A method and system for generating supervoxels from multimodal three-dimensional magnetic resonance imaging of the brain.

Abstract

This invention discloses a method and system for generating supervoxels in multimodal three-dimensional magnetic resonance imaging (MRI) of the brain, relating to the field of medical image processing technology. The method involves selecting a reference sequence to establish a unified real physical coordinate system, completing multimodal sequence registration and robust intensity preprocessing; extracting multi-dimensional structural features on a voxel-by-voxel basis and fusing soft anatomical prior probability vectors to construct boundary maps and local complexity maps; generating adaptive-scale initial seeds in real physical space; using anatomical perception combined with distance iteration to complete three-dimensional clustering; and outputting an integer-type three-dimensional supervoxel label map consistent with the reference space after connected component correction and selective merging based on anatomical constraints. This invention eliminates the need for manual annotation and can generate structure-preserving, boundary-consistent, and granular adaptive three-dimensional supervoxels, suitable for downstream medical image analysis scenarios such as brain region segmentation and lesion heterogeneity analysis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically to a method and system for generating supervoxels from multimodal three-dimensional magnetic resonance imaging of the brain. Background Technology

[0002] Currently, three-dimensional magnetic resonance imaging (MRI) typically includes multiple sequences such as T1, T1ce, T2, FLAIR, DWI, ADC, and SWI. Different sequences can reflect brain tissue contrast, lesion morphology, and signal characteristics from multiple dimensions. Supravoxel technology, by aggregating spatially adjacent voxels with similar image features into three-dimensional region units, can serve as a fundamental preprocessing unit for brain region segmentation, graph neural network modeling, weakly supervised learning, region merging, and lesion heterogeneity analysis, playing a crucial role in computer-aided diagnosis and medical image analysis.

[0003] Existing supervoxel generation techniques mainly fall into the following categories: First, simple linear iterative clustering methods, represented by SLIC, achieve local clustering by uniformly arranging seed points and combining gray-level differences with spatial distance weighted distances. This method has high computational efficiency, but it is mostly geared towards two-dimensional images or regular grid three-dimensional data, and is not optimized for the multimodal characteristics and anisotropic voxels of brain MRI. Second, three-dimensional supervoxel clustering methods for point clouds or general volume data emphasize three-dimensional connectivity and regional compactness, but they are not designed for brain tissue anatomy, lesion boundaries, and medical imaging characteristics, making it difficult to adapt to the tissue distribution and signal characteristics of brain MRI. Third, multimodal segmentation and probabilistic graph methods in the field of medical imaging can output the probability distribution of lesions or tissues, but they rely on training labels and specific diseases, and cannot directly output structure-preserving oversegmented units that can be used for general graph modeling.

[0004] In summary, existing technologies have the following defects in the generation of supervoxels in multimodal brain MRI: (1) Insufficient utilization of multimodal complementary information: Traditional methods mostly use single grayscale or a small number of local features as the basis for clustering, which cannot fully integrate the complementary expression capabilities of multiple sequences for different regions such as normal brain tissue, necrotic areas, enhanced areas, and peritumoral edema; (2) Spatial distortion caused by anisotropic voxels: Brain MRI often has inconsistencies between in-plane resolution and slice thickness. Directly using array index distance will cause physical scale distortion, leading to supervoxel shape distortion, inter-slice breaks, or incorrect cross-slice aggregation; (3) Regular seeds Unable to adapt to local complexity: Complex structural regions such as ventricles, cortical sulci and lesion edges require smaller supervoxel granularity, while homogeneous white matter regions can use larger units. Uniform seed layout is difficult to balance accuracy and efficiency; (4) Lack of boundary crossing constraints: Relying solely on local grayscale or gradient differences cannot prevent voxels with similar signals on both sides of the boundary from crossing the boundary, which easily disrupts the consistency of anatomical boundaries; (5) Lack of anatomical constraints in post-processing: Traditional small region merging is based solely on volume or grayscale similarity, which easily crosses the boundaries of white matter, gray matter, cerebrospinal fluid or lesion subregions, reducing the reliability of subsequent regional analysis.

[0005] Therefore, how to generate a three-dimensional connected, structure-preserving, boundary-consistent, and locally complex supervoxel-labeled map using multimodal three-dimensional brain MRI of the same subject without relying on real manual segmentation labels of the images to be processed is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for generating supervoxels of multimodal three-dimensional magnetic resonance imaging of the brain to solve the problems existing in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging of the brain, comprising: S1. Receive at least two three-dimensional magnetic resonance imaging (MRI) sequences of the same subject, read the spatial metadata of each sequence, select a reference sequence and generate reference spatial parameters, and output the original multimodal image set, reference spatial parameters and a list of modalities to be registered. S2. Based on the reference spatial parameters obtained in S1, each non-reference mode is registered to the real physical coordinate system of the reference sequence to obtain a spatially aligned multimodal image set and transformation parameters; S3. Based on the spatially aligned multimodal image set obtained in S2, bias field correction, brain region mask extraction and robust intensity normalization within the brain region are performed sequentially, and the preprocessed multimodal image, three-dimensional brain region mask and normalized parameters are output. S4. Based on the preprocessed multimodal images and three-dimensional brain region masks obtained in S3, multidimensional structural features are extracted voxel by voxel within the brain region mask area, and soft anatomical prior probability vectors are fused to output voxel-level multimodal structural feature sets, tissue or lesion probability vector sets and local heterogeneity maps. S5. Based on the multimodal structural feature set, tissue or lesion probability vector set and local heterogeneity map obtained in S4, the multimodal gradient, prior probability gradient and structural response are fused to generate a boundary map, and the local complexity map is calculated based on the boundary strength, prior uncertainty, lesion candidate probability and local heterogeneity. S6. Based on the obtained three-dimensional brain region mask, boundary map and local complexity map, calculate the real physical volume and basic physical scale of the brain region, generate voxel-level adaptive scale, sample and optimize the initial supervoxel seed center in the real physical space, and output the initial seed set and local scale parameters. S7. Based on the obtained multimodal structural feature set, tissue or lesion probability vector set, boundary map, initial seed set and local scale parameters, an adaptive physical search domain is established around various sub-centers. Voxel allocation and center update are performed iteratively through anatomical perception joint distance to generate an initial three-dimensional supervoxel label map. S8. Based on the initial three-dimensional hypervoxel label map, boundary map, tissue or lesion probability vector set and multimodal structural feature set, perform three-dimensional connected component analysis and retain the main connected components of each label. Based on anatomical constraints, selectively merge discrete small regions and output the corrected three-dimensional hypervoxel label map and region adjacency data. S9. Based on the corrected 3D hypervoxel label map, regional adjacency data and reference spatial parameters, generate an integer hypervoxel label map that is completely consistent with the spatial parameters of the reference sequence. After performing spatial consistency verification, output the final result and auxiliary statistical data.

