Brain image region label generation method based on standard space inverse deformation reflection

Abstract

The invention discloses a brain image area label generation method based on standard space inverse deformation reflection, and relates to the technical field of medical image processing and computer vision. The method comprises the following steps: performing skull stripping and resampling preprocessing on an individual space target brain image, performing rigid, affine and nonlinear cascade registration on the preprocessed brain image and a standard space template, and solving a forward deformation field and an inverse deformation field; calling an inverse deformation field to reversely propagate standard space anatomical atlas labels to an individual space, and resolving label conflicts to obtain an initial label graph; generating each modal coarse label graph through the registration matrix, and performing weighted fusion according to a confidence coefficient weight to obtain a consistent coarse label set; quality control is executed based on anatomy prior, an individualized label is obtained after topological repair, and finally a label, metadata and an uncertainty evaluation report are exported. According to the method, automation and standardization of brain image tag generation are realized, and the accuracy, consistency and generalization ability of tags are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing and computer vision, and in particular to a brain image region label generation method based on standard space inverse deformation back mapping. BACKGROUND

[0002] Early diagnosis and precise treatment of brain diseases are important research directions in the field of modern medicine. Clinical research and treatment of diseases such as Parkinson's disease, Alzheimer's disease, and brain glioma all rely on high-precision segmentation and consistent labeling of key nuclei and fine-grained regions of the whole brain. The brain has a complex anatomical structure and significant individual differences. The shape and position of key regions such as the basal ganglia, hippocampus, and thalamus change, which are core imaging indicators reflecting disease progression and assessing pathological damage. Precise labeling can provide core support for statistical analysis, radiomics feature extraction, and deep learning diagnostic model training, and is also the basis for constructing expert prior knowledge graph.

[0003] Current brain image region labeling mainly relies on two methods: manual region-by-region delineation or direct template segmentation. However, both methods have obvious limitations. Manual labeling not only requires hours or even days of time cost from medical experts, but is also easily affected by subjective experience, fatigue state, and other factors, resulting in poor consistency of labeling results from different labelers or different time periods of the same labeler, making it difficult to meet the needs of large-scale data set construction. Direct template segmentation can save time, but the inherent differences between standard templates and individual anatomical structures will cause label shift, and the imaging protocol differences across scanning devices and research centers will cause domain shift problems, severely affecting the generalization ability of the labeling results and making it impossible to provide unified standard labeling support for multi-center clinical research.

[0004] With the deep application of artificial intelligence technology in the field of medical imaging, the training of medical image large models and visual language models (VLM) increasingly requires large-scale and standardized labeling data. The pre-training, alignment, and retrieval enhancement process of such models requires a large amount of high-quality expert prior atlas as a supervision signal, and the existing labeling method cannot efficiently generate standardized pseudo-labeling data that meet the requirements. At the same time, the downstream tasks of multi-disease auxiliary diagnosis and longitudinal cohort study of brain diseases also require reliable region of interest (ROI) labeling as the basis for quantitative analysis, and the efficiency and accuracy bottleneck of existing labeling methods has become a key factor restricting the rapid progress of related research.

[0005] To solve the above technical problems, the industry urgently needs an automatic, high-precision, scalable brain image region label generation technology that can adapt to multiple regions of the whole brain, can be compatible with multi-modal brain image data such as MRI, PET, and CT, can effectively overcome the influence of individual anatomical differences and cross-center domain offset, and can generate individualized labels with high consistency and high reliability while reducing labor costs. SUMMARY

[0006] Based on the above technical problems, the present application discloses a brain image region label generation method based on standard space inverse deformation back mapping, which specifically includes:

[0007] Obtain a target brain image in individual space, perform preprocessing operations including skull stripping, resampling, and intensity standardization on the target brain image to obtain a preprocessed brain image;

[0008] Perform cascade registration on the preprocessed brain image as a fixed image and a standard space template as a moving image to complete rigid registration, affine precise registration, and nonlinear deformation registration to obtain a forward deformation field and an inverse deformation field ;

[0009] Call the inverse deformation field , perform inverse propagation on the label set in the standard space anatomical atlas through nearest neighbor interpolation, and resolve label conflicts of overlapping voxels in the propagation process using a structure scale priority rule to obtain an initial label map in individual space;

[0010] Establish a registration matrix of the target brain image and other modal brain images, and map the initial label map in individual space to the coordinate system of each modal brain image through the registration matrix to generate a coarse label map of the corresponding modal;

[0011] Configure confidence weights according to the recognition degree of each modal to anatomical structures, and perform weighted fusion on the coarse label maps of each modal to obtain a cross-modal consistent coarse label set;

[0012] Based on anatomical priors such as brain region volume range, left-right symmetry, and centroid anatomical position, perform quality control on the consistent coarse label set, complete topological repair through isolated connected domain rejection, boundary smoothing, and hole filling to obtain target individualized brain image region labels;

[0013] Export the target individualized brain image region labels, corresponding metadata, and uncertainty evaluation report.

[0014] Preferably, the specific preprocessing operations include: performing artifact removal on the target brain image after skull dissection to identify and eliminate strip-like and spot-like artifacts caused by scanning equipment or subject movement, with artifact identification based on the local standard deviation of voxel intensity; and performing intensity normalization on the brain image after artifact removal to map the intensity distribution of the brain image to a preset normal distribution. ,in The mean, The standard deviation is used, and the mapping method is a linear transformation: , For standardized voxels The intensity value, For original strength, , The mean and standard deviation of the original strength.

[0015] Preferably, the cascaded registration includes a rigid registration stage, an affine fine registration stage, and a nonlinear deformation registration stage, specifically:

[0016] In the rigid registration stage, mutual information is used as the similarity metric. A multi-resolution pyramid is constructed, and the rigid body transformation parameters are iteratively optimized at each resolution level until the change in the similarity metric is less than a preset threshold. ;

[0017] In the affine registration stage, based on the rigid registration results, affine transformation parameters are introduced, and multi-resolution pyramid iterative optimization is used until the change in the similarity metric is less than a preset threshold. ;

[0018] The nonlinear deformation registration stage adopts spline free deformation algorithm, setting the control point grid spacing to... , Matching the spatial scale of small brain structures, and using regularized mutual information as the objective function, the inverse deformation field is constrained to satisfy: ,in The positive definiteness threshold of the deformation field is used for iterative optimization until the change in the objective function is less than the preset threshold. .

[0019] Preferably, the forward deformation field With inverse deformation field The method for verifying reciprocity is as follows: randomly select individuals within the space Discrete voxel coordinates ,calculate and Spatial distance, if all distances are less than the voxel side length If 1 / 2 of the deformation field is obtained, then the deformation field is determined to satisfy the reciprocity constraint.

