Reliability-guided diffusion magnetic resonance imaging multi-scale registration method and system

Abstract

The application discloses a diffusion magnetic resonance imaging multi-scale registration method and system based on reliability guidance, and the method comprises the following steps: acquiring FA images and DTI tensor fields of fixed and mobile diffusion images; based on quantile statistics, the DTI tensor amplitude is adaptively adjusted and fused with the FA images to form a multi-channel diffusion representation; according to structural and directional feature differences, multi-scale structural mismatch maps and directional reliability maps are generated; the above reliability maps are used as guidance information, and features are adaptively modulated in multi-scale feature propagation to estimate an initial deformation field; the initial deformation field is corrected by coarse and fine scale residual displacement estimation to obtain a final deformation field, and spatial transformation and tensor reorientation are completed accordingly. The application can reduce the scale difference between FA and DTI, suppress the interference in a low reliability area, improve the anatomical structure registration accuracy, and maintain the diffusion direction consistency.

Description

Technical Field

[0001] This invention belongs to the field of magnetic resonance medical image processing, and relates to a reliability-guided multi-scale registration method and system for diffusion magnetic resonance imaging. Background Technology

[0002] Diffusion magnetic resonance imaging (DMRI) can reflect the microstructure and fiber orientation of brain white matter by detecting the diffusion motion of water molecules. Diffusion tensor imaging (DTI) uses a second-order tensor to describe local diffusion characteristics and can calculate indices such as fractional anisotropy, making it widely used in white matter structure analysis and imaging research of neurological diseases. DTI image registration aims to establish spatial correspondences between different subjects, different time points, or different imaging data, and is an important foundation for quantitative comparison of white matter microstructure and population image analysis. Unlike conventional structural image registration, DTI image registration requires not only accurate alignment of macroscopic anatomical structures but also the physical consistency of the diffusion tensor direction and shape.

[0003] Traditional registration methods often employ iterative optimization and tensor retargeting after registration, resulting in significant computational overhead and difficulty in directly incorporating directional information into deformation estimation. While existing deep learning methods have improved registration efficiency, they largely rely on scalar strength or implicit features, failing to adequately model local structural differences and directional reliability. Furthermore, the fractional anisotropy map and the diffusion tensor imaging tensor field exhibit significant scale differences, making direct joint modeling prone to modal bias, impacting network training stability and cross-data generalization ability. Summary of the Invention

[0004] The purpose of this invention is to provide a reliability-guided multi-scale registration method and system for diffusion magnetic resonance imaging, in order to solve the problems that existing methods cannot simultaneously ensure the alignment of macroscopic anatomical structures and the preservation of microscopic diffusion directions, as well as the interference of local low-reliability information with deformation estimation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a reliability-guided multi-scale registration method for diffusion magnetic resonance imaging, comprising the following steps: Acquire the fixed diffusion image and the moving diffusion image to be registered. Both the fixed diffusion image and the moving diffusion image include the fractional anisotropy FA map and the diffusion tensor imaging DTI tensor field. Based on the quantile statistics of the FA plot and the DTI tensor field, the relative scale relationship between the two is determined, and the amplitude of the DTI tensor field is adaptively adjusted according to the relative scale relationship. The adjusted DTI tensor field is then fused with the FA plot to form a multi-channel diffusion representation. Based on the structural feature differences between multi-channel diffusion characterizations, a multi-scale structural mismatch map is generated; based on the directional feature differences between DTI tensor fields, a multi-scale directional reliability map is generated. Using multi-scale structural mismatch maps and multi-scale directional reliability maps as guiding information, a multi-scale feature propagation process is introduced to adaptively modulate the features, and the initial deformation field is estimated based on the modulated features. The moving diffusion image is spatially transformed based on the initial deformation field to obtain a coarsely registered moving diffusion image. The residual displacement between the fixed diffusion image and the coarsely registered moving diffusion image is estimated at both coarse and fine scales. The initial deformation field is then corrected based on the residual displacement to obtain the final deformation field. Based on the final deformation field, spatial transformation and tensor redirection are performed on the FA map and DTI tensor field of the moving diffusion image to obtain the registered FA map and DTI tensor field.

[0006] Furthermore, the process of determining the relative scale relationship between the FA graph and the DTI tensor field based on the quantile statistics of the FA graph and the DTI tensor field, adaptively adjusting the amplitude of the DTI tensor field according to the relative scale relationship, and then fusing the adjusted DTI tensor field with the FA graph to form a multi-channel diffusion representation includes: Based on the FA amplitude and DTI tensor norm, the effective brain tissue region is determined; the amplitude quantiles of the FA map and the norm quantiles of the DTI tensor field are calculated within the effective brain tissue region respectively. Based on the magnitude quantiles and norm quantiles, the logarithmic scaling ratio between the FA plot and the DTI tensor field is determined as the relative scaling relationship; The logarithmic scaling ratio is input into a learnable bounded mapping function to generate adaptive scaling coefficients; these adaptive scaling coefficients are then used to uniformly apply to each component of the DTI tensor field to adjust its magnitude. The amplitude-adjusted DTI tensor field is spliced ​​with the FA plot to form a multi-channel diffusion characterization.