[0008] Optionally, the spatial metadata includes the spatial orientation, origin, affine matrix, voxel physical spacing, and matrix size of each sequence; the reference spatial parameters include the reference affine matrix, reference voxel spacing, and reference matrix size.

[0009] Optionally, it also includes the mapping relationship between voxel array indices and actual physical coordinates:

[0010] in, For the reference sequence affine matrix, For voxel array index, These are the actual physical coordinates. Optionally, the formula for calculating the robust strength normalization within the brain region is: No. The robust standardized representation of a modality within a brain region is as follows:

[0011] in, voxels In the The intensity of each mode, and These are the lower and upper quantiles of intensity within the brain region, respectively. The median value within the brain region. This represents the median absolute deviation. The value is positive. Optionally, the multi-dimensional structural features include modal normalization intensity, three-dimensional gradient features, local texture features, structural tensor response, and Hessian structural response; specifically represented as:

[0012] in, voxels In the Normalization strength of each mode Represents three-dimensional gradient features. Indicates local heterogeneity or texture features. Represents edge, sheet-like, tubular, or Hessian structural responses. This represents the prior probability vector of soft anatomy.

[0013] Optionally, the boundary map in S5 is calculated by fusing multimodal intensity gradient, anatomical prior probability gradient and structural response, and after normalization, it is used to characterize the probability that a voxel belongs to brain tissue or lesion boundary; the prior uncertainty is calculated based on the information entropy of the anatomical prior probability vector and is used to reflect the degree of uncertainty of voxel tissue attribution; the local complexity map is generated by fusing four indicators: boundary response intensity, prior uncertainty, lesion candidate probability and local heterogeneity.

[0014] Optionally, S7 specifically involves: establishing an adaptive physical search domain around each supervoxel center based on its corresponding local scale, and calculating the attribution relationship between the voxel and the center only within the search domain; Voxel allocation uses anatomically perceived joint distance as the criterion. This joint distance comprehensively considers four core constraints: multimodal feature differences, real physical spatial distance, boundary crossing cost on the path from the voxel to the center, and anatomical prior differences. Geodesic distance constraints are supplemented based on accuracy requirements. Each voxel is assigned to the supervoxel center with the smallest joint distance within the search domain. After each round of voxel allocation, the spatial position, feature mean, and prior mean of each supervoxel center are updated. This process is iterated until the maximum number of iterations is reached or the result converges, outputting an initial 3D supervoxel label map. Optionally, the region merging cost comprehensively considers the feature similarity of adjacent regions, differences in anatomical prior distribution, average strength of the common boundary, and volume shape constraints. Only when the common boundary strength of adjacent regions is lower than a preset boundary threshold, or the anatomical prior difference is lower than a preset prior threshold, are they allowed as merging candidates. Finally, the candidate region with the smallest merging cost is selected for merging. A multimodal three-dimensional brain magnetic resonance imaging supervoxel generation system includes a multimodal three-dimensional brain MRI acquisition module, a spatial registration and preprocessing module, a feature and anatomical prior construction module, a boundary map and complexity map calculation module, an adaptive seed initialization module, an anatomical perception three-dimensional clustering module, a connectivity correction and consistency merging module, and a result output and downstream interface module, which are connected sequentially. The multimodal three-dimensional brain MRI acquisition module is used to receive at least two three-dimensional brain magnetic resonance sequences of the same subject, read the spatial metadata corresponding to each sequence, select a reference sequence, and output the original multimodal image set, reference spatial parameters, and a list of modalities to be registered. The spatial registration and preprocessing module is used to register all non-reference modalities to the real physical coordinate system of the reference sequence, and sequentially perform bias field correction, brain region mask extraction and brain region intensity standardization on each spatially aligned modality, and output preprocessed multimodal images, three-dimensional brain region masks and standardized parameters. The feature and anatomical prior construction module is used to extract multi-dimensional structural features on a voxel-by-voxel basis within the coverage area of ​​the brain region mask, and to fuse soft anatomical prior probability vectors to output voxel-level multimodal structural feature sets, tissue or lesion probability vector sets and local heterogeneity maps. The boundary graph and complexity graph calculation module is used to fuse multimodal gradients, prior probability gradients and structural responses to generate a boundary graph, and to calculate a local complexity graph based on boundary strength, prior uncertainty, lesion candidate probability and local heterogeneity. The adaptive seed initialization module is used to calculate the real physical volume and basic physical scale of the brain region based on the obtained three-dimensional brain region mask, boundary map and local complexity map, generate voxel-level adaptive scale, sample and optimize the initial supervoxel seed center in the real physical space, and output the initial seed set and local scale parameters. The anatomical perception 3D clustering module is used to establish an adaptive physical search domain around various subcenters, and to perform voxel allocation and center update through anatomical perception joint distance iteration to generate an initial 3D supervoxel label map. The connectivity correction and consistency merging module is used to perform three-dimensional connectivity component analysis and retain the main connectivity components of each label based on the initial three-dimensional hypervoxel label map, boundary map, tissue or lesion probability vector set and multimodal structural feature set, and to perform selective merging of discrete small regions based on anatomical constraints, and output the corrected three-dimensional hypervoxel label map and region adjacency relationship data. The result output and downstream interface module is used to generate an integer-type three-dimensional hypervoxel label map that is completely consistent with the spatial parameters of the reference sequence, and simultaneously output the region adjacency map and region statistical attribute data.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for generating supervoxels of multimodal three-dimensional magnetic resonance imaging of the brain, which has the following beneficial effects: (1) Make full use of multimodal complementary information. This invention integrates the standardized intensity, gradient, texture and structural response of different MRI sequences into multimodal structural features, which helps to distinguish normal brain tissue, cerebrospinal fluid, ventricles, lesion core, enhancement area and edema area.

