Brain image data analysis method and system combined with large model

Abstract

The application provides a brain image data analysis method and system combined with a large model. The method comprises the following steps: acquiring multi-phase brain image data and corresponding acquisition parameters of a target object, and generating individual brain region image data; extracting voxel abnormality features, brain region structure features, brain connection disturbance features, and cross-time change features, and constructing a brain region evidence field based on the voxel abnormality features, the brain region structure features, the brain connection disturbance features, and the cross-time change features; generating an evidence-based prompt context according to the brain region evidence field, inputting the evidence-based prompt context into a large model for hierarchical reasoning, and generating an initial analysis result; performing evidence chain backtracking verification and conflict constraint verification on the initial analysis result, generating a modified prompt context according to an abnormal reference identifier obtained through the verification, inputting the modified prompt context into the large model for constraint modification, and generating a target analysis result for brain image data analysis. The application can improve evidence consistency, reduce conclusion bias, and enhance analysis traceability.

Description

Technical Field

[0001] This application relates to the field of brain imaging data analysis technology, and in particular to a brain imaging data analysis method and system that combines large models. Background Technology

[0002] Brain imaging data analysis is widely used in the auxiliary assessment of cerebrovascular diseases, neurodegenerative diseases, brain tumors, and neuropsychiatric disorders. With the development of examination methods such as magnetic resonance imaging, computed tomography, diffusion-weighted imaging, and functional imaging, clinicians can obtain multi-source brain imaging data containing information on structure, function, connectivity, and temporal changes, providing a data foundation for the identification of brain region abnormalities, analysis of lesion changes, and follow-up assessment.

[0003] Existing technologies typically first denoise, register, segment, and extract features from brain images, then use deep learning models to complete abnormal region detection, brain region classification, or image report generation. Some solutions further introduce large models, using the structured results or text descriptions output by the image model as prompts to generate brain image analysis conclusions or auxiliary reports, thereby improving the completeness of the analysis results.

[0004] However, existing methods still have shortcomings: brain images from different modalities and examination time points are difficult to form a unified evidentiary representation within the individual brain region space; there is a lack of traceable evidence chains between image features, brain region relationships, abnormal regions, and temporal changes; large models easily deviate from the original image evidence when generating conclusions, making it difficult to identify problems such as missing evidence citations, cross-modal conflicts, and inconsistent directions of change; at the same time, existing methods mostly rely on single model outputs or static threshold judgments, making it difficult to combine individual brain region structure and multi-temporal evolution information for constraint analysis. Therefore, there is an urgent need for a brain image data analysis method that can combine brain image features, large model reasoning, and evidence backtracking verification. Summary of the Invention

[0005] In view of this, embodiments of this application provide a brain imaging data analysis method and system that combines large models to solve the problems of inconsistent expression of multi-temporal brain imaging evidence, lack of evidentiary constraints on conclusions, and difficulty in retrospectively verifying conflicts in the prior art.

[0006] A first aspect of this application provides a brain imaging data analysis method combining a large model, comprising: acquiring multi-temporal brain imaging data of a target object and corresponding acquisition parameters; performing quality screening, deformation registration, intensity domain correction, and individual brain region segmentation on the multi-temporal brain imaging data; generating individual brain region imaging data bound to spatial coordinates, time markers, and quality markers; inputting the individual brain region imaging data into a three-dimensional image coding network and a brain region topology coding network; extracting voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and cross-temporal change representations; and based on voxel abnormality representations, brain region structural representations, and brain connectivity perturbation representations... A brain region evidence field is constructed based on trans-temporal change representations. Evidence-based contextualization, including spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains, is generated from this brain region evidence field. This contextualization is then input into a large model for hierarchical reasoning, generating initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships. Evidence chain backtracking and conflict constraint verification are performed on the initial analysis results. Based on the abnormal citation identifiers obtained from the verification, a correction contextualization is generated and input into the large model for constraint correction, generating target analysis results for brain imaging data analysis.

[0007] A second aspect of this application provides a brain imaging data analysis system combining a large model, comprising: an acquisition module for acquiring multi-temporal brain imaging data and corresponding acquisition parameters of a target object, performing quality screening, deformation registration, intensity domain correction, and individual brain region segmentation on the multi-temporal brain imaging data, and generating individual brain region imaging data bound to spatial coordinates, time markers, and quality markers; and an extraction module for inputting the individual brain region imaging data into a three-dimensional image coding network and a brain region topology coding network, extracting voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and cross-temporal change representations, and based on the voxel abnormality representations, brain region structural representations, and brain connectivity perturbation representations... The brain region evidence field is constructed by characterizing features and cross-temporal changes. The reasoning module generates an evidential cue context containing spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains based on the brain region evidence field. The evidential cue context is input into the large model for hierarchical reasoning, generating initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships. The generation module performs evidence chain backtracking verification and conflict constraint verification on the initial analysis results. Based on the abnormal citation identifiers obtained from the verification, a correction cue context is generated. The correction cue context is input into the large model for constraint correction, generating target analysis results for brain imaging data analysis.

[0008] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring multi-temporal brain imaging data and corresponding acquisition parameters of the target object, the multi-temporal brain imaging data is subjected to quality screening, deformation registration, intensity domain correction, and individual brain region segmentation to generate individual brain region imaging data bound with spatial coordinates, temporal markers, and quality markers. The individual brain region imaging data is input into a three-dimensional image coding network and a brain region topology coding network to extract voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and cross-temporal change representations. Based on the voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and cross-temporal change representations, a brain region evidence field is constructed. Based on the brain region evidence field, an evidentiary hint context containing spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains is generated. The evidentiary hint context is input into a large model for hierarchical reasoning to generate initial analysis results containing brain region analysis conclusions, abnormal region conclusions, change trend conclusions, and evidence citation relationships. The initial analysis results are subjected to evidence chain backtracking verification and conflict constraint verification. Based on the abnormal citation markers obtained from the verification, a correction hint context is generated and input into the large model for constraint correction to generate target analysis results for brain imaging data analysis. This application can improve the consistency of evidence, reduce the bias of conclusions, and enhance the traceability of analysis. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the brain imaging data analysis method combining a large model provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the brain imaging data analysis system combined with a large model provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] Brain imaging data analysis technology is widely used in the auxiliary assessment of cerebrovascular diseases, neurodegenerative diseases, brain tumors, and neuropsychiatric disorders. Current technologies typically perform structured processing of brain imaging data through image preprocessing, spatial registration, brain tissue segmentation, abnormal region detection, and feature extraction. Deep learning models are then used to identify lesions, classify brain regions, or generate image reports. With the development of large-scale modeling technology, some solutions are beginning to input the structured results or text descriptions output by image models into large-scale models, which then generate brain imaging analysis conclusions or auxiliary reports.

[0013] However, existing technologies still have significant shortcomings. On the one hand, brain imaging data from different examination times, scanning protocols, and modalities are difficult to form a unified evidentiary expression within the individual brain region space, resulting in a lack of stable correlations between voxel abnormalities, brain region structures, brain connectivity perturbations, and temporal changes. On the other hand, existing large-scale model analysis methods largely rely on textual reasoning based on the output results of image models, lacking explicit modeling of brain imaging evidence chains, cross-temporal evolutionary relationships, and conflict constraint relationships. This easily leads to problems such as missing conclusion citations, deviations in evidence direction, inconsistent changes, and the neglect of conflicting information. Therefore, existing technologies struggle to achieve reliable correspondence and retrospective verification between brain imaging analysis conclusions and original imaging evidence.