[0007] Furthermore, the generation of the multi-scale structural mismatch map includes: The multi-channel diffusion representations of the fixed diffusion image and the moving diffusion image are mapped to the feature space respectively, and the structural feature differences are calculated. The structural feature differences are sequentially convolved and activated by Sigmoid to generate a fine-scale structural mismatch map; the fine-scale structural mismatch map is then downsampled step by step to construct a multi-scale structural mismatch map. Generate multi-scale directional reliability maps, including: The fixed diffusion image and the moving diffusion image are described as DTI tensors, which are mapped to the feature space respectively, and the directional feature differences are calculated. The directional feature differences are then convolved and activated sequentially to generate fine-scale directional reliability maps. The fine-scale directional reliability maps are then downsampled step by step to construct multi-scale directional reliability maps.

[0008] Furthermore, the step of using multi-scale structural mismatch maps and multi-scale orientation reliability maps as guiding information to introduce a multi-scale feature propagation process and adaptively modulate the features includes: A multi-scale feature extraction network is used to encode fixed diffusion images and moving diffusion images to obtain features at different scales. At each scale of the decoding process, the skip connection features of the corresponding scale and the current decoded features are input into the reliability consistency gating unit. The reliability consistency gating unit generates a spatial attention response based on the skip connection characteristics and the current decoding characteristics, and generates an internal feature consistency response based on the difference between the two. The reliability consistency gating unit integrates the structural mismatch map and directional reliability map at the corresponding scale, and combines them with the spatial attention response and internal feature consistency response to generate feature gain coefficients. By utilizing the characteristic gain coefficient, the skip connection characteristics are modulated to enhance the characteristics of reliable regions and suppress interference in unreliable regions.

[0009] Furthermore, estimating the residual displacement between the fixed diffusion image and the coarse-registered moving diffusion image at both coarse and fine scales, and correcting the initial deformation field based on the residual displacement, includes: Extract coarse-scale and fine-scale features from the fixed diffusion image and the coarse-registered moving diffusion image; estimate the coarse-scale residual displacement based on the coarse-scale features; upsample the coarse-scale residual displacement and fuse it with the fine-scale features to estimate the fine-scale residual displacement. By merging coarse-scale and fine-scale residual displacements, a local residual displacement is formed; by combining the local residual displacement with the initial deformation field, the final deformation field is obtained.

[0010] Furthermore, it also includes a step of training the registration network constructed by the above steps, the training step including: Input fixed and moving diffusion image samples into the registration network to obtain the predicted deformation field, and the registered moving FA map and moving DTI tensor field based on the predicted deformation field; calculate the FA structural similarity loss based on the FA map of the fixed diffusion image samples and the registered moving FA map. Calculate the DTI tensor consistency loss based on the DTI tensor field of the fixed diffuse image samples and the registered and moved DTI tensor field; calculate the deformation field smoothing regularization term based on the predicted deformation field. A joint loss function is constructed by integrating FA structural similarity loss, DTI tensor consistency loss, and deformation field smoothing regularization term; the parameters of the registration network are optimized by minimizing the joint loss function.

[0011] Furthermore, the FA structural similarity loss is calculated based on multi-scale local normalized cross-correlation and is used to constrain the structural consistency between the fixed FA graph and the registered and moved FA graph. The DTI tensor consistency loss is calculated based on the tensor distance of the matrix logarithm and is calculated in the white matter region with reliable diffusion direction information. It is used to constrain the tensor orientation and shape consistency between the fixed DTI tensor field and the registered and moved DTI tensor field. The deformation field smoothing regularization term, calculated based on the bending energy of the predicted deformation field, is used to constrain the spatial continuity of the deformation field.

[0012] Secondly, the present invention provides a reliability-guided diffusion magnetic resonance imaging multi-scale registration system, the system comprising: The data receiving module is used to acquire the fixed diffusion image and the moving diffusion image to be registered. Both the fixed diffusion image and the moving diffusion image include the fractional anisotropy FA map and the diffusion tensor imaging DTI tensor field. The quantile feature normalization module is used to determine the relative scale relationship between the FA plot and the DTI tensor field based on the quantile statistics of the FA plot and the DTI tensor field, and adaptively adjust the amplitude of the DTI tensor field according to the relative scale relationship. The adjusted DTI tensor field is then fused with the FA plot to form a multi-channel diffusion representation. The reliability evidence extraction module is used to generate multi-scale structural mismatch maps based on the structural feature differences between multi-channel diffusion representations; and to generate multi-scale directional reliability maps based on the directional feature differences between DTI tensor fields. The deformation estimation module is used to introduce the multi-scale feature propagation process into the multi-scale feature propagation process by using the multi-scale structural mismatch map and multi-scale orientation reliability map as guiding information, adaptively modulate the features, and estimate the initial deformation field based on the modulated features. The residual refinement module is used to perform spatial transformation on the moving diffusion image based on the initial deformation field to obtain a coarse-registered moving diffusion image. It also estimates the residual displacement between the fixed diffusion image and the coarse-registered moving diffusion image at both coarse and fine scales, and corrects the initial deformation field based on the residual displacement to obtain the final deformation field. The registration output module is used to perform spatial transformation and tensor redirection on the FA map and DTI tensor field of the motion diffusion image based on the final deformation field, so as to obtain the registered FA map and DTI tensor field.