[0016] (2) All spatial calculations are based on real physical coordinates. This invention uses a reference sequence affine matrix and voxel physical spacing to define spatial distance, search domain, seed spacing and brain region volume, which can adapt to anisotropic images with inconsistent in-plane resolution and slice thickness.

[0017] (3) Seed density and local scale can vary with the complexity of brain structure. This invention adjusts the local scale by using boundary maps, prior uncertainty, lesion candidate probability and local heterogeneity, so that the lesion boundary, ventricle, cortical sulci and other regions can obtain finer-grained supervoxels, while homogeneous regions can maintain larger regional units.

[0018] (4) Explicit penalty for cross-boundary voxel assignment. This invention incorporates boundary crossing cost and anatomical prior difference into the joint distance, considering not only the local feature similarity between the voxel and the center, but also whether the path between them crosses strong tissue or lesion boundaries.

[0019] (5) Post-processing takes into account both three-dimensional connectivity and anatomical consistency. This invention performs 26-neighborhood connectivity analysis on the initial labels and introduces common boundary strength and prior difference threshold when merging discrete components or small regions to reduce the risk of crossing important anatomical boundaries in order to eliminate fragmentation.

[0020] (6) Soft anatomical priors improve deployability. This invention can use brain atlases, tissue segmentation, coarse lesion segmentation, multimodal clustering or self-supervised features to obtain probabilistic priors, but does not require the images to be processed to have artificial real labels, making it suitable for the generation of regional units in unsupervised, weakly supervised and multi-disease scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the method flow provided by the present invention; Figure 2 This is a schematic diagram of the system structure provided by the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0024] This invention discloses a method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging of the brain, such as... Figure 1 As shown, it includes: S1. Multimodal image acquisition and reference space determination: Receive at least two three-dimensional magnetic resonance imaging (MRI) sequences of the same subject, read the spatial metadata of each sequence, select a reference sequence and generate reference space parameters, and output the original multimodal image set, reference space parameters and a list of modalities to be registered; Specifically, the input data consists of at least two 3D brain MRI sequences of the same subject, the spatial orientation, origin, affine matrix, voxel spacing, and matrix size of each sequence. During processing, the system first checks whether each modality belongs to the same case, reads the spatial metadata of each modality, and selects one modality as the reference sequence; the reference sequence can be T1, T1ce, or other sequences with high spatial quality. The output data includes the original multimodal image set, the reference affine matrix, the reference voxel spacing, the reference matrix size, and a list of modalities to be registered.

[0025] The mapping from voxel array indices to actual physical coordinates can be represented as:

[0026] in, For the reference sequence affine matrix, For voxel array index, These are the actual physical coordinates. These coordinates are used consistently in subsequent calculations of spatial distance, search domain radius, seed spacing, and brain region volume.

[0027] S2, Multimodal Registration and Spatial Resampling: Based on the reference spatial parameters obtained in S1, each non-reference mode is registered to the real physical coordinate system of the reference sequence to obtain a spatially aligned multimodal image set and transformation parameters; Specifically, the input data consists of the original multimodal image output by S1, reference spatial information, and a list of modalities to be registered. During processing, each non-reference modality is registered to the physical coordinate system of the reference sequence. Registration can employ rigid, affine, and optional non-rigid transformations in sequence. Similarity metrics can be mutual information, normalized cross-correlation, or structural similarity. Trilinear or cubic interpolation is used for continuous intensity image resampling, while nearest-neighbor interpolation is used for discrete mask or label resampling. The output data includes a spatially aligned multimodal image set, the transformation matrix or deformation field of each modality to the reference space, and the registration quality control results.

[0028] No. The resampling result of each mode to the reference space can be expressed as:

[0029] in, For the first Spatial transformation from mode to reference space This is the spatially aligned modal image. If the registration quality does not meet the preset conditions, the system outputs a registration anomaly flag, or reverts to a more stable rigid or affine transformation result.

[0030] S3, Bias Field Correction, Brain Region Extraction and Intensity Standardization: Based on the spatially aligned multimodal image set obtained in S2, bias field correction, brain region mask extraction and robust intensity standardization within the brain region are performed sequentially, and the preprocessed multimodal image, three-dimensional brain region mask and standardized parameters are output. Specifically, the input data is the spatially aligned multimodal image output by S2. During processing, bias field correction is performed on each modality to reduce low-frequency intensity inhomogeneities; then, a brain region mask is generated, and maximum connected region preservation, hole filling, and necessary morphological corrections are performed on the brain region mask; finally, robust normalization of the intensity of each modality is performed within the brain region mask. The output data includes the preprocessed multimodal image, the 3D brain region mask, the brain region voxel set, and the normalized parameters for each modality.