[0020] Preferably, the structural scale priority rule is implemented as follows:

[0021] All labels in the standard space atlas are pre-divided into three levels: small structure region, spatial volume ≤ , subcortical region, spatial volume ∈ , and cortical region, spatial volume > 50 mm3. , wherein is a volume threshold value obtained based on standard template statistics;

[0022] For overlapping voxels in the reverse propagation , traverse all labels corresponding to the mapping in the standard space , extract the level to which each label belongs, and select the label corresponding to the highest priority level as the target label of ;

[0023] When the overlapping voxels correspond to multiple labels of the same level, a local majority voting method is used to resolve, and the formula is as follows: , wherein is the label set under the level, is the target label value of the voxel , is the three-dimensional coordinate of the voxel mapped to in the standard space, is an indicator function, is the original label value of the voxel in the standard space.

[0024] Preferably, the registration matrix establishment process is as follows:

[0025] With the preprocessed target brain image as the reference, rigid registration and affine registration are sequentially performed on other modality brain images. The rigid registration optimizes the translation and rotation parameters with mutual information as the measurement. The affine registration optimizes the scaling and shearing parameters based on the rigid result. For the registered other modality brain image, the normalized mutual information value with the target brain image is calculated, if , is a preset registration accuracy threshold, then the registration is re-executed until the condition is met. The transformation parameters obtained in the registration process are integrated into the inter-modality registration matrix , and the coordinates of the individual space initial label map are mapped to the coordinate system of the other modality brain image through to generate the corresponding modality coarse label map.

[0026] Preferably, the configuration and weighted fusion process of the confidence weight is as follows:

[0027] For each modality brain image, the label recognition degree of the preset set of anatomical structures is calculated, The calculation method is as follows: ,in Number of anatomical structures. , , The first The number of true positive, false positive, and false negative voxels in each structure; based on the discriminant value. Based on this, confidence weights are configured for each modality. ,satisfy , The number of modalities is such that the weight of structural modalities is not less than the minimum configuration confidence weight. For the corresponding voxels of each modality's coarse label map, perform weighted fusion: ,in For the first Each modality in voxels The coarse label value at that location.

[0028] Preferably, the quality control specifically includes:

[0029] For each brain region in the uniform coarse label set, calculate its volume. And check if it meets the requirements. ,in This represents the volume of the corresponding brain region in the standard spatial template. This is a preset lower limit volume ratio coefficient. This is a preset upper limit ratio coefficient for volume; if it is not met, the brain region is marked as abnormal; the volume ratio is calculated for the left and right brain regions with the same name. And check if it meets the requirements. ,in A preset symmetry threshold is used; if this threshold is not met, the brain regions are marked as abnormal. The centroid coordinates of each brain region are calculated. It also verifies whether the brain region is located within a preset neighborhood of the centroid of the corresponding brain region in the standard template, where the radius of the neighborhood is [missing information]. and The brain region is matched with the scale of the brain region. If the scale is exceeded, the brain region is marked as abnormal. For all marked abnormal brain regions, the label information of the corresponding brain region in the standard template is used for preliminary correction before entering the topology repair stage.

[0030] Preferably, the specific operations of the topology repair include:

[0031] For each brain region in the uniform coarse label set, extract all its connected components and remove voxels with a number less than [a certain value]. isolated connected components, is a threshold value set based on the brain region scale; a morphological opening and closing operation is used to smooth the brain region boundary; connectivity analysis is performed on the hollow voxels in the brain region; if there is a connectivity path between the hollow voxels and the brain region boundary voxels, and all voxels on the path belong to the brain region, then the hollow voxels are filled with the label of the brain region, and the filling rule satisfies:

[0032]

[0033] wherein is the boundary of the brain region , is the label of the brain region , is the original label of .

[0034] Preferably, the generation content of the uncertainty evaluation report includes:

[0035] Based on the Jacobian determinant calculation of the inverse deformation field, the uncertainty of each voxel in the individual space is calculated, and the formula is: , wherein is the Jacobian determinant value of the inverse deformation field at voxel , is the determinant calculation operator, is the gradient matrix of the inverse deformation field at voxel ; for high uncertainty region marking, if or , then is marked as a high uncertainty voxel, , , are preset determinant threshold values; a spatial distribution heat map containing the high uncertainty region, a proportion statistics of high uncertainty voxels in each brain region, a list of anatomical positions corresponding to the high uncertainty region and manual review suggestions for these regions are obtained.

[0036] Compared with the prior art, the technical scheme of the present application has the following technical effects:

[0037] The present application realizes efficient automation of brain image region label generation, significantly reduces the labor and time cost, and only needs a few minutes to complete the whole brain region label generation through standardized data preprocessing, cascade registration and inverse deformation propagation and other automatic steps, without relying on a large amount of manual intervention, greatly improves the label generation efficiency, eliminates the subjective differences of manual work, guarantees the stability and high reproducibility of label quality of different samples and different batches, and provides an efficient solution for large-scale data set construction.

[0038] The application realizes label standardization and high adaptability in cross-modal and cross-center scenes, through multi-modal registration and coordinate fusion taking T1WI as the reference mode, unifying the spatial reference of labels of each mode, combining with the mode confidence weighted fusion strategy, ensuring the consistency of multi-modal image labels such as MRI, PET and CT; at the same time, through standardized preprocessing and quality control, the influence of imaging protocol differences across devices and centers is weakened, the generated labels can be used in different research scenes, effectively improving the label generalization ability, and supporting multi-center clinical research of brain diseases.

[0039] The application realizes high-precision individualized label generation of whole brain fine-grained regions, provides reliable support for downstream multi-disease research, establishes accurate mapping of standard space atlas and individual image through nonlinear deformation registration and inverse deformation field solving, combines structure scale priority conflict resolution and topological repair, and ensures the boundary fit degree of small structure region labels such as basal ganglia and hippocampus; the generated labels meet the anatomical prior constraints such as brain volume and left-right symmetry, and can be used as the basis for ROI quantitative analysis of diseases such as PD, AD and brain glioma, and provide reliable data for disease auxiliary diagnosis and pathological mechanism research.

[0040] The application realizes effective empowerment of medical image large model and VLM training, the generated standardized labels can be used as expert prior atlas to provide high-quality supervision signal, support large model pre-training, instruction fine-tuning and visual language alignment; at the same time, QA question and answer pairs and image pairs containing modal recognition and regional relationship dimensions can be automatically generated, a large-scale training corpus is constructed, the problem of lack of high-quality labeled data in large model training is solved, and the performance improvement and clinical landing of medical image artificial intelligence model are assisted.

[0041] The above description is only a summary of the technical scheme of the present application, in order to more clearly understand the technical means of the present application, the contents of the description can be implemented, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following will be described in detail with the preferred embodiments of the present application and the accompanying drawings.

[0042] According to the detailed description of the specific embodiments of the present application in the following combined with the drawings, those skilled in the art will be more clear about the above and other purposes, advantages and characteristics of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual proportion.