[0013] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein when the processor executes the computer program, it implements the steps of the reliability-guided diffusion magnetic resonance imaging multiscale registration method.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of the reliability-guided diffusion magnetic resonance imaging multi-scale registration method.

[0015] Compared with the prior art, the present invention has the following technical effects: This invention employs deep learning technology to achieve end-to-end registration of diffuse magnetic resonance images. Compared with traditional iterative optimization methods, it can improve registration efficiency while taking into account the alignment of brain white matter anatomical structures and the preservation of diffusion direction, thereby enhancing the accuracy and reliability of registration results.

[0016] This invention fully considers the scale differences between the FA map and the DTI tensor field, as well as the reliability of local diffusion information. It reduces numerical deviations between different diffusion representations through quantile feature normalization and jointly constructs a structural mismatch pyramid and an orientation reliability pyramid to explicitly characterize the degree of structural misalignment and orientation reliability at different scales. Simultaneously, it enhances reliable region features and suppresses interference from low-reliability regions through reliability consistency gating and a coarse-to-fine residual refinement mechanism, improving the local registration accuracy of complex white matter regions. Furthermore, this invention combines FA structural similarity, DTI tensor consistency, and deformation field smoothness for joint optimization, and maintains the physical rationality of diffusion directions through tensor redirection. This provides reliable technical support for white matter fiber tract analysis, brain atlas construction, and imaging research of neurological diseases, and has significant application and promotion value. Attached Figure Description

[0017] Figure 1 The overall logic block diagram of this invention.

[0018] Figure 2 A schematic diagram of the overall network structure of this invention.

[0019] Figure 3 A schematic diagram of the quantile feature normalization module of this invention.

[0020] Figure 4 The present invention provides a schematic diagram of the structural mismatch pyramid and directional reliability pyramid structures.

[0021] Figure 5 A schematic diagram of the reliability consistency gating unit module of this invention.

[0022] Figure 6 A schematic diagram of the reliability constraint residual refinement module of this invention.

[0023] Figure 7 A schematic diagram comparing the registration results of this invention with other diffusion magnetic resonance image registration methods.

[0024] Figure 8Flowchart of this invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1, see Figure 1 and Figure 8 In one embodiment of the present invention, a reliability-guided multi-scale registration method for diffusion magnetic resonance imaging is provided for structural and orientation co-registration of diffusion magnetic resonance images from different subjects, at different time points, or from different imaging data.

[0028] Specifically, the reliability-guided multi-scale registration method for diffusion magnetic resonance imaging of the present invention includes the following steps: S1: Obtain the fixed diffusion image and the moving diffusion image to be registered. Both the fixed diffusion image and the moving diffusion image include a fractional anisotropy (FA) map and a diffusion tensor imaging (DTI) tensor field.

[0029] A fixed diffusion image serves as the reference target for registration, while a moving diffusion image represents the object to be deformed. Each diffusion image contains two types of information: a fractional anisotropy (FA) map, used to characterize the macroscopic anatomical structure of brain white matter; and a diffusion tensor imaging (DTI) tensor field, used to describe the diffusion direction and shape of local water molecules. These two types of information together constitute the raw input for all subsequent processing steps and are the data foundation for completing the structural and orientation co-registration.

[0030] S2: Based on the quantile statistics of the FA diagram and the DTI tensor field, determine the relative scale relationship between them, and adaptively adjust the amplitude of the DTI tensor field according to the relative scale relationship. Then, merge the adjusted DTI tensor field with the FA diagram to form a multi-channel diffusion representation.

[0031] Since FA values ​​are typically between 0 and 1, while the range of DTI tensor components is much wider, direct concatenation would cause network training to be biased towards one side, resulting in modality bias. This step uses statistical methods to calculate the feature scales of both, establishing a quantified relationship of the logarithmic scale ratio. Based on this, the DTI tensor is adaptively scaled to ensure its amplitude is within a comparable numerical range to the FA graph. The resulting multi-channel diffusion representation retains complete tensor information while avoiding scale mismatch issues.

[0032] S3: Generate a multi-scale structural mismatch map based on the structural feature differences between the multi-channel diffusion characterizations; generate a multi-scale directional reliability map based on the directional feature differences between the DTI tensor fields.

[0033] The structural mismatch map is generated by comparing the differences between two images in the multi-channel diffusion representation feature space. A larger response value indicates a higher degree of structural misalignment in the local region. The directional reliability map is generated by comparing the differences in DTI tensor field features. A larger response value indicates more consistent and reliable diffusion direction information in the region. By constructing a pyramid structure through stepwise downsampling, reliability evidence at multiple spatial scales can be obtained, providing guidance signals of corresponding granularity for network layers of different depths.