[0014] To address the aforementioned issues, this application acquires multi-temporal brain imaging data and corresponding acquisition parameters of the target subject. It then performs quality screening, deformation registration, intensity domain correction, and individual brain region segmentation on the multi-temporal brain imaging data, generating individual brain region imaging data bound to spatial coordinates, temporal markers, and quality markers. Furthermore, the individual brain region imaging data is input into a three-dimensional image coding network and a brain region topological coding network to extract voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and cross-temporal change representations. Based on these representations, a brain region evidence field is constructed, enabling the spatial abnormalities, brain region relationships, connectivity perturbations, and temporal evolution in brain imaging to be expressed in a unified evidence field.

[0015] Based on this, this application generates an evidential cue context comprising spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains according to the brain region evidence field. This evidential cue context is then input into a large model for hierarchical reasoning, generating initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships. Subsequently, the initial analysis results undergo evidence chain backtracking verification and conflict constraint verification. Based on the abnormal citation identifiers obtained from the verification, a corrected cue context is generated, and this corrected cue context is input into the large model for constraint correction, generating the target analysis results for brain imaging data analysis.

[0016] Through the above technical solutions, this application can convert multi-temporal brain imaging data into evidence-based data representations of individual brain regions, and subject the large-scale model reasoning process to spatial evidence chains, temporal evolution chains, and conflict constraint chains, thereby improving the consistency between brain imaging analysis results and imaging evidence. Simultaneously, this application can perform reverse evidence verification and constraint correction on the analytical conclusions generated by the large-scale model, reducing the probability of conclusions without evidence, conclusions deviating from evidence, and conflicting or omitted conclusions, thus enhancing the traceability and credibility of brain imaging data analysis results.

[0017] Before providing a detailed description of the specific embodiments of this application, it should be noted that the brain imaging data in the embodiments of this application includes all imaging data related to brain examination in the prior art, such as, but not limited to: electroencephalography (EEG), magnetoencephalography (MEG), brain CT, brain magnetic resonance imaging (MRI), transcranial Doppler ultrasound, etc.

[0018] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0019] Figure 1 This is a flowchart illustrating the brain imaging data analysis method combining a large model provided in an embodiment of this application. Figure 1 As shown, the method may specifically include: S101: Acquire multi-temporal brain imaging data and corresponding acquisition parameters of the target object; perform quality screening, deformation registration, intensity domain correction and individual brain region segmentation on the multi-temporal brain imaging data; and generate individual brain region imaging data bound with spatial coordinates, time labels and quality labels. S102, input individual brain region image data into a three-dimensional image coding network and a brain region topological coding network, extract voxel abnormality representation, brain region structural representation, brain connectivity perturbation representation and cross-temporal change representation, and construct a brain region evidence field based on voxel abnormality representation, brain region structural representation, brain connectivity perturbation representation and cross-temporal change representation; S103, Generate an evidentiary cue context containing spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains based on the brain region evidence field. Input the evidentiary cue context into the large model for hierarchical reasoning to generate initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships. S104, perform evidence chain backtracking verification and conflict constraint verification on the initial analysis results, generate correction prompt context based on the abnormal reference identifiers obtained from the verification, and input the correction prompt context into the large model for constraint correction to generate target analysis results for brain imaging data analysis.

[0020] In some embodiments, acquiring multi-temporal brain imaging data of the target object and corresponding acquisition parameters includes: Based on the examination identifiers of the target subject, retrieve the original brain images and image head information across examination time points from the brain imaging archiving system; The original brain images were subjected to object consistency verification and examination time point merging, and the acquisition parameters were extracted based on the image head information and bound to each examination time point. Based on the acquired parameters, protocol fingerprint matching is performed on the original brain images to determine the target image sequence under the same analysis task, and multi-temporal brain image data is generated according to the examination time point.

[0021] Specifically, when acquiring multi-temporal brain imaging data and corresponding acquisition parameters of a target subject, the brain imaging analysis platform first receives the examination identifier corresponding to the target subject. The examination identifier can be formed by combining the patient's visit number, examination serial number, image archive index, and analysis task number, and is used to locate brain imaging records of the same target subject at different examination time points in the brain imaging archive system. Based on the examination identifier, the system initiates a retrieval request, retrieving the original brain images of the target subject at the baseline examination time point, follow-up examination time point, and re-examination examination time point from the brain imaging archive system, while simultaneously reading the image head information bound to each original brain image. The image head information includes data describing the image acquisition conditions, such as scan location, acquisition time, sequence name, slice thickness, slice spacing, pixel spacing, matrix size, echo time, repetition time, flip angle, magnetic field strength, acquisition direction, and device identifier.

[0022] After retrieving the original brain images, the system performs object consistency verification. Object consistency verification does not simply determine if the examination identifiers are the same; rather, it jointly compares the object code, birth date hash value, gender identifier, examination institution identifier, and examination time range in the image header information, and combines this with the coarse registration results of brain structures to determine whether different examination records belong to the same target object. For example, if the system retrieves MRI brain images of the target object at three examination time points, and if the object code of one group of images is consistent but the coarse registration similarity of brain structures is below a preset threshold, the system marks this group of images as object data awaiting confirmation and does not directly include it in subsequent multi-temporal analysis. After passing the object consistency verification, the system merges the original brain images according to the acquisition time and examination task type, merging image records formed during the same examination due to rescanning, supplementary scanning, or segmented acquisition to the same examination time point, while retaining the corresponding acquisition order and quality markers.

[0023] Furthermore, the system extracts acquisition parameters bound to each examination time point based on image head information and encapsulates these parameters in a structured manner. For each examination time point, the system generates an acquisition parameter record by combining the sequence name, imaging orientation, spatial resolution, slice thickness, scan matrix, key temporal parameters, device parameters, and reconstruction parameters, and establishes a binding relationship between the acquisition parameter record and the corresponding original brain image. When the same examination time point contains multiple candidate image sequences, the system generates a protocol fingerprint that characterizes the scanning protocol based on the similarity and completeness between the acquisition parameters. The protocol fingerprint can be jointly determined by sequence category, spatial resolution level, slice thickness range, acquisition orientation, temporal parameter range, and device reconstruction method, and is used to identify whether different examination time points are comparable.

[0024] During protocol fingerprint matching, the system first determines the target protocol pattern corresponding to the current analysis task, and then matches the candidate image sequences at each examination time point with the target protocol pattern. If the target analysis task is to observe changes in brain region structure, the system prioritizes image sequences whose spatial resolution, slice thickness, and imaging orientation meet the requirements of structural analysis; if the target analysis task is to analyze changes in brain connectivity, the system selects functionally or diffusion-related image sequences with consistent acquisition orientation and time parameter range. The system generates a sequence availability identifier for successfully matched candidate image sequences, a deweighting identifier for candidate image sequences with significant protocol differences but still of reference value, and an exclusion identifier for candidate image sequences with missing acquisition parameters, incomplete slices, or mismatched protocol fingerprints.