[0031] No. The robust standardization of a modality within a brain region can be represented as:

[0032] in, and These are the lower and upper quantiles of intensity within the brain region, respectively. The median value within the brain region. This represents the median absolute deviation. It is a positive number. This step ensures that different modalities have comparable numerical scales before entering subsequent feature fusion.

[0033] S4. Construction of multimodal structural features and soft anatomical priors: Based on the preprocessed multimodal images and three-dimensional brain region masks obtained in S3, multidimensional structural features are extracted voxel by voxel within the brain region mask range, and soft anatomical prior probability vectors are fused to output voxel-level multimodal structural feature sets, tissue or lesion probability vector sets and local heterogeneity maps. Specifically, the input data consists of standardized multimodal images, brain region masks, and optional prior sources output by S3. During processing, standardized intensity, 3D gradient, local texture, structural tensor, or Hessian response for each modality are extracted voxel-by-voxel within the brain region, and these are fused with brain atlases, tissue segmentation, coarse lesion segmentation, multimodal clustering, or soft anatomical priors generated by self-supervised models. The output data includes voxel-level multimodal structural features, tissue or lesion probability vectors, local heterogeneity maps, and structural response maps.

[0034] The voxel-level multimodal structural features can be represented as:

[0035] in, Represents three-dimensional gradient features. Indicates local heterogeneity or texture features. Represents edge, sheet-like, tubular, or Hessian structural responses. This represents the soft-anatomical prior probability vector. This step transforms the multimodal grayscale data stream into a feature data stream that can be used for boundary judgment and cluster distance calculation.

[0036] S5. Boundary Graph and Local Complexity Graph Calculation: Based on the multimodal structural feature set, tissue or lesion probability vector set and local heterogeneity graph obtained in S4, the boundary graph is generated by fusing multimodal gradient, prior probability gradient and structural response, and the local complexity graph is calculated based on boundary strength, prior uncertainty, lesion candidate probability and local heterogeneity. Specifically, the input data consists of multimodal structural features, soft anatomical priors, local heterogeneity maps, and structural response maps output by S4. During processing, multimodal gradients, prior probability gradients, and structural responses are fused to obtain brain tissue or lesion boundary maps; then, boundary strength, prior uncertainty, lesion candidate probability, and local heterogeneity are fused to obtain local complexity maps. The output data includes boundary maps, prior uncertainty maps, lesion candidate probability maps, and local complexity maps.

[0037] The boundary diagram can be represented as:

[0038] Prior uncertainty can be expressed as:

[0039] The local complexity graph can be represented as:

[0040] in, Indicates the boundary response, This indicates prior uncertainty. Indicates the probability of lesion candidates. This indicates local heterogeneity. Higher complexity means denser subsequent seeds and smaller local scale.

[0041] S6. Adaptive Seed Initialization: Based on the obtained 3D brain region mask, boundary map and local complexity map, calculate the real physical volume and basic physical scale of the brain region, generate voxel-level adaptive scale, sample and optimize the initial supervoxel seed center in the real physical space, and output the initial seed set and local scale parameters. Specifically, the input data consists of the brain region mask obtained in S3, the local complexity map and boundary map obtained in S5, and the number of target supervoxels. During processing, the basic physical scale is first calculated based on the actual physical volume of the brain region and the number of target supervoxels, and then the adaptive scale at each voxel is determined based on the local complexity. Subsequently, variable-density Poisson disk sampling, weighted farthest-point sampling, or hierarchical grid sampling are used to generate an initial seed, which is then moved to a location with a lower boundary response within its neighborhood. The output data includes the initial set of supervoxel centers, the physical coordinates of each center, the initial feature mean, the prior mean, and the local scale.

[0042] The actual physical volume and fundamental physical scale of brain regions can be expressed as:

[0043] Voxel-level adaptive scaling can be expressed as:

[0044] Seed movement can be represented as:

[0045] in, Represents the average supervoxel physical scale. Indicates a local adaptive scale. This is the initial seed position. This represents the new seed position after the move. The purpose of moving the seed is to avoid placing the initial center on a strong boundary or in an unstable region.

[0046] S7. Anatomical perception 3D local iterative clustering: Based on the obtained multimodal structural feature set and tissue or lesion probability vector set, boundary map, initial seed set and local scale parameters, an adaptive physical search domain is established around various sub-centers. Voxel allocation and center update are performed through anatomical perception joint distance iteration to generate an initial 3D supervoxel label map. The input data consists of voxel features and soft anatomical priors output by S4, boundary maps output by S5, and seed centers and local scales output by S6. During processing, an adaptive physical search domain is established around each supervoxel center. Within this domain, the joint distance from the voxel to the center is calculated, and the voxel is assigned to the center with the smallest joint distance. After assignment, the spatial location, feature mean, prior mean, and local scale of the center are updated based on the voxels belonging to the same center, until the maximum number of iterations is reached or the label change rate is below a threshold. The output data includes the initial supervoxel label map, the updated center set, and statistical information for each supervoxel.

[0047] The adaptive physics search domain can be represented as:

[0048] The anatomical sensory joint distance can be expressed as:

[0049] Voxel tag assignment can be represented as:

[0050] A central update can be represented as:

[0051] Among them, feature differences in joint distance are used to maintain multimodal similarity, spatial distance is used to maintain regional compactness, boundary crossing cost is used to suppress cross-boundary assignment, anatomical prior differences are used to maintain prior consistency of tissues or lesions, and optional geodesic cost is used to further constrain clustering along weak boundaries or traversable paths.