[0034] S4: Using the multi-scale structural mismatch map and the multi-scale directional reliability map as guiding information, a multi-scale feature propagation process is introduced to adaptively modulate the features, and the initial deformation field is estimated based on the modulated features.

[0035] In the decoding process of a multi-scale feature extraction network, the detailed features (skip connection features) extracted during the encoding stage are modulated by a reliability consistency gating unit before being passed to the decoding layer. This unit uses the structural mismatch map and orientation reliability map of the corresponding scale to generate feature gain coefficients, which weight the feature map spatially and in channels. This modulation mechanism can amplify the response of effective features in structurally consistent and orientationally reliable regions, while suppressing potentially misleading interfering features in low-reliability regions, thereby guiding the network to focus on reliable information for deformation field estimation.

[0036] S5: Perform spatial transformation on the moving diffusion image based on the initial deformation field to obtain a coarse-registered moving diffusion image, and estimate the residual displacement between the fixed diffusion image and the coarse-registered moving diffusion image at both coarse and fine scales. Correct the initial deformation field based on the residual displacement to obtain the final deformation field.

[0037] This step is used to refine the initial deformation field generated in step S4. While the initial deformation field has achieved overall, macroscopic registration, local mismatches may still exist in complex white matter regions. This step first uses the initial deformation field to perform a spatial transformation to obtain a coarse registration result. Then, features are extracted and residual displacements are predicted at both coarse and fine scales. The coarse-scale branch is responsible for capturing large-scale residual deformation trends, while the fine-scale branch focuses on precise local boundary alignment. By fusing the residual displacements from both scales, the initial deformation field is fine-tuned, ultimately obtaining an accurate deformation field that considers both the overall and local conditions.

[0038] S6: Based on the final deformation field, perform spatial transformation and tensor redirection on the FA map and DTI tensor field of the moving diffusion image to obtain the registered FA map and DTI tensor field.

[0039] This step applies the final deformation field obtained in step S5 to the original moving-diffusion image, completing the final physical operation of registration. Spatial transformation refers to resampling the FA image and DTI tensor field based on the deformation field, mapping the moving image space to the fixed image space. Tensor redirection refers to adjusting the direction of the DTI tensor according to the local rotation components of the deformation field, ensuring that the direction of the registered tensor maintains physical consistency with the actual direction of white matter fibers. After this step, the final registered FA image and registered DTI tensor field are anatomically aligned with the fixed image and physically consistent in the diffusion direction.

[0040] This invention presents a reliability-guided multi-scale registration method for diffusion magnetic resonance imaging. Step S1 acquires raw data containing structural and orientational information, providing complete input for registration. Step S2 adaptively normalizes the DTI tensor amplitude based on quantile statistics and fuses it with the FA graph, effectively eliminating scale differences between modes and providing a numerically stable foundation for subsequent network processing. Step S3 explicitly generates multi-scale structural mismatch maps and orientation reliability maps, transforming abstract registration states and orientation reliability metrics into learnable and transferable intermediate representations, enabling the network to "see" where misalignment occurs and where reliability is achieved. Step S4 uses the aforementioned reliability evidence as prior guidance in the feature propagation process, employing gating mechanisms to selectively enhance and suppress features, ensuring that deformation estimation primarily relies on reliable information. Step S5, through coarse and fine-scale dual-stage residual estimation and correction, achieves hierarchical deformation field optimization from global to local levels. Step S6, through spatial transformation and tensor redirection, ultimately achieves the collaborative registration goal of macroscopic anatomical structure alignment and microscopic diffusion orientation preservation. Each step progresses sequentially and works in synergy to form an end-to-end, reliability-driven, high-precision diffusion magnetic resonance image registration scheme.

[0041] Example 2, please refer to Figures 1 to 8 A reliability-guided multi-scale registration method for diffusion magnetic resonance imaging includes: Obtain fixed-diffusion and moving-diffusion images containing FA plots and DTI tensor fields; A quantile feature normalization module is constructed to adaptively adjust the DTI tensor amplitude and integrate it with the FA graph to form a multi-channel diffusion representation. A multi-scale reliability evidence extraction module is constructed to generate a structure mismatch pyramid and an orientation reliability pyramid based on the structural and orientational differences between two images. A reliability-guided deformation estimation module is constructed, which uses multi-scale reliability information to modulate feature propagation and estimate the initial deformation field. A reliability constraint residual refinement module is constructed to estimate the local residual displacement from coarse to fine based on the coarse registration result, correct the initial deformation field and obtain the final deformation field. The above modules are connected in series to form a reliability-guided multi-scale registration network. The network parameters are trained by a joint loss function. The trained network is then used to perform spatial transformation and tensor redirection on the moving FA graph and the DTI tensor field to obtain the final registration result.

[0042] Specifically: An example of the diffusion magnetic resonance image registration method provided by this invention is as follows: Figure 1 As shown, the overall network structure is as follows: Figure 2 As shown, the specific steps include:

[0043] Data preprocessing: The data used to train the model was obtained from the diffusion magnetic resonance imaging (DMRI) dataset publicly released by the Human Connectome Project (HCP). This dataset mainly contains high spatial resolution DMRI images of healthy adult subjects, with a standardized acquisition process and high data quality, making it suitable as training data for the model. First, the diffusion images were spatially aligned and center-cropped to ensure that the data from all subjects had consistent spatial resolution, image size, and coordinate orientation. Then, diffusion tensor fitting was performed to obtain six independent DTI tensor components, and the FA (figurine image) map was further calculated. Finally, one FA channel was combined with the six DTI tensor components to form a seven-channel diffusion representation, which served as the input to the registration network.