[0025] For example, if a target subject undergoes brain examinations in March, June, and December 2025, the brain imaging archive system contains multiple sets of raw brain images for each examination time point. After retrieving all images using the examination identifier, the system first excludes images that fail the subject consistency check, and then merges rescan sequences from the same examination time point. Subsequently, the system generates a protocol fingerprint based on the image header information, filters out target image sequences with the same spatial orientation and similar resolution from the three examination time points, and arranges them in chronological order according to March, June, and December 2025 to generate multi-temporal brain imaging data. Each temporal brain imaging data carries corresponding acquisition parameters, examination time point, protocol fingerprint, sequence availability identifier, and data source index for subsequent quality screening, deformation registration, intensity domain correction, and individual brain region segmentation.

[0026] Through the processing method of this embodiment, the system can perform object-level verification, time-point merging, and protocol-level filtering on brain imaging data across examination time points before entering image encoding and evidence field construction. This reduces data deviations caused by the mixing of images of different objects, interference from repeated examination records, and differences in scanning protocols, providing a consistent and traceable input data foundation for subsequent multi-temporal brain imaging analysis.

[0027] In some embodiments, multi-temporal brain imaging data undergoes quality screening, deformation registration, intensity domain correction, and individual brain region segmentation to generate individual brain region imaging data bound to spatial coordinates, temporal markers, and quality markers, including: Based on the collected parameters and protocol fingerprints, artifact identification, layer integrity verification and signal availability evaluation are performed on multi-temporal brain imaging data to generate quality labels for each examination time point. Using the baseline temporal images of the target object as the registration base, and combining brain structure boundary constraints and cross-temporal deformation field estimation, the brain images at each examination time point are mapped to a unified individual space. Intensity domain distribution correction and individual brain region segmentation are performed based on a unified individual space. The corrected image voxels are associated with the corresponding brain regions, spatial coordinates, time markers and quality markers to generate individual brain region image data.

[0028] Specifically, after generating multi-temporal brain imaging data, the brain imaging analysis platform first reads the acquisition parameters and protocol fingerprints bound to each examination time point, and uses the protocol fingerprints as a reference benchmark for quality screening. The system performs consistency checks on the multi-temporal brain imaging data based on slice thickness, interslice spacing, pixel spacing, acquisition direction, sequence time parameters, and reconstruction parameters, identifying abnormal data caused by changes in scanning direction, missing slices, differences in reconstruction matrices, or protocol offsets. For each temporal image, the system further performs artifact identification, judging the presence of motion artifacts, metallic artifacts, or reconstruction artifacts by detecting abnormal head movement textures, edge ghosting, local signal abrupt changes, and extra-brain high-brightness interference. Simultaneously, it verifies slice integrity according to slice number, spatial location, and interslice spacing, and generates a signal usability evaluation based on the proportion of effective brain tissue signal, gray-white matter contrast, and background noise level. The system integrates these evaluation results into a quality identifier, which is bound to the examination time point, image sequence, and protocol fingerprint for weight control in subsequent registration and feature extraction processes.

[0029] After quality screening, the system determines the baseline temporal image from multi-temporal brain imaging data. The baseline temporal image can be selected from the examination time point with the highest quality label, complete slices, and spatial resolution meeting the analysis requirements, or the initial examination image can be used as the individual spatial basis for the target object. The system first performs coarse segmentation of brain tissue on the baseline temporal image, extracting the brain's outer contour, ventricular boundaries, cortical surface, and approximate regions of deep nuclei to form brain structural boundary constraints. Subsequently, using the baseline temporal image as the registration base, the system performs rigid registration, affine correction, and nonlinear deformation registration on brain images from other examination time points. During nonlinear deformation registration, brain structural boundary constraints are introduced to ensure that the boundaries of the ventricles, cortex, and major brain regions maintain structural correspondence after cross-temporal mapping. Based on the deformation parameters of each examination time point relative to the baseline temporal image, the system generates a cross-temporal deformation field and maps the brain images from each examination time point to a unified individual space.

[0030] Furthermore, the system performs intensity domain distribution correction within a unified individual space. Due to potential differences in equipment gain, coil sensitivity, and reconstruction parameters at different examination time points, the system establishes inter-phase intensity mapping relationships based on the intensity distribution characteristics within brain tissue regions, and performs downweighting correction on low-confidence regions in conjunction with quality labels. For image signals of the same brain region at different examination time points, the system corrects grayscale distribution, local contrast, and signal bias to the same intensity domain, ensuring comparability of subsequent voxel abnormality representations and cross-phase change representations. After completing intensity domain distribution correction, the system performs individual brain region partitioning on the baseline temporal image and propagates the partitioning results to each examination time point through a cross-phase deformation field, ensuring that each phase image voxel has a corresponding brain region label.

[0031] For example, the target subject generated three sets of brain imaging data in March 2025, June 2025, and December 2025. Based on protocol fingerprinting, the system detected mild head movement artifacts in the June 2025 image, but the slice was intact and the signal availability met the analysis requirements; therefore, a quality label lower than the baseline image was generated for this phase. The December 2025 image had the same spatial resolution as the baseline image, and was therefore used for cross-phase variation analysis. Using the March 2025 image as the baseline image, the system established individual spatial and brain structure boundary constraints, mapped the June 2025 and December 2025 images to the same space using cross-phase deformation fields, and performed distribution correction on the signal intensity of the three phases. Subsequently, the system associated each corrected image voxel with brain region labels, three-dimensional spatial coordinates, examination time markers, and quality labels to generate individual brain region imaging data.

[0032] Through the processing method of this embodiment, multi-temporal brain imaging data can complete quality grading, spatial unification, intensity normalization, and brain region association before entering the three-dimensional image coding network and brain region topology coding network, thereby improving the comparability of cross-temporal image data and reducing the interference caused by artifacts, protocol differences, and registration bias on the subsequent construction of brain region evidence fields.

[0033] In some embodiments, individual brain region imaging data are input into a three-dimensional image coding network and a brain region topological coding network to extract voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and trans-temporal change representations, including: Based on spatial coordinates, temporal markers, and quality markers, reliable voxel sampling and brain region anchor point construction are performed on individual brain region imaging data to generate brain region constrained image sequences. Brain region constraint image sequences are input into a three-dimensional image coding network. Constraints and brain region boundary consistency constraints are reconstructed using cross-temporal masks, and voxel abnormality representations and brain region structural representations are extracted. Brain region topology input is constructed based on brain region structural representation, and then input into brain region topology coding network. Brain connectivity perturbation representation is extracted based on brain region node state and brain region connectivity. By integrating voxel abnormality representation, brain region structural representation, and brain connectivity perturbation representation, a transtemporal change representation is generated.