[0052] S8. Connectivity Correction and Anatomical Consistency Merging: Based on the initial 3D hypervoxel label map, boundary map, tissue or lesion probability vector set and multimodal structural feature set, perform 3D connectivity component analysis and retain the main connectivity components of each label. Based on anatomical constraints, perform selective merging on discrete small regions and output the corrected 3D hypervoxel label map and region adjacency data. The input data consists of the initial supervoxel label map, supervoxel statistics, boundary map, soft anatomy priors, and voxel physical coordinates output by S7. During processing, a 26-neighborhood 3D connected component analysis is performed on each label, retaining the main connected components. For discrete components or small regions, candidate neighbors are searched in the region adjacency graph, and merging objects are selected based on regional feature differences, prior differences, common boundary strength, and volume / shape penalties. The output data consists of a connectivity-corrected supervoxel label map, a region adjacency graph, and a region-level statistical table.

[0053] The cost of merging regions can be expressed as:

[0054] The merger objective can be expressed as:

[0055] in, For the region Merge into adjacent areas The cost, The average intensity of the common boundary, For prior distribution differences, A penalty is imposed based on volume or shape. Candidate neighbors that cross strong boundaries or have excessively large prior differences are not considered for merging.

[0056] S9. Output Results and Spatial Consistency Verification: Based on the corrected 3D hypervoxel label map, regional adjacency data, and reference spatial parameters, generate an integer hypervoxel label map that is completely consistent with the spatial parameters of the reference sequence. After performing spatial consistency verification, output the final results and supplementary statistical data.

[0057] Specifically, the input data includes the corrected label map, reference spatial information, region adjacency graph, and region statistics table output by S8. During processing, the label map is written with the same spatial orientation, origin, voxel spacing, and matrix size as the reference sequence. The system checks whether all voxels within the brain region are labeled, whether the area outside the brain region is background, whether each label is 3D connected, and whether the number and volume distribution of labels meet preset ranges. The output data includes an integer-type 3D super-voxel label map, an optional region adjacency graph, a node feature table, an edge attribute table, and a quality control report.

[0058] The spatial consistency of the output label graph can be represented as:

[0059] A region adjacency graph can be represented as:

[0060] in, Each node corresponds to a supervoxel region, and node attributes may include multimodal mean, variance, texture, spatial location, and prior mean; Each edge corresponds to the common boundary between adjacent supervoxels, and edge attributes may include the common boundary area, average boundary strength, and regional feature differences.

[0061] The relevant parameter definitions are explained below: Image, voxel, and physical space symbols are shown in Table 1; Table 1. Image, Voxel, and Physical Space Symbols

[0062] The features, priors, and graphical symbols are shown in Table 2.

[0063] Table 2. Feature, Prior, and Graphical Symbols

[0064] The distance and weight symbols are shown in Table 3.

[0065] Table 3 Distance and Weight Symbols

[0066] In the affine matrix of the reference sequence is At that time, the transformation from voxel array indices to actual physical coordinates is as follows:

[0067] When the image orientation is aligned with the coordinate axes, it can be simplified to:

[0068] No. The robust normalization of each modality within the brain region mask is as follows:

[0069] in, and The lower and upper quantiles of intensity within the brain region. The median, This represents the median absolute deviation. It is a positive number.

[0070] The multimodal structural features can be represented as:

[0071] in, It is a three-dimensional gradient feature. It responds to edge, sheet, tubular, or Hessian structures.

[0072] Feature differences can be represented by Mahalanobis distance with regularization:

[0073] Boundary graphs can be constructed using the following formula:

[0074] The local complexity graph can be constructed using the following formula:

[0075] The basic physical scale of brain regions and the voxel-level adaptive scale are:

[0076] Seed movement can be performed by selecting a local low boundary location using the following formula:

[0077] The anatomical perception joint distance is:

[0078] in, , Can be adopted or , Available voxels To the center The maximum value, quantile, or mean of the response on the boundary graph along the physical straight path. The basic implementation can let... To reduce computational cost; high-precision implementation can approximate geodesic costs using the shortest path or fast travel method in the local search domain.

[0079] Key technical parameters are shown in Table 4; Table 4 Key Technical Parameters

[0080] A multimodal three-dimensional brain magnetic resonance imaging supervoxel generation system, such as Figure 2 As shown, it includes a multimodal 3D brain MRI acquisition module, a spatial registration and preprocessing module, a feature and anatomical prior construction module, a boundary map and complexity map calculation module, an adaptive seed initialization module, an anatomical perception 3D clustering module, a connectivity correction and consistency merging module, and a result output and downstream interface module, which are connected sequentially. The multimodal 3D brain MRI acquisition module is used to receive at least two 3D brain MRI sequences of the same subject, read the spatial metadata corresponding to each sequence, select a reference sequence, and output the original multimodal image set, reference spatial parameters, and a list of modalities to be registered. This module is also used to receive multiple 3D brain MRI sequences of the same subject and read information such as the device protocol, spatial orientation, origin, matrix size, and voxel physical spacing corresponding to each sequence. The function of this module is to confirm whether the input images belong to the same case and whether they possess metadata suitable for spatial unification, and to provide a unified data entry point for subsequent registration, real physical coordinate calculation, and label map output.

[0081] The spatial registration and preprocessing module is used to register all non-reference modalities to the real physical coordinate system of the reference sequence, and sequentially perform bias field correction, brain region mask extraction and brain region intensity standardization on each spatially aligned modality, and output preprocessed multimodal images, three-dimensional brain region masks and standardized parameters. This module unifies brain MRI images from different modalities into the spatial coordinate system of a reference sequence and performs bias field correction, brain region extraction, denoising, and intensity normalization within each modality. Its function is to eliminate spatial misalignment and intensity scale differences between different sequences, enabling subsequent feature extraction and supravoxel clustering to be performed in the same physical space and within a comparable intensity range.

[0082] The feature and anatomical prior construction module is used to extract multi-dimensional structural features voxel-by-voxel within the brain region mask coverage area, and fuse soft anatomical prior probability vectors to output a voxel-level multimodal structural feature set, tissue or lesion probability vector set, and local heterogeneity map. This module is used to construct the multimodal structural representation of each voxel within the brain region. Its functions include extracting the normalized intensity, three-dimensional gradient, local texture, Hessian or structural tensor response of each modality, and fusing brain tissue probability, lesion candidate probability, or other soft anatomical priors. The features and priors output by this module are used to describe the image representation, tissue attributes, and lesion correlation of voxels.