[0044] For clinical-quality diffusion MRI data, the Alzheimer's Disease Neuroimaging Project (ADNI) and the Parkinson's Disease Progression Biomarkers Project (PPMI) datasets were primarily used. Both ADNI and PPMI are multi-center clinical imaging databases, and their data differ significantly in terms of scanning equipment, acquisition protocols, spatial resolution, and signal-to-noise ratio. First, the diffusion images were denoised and corrected for Gibbs ringing artifacts, and then resampled to 1.25 mm isotropic voxel resolution. Next, nearest-neighbor interpolation was used to map the brain mask in T1 space to the diffusion space, and image size and data format were standardized. Finally, diffusion tensor fitting and FA map calculation were performed, combining one FA channel with six DTI tensor components to form a seven-channel diffusion representation, which served as the input to the trained registration network.

[0045] Construct a quantile feature normalization module: Quantile feature normalization module, such as Figure 3 As shown, this module is placed at the input of the registration network to reduce the numerical scale difference between the FA image and the DTI tensor field. The module takes the FA image and DTI tensor field corresponding to the fixed diffusion image and the moving diffusion image as input, where the DTI tensor at voxel (x) consists of six independent components. First, the effective brain tissue region is determined based on the FA amplitude and the Frobenius norm of the DTI tensor, and the preset quantiles of the FA amplitude and the DTI tensor norm are calculated within this region to obtain the feature scale. and In this embodiment, the 90th percentile is used to reduce the impact of outliers on scale estimation. The relative scale between the FA plot and the DTI tensor field is expressed as follows:

[0046] Where r represents the logarithmic scale ratio, This represents a small positive number used to ensure numerical stability. Subsequently, adaptive scaling coefficients are generated using a learnable bounded mapping function, and the DTI tensor is normalized.

[0047]

[0048] in, and For learnable parameters, and These represent the lower and upper limits of the scaling factor, respectively. Since the same scaling factor applies uniformly to all six components of the DTI tensor, only the tensor amplitude is adjusted, without changing its diffusion direction. Finally, the normalized six-channel DTI tensor field is stitched together with the corresponding single-channel FA plot to form a seven-channel diffusion representation, which serves as the input for the subsequent multi-scale reliability evidence extraction module and reliability-guided deformation estimation module.

[0049] Quantile feature normalization module, such as Figure 3 As shown, this module, located at the input of the registration network, reduces the numerical scale difference between the FA map and the DTI tensor field. The module takes the FA map and DTI tensor field corresponding to the fixed and moving diffusion images as input. It determines the effective brain tissue region based on the FA amplitude and DTI tensor norm, and calculates preset quantile statistics for the FA amplitude and DTI tensor norm within the effective region to obtain the characteristic scale of the two types of diffusion representations. Subsequently, it calculates the relative scale relationship between the FA map and the DTI tensor field based on the quantile statistics, and generates adaptive scaling coefficients through a learnable bounded mapping function. These scaling coefficients uniformly act on the six independent components of the DTI tensor, adjusting the tensor amplitude without changing its directional information. Finally, the normalized six-channel DTI tensor field is concatenated with the corresponding single-channel FA map to form a seven-channel diffusion representation. This module reduces the impact of outliers and cross-data distribution differences on the characteristic scale, and its output serves as the input for subsequent reliability evidence extraction and deformation estimation modules.

[0050] Construct a multi-scale reliability evidence extraction module: Multi-scale reliability evidence extraction module, such as Figure 4As shown, this module consists of two parallel branches: a structure mismatch pyramid and a direction reliability pyramid. These branches characterize the local structural differences and the reliability of diffusion direction information between the fixed and moving diffusion images, respectively. Both branches employ a structurally consistent lightweight feature extraction method and generate complementary reliability evidence at multiple spatial scales. The structure mismatch pyramid takes the seven-channel diffusion representations of the fixed and moving diffusion images as input, maps them to a common feature space through convolutional layers, calculates the feature differences at corresponding locations, and then performs convolution, LeakyReLU activation, and Sigmoid activation to generate a fine-scale structure mismatch map. The direction reliability pyramid takes the six-channel DTI tensor fields of the fixed and moving diffusion images as input, calculates the directional feature differences in the same way, and generates a fine-scale direction reliability map. Subsequently, the fine-scale structure mismatch map and direction reliability map are subjected to progressive average pooling to obtain structure mismatch maps and direction reliability maps at multiple scales. A larger response value in the structure mismatch map indicates a higher degree of local structural misalignment in the corresponding region; a larger response value in the direction reliability map indicates a higher degree of consistency and reliability of the local diffusion direction. The multi-scale reliability evidence output by this module will serve as external prior information for the reliability consistency gating unit.