[0044] According to the description of the drawings in the document and the corresponding technical content, the titles of the drawings are as follows:

[0045] Figure 1 . The flowchart of the brain image region label generation method based on standard space inverse deformation remapping;

[0046] Figure 2 . The flowchart of processing individual space multi-modal brain image to cross-modal precise label;

[0047] Figure 3 . The module architecture diagram of the brain image region label generation system based on standard space inverse deformation remapping;

[0048] Figure 4 . The image comparison diagram before and after the skull stripping of brain image;

[0049] Figure 5 . The image comparison diagram before and after the intensity standardization and bias field correction of brain image;

[0050] Figure 6 . The distribution diagram of registration deformation field in three-dimensional space and multi-view;

[0051] Figure 7 . The three-dimensional space distribution comparison diagram of forward deformation field and inverse deformation field;

[0052] Figure 8 . The superimposed display diagram of the coarse label generation result corresponding to multi-modal brain image;

[0053] Figure 9 . The comparison diagram of original label image, coarse label image and precise label image after topology repair;

[0054] Figure 10 . The three-dimensional space coordinate system and slice extraction method diagram of brain image volume data;

[0055] Figure 11 . The flowchart of nnU-Net-based putamen and globus pallidus segmentation and PD and MSA classification;

[0056] Figure 12. Multi-disease multi-modal brain image corresponding caudate-putamen shell label and 3D visualization schematic diagram. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the following will be combined with the drawings in the embodiments of the present application to make a clear and complete description of the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. In the following description, specific details such as specific configurations and components are provided only to help a comprehensive understanding of the embodiments of the present application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, descriptions of known functions and structures are omitted in the embodiments for clarity and conciseness.

[0058] It should be understood that the term "one embodiment" or "the embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "one embodiment" or "the embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner.

[0059] In addition, reference numerals and / or letters can be repeated in different examples in the present application. Such repetition is for the purpose of simplification and clarity, and does not itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0060] The term "and / or" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, B exists alone, and A and B exist simultaneously. The term "and" herein is a description of another association relationship of the associated objects, which means that there can be two relationships, for example, A and B can mean that A exists alone and A and B exist simultaneously. In addition, the character " / " herein generally means that the associated objects before and after are in an "or" relationship.

[0061] The term "at least one" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, at least one of A and B can mean that A exists alone, A and B exist simultaneously, and B exists alone.

[0062] It is also need to make clear that, in this article, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "include", "contain" or any other variant thereof is intended to cover non-exclusive inclusion.

[0063] Embodiment 1

[0064] This embodiment mainly describes a brain image region label generation method based on standard space inverse deformation reflection mapping, as shown in Figure 1 , specifically comprising:

[0065] Obtaining a target brain image of an individual space, performing a preprocessing operation including skull stripping, resampling and intensity standardization on the target brain image to obtain a preprocessed brain image;

[0066] Performing cascade registration on the preprocessed brain image as a fixed image and a standard space template as a moving image to complete rigid registration, affine fine registration and non-linear deformation registration to obtain a forward deformation field and an inverse deformation field ;

[0067] Calling the inverse deformation field , performing inverse propagation on a label set in a standard space anatomical atlas through nearest neighbor interpolation, and using a structure scale priority rule to resolve label conflicts of overlapping voxels in the propagation process to obtain an initial label map of the individual space;

[0068] Establishing a registration matrix of the target brain image and other modal brain images, mapping the initial label map of the individual space to each modal brain image coordinate system through the registration matrix to generate a coarse label map of the corresponding modal;

[0069] Configuring a confidence weight according to the recognition degree of each modal to the anatomical structure, and performing weighted fusion on the coarse label map of each modal to obtain a cross-modal consistent coarse label set;

[0070] Performing quality control on the consistent coarse label set based on anatomical priors of brain region volume range, left-right symmetry degree and centroid anatomical position, and completing topological repair through isolated connected domain rejection, boundary smoothing and hole filling to obtain a target individualized brain image region label;

[0071] Exporting the target individualized brain image region label, corresponding metadata and uncertainty evaluation report.

[0072] Furthermore, the specific preprocessing operations also include performing artifact removal on the target brain image after skull dissection, identifying and eliminating strip-like and spot-like artifacts caused by scanning equipment or subject movement. Artifact identification is based on the local standard deviation of voxel intensity; if the local standard deviation is greater than... ( If a preset artifact detection threshold is set, the region is identified as an artifact region. When performing intensity normalization on the brain image after artifact removal, the intensity distribution of the brain image is mapped to a preset normal distribution. ,in Represents a normal distribution. Let be the mean of this normal distribution. The standard deviation of this normal distribution is represented by a linear transformation. ,in For standardized voxels The intensity value, voxels The original intensity value, This represents the mean intensity of all voxels in the original brain image. The standard deviation of the intensity of all voxels in the original brain image; the resampling refers to resampling the target brain image to an isotropic voxel size. ( (Preset voxel side lengths) to ensure consistency of spatial information during registration.

[0073] Furthermore, cascaded registration includes a rigid registration stage, an affine fine registration stage, and a nonlinear deformation registration stage, specifically as follows:

[0074] In the rigid registration stage, mutual information is used as the similarity metric. A multi-resolution pyramid is constructed, and the rigid body transformation parameters are iteratively optimized at each resolution level until the change in the similarity metric is less than a preset threshold. ;

[0075] In the affine registration stage, based on the rigid registration results, affine transformation parameters are introduced, and multi-resolution pyramid iterative optimization is used until the change in the similarity metric is less than a preset threshold. ;

[0076] The nonlinear deformation registration stage adopts spline free deformation algorithm, setting the control point grid spacing to... , Matching the spatial scale of small brain structures, and using regularized mutual information as the objective function, the inverse deformation field is constrained to satisfy: ,in The positive definiteness threshold of the deformation field is used for iterative optimization until the change in the objective function is less than the preset threshold. .

[0077] Furthermore, the forward deformation field With inverse deformation field The method for verifying reciprocity is as follows: randomly select individuals within the space Discrete voxel coordinates ,calculate and Spatial distance, if all distances are less than the voxel side length If 1 / 2 of the deformation field is obtained, then the deformation field is determined to satisfy the reciprocity constraint.

[0078] Furthermore, the specific implementation of the structural scale priority rule is as follows: all labels in the standard spatial anatomical atlas are pre-divided into three levels: small structural regions, spatial volume ≤ Subcortical region, spatial volume ∈ ), cortical region, spatial volume > ,in The volume threshold is obtained statistically based on a standard template;

[0079] Overlapping voxels during backpropagation Iterate through all the labels that correspond to its mapping in the standard space. Extract the hierarchy to which each tag belongs, and select the tag corresponding to the highest priority hierarchy as... The target label;

[0080] When overlapping voxels correspond to multiple labels at the same level, a local majority voting method is used for resolution, with the following formula: ,in This is the set of tags at this level. voxels The target label value, Mapped to standard space The three-dimensional coordinates of the voxels For indicator functions, Standard space voxels The original label value.