[0083] The boundary map and complexity map calculation module is used to fuse multimodal gradients, prior probability gradients, and structural responses to generate a boundary map, and to calculate a local complexity map based on boundary strength, prior uncertainty, lesion candidate probability, and local heterogeneity. This module is used to determine which locations may belong to the boundaries of brain tissue, ventricles, cerebrospinal fluid, or lesion subregions, and to assess the local structural complexity. The boundary map is used to suppress supravoxels from crossing strong tissue boundaries or lesion boundaries, and the complexity map is used to determine whether a denser seed and a smaller regional scale are needed locally. This module enables the supravoxel generation process to adapt to complex regions such as ventricles, sulci, and lesion edges.

[0084] The adaptive seed initialization module is used to calculate the true physical volume and basic physical scale of the brain region based on the obtained 3D brain region mask, boundary map, and local complexity map, generate a voxel-level adaptive scale, sample and optimize the initial supervoxel seed centers in the real physical space, and output the initial seed set and local scale parameters. This module is used to initialize the supervoxel centers in the real physical space. Its function is to determine the seed density and initial scale according to the local complexity, placing finer seeds in regions with complex structures, uncertain priors, or high lesion candidate probabilities, and sparser seeds in homogeneous brain tissue regions. This module also moves the initial seeds to positions with lower boundary responses within the neighborhood, reducing the probability of the center falling on the tissue boundary.

[0085] The anatomically-aware 3D clustering module is used to establish an adaptive physical search domain around various sub-centers. It iteratively performs voxel allocation and center updates through anatomically-aware joint distance to generate an initial 3D hypervoxel label map. This module performs the core 3D hypervoxel clustering. Its function is to establish an adaptive physical search domain around each hypervoxel center and perform voxel allocation by integrating multimodal feature differences, real spatial distance, boundary crossing costs, anatomical prior differences, and optional geodesic costs. Through multiple rounds of allocation and center updates, this module generates an initial 3D hypervoxel label map.

[0086] The connectivity correction and consistency merging module is used to perform three-dimensional connectivity component analysis based on the initial 3D ultravoxel label map, boundary map, tissue or lesion probability vector set, and multimodal structural feature set. It retains the main connected components of each label and selectively merges discrete small regions based on anatomical constraints, outputting a corrected 3D ultravoxel label map and region adjacency data. This module corrects discrete components, small fragments, and unreasonable connections that may appear in the initial label map. Its function is to perform 3D connectivity analysis, retain the main connected regions, and select merging objects based on the feature similarity of adjacent regions, prior consistency, common boundary strength, and volume shape constraints. This module prohibits merging regions that cross strong boundaries or have excessively large anatomical prior differences.

[0087] The output and downstream interface module is used to generate an integer-type 3D ultravoxel label map that is completely consistent with the spatial parameters of the reference sequence, and simultaneously outputs a region adjacency map and region statistical attribute data. This module outputs an integer-type 3D ultravoxel label map consistent with the spatial parameters of the reference sequence, and can simultaneously output a region adjacency map, a region feature table, and a boundary attribute table. Its function is to ensure that the ultravoxel results can be directly used by subsequent segmentation, graph modeling, weakly supervised annotation propagation, lesion heterogeneity analysis, or manual interactive annotation tools.

[0088] The following specific embodiments are introduced to further illustrate the method of the present invention.

[0089] Example 1: Quadratic Brain Tumor MRI Supravoxel Generation

[0090] Input four 3D sequences: T1, T1ce, T2, and FLAIR, for the same subject. Set the T1ce or T1 sequence as the reference sequence and read the affine matrix and voxel spacing of each NIfTI file. If the four sequences are already aligned with the same matrix size and physical coordinate system, verify the consistency between the affine matrix and key brain structures; otherwise, perform rigid registration first, then affine registration, and add non-rigid registration if necessary.

[0091] N4 bias field correction was performed on each of the four modalities. Brain region masks were obtained using a skull dissection model or brain atlas registration, and the masks were subjected to maximum connected component preservation, hole filling, and morphological closure operations. Quantile truncation and robust normalization were performed on each modality within the brain region. If a modality was missing, the corresponding channel was removed from the feature tensor, and the weights of the remaining modalities were renormalized.

[0092] Construct four-modal normalized intensity, three-dimensional gradient magnitude, local variance, local entropy, and three-dimensional structural tensor or Hessian response for each modality. Utilize a coarse lesion segmentation model to generate probabilities for non-tumor, peritumoral edema, tumor core, and enhanced regions; probabilities for white matter, gray matter, and cerebrospinal fluid tissue can also be incorporated simultaneously. These probabilities are only used as soft priors.

[0093] Construct a boundary graph based on multimodal gradients and probability gradients. The complexity graph weights can be initially set to... , , , Then adjust based on the development set. Target number of supervoxels. The basic scale will be determined based on subsequent tasks. It is automatically calculated from the actual physical volume of the brain region.

[0094] The seed is initialized using variable density farthest point sampling, and the seed is placed in a radius of... Move to the lower boundary position within the physical neighborhood. The joint distance weight can be initially set to... , , , , Search domain multiplier The maximum number of iterations can be set to 2, and the maximum number of iterations can be set to 10. The process will stop early when the maximum physical displacement of the center is less than 0.2 mm or the tag change rate is lower than a preset threshold. The above parameters are for implementation examples and do not limit the scope of protection.

[0095] After clustering, connected components are analyzed using a 26-neighborhood approach. The minimum region volume threshold can be set to 0.1 to 0.3 times the average supervoxel volume. Merging is only allowed when the average strength and prior divergence of the common boundary are both below the threshold. The output label map is an integer-type 3D matrix with the same spatial orientation, origin, voxel spacing, and matrix size as the reference sequence.

