auxiliary features, the target contrast features and alignment auxiliary features are processed according to their corresponding spatial regions, so that a region in the target space can form a correspondence with the corresponding region in the spatially corrected auxiliary space, thereby obtaining multiple corresponding region feature pairs. Each corresponding region feature pair contains both target region features and auxiliary region features.

[0042] For each corresponding feature pair, the correlation between the target region features and the auxiliary region features is determined to obtain region correlation information. This region correlation information represents the degree of matching between the structural information in the auxiliary region and the structural information of the target region itself after dynamic spatial correction. Since different magnetic resonance contrast ratios exhibit different signal representations for the same tissue, reliability judgment does not require the two contrast ratios to have identical grayscale representations. Instead, it determines whether the tissue structural information carried by the auxiliary region can effectively correspond to the target region.

[0043] Based on the regional correlation information corresponding to each spatial region, the reliability of the corresponding auxiliary information is determined, thus obtaining the reliability information of the auxiliary information. In one implementation, the reliability judgment can be completed through regional feature correlation analysis and attention-guided methods, enabling the reliability evaluation to dynamically change according to the actual cross-contrast corresponding state of different regions, rather than setting a uniform auxiliary information contribution level for all auxiliary regions.

[0044] Based on the reliability information of the auxiliary information, the alignment auxiliary features are adaptively adjusted. For auxiliary region features with high structural consistency with the target contrast features, their feature responses are enhanced, enabling them to provide more shared anatomical structural information in the subsequent target image restoration process. For auxiliary region features that still have obvious misalignment, deformation, or structural inconsistency, their feature responses are suppressed to reduce the possibility of the auxiliary information in that region causing erroneous constraints on the reconstruction results. Reliable auxiliary features are obtained after processing.

[0045] Through the above processing, it is possible to further determine whether the corrected auxiliary information is worth using after spatial position correction is completed, so that spatial alignment and information reliability judgment form two continuous but functionally different processing levels. As a result, auxiliary information can be dynamically filtered according to the cross-contrast correlation differences of different spatial regions, enhancing the structural information that is truly beneficial to the restoration of the target image and suppressing potential erroneous information, thereby reducing local structural distortion caused by misuse of auxiliary information and improving the stability and robustness of multi-contrast magnetic resonance reconstruction under complex scanning conditions.

[0046] In this embodiment, a reliability assessment is performed on the target contrast features and alignment auxiliary features to obtain the reliability assessment results. The reliability assessment results are used to reflect the usability of auxiliary features in different spatial regions for target image restoration, and can be used to subsequently determine the degree of participation of auxiliary information in each region.

[0047] Based on the reliability assessment results, the degree of auxiliary feature fusion for different spatial regions is determined, resulting in regional fusion weights. These regional fusion weights are used to adjust the contribution of auxiliary information at different spatial locations, ensuring that different regions no longer employ a fixed, uniform auxiliary information fusion strength. For spatial regions with high reliability, a higher level of auxiliary information participation can be set; for spatial regions with low reliability, the participation level is correspondingly reduced.

[0048] The alignment auxiliary features are weighted and modulated according to the region fusion weights to obtain reliable auxiliary features. Thus, reliable auxiliary features retain both the structural information corresponding to the target space after dynamic spatial alignment and reflect the differences in reliability between different spatial regions.

[0049] Subsequently, reliable auxiliary features are fused with target contrast features across contrast ratios. Under different network implementations, fusion can employ methods such as feature concatenation, weighted modulation, or attention-guided fusion. Feature concatenation combines target and auxiliary information along the feature dimension; weighted modulation adjusts the influence of auxiliary information on target features based on its reliability; and attention-guided fusion dynamically selects auxiliary features according to the current information state required by the target region. The resulting cross-contrast fused features are then used for subsequent image restoration.

[0050] When the input includes multiple auxiliary contrast images, the auxiliary contrast features corresponding to each auxiliary contrast can be extracted separately, and dynamic spatial correction and reliability assessment can be performed separately. This ensures that the information of different auxiliary contrasts undergoes spatial relationship judgment and reliability adjustment with the target contrast before entering the target image reconstruction process, thus maintaining the processing logic of "alignment and judgment first, then fusion".