[0081] Furthermore, the registration matrix is ​​established as follows: using the preprocessed target brain image as a reference, rigid registration and affine registration are performed sequentially on other modal brain images. Rigid registration optimizes translation and rotation parameters using mutual information as a metric; affine registration optimizes scaling and shearing parameters based on the rigid registration result; for the registered other modal brain images, the normalized mutual information value with the target brain image is calculated. ,like , If the preset registration accuracy threshold is not met, registration is re-executed until the condition is satisfied; the transformation parameters obtained during the registration process are integrated into an inter-modal registration matrix. ,pass Mapping the coordinates of the individual spatial initial label map to the coordinate system of the brain image of other modalities to generate the coarse label map of the corresponding modality.

[0082] Further, the configuration and weighting fusion process of the confidence weight is: for the brain image of each modality, calculate the label recognition degree of the preset anatomical structure set , The calculation method is: , wherein is the number of anatomical structures, , , , respectively, the true positive, false positive, and false negative voxel number of the first structure; based on the recognition degree , configure the confidence weight of each modality, satisfying , is the number of modalities, and the weight proportion of the structure class modality is not less than the minimum configured confidence weight ; for the corresponding voxel of each modality coarse label map, perform weighted fusion: , wherein is the coarse label value of the first modality at voxel .

[0083] Further, the specific quality control is: for each brain region in the consistent coarse label set, calculate its volume and check whether it satisfies , wherein is the volume of the corresponding brain region in the standard space template, is a preset volume lower limit proportion coefficient, is a preset volume upper limit proportion coefficient, if not satisfied, mark the brain region as abnormal; calculate the volume ratio of the left and right homonymic brain regions and check whether it satisfies , wherein is a preset symmetry threshold, if not satisfied, mark the brain region as abnormal; calculate the centroid coordinates of each brain region and check whether it is within the preset neighborhood range of the centroid of the corresponding brain region in the standard template, the neighborhood radius is and matches the brain region scale, if it exceeds, mark the brain region as abnormal; for all marked abnormal brain regions, call the label information of the corresponding brain region in the standard template for preliminary correction, and then enter the topology repair link.

[0084] Further, the specific operation of topology repair includes: for each brain region in the consistent coarse label set, extract all connected domains thereof, and remove isolated connected domains with a voxel number less than , A threshold is set based on brain region scale; morphological opening and closing operations are used to smooth brain region boundaries; connectivity analysis is performed on void voxels within brain regions: if a void voxel has a connected path to a brain region boundary voxel and all voxels on the path belong to that brain region, then the void voxel is filled with the label of that brain region, and the filling rule satisfies:

[0085]

[0086] in brain region The boundary, brain region The tag, for The original tag.

[0087] Furthermore, the uncertainty assessment report includes the following content:

[0088] Based on the Jacobian determinant calculation of the inverse deformation field, for each voxel in the individual space The calculation formula is as follows: ,in For the inverse deformation field in voxels The Jacobian determinant value at that location, Operators for calculating determinants, For the inverse deformation field in voxels The gradient matrix at the location; for marking high uncertainty regions, if or Then Marked as high uncertainty voxels, , These are preset matrix thresholds; the system obtains a spatial distribution heatmap containing high uncertainty regions, a statistical analysis of the proportion of high uncertainty voxels in each brain region, a list of anatomical locations corresponding to high uncertainty regions, and manual review suggestions for these regions.

[0089] This implementation details how standardized preprocessing, cascaded registration, and inverse deformation propagation automate the entire process, effectively eliminating subjective differences in manual annotation and ensuring the stability and reproducibility of label quality. By leveraging multimodal registration fusion and anatomical prior quality control, it overcomes individual anatomical differences and cross-central region offset issues, improving the consistency and generalization ability of labels across multimodal data. The generated individualized labels accurately fit fine-grained regions of the whole brain, providing reliable ROIs for brain disease auxiliary diagnosis and radiomics analysis, while also meeting the training needs of large-scale medical imaging models and providing high-quality supervision signals.

[0090] Example 2

[0091] The embodiment describes in detail a brain image region label generation system based on standard space inverse deformation reflection mapping, which is used to implement the brain image region label generation method based on standard space inverse deformation reflection mapping. The system includes a data preprocessing unit, a template registration unit, an inverse deformation field solving unit, a label propagation unit, a cross-modal mapping unit, a multi-modal fusion unit, a quality control unit, a topology repair unit, an uncertainty evaluation unit, and an export unit. Specifically, the system includes:

[0092] The data preprocessing unit is used to obtain a target brain image in an individual space, perform skull stripping on the target brain image to generate a brain mask, perform resampling on the brain mask and the target brain image to a preset isotropic voxel size, and simultaneously perform intensity standardization and artifact removal operations, and output a preprocessed brain image.

[0093] The template registration unit takes the preprocessed brain image as a fixed image and a standard space template as a moving image, and sequentially performs rigid registration, affine fine registration, and non-linear deformation registration. In the rigid registration, mutual information is used as a similarity measure to construct a multi-resolution pyramid to iteratively optimize rigid transformation parameters. In the affine fine registration, affine transformation parameters are introduced and optimized based on the results of the rigid registration. In the non-linear deformation registration, a B-spline free deformation algorithm is used and the deformation field is constrained to be reversible. The forward deformation field is solved and output. and the inverse deformation field , and has the function of performing mutual inverse verification on the two deformation fields.

[0094] The inverse deformation field solving unit receives the forward deformation field output by the template registration unit , and uses a fixed point iteration method or a gradient method to numerically invert it. In the iteration process, a preset convergence threshold is used to control the iteration stop. The inverse deformation field with the required precision is output to the label propagation unit.

[0095] The label propagation unit calls the inverse deformation field output by the inverse deformation field solving unit , and performs inverse propagation along the inverse deformation field by nearest neighbor interpolation on the label set in the standard space anatomical atlas. For overlapping voxel label conflicts that occur during propagation, the structure scale priority rule of small structure region→subcortical region→cortical region is used to resolve them. If there are multiple labels at the same level, a local majority voting method is used for further processing. The individual space initial label map is output as the target.

[0096] The cross-modal mapping unit takes the preprocessed brain image as a reference, and sequentially performs rigid registration and affine registration on other modal brain images to establish a registration matrix. The individual space initial label map output by the label propagation unit is mapped to the coordinate system of each modal brain image through the registration matrix. The label part that exceeds the field of view of the modal is automatically cropped and the proportion of valid voxels is recorded. The corresponding coarse label map of each modal is generated and output.