[0096] Example 2: Multimodal brain MRI supravoxel generation of a lesion-free coarse segmentation model

[0097] In the absence of a coarse segmentation network for lesions, T1, T2, and FLAIR trimodal sequences can be input. Probabilities of white matter, gray matter, and cerebrospinal fluid are obtained using brain atlases or tissue segmentation models; lesion candidate probabilities are also calculated. It can be obtained by fusing the anomaly probabilities of FLAIR high signal, T1 low signal, and multimodal Gaussian mixture model. If lesion candidate probabilities are not constructed, then let the complexity graph... And in the seed move item, let .

[0098] This embodiment illustrates that the present invention does not rely on real artificial segmentation labels, nor on precise segmentation networks specific to certain diseases. In scenarios such as stroke, white matter hyperintensity, demyelinating lesions, or postoperative changes, the method of constructing lesion candidate probabilities can be replaced, while keeping the real physical space, adaptive seeds, joint distance, and connectivity correction process unchanged.

[0099] Example 3: Anisotropic and Large-Volume Image Processing

[0100] For anisotropic images with voxel spacings of 0.5 mm, 0.5 mm, and 3.0 mm, resampling to isotropic is not mandatory. Spatial distances are directly calculated using real physical coordinates, the search domain array radius is converted according to the voxel spacing in each direction, and fewer indexing steps are used in the layer thickness direction to maintain physical coverage consistent with the in-plane direction. If necessary, temporary isotropic feature maps can be generated only for gradient and texture calculations, and the final labels are still mapped to the original mesh of the reference sequence.

[0101] For large-volume 3D images that cannot be loaded into video memory or RAM at once, brain regions can be divided into 3D blocks with overlapping boundaries based on their physical location. Each supervoxel center calculates candidate distances only within blocks intersecting its physical search domain, and overlapping voxels retain the minimum distance result from all candidate centers. Center updates use globally accumulated features and coordinate statistics to reduce the influence of block boundaries on supervoxel boundaries.

[0102] Example 4: Output Verification and Downstream Applications

[0103] After generation, the following conditions can be checked: all brain region voxels have non-zero labels, and labels outside brain regions are zero; each label is a single connected region in a 26-neighborhood; the affine matrix of the label map is consistent with the reference sequence; the number of supervoxels, the average volume, the volume distribution, and the proportion of abnormal small regions are within the preset range.

[0104] For research-annotated development data, edge recall, undersegmentation error, anatomical purity, reachability Dice, and connectivity can be used to evaluate hypervoxel quality. These annotations and metrics are for research validation only and should not be used as inputs for generating hypervoxels during clinical deployment.

[0105] The output supervoxels can be used to construct a region adjacency graph. Each supervoxel serves as a node, and node features include modal mean, variance, texture, location, and prior mean. Edges are established between supervoxels sharing a common surface, and edge attributes include common boundary area, average boundary strength, and feature differences. This adjacency graph can be used for graph neural networks, region merging, weakly supervised annotation propagation, or lesion heterogeneity analysis.

[0106] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0107] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging of the brain, characterized in that, include: S1. Receive at least two three-dimensional magnetic resonance imaging (MRI) sequences of the same subject, read the spatial metadata of each sequence, select a reference sequence and generate reference spatial parameters, and output the original multimodal image set, reference spatial parameters and a list of modalities to be registered. S2. Based on the reference spatial parameters obtained in S1, each non-reference mode is registered to the real physical coordinate system of the reference sequence to obtain a spatially aligned multimodal image set and transformation parameters; S3. Based on the spatially aligned multimodal image set obtained in S2, bias field correction, brain region mask extraction and robust intensity normalization within the brain region are performed sequentially, and the preprocessed multimodal image, three-dimensional brain region mask and normalized parameters are output. S4. Based on the preprocessed multimodal images and three-dimensional brain region masks obtained in S3, multidimensional structural features are extracted voxel by voxel within the brain region mask area, and soft anatomical prior probability vectors are fused to output voxel-level multimodal structural feature sets, tissue or lesion probability vector sets and local heterogeneity maps. S5. Based on the multimodal structural feature set, tissue or lesion probability vector set and local heterogeneity map obtained in S4, the multimodal gradient, prior probability gradient and structural response are fused to generate a boundary map, and the local complexity map is calculated based on the boundary strength, prior uncertainty, lesion candidate probability and local heterogeneity. S6. Based on the obtained three-dimensional brain region mask, boundary map and local complexity map, calculate the real physical volume and basic physical scale of the brain region, generate voxel-level adaptive scale, sample and optimize the initial supervoxel seed center in the real physical space, and output the initial seed set and local scale parameters. S7. Based on the obtained multimodal structural feature set, tissue or lesion probability vector set, boundary map, initial seed set and local scale parameters, an adaptive physical search domain is established around various sub-centers. Voxel allocation and center update are performed iteratively through anatomical perception joint distance to generate an initial three-dimensional supervoxel label map. S8. Based on the initial three-dimensional hypervoxel label map, boundary map, tissue or lesion probability vector set and multimodal structural feature set, perform three-dimensional connected component analysis and retain the main connected components of each label. Based on anatomical constraints, selectively merge discrete small regions and output the corrected three-dimensional hypervoxel label map and region adjacency data. S9. Based on the corrected 3D hypervoxel label map, regional adjacency data and reference spatial parameters, generate an integer hypervoxel label map that is completely consistent with the spatial parameters of the reference sequence. After performing spatial consistency verification, output the final result and auxiliary statistical data.

2. The method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging (MRI) of the brain according to claim 1, characterized in that, The spatial metadata includes the spatial orientation, coordinate origin, affine matrix, voxel physical spacing, and matrix size of each sequence; The reference space parameters include the reference affine matrix, the reference voxel spacing, and the reference matrix size.