[0051] Through the above processing, the auxiliary contrast information is not directly added to the target reconstruction with uniform weights or in an unfiltered form. Instead, its degree of fusion is determined according to the actual reliability of each spatial region. This allows the reliable auxiliary structure to effectively supplement the missing information in the undersampled target data, while reducing the propagation of erroneous information caused by misalignment, deformation, and cross-contrast inconsistency regions. This, in turn, improves the utilization efficiency of multi-contrast complementary information and the accuracy of target contrast image restoration.

[0052] like Figure 2 As shown, after obtaining the cross-contrast fusion feature formed by fusing the target contrast feature and reliable auxiliary features, multi-scale feature recovery is performed on the cross-contrast fusion feature. Scale reconstruction features adapted to the corresponding image recovery state are gradually formed at different reconstruction stages, thus obtaining scale reconstruction features corresponding to multiple reconstruction scales.

[0053] For each reconstruction scale, instead of directly using the cross-contrast dynamic alignment information formed in the previous stage as the fixed spatial relationship for the current scale, the cross-contrast dynamic alignment information is updated based on the spatial offset relationship between the scale reconstruction features and reliable auxiliary features corresponding to the current scale, thus obtaining the scale alignment information corresponding to the current reconstruction scale. As the tissue contours, local edges, and image details of the target image are gradually restored, the target features themselves can provide richer spatial structural information. Therefore, the spatial correspondence between the target contrast and the auxiliary contrast can be further adjusted using the current restoration state.

[0054] Based on the scale alignment information, the reliable auxiliary features at the corresponding reconstruction scale are spatially adjusted again to ensure that the reliable auxiliary features correspond to the target structure that has been restored at the current scale. Then, the spatially adjusted reliable auxiliary features are fused with the scale reconstruction features corresponding to that reconstruction scale to obtain the scale-updated reconstruction features. Thus, the reliable auxiliary features at each scale do not simply use the initial alignment results, but can be readjusted according to the current image restoration state.

[0055] Progressive decoding is performed on the obtained scale-up reconstructed features, and the decoding result of the current reconstruction scale is used for feature recovery at subsequent reconstruction scales. Upon entering a subsequent reconstruction scale, the scale alignment information is continuously updated based on the spatial offset relationship between the newly formed scale-up reconstructed features and reliable auxiliary features. As the feature recovery and decoding process at multiple scales progresses, the spatial correspondence between different contrast ratios is also optimized synchronously, allowing the coarse structural correspondence to gradually serve the recovery of more detailed structures and image details. After completing multi-scale progressive decoding, the target contrast reconstructed image can be obtained. In an implementation that further performs magnetic resonance physical consistency correction, this multi-scale output can also be used as the target contrast image to be corrected in subsequent processing.

[0056] This processing procedure corresponds to Figure 2 The progressive implicit alignment decoding idea shown is that cross-contrast space alignment is not a preprocessing result independent of the reconstruction process, but is embedded in multiple reconstruction scales and adjusted synchronously with the target image restoration.

[0057] Through the above processing, the cross-contrast spatial correspondence can be continuously corrected during the gradual recovery of the target image from the initial undersampled state, avoiding the continuous accumulation of fixed registration or single dynamic alignment errors in the subsequent reconstruction process. This enables spatial correction, auxiliary information fusion and target image recovery to work together, thereby further improving the recovery capability of tissue edges, local structures and image details, and improving the reconstruction adaptability under different scanning protocols, different contrast combinations and different spatial variation conditions.

[0058] In the multi-scale reconstruction process, the current reconstruction scale is determined based on the current image restoration state, and the scale reconstruction features corresponding to the current reconstruction scale are obtained. Subsequently, based on the spatial offset relationship between the scale reconstruction features and reliable auxiliary features, the existing scale alignment information is updated to obtain the current scale alignment information. The current scale alignment information reflects the spatial correspondence between the auxiliary information and the target structure in the current reconstruction state.

[0059] Based on the current scale alignment information, the reliable auxiliary features involved in the current scale reconstruction are spatially corrected so that the auxiliary information can be further adapted to the target organizational structure that has been recovered at the current scale, thus obtaining the current scale aligned auxiliary features.

[0060] The current scale alignment auxiliary features are fused with the scale reconstruction features corresponding to the current reconstruction scale to obtain the current scale updated reconstruction features, which are then decoded. The decoding results are used to form the feature recovery basis required for subsequent reconstruction scales, enabling the later stage to continue feature recovery and spatial correspondence adjustment based on the target structure already recovered in the previous stage.