[0097] The multi-modal fusion unit calculates the label recognition degree of each modality brain image on the preset anatomical structure set, configures the confidence weight of each modality (the structure class modality weight proportion is not less than a preset lower limit) based on the recognition degree, performs weighted fusion on the coarse label set of each modality output by the cross-modality mapping unit, and outputs the cross-modality consistent coarse label set;

[0098] The quality control unit loads the volume, centroid coordinates and other anatomical prior data of each brain region in the standard space template, verifies whether the volume of each brain region in the consistent coarse label set output by the multi-modal fusion unit is within the preset proportion range of the template volume, whether the volume ratio of the left and right homonymous brain regions meets the symmetry requirement, and whether the centroid coordinates are located within the preset anatomical neighborhood, marks the brain regions that do not meet the conditions as abnormal, and calls the standard template label for preliminary correction;

[0099] The topology repair unit receives the coarse label set corrected by the quality control unit, extracts the connected domain of each brain region and removes the isolated connected domain with a number of voxels less than a preset threshold, performs morphological opening and closing operation on the brain region boundary for smoothing processing, and simultaneously performs connectivity analysis on the hollow voxels in the brain region to fill the hollows that meet the conditions of existing connectivity path with the brain region boundary and consistent path voxel label, and outputs the target individualized brain image regional label;

[0100] The uncertainty evaluation unit receives the inverse deformation field output by the inverse deformation field solving unit , calculates the Jacobian determinant of the inverse deformation field for each voxel in the individual space, marks the voxels with a determinant value exceeding a preset threshold range as high-uncertainty voxels and generates a high-uncertainty mask, and generates an uncertainty evaluation report in combination with the proportion statistics of high-uncertainty voxels in each brain region;

[0101] The export unit receives the target individualized brain image regional label output by the topology repair unit, the evaluation report output by the uncertainty evaluation unit, and the metadata (including label-name-color dictionary and brain region volume statistical value) generated by each unit, exports the label in NIfTI / NRRD format, the metadata in CSV / JSON format, and the evaluation report in PDF / HTML format.

[0102] This embodiment describes in detail that each link is efficiently coordinated through modular design, the inverse deformation field solving and label propagation unit guarantees accurate mapping of the standard atlas to the individual space, and the multi-modal fusion and topology repair unit further optimizes the label accuracy. The quality control and uncertainty evaluation unit forms a closed-loop quality control to ensure that the label meets the anatomical constraints. The system can generate standardized labels on a large scale, greatly reducing labor costs, supporting multi-center clinical research and large model corpus construction, and its flexible adaptability can be compatible with different modal brain image data, providing stable support for downstream multi-disease research and artificial intelligence model landing.

[0103] Based on Embodiment 1 or 2, this embodiment details the automatic construction of image pairs and question-answer pairs for VLM training corpus, using individual spatial labels obtained by inverse deformation of standardized whole brain anatomical atlas (MNI standard space) to automatically generate three types of corpus suitable for VLM training needs: single-image QA, paired-image QA and discriminative QA. The specific implementation details are optimized around the corpus construction needs, specifically:

[0104] In the data preprocessing stage, the original brain image data is uniformly converted to NIfTI format, the direction matrix (such as RAS or LPS) and voxel spacing metadata of the image are retained, PHI processing is completed to protect data privacy, and deface operation can also be selected; then skull stripping is performed on T1WI modality brain image, brain mask is obtained by intensity threshold combined with morphological method or learning-based segmentation algorithm, and T1WI is masked to suppress non-brain tissue signal. The process effect is shown in Figure 4 Figure 4 The left side is the original brain image before skull stripping, and the non-brain tissue region signal such as skull and scalp is obvious, Figure 4 The right side is the result after skull stripping, only the brain parenchyma region is retained, effectively eliminating the interference of non-brain tissue to the subsequent process; then spatial and intensity standardization is performed, resampling T1WI to isotropic 0.8–1.2mm³ (preferably 1.0mm³), adjusting the intensity distribution by z-score standardization, and performing low-frequency bias field correction by polynomial or spline field estimation, as shown in Figure 5 Figure 5 The left side is the image before standardization, there are obvious intensity unevenness and spatial resolution difference, Figure 5 The right side is the result after standardization, the image intensity distribution is more uniform and the spatial details are clearer, effectively alleviating the artifacts caused by uneven receiving field; for other modalities of brain images such as T2, FLAIR and PET, maintain their original spatial relationship with T1WI, and do not perform cross-modality registration for the time being, waiting for subsequent unified processing.

[0105] In the MNI standard space template and individual T1WI registration link, a cascaded rigid registration and affine fine registration is adopted, taking the preprocessed T1WI as the fixed image and the MNI standard space template as the moving image, constructing a 3–5 layer multi-resolution pyramid, combining normalized mutual information (NMI) as the similarity measure, estimating 6 degrees of freedom rigid parameters first, and then optimizing 12 degrees of freedom affine parameters, with 300–1000 iterations per layer and step and convergence threshold decreasing with level; if higher accuracy is required, further non-linear deformation registration can be performed, using BSpline basis function deformation or differentiable deformation field algorithm to estimate smooth reversible displacement field, controlling the range of Jacobian determinant through regularization constraint, and finally outputting the forward deformation field ​​And the affine matrix, the effect of the forward deformation field is as follows Figure 6 As shown, Figure 6 The upper part is a three-dimensional schematic diagram of the deformation field, which allows for a direct observation of the overall deformation trend from the individual space to the MNI standard space. Figure 6 The lower half, from left to right, presents the two-dimensional deformation field in the axial view, coronal view, and sagittal view. The gradient distribution of different colors clearly reflects the local deformation differences of each anatomical region, providing an intuitive deformation reference for subsequent inverse deformation field solution.

[0106] In the inverse deformation field solution and the backpropagation stage of the map label, the forward deformation field is first solved. Numerical inversion is performed using a fixed-point iterative method or gradient method (50–200 iterations, convergence threshold). ), to obtain the inverse deformation field The contrast between the forward and inverse deformation fields in this process is as follows: Figure 7 As shown, Figure 7 The lower half is the forward deformation field. A three-dimensional illustration, Figure 7 The upper part is the inverse deformation field The three-dimensional illustration shows that the two are inversely related in spatial distribution, verifying the reversibility of the deformation field; then, standardized whole-brain anatomical atlases under the MNI standard space (such as Neuromorphometrics, Harvard–Oxford, AAL, etc.) are selected as label sources, and nearest-neighbor interpolation is performed along the inverse deformation field. The labels Li are inversely mapped to the individual space. For overlapping voxels, conflicts are resolved using a small structure → subcortical → cortical priority table or a local majority voting method. Finally, an integer label graph with metadata is output, providing a label basis for subsequent corpus generation.