[0051] Construct a reliability-guided deformation estimation module: The reliability consistency gating unit in the reliability-guided deformation estimation module, such as Figure 5 As shown, this module encodes the seven-channel diffusion representations of fixed and moving diffusion images through a multi-scale feature extraction and decoding structure, obtaining structural and directional features at different scales and recovering spatial information step by step. At each decoding scale, the skip connection features from the encoding stage and the current decoding features are input into a reliability consistency gating unit. The gating unit maps and fuses the two types of features to generate a spatial attention response; simultaneously, it generates an internal consistency response based on the feature differences between the two, used to measure the degree of matching between the encoded features and the decoding context. Subsequently, the structural mismatch map and directional reliability map at the corresponding scale are used as external reliability priors, fused with the above responses to generate bounded feature gain coefficients, and adaptively modulated the skip connection features to enhance the effective features of structurally consistent and directionally reliable regions and suppress interference in low-reliability regions. After multi-scale modulation and step-by-step decoding, the features at each scale are fused and the initial deformation field is estimated, providing an overall registration basis for subsequent residual refinement.

[0052] Construct a reliability constraint residual refinement module: Reliability constraint residual refinement module, such as Figure 6As shown, this module uses the initial deformation field as a priori for overall registration, predicting only local residual displacements to correct local structural mismatches that still exist after coarse registration. First, the moving-diffusion image is spatially transformed based on the initial deformation field to obtain a coarse-registered moving-diffusion image. Then, the fixed-diffusion image and the coarse-registered moving-diffusion image are input into a feature extraction branch with consistent structure to obtain coarse-scale and fine-scale features. In the coarse-scale stage, deep features are fused to predict coarse-scale residual displacements and upsampled. In the fine-scale stage, the upsampled results are fused with shallow features to further predict fine-scale residual displacements. Finally, the coarse and fine-scale residual displacements are fused to form local residual displacements, which are then combined with the initial deformation field to obtain the final deformation field. This module can improve the local registration accuracy in complex white matter regions while maintaining the overall registration relationship.

[0053] Model training: The reliability-guided multi-scale registration network consists of quantile feature normalization, multi-scale reliability modeling, reliability gating, and residual refinement modules. During training, the FA maps and DTI tensor fields of the fixed and moving images are used as inputs. The network predicts the deformation field from the moving image space to the fixed image space, and the joint loss function consisting of FA structural similarity loss, DTI tensor consistency loss, and deformation field smoothing regularization term is used as the training signal. By minimizing the structural and orientation differences between the registered moving and fixed images, the network parameters are optimized, thereby improving the structural alignment accuracy and orientation preservation capability of diffuse magnetic resonance images.

[0054] The model settings in this embodiment of the invention are as follows: 1. Model Implementation: A reliability-guided multi-scale registration network is built using the PyTorch framework and trained end-to-end on a GPU. Automatic mixed precision is used during training to improve computational efficiency and reduce GPU memory usage.

[0055] 2. Loss Function: A joint loss function is constructed using FA structural similarity loss, DTI tensor consistency loss, and deformation field smoothing regularization term.

[0056] in, For FA structural similarity loss based on multi-scale local normalized cross-correlation, The DTI tensor consistency loss is based on Log-Euclidean distance. This is a deformation field smoothing regularization term based on bending energy; , and In this embodiment, the corresponding weights are set to 10.0, 2.0, and 1.0, respectively.

[0057] 3. Training Details: The Adam optimizer is used to update model parameters, with the initial learning rate and weight decay set to [value missing]. The learning rate is adjusted using a cosine annealing strategy. The model is trained for a maximum of 1000 epochs, and an early stopping strategy is adopted based on the validation loss. Gradient norm pruning is also used to improve training stability.

[0058] Another embodiment of the present invention provides a reliability-guided multi-scale registration system for diffusion magnetic resonance imaging (DMRI) to implement the above-described diffusion magnetic resonance image registration method. Specifically, the system includes: The data receiving module is used to acquire the fixed diffusion image and the moving diffusion image to be registered. Both the fixed diffusion image and the moving diffusion image include the fractional anisotropy FA map and the diffusion tensor imaging DTI tensor field. The quantile feature normalization module is used to determine the relative scale relationship between the FA plot and the DTI tensor field based on the quantile statistics of the FA plot and the DTI tensor field, and adaptively adjust the amplitude of the DTI tensor field according to the relative scale relationship. The adjusted DTI tensor field is then fused with the FA plot to form a multi-channel diffusion representation. The reliability evidence extraction module is used to generate multi-scale structural mismatch maps based on the structural feature differences between multi-channel diffusion representations; and to generate multi-scale directional reliability maps based on the directional feature differences between DTI tensor fields. The deformation estimation module is used to introduce the multi-scale feature propagation process into the multi-scale feature propagation process by using the multi-scale structural mismatch map and multi-scale orientation reliability map as guiding information, adaptively modulate the features, and estimate the initial deformation field based on the modulated features. The residual refinement module is used to perform spatial transformation on the moving diffusion image based on the initial deformation field to obtain a coarse-registered moving diffusion image. It also estimates the residual displacement between the fixed diffusion image and the coarse-registered moving diffusion image at both coarse and fine scales, and corrects the initial deformation field based on the residual displacement to obtain the final deformation field. The registration output module is used to perform spatial transformation and tensor redirection on the FA map and DTI tensor field of the motion diffusion image based on the final deformation field, so as to obtain the registered FA map and DTI tensor field.