[0034] Specifically, after obtaining individual brain region imaging data, the brain imaging analysis platform first reads the spatial coordinates, temporal markers, and quality markers associated with each image voxel, and establishes a mapping relationship between voxels and brain regions based on the brain region segmentation results. The system uses the quality marker as the sampling weight to perform reliable voxel sampling at different examination time points. Voxels corresponding to artifact regions, low-signal regions, and unstable registration regions are assigned low sampling weights, while voxels with clear brain region boundaries, continuous signals, and stable positions across time phases are assigned high sampling weights. Subsequently, the system determines brain region anchor points within each brain region based on the spatial density, local signal distribution, and cross-time phase stability of the reliable voxels. These anchor points represent the local structural center and feature convergence location of the corresponding brain region in a unified individual space. Using these anchor points as indexes, the system organizes reliable voxel blocks of the same brain region at different examination time points in chronological order, generating a brain region-constrained image sequence.

[0035] After generating brain region constraint image sequences, the system inputs these sequences into a 3D image coding network. The 3D image coding network uses 3D voxel blocks as input units and simultaneously introduces cross-temporal mask reconstruction constraints and brain region boundary consistency constraints during the encoding process. Cross-temporal mask reconstruction constraints refer to masking certain temporal phases, brain regions, or voxel blocks during training or inference adaptation, allowing the network to reconstruct the masked regions based on the image context of adjacent temporal phases and brain regions, thereby obtaining voxel-based abnormality representations that reflect the degree of abnormal deviation. Brain region boundary consistency constraints refer to preserving brain region boundaries, sulcus and gyri edges, and ventricular boundaries during feature encoding, avoiding excessive mixing of features from different brain regions during deep encoding, thereby extracting brain region structural representations that characterize brain region morphology, signal distribution, and local structural states.

[0036] Furthermore, the system constructs brain region topological input based on brain region structural representations. The brain region topological input uses the brain region corresponding to the anchor point as a node, and uses the spatial adjacency, structural pathway relationships, and cross-temporal signal coordination relationships between brain regions as the basis for node connections. Brain region structural representations, quality indicators, and temporal indicators are used as node states. The brain region topological coding network performs message passing and relationship updates on the brain region topological input. During the encoding process, it determines local perturbations in brain regions based on changes in node states and determines inter-brain region coordination anomalies based on changes in node connection relationships, thereby extracting brain connectivity perturbation representations. Brain connectivity perturbation representations do not merely represent the abnormal intensity of a single brain region, but rather represent connectivity changes between brain regions at the levels of spatial adjacency, structural association, or temporal coordination.

[0037] For example, the target subject has individual brain region imaging data corresponding to March 2025, June 2025, and December 2025. The system constructs brain region anchors in the hippocampus, medial frontal lobe, and adjacent areas of the lateral ventricle, and organizes reliable voxel blocks corresponding to the same brain region anchors at the three examination time points into a brain region-constrained image sequence. A 3D image coding network performs cross-temporal mask reconstruction on some voxel blocks from June 2025 and compares them with corresponding voxel blocks from March and December 2025, extracting voxel anomaly representations and brain region structural representations formed by local signal enhancement, structural volume changes, and boundary morphology alterations. The brain region topology coding network further analyzes the node state changes between the hippocampus and adjacent brain regions, extracting brain connectivity perturbation representations reflecting coordinated changes in brain region connectivity. Finally, the system integrates voxel anomaly representations, brain region structural representations, and brain connectivity perturbation representations, combining them with the examination time point sequence to generate cross-temporal change representations.

[0038] Through the processing method of this embodiment, individual brain region imaging data can be transformed from voxel-level imaging information into brain region-level, connectivity-level, and temporal evolution-level feature expressions, enabling abnormal regions, brain region structures, and changes in brain connectivity to be expressed in a unified coding framework, thereby improving the feature completeness and temporal consistency of subsequent brain region evidence field construction.

[0039] In some embodiments, a brain region evidence field is constructed based on voxel anomaly representation, brain region structural representation, brain connectivity perturbation representation, and trans-temporal change representation, including: Using brain region anchors as the aggregation benchmark, voxel abnormalities are mapped to corresponding brain regions, and evidence credibility weights are generated based on quality labels. The brain region structural representations, brain connectivity perturbation representations, and transtemporal change representations are linked according to spatial adjacency, connectivity propagation, and temporal evolution to form multi-scale evidence units; Based on the evidence credibility weight, uncertainty propagation and conflict constraint labeling are performed on multi-scale evidence units to generate a brain region evidence field that includes evidence strength, evidence source and evolution direction.

[0040] Specifically, after obtaining voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and trans-temporal change representations, the brain imaging analysis platform uses brain region anchors as a unified aggregation benchmark to integrate image features from different spatial locations, different examination times, and different quality levels within the same brain region. The system first determines the brain region attribution corresponding to the voxel abnormality representation based on the spatial distance between the brain region anchor and the credible voxel, local signal similarity, and boundary attribution relationships, and then maps the voxel abnormality representation to the abnormality evidence slot within the corresponding brain region. For abnormal voxels that cross brain region boundaries, the system combines brain region boundary consistency constraints and the response intensity of adjacent brain region anchors to determine the primary and associated brain regions, avoiding the misclassification of single voxel abnormalities into adjacent brain regions.

[0041] During the mapping process, the system generates evidence credibility weights based on quality labels. Examination time points and image regions with higher quality labels correspond to higher evidence credibility weights, while examination time points with mild artifacts, unstable registration, or insufficient signal availability correspond to lower evidence credibility weights. Evidence credibility weights, along with voxel abnormality intensity, brain region structural stability, and cross-temporal consistency, participate in the calculation of evidence strength. For example, if the target subject has images from three examination time points in March 2025, June 2025, and December 2025, and the June 2025 image shows mild head movement artifacts, the system converts the quality label of this time point into a lower evidence credibility weight, reducing the impact of the local abnormality intensity at that time point on the overall evidence strength when aggregating hippocampal abnormality evidence.

[0042] Furthermore, the system correlates brain region structural representations, brain connectivity perturbation representations, and trans-temporal change representations according to spatial adjacency, connectivity propagation, and temporal evolution relationships to form multi-scale evidence units. Spatial adjacency relationships describe the spatial correspondence between the same anomalous region and adjacent brain regions, ventricular boundaries, or cortical boundaries; connectivity propagation relationships describe the association changes between brain regions caused by structural pathways, functional synergy, or signal perturbations; and temporal evolution relationships describe the direction and magnitude of change of the same brain region or the same anomalous region across multiple examination time points. Each multi-scale evidence unit is bound to a brain region anchor point, spatial coordinate range, examination time point, evidence source, evidence credibility weight, and feature summary, enabling subsequent large-scale models to perform hierarchical reasoning according to evidence source and evolutionary direction.

[0043] After forming multi-scale evidence units, the system performs uncertainty propagation based on evidence credibility weights. If the anomalous evidence for a certain brain region mainly originates from low-quality time points, the system propagates the uncertainty of the corresponding brain region evidence unit to adjacent spatial and temporal evolution evidence, reducing the support strength of conclusions formed by a single low-credibility evidence. If the brain region structural representation shows abnormal enhancement, but the brain connectivity perturbation representation does not show corresponding changes, or the direction of change of cross-temporal change representation is inconsistent with the direction of change of voxel abnormal representation, the system generates conflict constraint markers and records the brain region, time point, evidence source, and conflict type where the conflict occurs. Subsequently, the system writes the evidence strength, evidence source, evolution direction, uncertainty level, and conflict constraint markers into the same brain region evidence field.