[0107] In the coarse label generation stage for each modality's brain region, the rigidity / affine matrix of other modalities relative to T1WI is first estimated. The T1 space label map is then resampled to the coordinate system of each modality to generate coarse label maps. For modalities with incomplete vision (such as parts of FLAIR and PET), the areas exceeding the field of view are automatically cropped and the coverage is recorded. At the same time, the visualization effect is rendered based on the region ID (α = 0.3–0.6 transparent fill + 2–6px outline). In the cross-modal coordinate fusion and unification stage, the coarse label maps of each modality are back-projected to T1 space to achieve pixel-level alignment. The integrated labels are obtained by weighted fusion according to modal confidence, and the effect is as follows: Figure 8 As shown, Figure 8The coarse label generation results for five modalities—T1W, T2W, FLAIR, PET-CFT, and PET-FDG—are presented sequentially from left to right and top to bottom. The labels for each modality are precisely superimposed on the corresponding brain images, and the outlines of the labels for different anatomical regions are clearly distinguishable, intuitively demonstrating the consistency and accuracy of cross-modal labels. In terms of slice selection, the slice with the largest axial area in each brain region is taken as the central layer k, and k±1 or k±2 layers are derived as candidate slices for image pairs.

[0108] In the label quality control and topology repair stage, label quality is verified based on prior dissection (volume 50%–150% of template volume, left-right volume ratio 0.5–2.0, centroid position compliance). Isolated clusters below 50–200 voxels are removed through connected component analysis. Boundaries are smoothed using hole filling and morphological opening / closing operations. If the effect is unsatisfactory, a segmentation-joint smoothing algorithm is used to eliminate burrs. Figure 9 As shown, Figure 9 From left to right, the images are the original label image, the coarse label image, and the repaired fine label image. The original label image has boundary spikes and small isolated clusters. The coarse label image retains the main areas but still has some detail defects. The repaired fine label image has smoother boundaries and higher regional integrity, effectively avoiding label leakage. At the same time, a high uncertainty mask is generated based on Jacobian determinant or registration residual, which prioritizes manual review to ensure that the label quality meets the requirements of the VLM corpus.

[0109] During the evaluation and export phase, volume difference, symmetry index, and coverage were used as indirect evaluation indicators. Subsequently, the corpus data was exported according to rules, including original slice images, masked highlighted slice images, and JSON / JSONL format dialogue entries (including single-image QA, paired-image QA, and discriminative QA); as shown in the table below. Figure 10 As shown:

[0110] Image Question Answer Figure 10 (a) "image": "images / Patientl_Axial_Tl_Slice6.png" "conversations": [{"from":"human", "value": " What's in the image?"}, {"from": "gpt", "value": "It's an axial T1-weighted MRI of the brain."}"meta": {"plane": "axial", "modality": "MRI-T1", "spacing_mm":[1.0,1.0,1.0]} Figure 10 (b) "image": "images / Patientl_Axial_Tl_Slice6_L35.png" "conversations": [{"from":"human", "value": " What's therelationship between thefirst and second image?"}, {"from": "gpt", "value": "The second imagehighlights the Brainstem region derived fromthe first image via atlas-based inversewarping."}"meta": {"region_id": 35, "region_name": "Brainstem", "render": "filled+contour","alpha": 0.45}

[0111] The file names are in the format PatientID_Plane_Modality_Slice{idx}_L{region}.png. The metadata includes information such as region_id, hemisphere, and area_px, ensuring that the corpus is traceable and reusable and can be directly used for VLM pre-training and alignment, providing high-quality supervision signals for VLM to understand brain image characteristics and anatomical relationships.

[0112] Based on Embodiment 1 or 2, this embodiment details the nnU-Net-based segmentation algorithm for the putamen and caudate nucleus, and its application verification for differentiating early symptoms of Parkinson's disease (PD) and multiple system atrophy (MSA). High-precision individualized labels are generated using standard spatial inverse deformation mapping technology, providing high-quality training data for the nnU-Net segmentation model. The resulting ROIs (Regions of Interest) in the basal ganglia, such as the putamen and caudate nucleus, are then used to help differentiate early symptoms of PD and MSA. The overall process is consistent with the whole-brain region label generation method, and key parameters and steps have been optimized for precise segmentation of the basal ganglia. Each core step is illustrated with diagrams to visually demonstrate the technical effects. The core objective of this application is to automatically generate individual (patient) spatial whole-brain region labels using inverse deformation based on the MNI standard spatial anatomical atlas within a multi-center MRI / PET dataset. The focus is on the putamen, caudate nucleus, and other basal ganglia regions closely related to the pathological mechanisms of PD and MSA, providing reliable anatomical references for subsequent disease classification and quantitative analysis. The overall technical framework is as follows: Figure 2 As shown in the flowchart (brain image segmentation), from Figure 11 As can be clearly seen, multimodal brain images (including MRI series T1WI, T2WI, FLAIR and PET series CFT, FDG, etc.) are first preprocessed and registered in two stages. Region labels are then created by combining the inverse matrix of MNI transformation. The data is then input into the nnU-Net model to complete the segmentation of the caudate nucleus and putamen. The segmentation results are combined with the multi-slice extraction, multimodal fusion and voting mechanism of the ViT model to achieve the classification of PD and MSA. At the same time, heat map visualization results are output, which fully presents the entire link technology logic from data input to disease classification.

[0113] In the brain imaging dataset labeling process, the data preprocessing stage uses T1WI as the core reference modality. First, a skull dissection operation is performed to remove interference from non-brain tissues such as the skull and scalp. Then, the T1WI is resampled to an isotropic voxel size of 0.8–1.2 mm³ (preferably 1.0 mm³) to ensure that the spatial resolution meets the segmentation requirements of small structures in the basal ganglia. Subsequently, z-score intensity standardization is performed to eliminate intensity deviations caused by different scanning devices and scanning parameters. If there is low-frequency intensity unevenness in the image, N4 Bias-Field correction can also be performed to further improve image quality. The registration process employs a "individual T1WI (fixed image) ↔ MNI standard spatial template (moving image)" registration strategy, constructing a multi-resolution pyramid (3–5 layers) to balance registration accuracy and efficiency. First, a rigid initial registration of 6 degrees of freedom (maximum step size 2.0, minimum step size 0.1) is performed using mutual information metric to preliminarily align the individual brain image with the overall anatomical structure of the standard template. Then, a 12-degree-of-freedom affine fine registration (maximum step size 0.05, minimum step size 0.001) is performed using mutual information as the metric, optimizing scaling and cropping parameters to better reflect individual anatomical differences. To further improve the registration accuracy of small structures (putamen, caudate nucleus), nonlinear registration can be performed using SyN or B-Spline. The FFD algorithm sets the control point spacing to 3–6 mm (preferably 4 mm), the regularization parameter λ = 0.1–1.0 (preferably 0.5), and the gradient step size to 0.1–0.5 (preferably 0.2). The smooth regularization constraint ensures the reversibility and continuity of the deformation field. Finally, the forward deformation field Phi(T1→MNI) and the inverse deformation field Phi(MNI→T1) are obtained.