[0061] Upon entering subsequent reconstruction scales, the scale alignment information is updated again based on the spatial offset relationship between the scale reconstruction features and reliable auxiliary features at the current reconstruction scale. Auxiliary feature spatial correction and feature fusion are then performed based on the updated scale alignment information. This process of updating scale alignment information, correcting auxiliary feature spatial correction, feature fusion, and decoding is executed progressively at each reconstruction scale until progressive decoding across multiple reconstruction scales is completed, yielding the target contrast prediction image.

[0062] This progressive processing does not limit the use of a fixed single-scale processing method. Instead, it emphasizes that during the gradual image restoration process, the decoding results of the current scale are used for feature restoration at the subsequent reconstruction scale. The newly formed scale reconstruction features continue to participate in the determination of the cross-contrast spatial offset relationship and the updating of scale alignment information, so that the spatial correspondence state of the auxiliary information changes synchronously with the target image restoration process.

[0063] Through the above processing, the restoration results of each stage of the target contrast image can continue to have a feedback effect on subsequent cross-contrast space alignment, so that dynamic alignment is no longer limited to a single correction before reconstruction begins, but runs through the multi-scale decoding process. This can reduce the impact of the initial spatial estimation error on subsequent detail restoration, and enable cross-contrast structural information to be used more accurately as the target image is gradually restored, thereby improving the restoration quality of local edges, textures and fine anatomical structures.

[0064] In one implementation, a target contrast image to be corrected is obtained through the aforementioned multi-scale reconstruction process. This target contrast image has already fused undersampled target contrast information and auxiliary contrast information after dynamic alignment and reliability screening. However, in order to further ensure that the image reconstruction result conforms to the actual sampling rules of magnetic resonance, k-space data consistency correction is performed on it.

[0065] First, a Fourier transform is performed on the contrast image of the target to be corrected to convert the reconstruction result in the image space to the k space, thus obtaining the predicted k space data.

[0066] Subsequently, the sampling mask corresponding to the aforementioned undersampled k-space data of the target contrast is read. The sampling mask is used to distinguish between sampled and unsampled locations in the predicted k-space data. For sampled locations, the corresponding data in the predicted k-space data is replaced with the actual measurement data from the undersampled k-space data of the target contrast, thus preserving the observations obtained from the actual magnetic resonance scans. For unsampled locations, the predicted data generated by the reconstruction network is retained to supplement the actual unacquired frequency domain data using cross-contrast auxiliary information. The resulting corrected k-space data is then obtained.

[0067] An inverse Fourier transform is performed on the corrected k-space data, and the frequency domain data after real sampling constraints and prediction completion is converted back to the image space to obtain the target contrast reconstructed image.

[0068] After obtaining the target contrast reconstructed image, in performance evaluation scenarios with a reference high-quality target contrast image, the reconstructed image can be compared with the reference high-quality image, and the reconstruction quality can be evaluated using metrics such as PSNR, SSIM, and NMSE. Furthermore, reconstruction performance can be compared and analyzed for different undersampling rates, different sampling modes, and different combinations of target contrast and auxiliary contrast to evaluate the adaptability of cross-contrast dynamic alignment under different fast magnetic resonance scanning conditions.

[0069] The resulting target contrast reconstructed image can be used for image browsing and assisted diagnosis after rapid MRI scans, as well as for multi-contrast medical image analysis, lesion detection, and automated image reconstruction and quality control in intelligent MRI systems. In specific applications, the aforementioned quality assessment is an optional process and does not affect the execution of the aforementioned target contrast reconstruction procedure.

[0070] Through the above processing, while supplementing the undersampled target data with auxiliary contrast structure information, the reconstruction results can be reconstrained with the actual acquired target contrast k-space data. This prevents the network from arbitrarily changing the frequency domain information that has already been acquired, and only uses the prediction results to supplement the unsampled positions. As a result, the final image simultaneously meets the requirements of cross-contrast structure restoration and the constraints of actual magnetic resonance sampling data, improving the authenticity, data consistency and clinical reliability of the target contrast reconstruction results.

[0071] In one embodiment, a magnetic resonance imaging (MRI) image reconstruction system with cross-contrast dynamic alignment is provided. This system can be deployed in an image reconstruction workstation of an MRI scanner, a medical image processing server, or a computing device capable of performing MRI image reconstruction tasks. Figure 2 As shown, its functional processing can be compared with the aforementioned cross-contrast dynamic alignment and progressive reconstruction framework.