[0044] Through the processing method of this embodiment, the system can integrate scattered voxel-level abnormalities, brain region-level structures, connectivity-level perturbations, and temporal-level changes into a unified brain region evidence field. This enables subsequent evidence-based contextualization to obtain clear evidence sources, credibility, and evolutionary relationships, reducing interference from low-quality image evidence or cross-feature conflicts on the analysis results of large models, and improving the consistency and traceability of brain image analysis results.

[0045] In some embodiments, generating an evidentiary cue context comprising a spatial evidence chain, a modal evidence chain, a temporal evolution chain, and a conflict constraint chain based on a brain region evidence field includes: Evidence role analysis is performed on evidence units in the brain region evidence field to determine the brain region location, image source, temporal state and conflict attributes corresponding to each evidence unit; Based on the spatial adjacency of brain regions, cross-modal support relationships, cross-temporal evolution relationships, and conflict constraint relationships, the evidence units are chained together to generate a multi-dimensional evidence chain; Based on multidimensional evidence chains and evidence credibility weights, an evidence-based hinting context is generated that has a limited scope of reasoning, evidence citation rules, and conflict retention rules.

[0046] Specifically, after generating the brain region evidence field, the brain imaging analysis platform performs evidence role analysis on each evidence unit based on its strength, source, evolutionary direction, uncertainty level, and conflict constraint markers. Evidence role analysis determines the boundary of the evidence unit's role in subsequent large-scale model reasoning. The system first reads the brain region anchor points and spatial coordinate range associated with the evidence unit to determine its corresponding brain region location; then it reads the image sequence, acquisition protocol fingerprint, and inspection time point from which the evidence unit originates to determine the image source and temporal state; simultaneously, it reads the conflict constraint markers to determine whether the evidence unit belongs to primary evidence, auxiliary evidence, altered evidence, conflicting evidence, or low-confidence evidence.

[0047] During the evidence role analysis process, the system does not directly convert the brain region evidence field into ordinary text. Instead, it forms evidence role descriptions based on the binding relationships between evidence units and brain region structures, image sources, examination time points, and conflict markers. For example, if the target object has hippocampal-related evidence units in March 2025, June 2025, and December 2025, with enhanced voxel abnormalities in December 2025, a decrease in local volume in brain region structure representation, and a weakening of connectivity perturbation representation showing weakened connectivity with adjacent memory-related brain regions, the system will analyze these evidence units as spatial abnormality evidence, structural change evidence, connectivity perturbation evidence, and temporal evolution evidence, respectively. The abnormal evidence formed by low-quality images in June 2025 will be marked as low-confidence auxiliary evidence.

[0048] Furthermore, the system chains evidence units according to spatial adjacency relationships, cross-modal support relationships, cross-temporal evolution relationships, and conflict constraint relationships. The spatial evidence chain starts with the brain region anchor point and spatial coordinate range, connecting evidence units of abnormal voxels, brain region boundaries, and adjacent brain regions in spatial order of attribution; the modal evidence chain uses image source and protocol fingerprint as indexes, connecting evidence units from the same examination time point but different image sources according to support relationships; the temporal evolution chain uses the examination time point as the order, connecting evidence units of the same brain region or the same abnormal region in multiple temporal phases according to the direction of change; the conflict constraint chain separately arranges evidence units with inconsistent evidence directions, low credibility of evidence sources, or insufficient modal support relationships, and retains the conflict reasons and attributes to be reviewed.

[0049] After generating multidimensional evidence chains, the system prunes and prioritizes these chains based on their credibility weights. For evidence chains with high credibility weights and directly relevant to the analysis task, the system designates them as core evidence that the large model must cite. For evidence chains with lower credibility weights but related to trends, the system designates them as auxiliary evidence. For evidence chains with conflict constraint markers, the system designates them as conflicting evidence that cannot be ignored. The system generates an evidence-based prompting context based on these chain attributes. This context includes limiting the scope of reasoning, evidence citation rules, and conflict retention rules. Limiting the scope of reasoning constrains the large model to analyze only the generated brain region evidence field. Evidence citation rules require analysis conclusions to be bound to corresponding evidence units. Conflict retention rules require the large model to retain uncertainties when evidence is insufficient or contradictory.

[0050] For example, in the hippocampal follow-up analysis task, the system uses the abnormal voxel enhancement, local structural changes, and connectivity perturbations in the hippocampus in December 2025 as the core evidence chain, the abnormal indications from low-quality images in June 2025 as auxiliary evidence chains, and the situation where structural changes and some modalities are not synchronously supported as conflict constraint chains written into the evidence context. After receiving this evidence context, the large model can only perform hierarchical reasoning around the hippocampus and spatially adjacent brain regions, corresponding image sources, cross-temporal change directions, and conflict constraint relationships, and synchronously cites evidence units when generating analysis conclusions.

[0051] Through the processing method of this embodiment, the brain region evidence field can be transformed into an evidentiary cue context with reasoning boundaries, citation rules, and conflict retention requirements. This makes the large model analysis process subject to the joint constraints of spatial evidence, modal evidence, temporal evidence, and conflicting evidence, reducing the probability of generating unfounded conclusions and conflicting omissions, and improving the evidentiary consistency and traceability of brain imaging analysis results.

[0052] In some embodiments, the evidential cue context is input into the large model for hierarchical reasoning, generating initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships, including: Input the evidence-based context into the large model, and perform evidence node retrieval and reasoning task decomposition on the multi-dimensional evidence chain based on the limited reasoning scope. According to the evidence citation rules, the large model is driven to generate corresponding candidate conclusions according to the brain region layer, abnormal region layer and temporal evolution layer, and a citation mapping between candidate conclusions and evidence units is established simultaneously. Based on the conflict retention rule, candidate conclusions are constrained for consistency and marked with uncertainty to generate initial analysis results that include brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships.

[0053] Specifically, after forming the evidential cue context, before inputting it into the large model, the brain imaging analysis platform first structurally encapsulates the limited inference scope, evidence citation rules, and conflict retention rules within the context. This enables the large model to identify the scope of evidence, inference level, and output constraints allowed for the current analysis task. The limited inference scope can be jointly determined by the target brain region, adjacent brain regions, target examination time point, related image sources, and multidimensional evidence chains. This is used to prevent the large model from using information outside the brain region's evidence field to generate unfounded analysis content. The system converts the evidence units in the multidimensional evidence chain into searchable evidence nodes. Each evidence node carries a brain region anchor point, spatial coordinate range, examination time point, evidence strength, evidence source, evolutionary direction, and conflict attributes.

[0054] After receiving the evidence-based context, the large model first performs evidence node retrieval based on a limited reasoning scope, and then decomposes the retrieved evidence nodes into reasoning tasks according to the analysis task. The system breaks down the overall brain imaging analysis task into brain region-level reasoning, abnormal region-level reasoning, and temporal evolution-level reasoning. Brain region-level reasoning is used to determine the evidence status of the target brain region and adjacent brain regions based on brain region structural representations and brain connectivity perturbation representations; abnormal region-level reasoning is used to determine the location, extent, and local evidence strength of abnormal regions based on voxel abnormal representations and spatial evidence chains; and temporal evolution-level reasoning is used to determine the direction and magnitude of change of the same brain region or abnormal region between different examination time points based on the temporal evolution chain. Each level of reasoning uses evidence nodes as input boundaries and outputs candidate conclusions for the corresponding level.