3. The method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging (MRI) of the brain according to claim 1, characterized in that, It also includes the mapping relationship between voxel array indices and actual physical coordinates: in, For the reference sequence affine matrix, For voxel array index, These are the actual physical coordinates.

4. The method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging (MRI) of the brain according to claim 1, characterized in that, The formula for calculating the standardized robustness of the brain region is as follows: No. The robust standardized representation of a modality within a brain region is as follows: in, voxels In the The intensity of each mode, and These are the lower and upper quantiles of intensity within the brain region, respectively. The median value within the brain region. This represents the median absolute deviation. It is a positive number.

5. The method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging (MRI) of the brain according to claim 1, characterized in that, The multi-dimensional structural features include modal normalization intensity, three-dimensional gradient features, local texture features, structural tensor response, and Hessian structural response; specifically represented as: in, voxels In the Normalization strength of each mode Represents three-dimensional gradient features. Indicates local heterogeneity or texture features. Represents edge, sheet-like, tubular, or Hessian structural responses. This represents the prior probability vector of soft anatomy.

6. The method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging (MRI) of the brain according to claim 1, characterized in that, The boundary map in S5 is calculated by fusing multimodal intensity gradient, anatomical prior probability gradient and structural response. After normalization, it is used to characterize the probability that a voxel belongs to brain tissue or lesion boundary. The prior uncertainty is calculated based on the information entropy of the anatomical prior probability vector and is used to reflect the degree of uncertainty of voxel tissue attribution. The local complexity map is generated by fusing four indicators: boundary response intensity, prior uncertainty, lesion candidate probability and local heterogeneity.

7. The method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging (MRI) of the brain according to claim 1, characterized in that, Specifically, S7 involves: establishing an adaptive physical search domain around each supervoxel center based on its corresponding local scale, and calculating the attribution relationship between the voxel and the center only within the search domain; Voxel assignment uses anatomical perception joint distance as the criterion. The joint distance comprehensively considers four core constraints: multimodal feature differences, real physical space distance, boundary crossing cost on the path from the voxel to the center, and anatomical prior differences. Geodesic distance constraints are added according to accuracy requirements. Each voxel is assigned to the supervoxel center with the smallest joint distance in the search domain. After each round of voxel assignment, the spatial position, feature mean, and prior mean of each supervoxel center are updated. The process is iterated until the maximum number of iterations is reached or the result converges, and the initial three-dimensional supervoxel label map is output.

8. The method for generating supervoxels from multimodal three-dimensional magnetic resonance imaging (MRI) of the brain according to claim 1, characterized in that, The region merging cost comprehensively considers the feature similarity of adjacent regions, the difference in anatomical prior distribution, the average strength of the common boundary, and volume shape constraints. Only when the common boundary strength of adjacent regions is lower than the preset boundary threshold and the anatomical prior difference is lower than the preset prior threshold, are they allowed to be merged as candidates. Finally, the candidate region with the lowest merging cost is selected for merging.

9. A multimodal three-dimensional brain magnetic resonance imaging supervoxel generation system, characterized in that, The method for generating supervoxels of multimodal three-dimensional brain magnetic resonance imaging according to any one of claims 1-8 includes a multimodal three-dimensional brain MRI acquisition module, a spatial registration and preprocessing module, a feature and anatomical prior construction module, a boundary map and complexity map calculation module, an adaptive seed initialization module, an anatomical perception three-dimensional clustering module, a connectivity correction and consistency merging module, and a result output and downstream interface module, which are connected in sequence. The multimodal three-dimensional brain MRI acquisition module is used to receive at least two three-dimensional brain magnetic resonance sequences of the same subject, read the spatial metadata corresponding to each sequence, select a reference sequence, and output the original multimodal image set, reference spatial parameters, and a list of modalities to be registered. The spatial registration and preprocessing module is used to register all non-reference modalities to the real physical coordinate system of the reference sequence, and sequentially perform bias field correction, brain region mask extraction and brain region intensity standardization on each spatially aligned modality, and output preprocessed multimodal images, three-dimensional brain region masks and standardized parameters. The feature and anatomical prior construction module is used to extract multi-dimensional structural features on a voxel-by-voxel basis within the coverage area of ​​the brain region mask, and to fuse soft anatomical prior probability vectors to output voxel-level multimodal structural feature sets, tissue or lesion probability vector sets and local heterogeneity maps. The boundary graph and complexity graph calculation module is used to fuse multimodal gradients, prior probability gradients and structural responses to generate a boundary graph, and to calculate a local complexity graph based on boundary strength, prior uncertainty, lesion candidate probability and local heterogeneity. The adaptive seed initialization module is used to calculate the real physical volume and basic physical scale of the brain region based on the obtained three-dimensional brain region mask, boundary map and local complexity map, generate voxel-level adaptive scale, sample and optimize the initial supervoxel seed center in the real physical space, and output the initial seed set and local scale parameters. The anatomical perception 3D clustering module is used to establish an adaptive physical search domain around various subcenters, and to perform voxel allocation and center update through anatomical perception joint distance iteration to generate an initial 3D supervoxel label map. The connectivity correction and consistency merging module is used to perform three-dimensional connectivity component analysis and retain the main connectivity components of each label based on the initial three-dimensional hypervoxel label map, boundary map, tissue or lesion probability vector set and multimodal structural feature set, and to perform selective merging of discrete small regions based on anatomical constraints, and output the corrected three-dimensional hypervoxel label map and region adjacency relationship data. The result output and downstream interface module is used to generate an integer-type three-dimensional hypervoxel label map that is completely consistent with the spatial parameters of the reference sequence, and simultaneously output the region adjacency map and region statistical attribute data.

Images