[0072] The system includes a feature construction module. This module acquires target contrast undersampled k-space data and at least one auxiliary contrast image of the same subject, and constructs target contrast features and auxiliary contrast features respectively. For the target contrast undersampled k-space data, the feature construction module sequentially performs data format conversion, multi-coil correction, normalization, and initial reconstruction, and extracts target contrast features from the initial target image; for the auxiliary contrast image, it extracts auxiliary contrast features to provide tissue structure and cross-contrast complementary information.

[0073] The system also includes a dynamic alignment module. This module determines the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region based on the target contrast features and auxiliary contrast features. It then determines the offset relationship between different spatial regions based on the spatial correspondence, obtaining cross-contrast dynamic alignment information. Finally, it performs spatial correction on the auxiliary contrast features based on this cross-contrast dynamic alignment information to obtain aligned auxiliary features. In one specific implementation, the dynamic alignment module may further include cross-contrast spatial relationship modeling and dynamic offset prediction functions. It determines the spatial correspondence state through the structural relationship between the target features and auxiliary features, and obtains dynamic displacement information using motion estimation or deformation field prediction to guide the auxiliary features to adjust towards the target space.

[0074] The system also includes a reliability adjustment module. This module assesses and adaptively adjusts the alignment auxiliary features based on target contrast characteristics to obtain reliable auxiliary features. The module can determine the reliability of auxiliary information in different spatial regions based on the correlation between corresponding regions, and enhance the feature response of effective auxiliary regions according to the reliability level, while suppressing the auxiliary feature responses of misaligned, locally deformed, or structurally inconsistent regions.

[0075] The system also includes a fusion reconstruction module. This module fuses target contrast features and reliable auxiliary features. During multi-scale reconstruction, it updates cross-contrast dynamic alignment information based on the spatial offset relationship between the reconstructed features and reliable auxiliary features at the current reconstruction scale. It then performs spatial correction on the reliable auxiliary features based on the updated cross-contrast dynamic alignment information, fuses the spatially corrected reliable auxiliary features with the reconstructed features at the current reconstruction scale, and uses the fused decoding result for subsequent reconstruction scales. This update, spatial correction, fusion, and decoding process is repeated to obtain the target contrast reconstructed image. The fusion reconstruction module can achieve the fusion of target information and reliable auxiliary information through feature stitching, weighted modulation, or attention-guided methods. Furthermore, through multi-scale feature recovery and progressive implicit alignment decoding, the cross-contrast spatial relationship is continuously adjusted as the target image is restored.

[0076] When further magnetic resonance physical constraints are required, the target contrast image output by the fusion reconstruction module can also enter the data consistency processing process, which converts it to k-space, and retains the real measurement data at the sampled positions and the predicted data at the unsampled positions according to the sampling mask corresponding to the undersampled data of the target contrast. Then it is converted to the image space to obtain the final target contrast reconstruction image.

[0077] Each module can be arranged according to Figure 2 The data processing relationships shown are executed sequentially, and can also form interconnected functional structures within a unified deep learning magnetic resonance reconstruction network. The dynamic alignment process, reliable information filtering process, and multi-scale reconstruction process can be executed jointly, so that the auxiliary contrast image is no longer directly input into the target reconstruction as unjudged static prior information, but participates in image restoration after being processed at both the spatial location and regional reliability levels.

[0078] Through the above system structure, the undersampled observation information of the target contrast itself, the complementary tissue structure information provided by the auxiliary contrast, the cross-contrast dynamic spatial correspondence, and the reliability of regional auxiliary information can be unified into the magnetic resonance image reconstruction process. The spatial alignment relationship is continuously updated according to the spatial offset relationship between the reconstruction features and reliable auxiliary features at the current reconstruction scale, so that the decoding results at the current scale can participate in the feature recovery of subsequent reconstruction scales. This can reduce the cross-contrast spatial mismatch caused by patient motion, sequence parameter differences, and local tissue deformation, while suppressing the continuous propagation of unreliable auxiliary information in multiple reconstruction scales. Under the condition of reducing the amount of target contrast data acquisition, the structural consistency, detail recovery ability, stability, and clinical applicability of the reconstructed image are improved.