[0055] When generating candidate conclusions, the system drives the large model to synchronously establish a reference mapping between candidate conclusions and evidence units according to evidence citation rules. For each brain region analysis conclusion, the large model needs to associate brain region structural evidence, connection perturbation evidence, or corresponding spatial evidence nodes; for each abnormal region conclusion, the large model needs to associate abnormal voxel evidence, spatial coordinate range, and image source; for each trend conclusion, the large model needs to associate temporal evolution evidence between different examination time points. If a candidate conclusion cannot be matched with an evidence unit that meets the credibility weight requirement, the system marks the candidate conclusion as a low-evidence-support conclusion and restricts the candidate conclusion from entering the deterministic analysis results.

[0056] For example, the target subject had hippocampal follow-up images in March, June, and December 2025. Evidence-based contextualization limited the large model to analyze only the hippocampus, parahippocampal gyrus, and related connectivity regions. The large model first retrieved the spatial evidence chain of the hippocampus, identifying that the intensity of local abnormal voxels in the hippocampus in December 2025 was higher than the previous two examination time points. It then further retrieved the temporal evolution chain, determining that the direction of local abnormality change was gradually increasing. Subsequently, the large model generated candidate conclusions on the structural state of the hippocampus based on brain region layer reasoning, candidate conclusions on medial hippocampal abnormal regions based on abnormal region layer reasoning, and candidate conclusions on cross-temporal change trends based on temporal evolution layer reasoning, and then bound these candidate conclusions to their corresponding evidence units.

[0057] Furthermore, the system applies consistency constraints and uncertainty markers to candidate conclusions based on conflict retention rules. If both the spatial evidence chain and the temporal evolution chain support the same direction of change, but some image sources in the modal evidence chain lack synchronous support, the system will not directly delete conflicting evidence. Instead, it will add uncertainty markers to the corresponding candidate conclusions and retain the citation relationships of the conflicting evidence units. If there is a spatial orientation deviation between the brain region layer conclusions and the abnormal region layer conclusions, the system will rank the candidate conclusions according to the strength of evidence and the credibility weight of evidence, and mark the candidate conclusions with larger deviations as requiring backtracking verification. After completing the consistency constraints, the system generates initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships.

[0058] Through the processing method of this embodiment, the reasoning process of the large model can be confined within the evidence boundary formed by the brain region evidence field, and evidence citation relationships can be established simultaneously when generating conclusions related to brain regions, abnormal regions, and changing trends. This reduces conclusions without evidence, conclusions with mismatched evidence, and conclusions with conflicting omissions, and improves the evidence consistency, hierarchical integrity, and subsequent verifiability of the initial analysis results.

[0059] In some embodiments, the initial analysis results are subjected to evidence chain backtracking verification and conflict constraint verification, and a correction prompt context is generated based on the abnormal reference identifier obtained from the verification, including: The analysis results are analyzed to identify the conclusion objects, abnormality types, direction of change, and evidence citation relationships in the initial analysis results, and the analysis results are back-matched to the corresponding evidence chains in the brain region evidence field. Based on the backtracking matching results, abnormal citation markers are generated for cases where conclusions are missing, evidence points are deviating, evolution directions are inconsistent, and conflict constraints are not preserved. The abnormal reference identifier, the corresponding evidence chain, and the conflict constraint relationship are associated and encapsulated to generate a correction prompt context for constraint correction.

[0060] Specifically, after the large model generates initial analysis results, the brain imaging analysis platform performs structured analysis on these results, first identifying the conclusion object, abnormality type, direction of change, and evidence citation relationship corresponding to each analysis conclusion. The conclusion object can be a specific brain region, an abnormal region within a brain region, a cross-brain region connectivity relationship, or an object changing across time phases; the abnormality type can be formed by local signal abnormalities, structural morphological changes, connectivity perturbations, or shifts in evolutionary trends; the direction of change can characterize the enhancement, weakening, expansion, shrinkage, or stable state of the same conclusion object at different examination time points. During the analysis process, the system simultaneously reads the evidence unit number, evidence chain type, evidence credibility weight, and conflict attributes cited for each conclusion, forming a conclusion analysis record.

[0061] After generating the conclusion analysis record, the system backtracks and matches it to the corresponding evidence chain in the brain region evidence field. This backtracking matching is not merely checking the existence of evidence numbers; rather, it involves joint matching based on the spatial relationship between the conclusion object and the brain region anchor point, the characteristic relationship between the abnormality type and the evidence strength, the directional relationship between the direction of change and the temporal evolution chain, and the binding relationship between evidence citation relationships and the multidimensional evidence chain. If a conclusion points to an abnormal region in the medial hippocampus, the system needs to find a spatial evidence chain in the brain region evidence field that matches the hippocampal anchor point, the corresponding spatial coordinate range, the target examination time point, and the abnormal evidence slot. If a conclusion indicates that the abnormal region has expanded compared to the previous examination time point, the system also needs to backtrack the temporal evolution chain to confirm whether the voxel abnormality range and the direction of change in evidence strength of the corresponding abnormal region are consistent between adjacent examination time points.

[0062] Furthermore, the system generates anomaly citation markers based on the backtracking matching results. If the analysis conclusion is not bound to any evidence unit, or the bound evidence unit is below the evidence credibility weight threshold, the system generates a conclusion citation missing marker; if the brain region object in the analysis conclusion is inconsistent with the brain region to which the cited evidence belongs, or the coordinates of the abnormal region deviate from the spatial range of the cited evidence, the system generates an evidence pointing deviation marker; if the direction of change in the analysis conclusion is opposite to the direction of evolution recorded in the time evolution chain, or the magnitude of change does not match the strength of evidence, the system generates an evolution direction inconsistency marker; if the initial analysis result does not retain the marked conflict constraint relationships in the brain region evidence field, or uses low-credibility conflict evidence as the basis for a definitive conclusion, the system generates a conflict constraint not retained marker.

[0063] For example, the target subject had multi-temporal brain imaging data generated in March, June, and December 2025. The initial analysis results included the conclusion that the abnormal region in the hippocampus was continuously expanding. After system analysis, it was found that this conclusion cited spatial evidence units in the hippocampus in December 2025, but did not cite the temporal evolution chain between March and June 2025. The system further traced back the brain region evidence field and found low-quality markers in the June 2025 images, and uncertain markers for the corresponding changes in the abnormal range. Therefore, the conclusion lacked complete cross-temporal evidence. Based on this, the system generated missing conclusion citation markers and unretained conflict constraint markers, and recorded the relevant spatial evidence chain, temporal evolution chain, and low-quality conflict constraint relationships.