[0059] To verify the effectiveness of this invention, comparative experiments were conducted on the HCP, ADNI, and PPMI public datasets using the method of this invention and diffusion magnetic resonance image registration methods such as FSL, SyN, VoxelMorph, and DDMReg. Qualitative results are as follows: Figure 7 As shown, the present invention exhibits better structural alignment and orientation retention effects in white matter structural boundaries, deep white matter, and complex local regions.

[0060] In the quantitative experiments, the Pearson correlation coefficient and Dice coefficient of the FA plot were used to evaluate scalar structure consistency, the main diffusion direction angle error and cosine similarity were used to evaluate directional consistency, and the Dice coefficient of white matter fiber bundles was used to evaluate the degree of structural overlap. Experimental results show that this invention achieves superior performance on the main evaluation metrics of the three datasets, can improve the registration accuracy of anatomical structures while maintaining diffusion direction information, and has good cross-dataset adaptability.

[0061] Table 1. Quantitative comparison of dMRI registration methods on the HCP dataset.

[0062] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a reliability-guided diffusion magnetic resonance imaging multi-scale registration method.

[0063] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the reliability-guided diffusion magnetic resonance imaging multi-scale registration method in the above embodiments.

[0064] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0066] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A reliability-guided multi-scale registration method for diffusion magnetic resonance imaging, characterized in that, Includes the following steps: Acquire the fixed diffusion image and the moving diffusion image to be registered. Both the fixed diffusion image and the moving diffusion image include the fractional anisotropy FA map and the diffusion tensor imaging DTI tensor field. Based on the quantile statistics of the FA plot and the DTI tensor field, the relative scale relationship between the two is determined, and the amplitude of the DTI tensor field is adaptively adjusted according to the relative scale relationship. The adjusted DTI tensor field is then fused with the FA plot to form a multi-channel diffusion representation. Based on the structural feature differences between multi-channel diffusion characterizations, a multi-scale structural mismatch map is generated; based on the directional feature differences between DTI tensor fields, a multi-scale directional reliability map is generated. Using multi-scale structural mismatch maps and multi-scale directional reliability maps as guiding information, a multi-scale feature propagation process is introduced to adaptively modulate the features, and the initial deformation field is estimated based on the modulated features. The moving diffusion image is spatially transformed based on the initial deformation field to obtain a coarsely registered moving diffusion image. The residual displacement between the fixed diffusion image and the coarsely registered moving diffusion image is estimated at both coarse and fine scales. The initial deformation field is then corrected based on the residual displacement to obtain the final deformation field. Based on the final deformation field, spatial transformation and tensor redirection are performed on the FA map and DTI tensor field of the moving diffusion image to obtain the registered FA map and DTI tensor field.

2. The reliability-guided multi-scale registration method for diffusion magnetic resonance imaging according to claim 1, characterized in that, The process involves determining the relative scale relationship between the FA plot and the DTI tensor field based on the quantile statistics, adaptively adjusting the amplitude of the DTI tensor field according to the relative scale relationship, and then fusing the adjusted DTI tensor field with the FA plot to form a multi-channel diffusion representation, including: Based on the FA amplitude and DTI tensor norm, the effective brain tissue region is determined; the amplitude quantiles of the FA map and the norm quantiles of the DTI tensor field are calculated within the effective brain tissue region respectively. Based on the magnitude quantiles and norm quantiles, the logarithmic scaling ratio between the FA plot and the DTI tensor field is determined as the relative scaling relationship; The logarithmic scaling ratio is input into a learnable bounded mapping function to generate adaptive scaling coefficients; these adaptive scaling coefficients are then used to uniformly apply to each component of the DTI tensor field to adjust its magnitude. The amplitude-adjusted DTI tensor field is spliced ​​with the FA plot to form a multi-channel diffusion characterization.

3. The reliability-guided multi-scale registration method for diffusion magnetic resonance imaging according to claim 1, characterized in that, The generation of the multi-scale structural mismatch map includes: The multi-channel diffusion representations of the fixed diffusion image and the moving diffusion image are mapped to the feature space respectively, and the structural feature differences are calculated. The structural feature differences are sequentially convolved and activated by Sigmoid to generate a fine-scale structural mismatch map; the fine-scale structural mismatch map is then downsampled step by step to construct a multi-scale structural mismatch map. Generate multi-scale directional reliability maps, including: The fixed diffusion image and the moving diffusion image are described as DTI tensors, which are mapped to the feature space respectively, and the directional feature differences are calculated. The directional feature differences are then convolved and activated sequentially to generate fine-scale directional reliability maps. The fine-scale directional reliability maps are then downsampled step by step to construct multi-scale directional reliability maps.