[0079] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A magnetic resonance image reconstruction method with cross-contrast dynamic alignment, characterized in that, The method includes: Acquire target contrast undersampled k-space data and at least one auxiliary contrast image of the same subject, and construct target contrast features and auxiliary contrast features respectively; Based on the target contrast feature and the auxiliary contrast feature, the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region is determined. Based on the spatial correspondence, the offset relationship between different spatial regions is determined to obtain cross-contrast dynamic alignment information. Based on the cross-contrast dynamic alignment information, the auxiliary contrast feature is spatially corrected to obtain alignment auxiliary features. Based on the target contrast features, the alignment auxiliary features are reliably evaluated and adaptively adjusted to obtain reliable auxiliary features; The target contrast features and the reliable auxiliary features are fused together. During the multi-scale reconstruction process, the cross-contrast dynamic alignment information is updated according to the spatial offset relationship between the reconstruction features at the current reconstruction scale and the reliable auxiliary features. The reliable auxiliary features are spatially corrected according to the updated cross-contrast dynamic alignment information. The spatially corrected reliable auxiliary features are then fused with the reconstruction features at the current reconstruction scale. The fused decoding result is used for subsequent reconstruction scales. The update, spatial correction, fusion, and decoding processes are repeated to obtain the target contrast reconstructed image.

2. The magnetic resonance image reconstruction method with cross-contrast dynamic alignment according to claim 1, characterized in that, The construction of target contrast features and auxiliary contrast features respectively includes: The target contrast undersampled k-space data is subjected to data format conversion, multi-coil correction and normalization to obtain preprocessed target k-space data; Perform an inverse Fourier transform on the preprocessed target k-space data to obtain the initial image of the target contrast, and retain the sampling mask corresponding to the undersampled k-space data of the target contrast. Feature extraction is performed on the initial image of the target contrast to obtain the target contrast features; Feature extraction is performed on the auxiliary contrast image to obtain the auxiliary contrast features.

3. The magnetic resonance image reconstruction method with cross-contrast dynamic alignment according to claim 1, characterized in that, The step of determining the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region, and determining the offset relationship between different spatial regions based on the spatial correspondence, includes: Perform cross-contrast structure correlation analysis on the target contrast feature and the auxiliary contrast feature to obtain the cross-contrast structure correlation results; Based on the cross-contrast structure association results, the spatial correspondence between the target contrast space region and the auxiliary contrast space region is determined, and the cross-contrast space correspondence results are obtained. Based on the cross-contrast space correspondence results, the offset relationship between different spatial regions is determined, and the cross-contrast dynamic alignment information is obtained.

4. The magnetic resonance image reconstruction method with cross-contrast dynamic alignment according to claim 3, characterized in that, The step of determining the offset relationship between different spatial regions based on the cross-contrast space correspondence result to obtain the cross-contrast dynamic alignment information includes: Based on the cross-contrast space correspondence results, motion estimation or deformation field prediction is performed on the target contrast features and the auxiliary contrast features to obtain pixel-level or region-level spatial offset information. A cross-contrast dynamic displacement field is constructed based on the spatial offset information, and the cross-contrast dynamic displacement field is used as the cross-contrast dynamic alignment information. Based on the cross-contrast dynamic displacement field, the auxiliary contrast feature is spatially mapped to obtain the alignment auxiliary feature corresponding to the target contrast feature.

5. The magnetic resonance image reconstruction method with cross-contrast dynamic alignment according to claim 1, characterized in that, The reliability assessment and adaptive adjustment of the alignment auxiliary features based on the target contrast features include: The target contrast feature and the alignment auxiliary feature are divided into multiple corresponding spatial regions to obtain multiple corresponding region feature pairs. The correlation between the target region features and the auxiliary region features in each corresponding region feature pair is determined to obtain region correlation information; Based on the regional correlation information, the reliability of the auxiliary information corresponding to each spatial region is determined to obtain the reliability information of the auxiliary information. The reliable auxiliary features are obtained by enhancing the auxiliary region features that are consistent with the target contrast feature structure based on the reliability information of the auxiliary information, and suppressing the auxiliary region features that are misaligned or structurally inconsistent with the target contrast feature.

6. The magnetic resonance image reconstruction method with cross-contrast dynamic alignment according to claim 1, characterized in that, The step of performing reliability assessment and adaptive adjustment of the alignment auxiliary feature based on the target contrast feature, and fusing the target contrast feature and the reliable auxiliary feature, includes: The reliability of the target contrast feature and the alignment auxiliary feature is evaluated to obtain the reliability evaluation result. Based on the reliability assessment results, the degree of auxiliary feature fusion corresponding to different spatial regions is determined, and the region fusion weight is obtained. The alignment auxiliary features are weighted and modulated according to the region fusion weights to obtain the reliable auxiliary features; The reliable auxiliary features are combined with the target contrast features through feature splicing, weighted modulation, or attention-guided fusion to obtain cross-contrast fusion features.