[0064] After generating anomalous reference identifiers, the system associates and encapsulates the anomalous reference identifiers, corresponding evidence chains, and conflict constraint relationships to generate a correction prompt context for constraint correction. The correction prompt context includes the conclusion to be corrected, the type of anomalous reference, the allowed evidence chains, the evidence boundaries prohibiting expanded reasoning, the uncertainties to be retained, and the correction output requirements. For conclusions where the evidence points to a different point, the correction prompt context restricts the large model to only regenerate the conclusion based on the correct brain region evidence chain; for conclusions with inconsistent evolutionary directions, the correction prompt context requires the large model to use the direction of change in the temporal evolution chain as the standard; for conclusions where conflict constraints are not retained, the correction prompt context requires the large model to retain the source of conflict and uncertainty markers in the correction result.

[0065] Through the processing method of this embodiment, the system can reverse verify the correspondence between the analysis conclusions and the brain region evidence field after the large model generates the initial analysis results. It can also transform problems such as missing references, deviations in pointing, inconsistencies in direction, and omissions of conflicts into correction prompts that can be processed again by the large model. This improves the consistency between the brain imaging analysis results and the original evidence, reduces the bias of the conclusions, and enhances the traceability of the analysis process.

[0066] In some embodiments, the large model input with corrected contextual prompts is constrained to generate target analysis results for brain imaging data analysis, including: Input the correction prompt context into the large model, and determine the conclusion to be corrected and the corresponding correction constraints based on the abnormal reference identifier; Based on the corresponding evidence chain and conflict constraint relationship, the conclusion to be revised is supplemented with evidence, reconstructed with citations, and remarked with uncertainty to generate the revised analysis results. The consistency between the corrected analysis results and the brain region evidence field is checked, and the target analysis results for brain imaging data analysis are generated after the check is passed.

[0067] Specifically, after generating the correction prompt context, the brain imaging analysis platform inputs the correction prompt context into the large model, and before inputting it, organizes the conclusion to be corrected, abnormal citation markers, citationable evidence chains, and conflict constraint relationships into a structured correction task. The system first determines the correction type of the conclusion to be corrected based on the abnormal citation markers. If the abnormal citation markers indicate that the conclusion citation is missing, then supplementary evidence citations are used as correction constraints; if the abnormal citation markers indicate that the evidence is deviating, then brain region anchors, spatial coordinate ranges, and correct evidence chains are used as correction boundaries; if the abnormal citation markers indicate that the evolutionary direction is inconsistent, then the change direction recorded in the temporal evolution chain is used as a priority constraint; if the abnormal citation markers indicate that conflict constraints are not preserved, then conflict sources, uncertainty levels, and low-confidence evidence markers are required to be preserved.

[0068] After receiving the correction prompt context, the large model does not regenerate a complete analysis conclusion. Instead, it performs local constraint corrections around the conclusion to be corrected. For conclusions lacking evidence citations, the large model retrieves the spatial evidence chain, modal evidence chain, and temporal evolution chain corresponding to the conclusion object from the correction prompt context and supplements the conclusion with evidence units that meet the evidence credibility weight requirements. For conclusions with deviated evidence pointing in the wrong direction, the large model redetermines the conclusion object based on brain region anchor points and spatial coordinate ranges, ensuring that the corrected conclusion is consistent with the target brain region, abnormal region, and image source in the brain region evidence field. For conclusions with inconsistent evolutionary directions, the large model regenerates trend descriptions based on the change direction and evidence strength changes between adjacent examination time points. For conclusions where conflict constraints are not retained, the large model writes conflicting evidence and uncertainty markers into the correction results to avoid expressing conflicting content as a definitive conclusion.

[0069] For example, the initial analysis concluded that the anomalous area in the hippocampus region was continuously expanding. However, backtesting of the evidence chain revealed that this conclusion only cited spatial evidence units from December 2025, neglecting the temporal evolution chain between March and June 2025, and that the June 2025 imagery contained low-quality markers. Based on this, the system generated a correction prompt context and input the spatial evidence chain, temporal evolution chain, low-quality conflict constraints, and the conclusion to be corrected into the large model. The large model supplemented the evidence references between March, June, and December 2025 based on the temporal evolution chain, and simultaneously incorporated the uncertainty markers of the low-quality June 2025 imagery into the correction analysis results, ensuring that the corrected trend description was constrained by complete cross-temporal evidence.

[0070] After generating the corrected analysis results, the system further verifies their consistency with the brain region evidence field. The system analyzes the conclusion objects, evidence citations, directions of change, and uncertainty markers in the corrected analysis results, and then re-matches them with the spatial evidence chain, modal evidence chain, temporal evolution chain, and conflict constraint chain in the brain region evidence field. If the conclusion objects in the corrected analysis results are consistent with the brain region anchor points, the evidence citations can be traced back to the corresponding evidence units, the direction of change conforms to the temporal evolution chain, and the conflict constraint relationships are preserved, the verification passes. If there are still missing citations, evidence mismatches, or conflict omissions, the system continues to retain the abnormal citation markers and marks the corresponding conclusions as content awaiting manual verification. After the verification passes, the system generates target analysis results for brain imaging data analysis. These target analysis results include the corrected brain region analysis conclusions, abnormal region conclusions, trend conclusions, evidence citation relationships, uncertainty markers, and verification status.

[0071] Through the processing method of this embodiment, the system can transform the missing references, deviations in direction, trend errors, and conflict omissions in the initial analysis results into structured tasks that can be constrained and corrected. After correction, a consistency review is performed again to ensure that the target analysis results maintain a stable correspondence with the brain region evidence field, thereby improving the evidence integrity, directional consistency, and retrospective reliability of brain imaging analysis conclusions.

[0072] The following are system embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0073] Figure 2 This is a schematic diagram of the structure of the brain imaging data analysis system combining a large model provided in an embodiment of this application. Figure 2 As shown, the system includes: The acquisition module 201 is used to acquire multi-temporal brain image data and corresponding acquisition parameters of the target object, perform quality screening, deformation registration, intensity domain correction and individual brain region segmentation on the multi-temporal brain image data, and generate individual brain region image data bound with spatial coordinates, time labels and quality labels. The extraction module 202 inputs individual brain region image data into a three-dimensional image coding network and a brain region topological coding network to extract voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and trans-temporal change representations, and constructs a brain region evidence field based on voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and trans-temporal change representations. The reasoning module 203 generates an evidential cue context containing spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains based on the brain region evidence field. The evidential cue context is then input into the large model for hierarchical reasoning, generating initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships. The generation module 204 performs evidence chain backtracking verification and conflict constraint verification on the initial analysis results, generates a correction prompt context based on the abnormal reference identifiers obtained from the verification, and inputs the correction prompt context into the large model for constraint correction, generating the target analysis results for brain imaging data analysis.

[0074] In some embodiments, Figure 2 The acquisition module 201 retrieves original brain images and image head information across examination time points from the brain image archiving system based on the examination identifier of the target object; performs object consistency verification and examination time point merging on the original brain images; extracts the acquisition parameters bound to each examination time point based on the image head information; performs protocol fingerprint matching on the original brain images based on the acquisition parameters; determines the target image sequence under the same analysis task; and generates multi-temporal brain image data according to the examination time points.