4. The reliability-guided multi-scale registration method for diffusion magnetic resonance imaging according to claim 1, characterized in that, The step of using multi-scale structural mismatch maps and multi-scale orientation reliability maps as guiding information to introduce a multi-scale feature propagation process and adaptively modulate the features includes: A multi-scale feature extraction network is used to encode fixed diffusion images and moving diffusion images to obtain features at different scales. At each scale of the decoding process, the skip connection features of the corresponding scale and the current decoded features are input into the reliability consistency gating unit. The reliability consistency gating unit generates a spatial attention response based on the skip connection characteristics and the current decoding characteristics, and generates an internal feature consistency response based on the difference between the two. The reliability consistency gating unit integrates the structural mismatch map and directional reliability map at the corresponding scale, and combines them with the spatial attention response and internal feature consistency response to generate feature gain coefficients. By utilizing the characteristic gain coefficient, the skip connection characteristics are modulated to enhance the characteristics of reliable regions and suppress interference in unreliable regions.

5. The reliability-guided multi-scale registration method for diffusion magnetic resonance imaging according to claim 1, characterized in that, The estimation of residual displacement between the fixed diffusion image and the coarsely registered moving diffusion image at both coarse and fine scales, and the correction of the initial deformation field based on the residual displacement, includes: Extract coarse-scale and fine-scale features from the fixed diffusion image and the coarse-registered moving diffusion image; estimate the coarse-scale residual displacement based on the coarse-scale features; upsample the coarse-scale residual displacement and fuse it with the fine-scale features to estimate the fine-scale residual displacement. By merging coarse-scale and fine-scale residual displacements, a local residual displacement is formed; by combining the local residual displacement with the initial deformation field, the final deformation field is obtained.

6. The reliability-guided multi-scale registration method for diffusion magnetic resonance imaging according to claim 1, characterized in that, It also includes a step of training the registration network constructed by the above steps, the training step including: Input fixed and moving diffusion image samples into the registration network to obtain the predicted deformation field, and the registered moving FA map and moving DTI tensor field based on the predicted deformation field; calculate the FA structural similarity loss based on the FA map of the fixed diffusion image samples and the registered moving FA map. Calculate the DTI tensor consistency loss based on the DTI tensor field of the fixed diffuse image samples and the registered and moved DTI tensor field; calculate the deformation field smoothing regularization term based on the predicted deformation field. A joint loss function is constructed by integrating FA structural similarity loss, DTI tensor consistency loss, and deformation field smoothing regularization term; the parameters of the registration network are optimized by minimizing the joint loss function.

7. The reliability-guided multi-scale registration method for diffusion magnetic resonance imaging according to claim 6, characterized in that, The FA structural similarity loss is calculated based on multi-scale local normalized cross-correlation and is used to constrain the structural consistency between the fixed FA graph and the registered and moved FA graph. The DTI tensor consistency loss is calculated based on the tensor distance of the matrix logarithm and is calculated in the white matter region with reliable diffusion direction information. It is used to constrain the tensor orientation and shape consistency between the fixed DTI tensor field and the registered and moved DTI tensor field. The deformation field smoothing regularization term, calculated based on the bending energy of the predicted deformation field, is used to constrain the spatial continuity of the deformation field.

8. A reliability-guided diffusion magnetic resonance imaging multi-scale registration system, characterized in that, The system for performing the method according to any one of claims 1 to 7, the system comprising: The data receiving module is used to acquire the fixed diffusion image and the moving diffusion image to be registered. Both the fixed diffusion image and the moving diffusion image include the fractional anisotropy FA map and the diffusion tensor imaging DTI tensor field. The quantile feature normalization module is used to determine the relative scale relationship between the FA plot and the DTI tensor field based on the quantile statistics of the FA plot and the DTI tensor field, and adaptively adjust the amplitude of the DTI tensor field according to the relative scale relationship. The adjusted DTI tensor field is then fused with the FA plot to form a multi-channel diffusion representation. The reliability evidence extraction module is used to generate multi-scale structural mismatch maps based on the structural feature differences between multi-channel diffusion representations; and to generate multi-scale directional reliability maps based on the directional feature differences between DTI tensor fields. The deformation estimation module is used to introduce the multi-scale feature propagation process into the multi-scale feature propagation process by using the multi-scale structural mismatch map and multi-scale orientation reliability map as guiding information, adaptively modulate the features, and estimate the initial deformation field based on the modulated features. The residual refinement module is used to perform spatial transformation on the moving diffusion image based on the initial deformation field to obtain a coarse-registered moving diffusion image. It also estimates the residual displacement between the fixed diffusion image and the coarse-registered moving diffusion image at both coarse and fine scales, and corrects the initial deformation field based on the residual displacement to obtain the final deformation field. The registration output module is used to perform spatial transformation and tensor redirection on the FA map and DTI tensor field of the motion diffusion image based on the final deformation field, so as to obtain the registered FA map and DTI tensor field.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the reliability-guided diffusion magnetic resonance imaging multiscale registration method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the reliability-guided diffusion magnetic resonance imaging multiscale registration method as described in any one of claims 1 to 7.

Images