7. The magnetic resonance image reconstruction method with cross-contrast dynamic alignment according to claim 1, characterized in that, Updating the cross-contrast dynamic alignment information during multi-scale reconstruction includes: Multi-scale feature recovery is performed on the cross-contrast fusion feature formed by fusing the target contrast feature and the reliable auxiliary feature to obtain scale reconstruction features corresponding to multiple reconstruction scales; For each of the reconstruction scales, the cross-contrast dynamic alignment information is updated based on the spatial offset relationship between the scale reconstruction features and the reliable auxiliary features at the corresponding scale to obtain scale alignment information; For each reconstruction scale, the reliable auxiliary features are spatially adjusted according to the corresponding scale alignment information, and the spatially adjusted reliable auxiliary features are fused with the scale reconstruction features corresponding to the reconstruction scale to obtain scale-up reconstruction features; The scale-up reconstruction features are progressively decoded, and the scale alignment information is continuously updated during the decoding process to obtain the target contrast reconstruction image.

8. The magnetic resonance image reconstruction method with cross-contrast dynamic alignment according to claim 7, characterized in that, The progressive decoding of the scale-up reconstructed features includes: Determine the current reconstruction scale, and update the scale alignment information according to the spatial offset relationship between the scale reconstruction features corresponding to the current reconstruction scale and the reliable auxiliary features to obtain the current scale alignment information; Based on the current scale alignment information, the reliable auxiliary features are spatially corrected to obtain the current scale alignment auxiliary features; The current scale alignment auxiliary feature is fused with the scale reconstruction feature corresponding to the current reconstruction scale to obtain the current scale updated reconstruction feature; The current scale-up reconstruction features are decoded, and the decoding results are used for feature recovery at subsequent reconstruction scales. The scale alignment information update, spatial correction, fusion, and decoding processes are repeated until the progressive decoding of each reconstruction scale is completed, resulting in a target contrast prediction image.

9. The magnetic resonance image reconstruction method with cross-contrast dynamic alignment according to claim 1, characterized in that, The process of obtaining the target contrast reconstructed image includes: The contrast image of the target to be corrected is obtained through the multi-scale reconstruction process. Perform a Fourier transform on the contrast image of the target to be corrected to obtain the predicted k-space data; Based on the sampling mask corresponding to the target contrast undersampled k-space data, the data at the sampled positions in the predicted k-space data are replaced with the actual measurement data in the target contrast undersampled k-space data, while the predicted data at the unsampled positions are retained, to obtain the corrected k-space data. The target contrast reconstruction image is obtained by performing an inverse Fourier transform on the corrected k-space data.

10. A magnetic resonance image reconstruction system with dynamic alignment across contrast, characterized in that, The system includes: The feature construction module is used to acquire the target contrast undersampled k-space data and at least one auxiliary contrast image of the same inspected object, and to construct the target contrast features and auxiliary contrast features respectively. The dynamic alignment module is used to determine the spatial correspondence between the target contrast spatial region and the auxiliary contrast spatial region based on the target contrast feature and the auxiliary contrast feature, determine the offset relationship between different spatial regions based on the spatial correspondence, obtain cross-contrast dynamic alignment information, and perform spatial correction on the auxiliary contrast feature based on the cross-contrast dynamic alignment information to obtain alignment auxiliary features. A reliability adjustment module is used to perform reliability assessment and adaptive adjustment of the alignment auxiliary feature based on the target contrast feature to obtain a reliable auxiliary feature. The fusion reconstruction module is used to fuse the target contrast features and the reliable auxiliary features. During the multi-scale reconstruction process, it updates the cross-contrast dynamic alignment information according to the spatial offset relationship between the reconstruction features at the current reconstruction scale and the reliable auxiliary features. It then performs spatial correction on the reliable auxiliary features based on the updated cross-contrast dynamic alignment information and fuses the spatially corrected reliable auxiliary features with the reconstruction features at the current reconstruction scale. The fused decoding result is used for subsequent reconstruction scales. The update, spatial correction, fusion, and decoding processes are repeated to obtain the target contrast reconstructed image.

Images