[0114] In the reverse deformation and label backpropagation process, the obtained reverse deformation field Phi(MNI→T1) is called to perform backpropagation on each label (especially the labels corresponding to the putamen and caudate nucleus) in the standardized whole-brain anatomical atlas (such as the Neuromorphometrics atlas, which has a more refined division of the basal ganglia region) under the MNI standard space through nearest neighbor interpolation, mapping them to the individual T1WI space. To address the potential issue of voxel label overlap during propagation, a conflict resolution rule of "subcortical region first, followed by cortical region" is adopted. Since the putamen and caudate nucleus are both subcortical basal ganglia structures, their label information is retained first when there is a label conflict, ensuring the integrity and accuracy of the basal ganglia region labels, and finally forming a single integer label map. In the coarse label generation stage for each modality, a rigid or affine registration matrix is ​​first established between the individual's T1WI and other modalities (T2WI, FLAIR, PET-CFT, PET-FDG, etc.). Then, the label map in the T1WI space is mapped to each modal space through this matrix, and nearest neighbor interpolation is used to keep the label encoding unchanged. For modalities where the field of view cannot completely cover the whole brain (such as some PET modalities), the label part that exceeds the field of view is automatically cropped during the mapping process, and the effective label voxel ratio is recorded. This ratio can be used as an important indicator for subsequent label quality control.

[0115] The cross-modal unification process uses the T1WI coordinate system as a unified reference, resampling or backprojecting the coarse label images generated by each modality to the T1WI space to ensure that all modal label images have the same voxel spacing, matrix size, and orientation conventions (such as RAS), thus achieving cross-modal coordinate alignment. Subsequently, confidence weights are assigned according to the modality's recognition of the basal ganglia region. For example, T1WI displays anatomical structures more clearly and has higher recognition of the putamen and caudate nucleus, so it is assigned a higher weight. PET modalities (such as PET-CFT, which can reflect the distribution of dopaminergic neurons) are more sensitive to pathological changes, so they are assigned a medium weight. FLAIR displays lesion areas better, so it is assigned an auxiliary weight. Through weighted fusion, a cross-modal unified whole-brain coarse label set is obtained, especially ensuring the consistency of putamen and caudate nucleus labels. In the label quality control and topology repair stages, rigorous quality control is performed on the standardized coarse label set based on anatomical priors: In terms of volume, the volumes of the putamen and caudate nucleus are constrained to be within 50%–150% of the corresponding region volume in the MNI standard template to avoid label volume abnormalities caused by registration deviations; in terms of symmetry, the volume ratio of the corresponding basal ganglia regions on the left and right sides (left caudate nucleus-right caudate nucleus, left putamen-right putamen) is checked to ensure it is within a reasonable range of 0.5–2.0, conforming to the symmetry characteristics of normal brain anatomy; in terms of location, the centroid coordinates of the putamen and caudate nucleus are verified to be within the preset anatomical quadrant of the basal ganglia region of the cerebral hemispheres, ensuring the accuracy of label position. In the topology repair stage, considering the characteristics of small structures in the basal ganglia, an isolated connected domain threshold (preferably 100 voxels) is set to remove isolated voxel clusters with excessively small volumes in the label image, avoiding noise interference; simultaneously, cavity filling and morphological opening and closing operations are performed to smooth label boundaries, eliminate label spurs, and ensure the topological integrity of the putamen and caudate nucleus labels.

[0116] In the evaluation and export phase, if the dataset contains manually drawn gold standard labels for the putamen and caudate nucleus, indicators such as the Dice similarity coefficient (measuring label overlap), HD95 (95% Hausdorff distance, measuring boundary differences), and average surface distance (ASD, measuring overall boundary closeness) can be calculated to quantitatively evaluate label accuracy. In multi-center datasets lacking gold standards, volume deviation from the MNI template, left-right symmetry index, and mean / standard deviation of signals within the label region are used as indirect evaluation indicators to ensure the reliability of label quality. Finally, the generated individualized whole-brain labels (primarily including putamen and caudate nucleus labels) are exported in NIfTI or NRRD format, along with a "label-name-color" dictionary file (clearly indicating the label codes and display colors corresponding to the left / right caudate nucleus and left / right putamen), a CSV / JSON file of volume statistics for each region (recording the putamen and caudate nucleus volume data for each patient, providing a basis for subsequent quantitative pathological analysis), and optionally, a quality control report can be generated, indicating abnormal label areas and treatment suggestions. The label generation results are as follows: Figure 12 As shown in the diagram (showing how to create a segmentation label), from...Figure 3 As can be seen intuitively, accurate 3D labels were generated for different patients (patients 1 to 4) under the three modalities of T1WI, T2WI and FLAIR. The labels of the putamen and caudate nucleus regions are highly consistent with the anatomical structures of the corresponding modal brain images. The differences in labels among different patients also clearly reflect individual anatomical differences, providing high-quality data support for subsequent model training.

[0117] To further ensure the accuracy of the labels, the generated putamen and caudate nucleus labels will be submitted to two medical professionals with five or more years of experience in neuroimaging diagnosis for review and correction. Doctors will combine multimodal imaging (such as T1WI to observe anatomical structures, FLAIR to exclude lesion interference, and PET to observe functional changes) to determine whether the labels accurately cover the boundaries of the putamen and caudate nucleus, whether there are any label leakage or missing issues, and manually adjust labels with large deviations to ensure that the labels used for model training meet clinical anatomical standards. These precise putamen and caudate nucleus segmentation labels serve two purposes. First, they act as supervisory signals input into the nnU-Net segmentation model for training and parameter optimization, enabling the model to automatically segment key basal ganglia structures in multi-center, multimodal data. Second, the segmented putamen and caudate nucleus regions of interest (ROIs) can be used to extract radiomics features (such as volume, shape, and grayscale texture features) or combine them with PET functional imaging features (such as dopamine transporter uptake rate). These features can then be input into classification models such as ViT. By comparing the anatomical differences (e.g., more significant putamen atrophy in MSA patients) and functional metabolic differences in the putamen and caudate nucleus regions between PD and MSA patients, early symptom differentiation between the two diseases can be achieved. This provides objective imaging evidence for clinical diagnosis, especially in scenarios where early symptoms are similar and clinical differentiation is difficult. This technology can significantly improve the accuracy and efficiency of diagnosis.

[0118] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A brain image region label generation method based on standard spatial inverse deformation inverse mapping, characterized in that, include: A target brain image of an individual space is acquired, and a preprocessing operation including skull dissection, resampling and intensity normalization is performed on the target brain image to obtain a preprocessed brain image. Using the preprocessed brain images as stationary images and the standard spatial template as moving images, cascaded registration was performed to complete rigid registration, affine registration, and nonlinear deformation registration, resulting in the forward deformation field. With inverse deformation field ; Invoking the inverse deformation field The label set in the standard spatial anatomy atlas is backpropagated through nearest neighbor interpolation. Label conflicts of overlapping voxels during the propagation process are resolved using structural scale priority rules to obtain the initial label map of the individual space. Establish a registration matrix between the target brain image and other modal brain images, and map the initial spatial label map of the individual to the coordinate system of each modal brain image through the registration matrix to generate a coarse label map of the corresponding modality; Confidence weights are assigned according to the recognition of anatomical structures by each modality, and weighted fusion is performed on the coarse label maps of each modality to obtain a cross-modal consistent coarse label set. Based on the anatomical priors of brain region volume range, left-right symmetry and centroid anatomical location, quality control is performed on the uniform coarse label set. Topological repair is completed through isolated connected domain removal, boundary smoothing and cavity filling to obtain target individualized brain image region labels. Export the target individualized brain imaging region labels, corresponding metadata, and uncertainty assessment report.

2. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 1, characterized in that, The specific preprocessing operations include: performing artifact removal on the target brain image after skull dissection, identifying and eliminating strip-like and speckled artifacts caused by scanning equipment or subject movement, with artifact identification based on the local standard deviation of voxel intensity; and performing intensity standardization on the brain image after artifact removal, mapping the intensity distribution of the brain image to a preset normal distribution. ,in The mean, The standard deviation is used, and the mapping method is a linear transformation: , For standardized voxels The intensity value, For original strength, , The mean and standard deviation of the original strength.

3. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 1, characterized in that, The cascaded registration includes a rigid registration stage, an affine fine registration stage, and a nonlinear deformation registration stage, specifically: In the rigid registration stage, mutual information is used as the similarity metric. A multi-resolution pyramid is constructed, and the rigid body transformation parameters are iteratively optimized at each resolution level until the change in the similarity metric is less than a preset threshold. ; In the affine registration stage, based on the rigid registration results, affine transformation parameters are introduced, and multi-resolution pyramid iterative optimization is used until the change in the similarity metric is less than a preset threshold. ; The nonlinear deformation registration stage adopts spline free deformation algorithm, setting the control point grid spacing to... , Matching the spatial scale of small brain structures, and using regularized mutual information as the objective function, the inverse deformation field is constrained to satisfy: ,in The positive definiteness threshold of the deformation field is used for iterative optimization until the change in the objective function is less than the preset threshold. .

4. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 3, characterized in that, The forward deformation field With inverse deformation field The method for verifying reciprocity is as follows: randomly select individuals within the space Discrete voxel coordinates ,calculate and Spatial distance, if all distances are less than the voxel side length If 1 / 2 of the deformation field is obtained, then the deformation field is determined to satisfy the reciprocity constraint.

5. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 1, characterized in that, The specific implementation of the structural scale priority rule is as follows: All labels in the standard spatial anatomical atlas are pre-divided into three levels: small structural regions, spatial volume ≤ Subcortical region, spatial volume ∈ ), cortical region, spatial volume > ,in The volume threshold is obtained statistically based on a standard template; Overlapping voxels during backpropagation Iterate through all the labels that correspond to its mapping in the standard space. Extract the hierarchy to which each tag belongs, and select the tag corresponding to the highest priority hierarchy as... The target label; When overlapping voxels correspond to multiple labels at the same level, a local majority voting method is used for resolution, with the following formula: ,in This is the set of tags at this level. voxels The target label value, Mapped to standard space The three-dimensional coordinates of the voxels For indicator functions, Standard space voxels The original label value.

6. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 1, characterized in that, The process of establishing the registration matrix is ​​as follows: Using the preprocessed target brain image as a reference, rigid registration and affine registration are performed sequentially on other modalities of brain images. Rigid registration optimizes translation and rotation parameters using mutual information as a metric; affine registration optimizes scaling and shearing parameters based on the rigid registration result. For the registered other modalities of brain images, the normalized mutual information value with the target brain image is calculated. ,like , If the preset registration accuracy threshold is not met, registration is re-executed until the condition is satisfied; the transformation parameters obtained during the registration process are integrated into an inter-modal registration matrix. ,pass The coordinates of the initial label map of an individual space are mapped to the coordinate system of other modal brain images to generate coarse label maps for the corresponding modalities.

7. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 1, characterized in that, The configuration and weighted fusion process of the confidence weights is as follows: For each modality of brain image, calculate its label recognition accuracy against a predefined set of anatomical structures. , The calculation method is as follows: ,in Number of anatomical structures. , , The first The number of true positive, false positive, and false negative voxels in each structure; based on the discriminant value. Based on this, confidence weights are configured for each modality. ,satisfy , The number of modalities is such that the weight of structural modalities is not less than the minimum configuration confidence weight. For the corresponding voxels of each modality's coarse label map, perform weighted fusion: ,in For the first Each modality in voxels The coarse label value at that location.

8. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 1, characterized in that, The specific aspects of the quality control are as follows: For each brain region in the uniform coarse label set, calculate its volume. And check if it meets the requirements. ,in This represents the volume of the corresponding brain region in the standard spatial template. This is a preset lower limit volume ratio coefficient. This is a preset upper limit ratio coefficient for volume; if it is not met, the brain region is marked as abnormal. Calculate the volume ratio of the left and right corresponding brain regions. And check if it meets the requirements. ,in The preset symmetry threshold is used; if it is not met, the brain regions in this group are marked as abnormal. Calculate the centroid coordinates of each brain region. It also verifies whether the brain region is located within a preset neighborhood of the centroid of the corresponding brain region in the standard template, where the radius of the neighborhood is [missing information]. and Match the brain region scale; if it exceeds the scale, the brain region is marked as abnormal. For all marked abnormal brain regions, the label information of the corresponding brain regions in the standard template is used for preliminary correction before proceeding to the topology repair stage.

9. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 8, characterized in that, The specific operations of the topology repair include: For each brain region in the uniform coarse label set, extract all its connected components and remove voxels with a number less than [a certain value]. isolated connected components, A threshold is set based on brain region scale; morphological opening and closing operations are used to smooth brain region boundaries; connectivity analysis is performed on void voxels within brain regions: if a void voxel has a connected path to a brain region boundary voxel and all voxels on the path belong to that brain region, then the void voxel is filled with the label of that brain region, and the filling rule satisfies: in brain region The boundary, brain region The tag, for The original tag.

10. The brain image region label generation method based on standard spatial inverse deformation inverse mapping according to claim 1, characterized in that, The uncertainty assessment report includes the following content: Based on the Jacobian determinant calculation of the inverse deformation field, for each voxel in the individual space The calculation formula is as follows: ,in For the inverse deformation field in voxels The Jacobian determinant value at that location, Operators for calculating determinants, For the inverse deformation field in voxels The gradient matrix at the location; for marking high uncertainty regions, if or Then Marked as high uncertainty voxels, , These are preset matrix thresholds; the system obtains a spatial distribution heatmap containing high uncertainty regions, a statistical analysis of the proportion of high uncertainty voxels in each brain region, a list of anatomical locations corresponding to high uncertainty regions, and manual review suggestions for these regions.

Images