[0075] In some embodiments, Figure 2 The acquisition module 201 performs artifact identification, layer integrity verification, and signal availability evaluation on multi-temporal brain imaging data based on the acquisition parameters and protocol fingerprint, and generates quality labels for each examination time point. Using the reference temporal image of the target object as the registration base, combined with brain structure boundary constraints and cross-temporal deformation field estimation, the brain images of each examination time point are mapped to a unified individual space. Based on the unified individual space, intensity domain distribution correction and individual brain region segmentation are performed, and the corrected image voxels are associated with the corresponding brain regions, spatial coordinates, time labels, and quality labels to generate individual brain region image data.

[0076] In some embodiments, Figure 2 The extraction module 202, based on spatial coordinates, temporal markers, and quality markers, performs reliable voxel sampling and brain region anchor point construction on individual brain region image data to generate a brain region constrained image sequence. This brain region constrained image sequence is then input into a three-dimensional image coding network. Constraints and brain region boundary consistency constraints are reconstructed using a cross-temporal mask to extract voxel anomaly representations and brain region structural representations. A brain region topology input is constructed based on the brain region structural representations and input into the brain region topology coding network. Brain connectivity perturbation representations are extracted based on brain region node states and brain region connectivity relationships. Finally, voxel anomaly representations, brain region structural representations, and brain connectivity perturbation representations are fused to generate a cross-temporal change representation.

[0077] In some embodiments, Figure 2The extraction module 202 uses brain region anchors as the aggregation benchmark to map voxel abnormalities to corresponding brain regions and generates evidence credibility weights based on quality labels. It associates brain region structural representations, brain connectivity perturbation representations, and trans-temporal change representations according to spatial adjacency, connectivity propagation, and temporal evolution to form multi-scale evidence units. Based on the evidence credibility weights, it performs uncertainty propagation and conflict constraint labeling on the multi-scale evidence units to generate a brain region evidence field containing evidence strength, evidence source, and evolution direction.

[0078] In some embodiments, Figure 2 The reasoning module 203 performs evidence role analysis on the evidence units in the brain region evidence field, and determines the brain region location, image source, temporal state and conflict attributes corresponding to each evidence unit; according to the spatial adjacency relationship of the brain region, cross-modal support relationship, cross-temporal evolution relationship and conflict constraint relationship, the evidence units are chained to generate a multi-dimensional evidence chain; based on the multi-dimensional evidence chain and evidence credibility weight, an evidence-based prompting context with limited reasoning scope, evidence citation rules and conflict retention rules is generated.

[0079] In some embodiments, Figure 2 The reasoning module 203 inputs the evidence-based context into the large model, and performs evidence node retrieval and reasoning task decomposition on the multidimensional evidence chain based on the limited reasoning scope. According to the evidence citation rules, it drives the large model to generate corresponding candidate conclusions according to the brain region layer, abnormal region layer and temporal evolution layer, and simultaneously establishes the citation mapping between candidate conclusions and evidence units. Based on the conflict retention rule, it applies consistency constraints and uncertainty marking to the candidate conclusions, and generates initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions and evidence citation relationships.

[0080] In some embodiments, Figure 2 The generation module 204 parses the conclusion object, anomaly type, direction of change, and evidence citation relationship in the initial analysis results, and backtracks the analysis results to the corresponding evidence chain in the brain region evidence field; based on the backtracking matching results, it generates anomaly citation identifiers for cases where conclusion citations are missing, evidence points are deviated, evolutionary directions are inconsistent, and conflict constraints are not preserved; it associates and encapsulates the anomaly citation identifiers, corresponding evidence chains, and conflict constraint relationships to generate a correction prompt context for constraint correction.

[0081] In some embodiments, Figure 2 The generation module 204 inputs the correction prompt context into the large model, determines the conclusion to be corrected and the corresponding correction constraints based on the abnormal reference identifier; according to the corresponding evidence chain and conflict constraint relationship, it performs evidence completion, reference reconstruction and uncertainty re-marking on the conclusion to be corrected, and generates the correction analysis result; it performs consistency verification between the correction analysis result and the brain region evidence field, and generates the target analysis result for brain imaging data analysis after the verification is passed.

[0082] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0083] Figure 3 This is a schematic diagram of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various system embodiments described above.

[0084] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0085] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0086] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0088] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which may be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0089] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for analyzing brain imaging data using a large model, characterized in that, include: Acquire multi-temporal brain imaging data and corresponding acquisition parameters of the target object, perform quality screening, deformation registration, intensity domain correction and individual brain region segmentation on the multi-temporal brain imaging data, and generate individual brain region imaging data bound with spatial coordinates, time markers and quality markers; The individual brain region image data are input into a three-dimensional image coding network and a brain region topological coding network to extract voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations and transtemporal change representations, and a brain region evidence field is constructed based on the voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations and transtemporal change representations. Based on the brain region evidence field, an evidentiary cue context is generated, which includes spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains. The evidentiary cue context is then input into a large model for hierarchical reasoning, generating initial analysis results that include brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships. The initial analysis results are subjected to evidence chain backtracking verification and conflict constraint verification. Based on the abnormal reference identifiers obtained from the verification, a correction prompt context is generated, and the correction prompt context is input into the large model for constraint correction, generating target analysis results for brain imaging data analysis.

2. The method according to claim 1, characterized in that, The acquisition of multi-temporal brain imaging data and corresponding acquisition parameters of the target object includes: Based on the examination identifiers of the target subject, retrieve the original brain images and image head information across examination time points from the brain imaging archiving system; The original brain images are subjected to object consistency verification and inspection time point merging, and the acquisition parameters bound to each inspection time point are extracted based on the image head information. Based on the acquired parameters, the original brain images are matched using protocol fingerprints to determine the target image sequence under the same analysis task, and multi-temporal brain image data is generated according to the examination time point.

3. The method according to claim 2, characterized in that, The process of performing quality screening, deformation registration, intensity domain correction, and individual brain region segmentation on the multi-temporal brain imaging data to generate individual brain region imaging data bound with spatial coordinates, temporal markers, and quality markers includes: Based on the acquisition parameters and protocol fingerprint, artifact identification, layer integrity verification and signal availability evaluation are performed on the multi-temporal brain imaging data to generate quality labels for each examination time point; Using the baseline temporal images of the target object as the registration base, and combining brain structure boundary constraints and cross-temporal deformation field estimation, the brain images at each examination time point are mapped to a unified individual space. Based on the unified individual space, intensity domain distribution correction and individual brain region segmentation are performed. The corrected image voxels are associated with the corresponding brain regions, spatial coordinates, time markers and quality markers to generate individual brain region image data.

4. The method according to claim 1, characterized in that, The step of inputting the individual brain region imaging data into a three-dimensional image coding network and a brain region topological coding network to extract voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and trans-temporal change representations includes: Based on the spatial coordinates, time markers, and quality markers, the individual brain region image data is subjected to reliable voxel sampling and brain region anchor point construction to generate a brain region constrained image sequence. The brain region constraint image sequence is input into the three-dimensional image coding network, and the constraints and brain region boundary consistency constraints are reconstructed using cross-temporal masking to extract voxel abnormality representations and brain region structural representations. Based on the brain region structural representation, a brain region topology input is constructed, and the brain region topology input is input into the brain region topology coding network. Based on the brain region node state and brain region connectivity, brain connectivity perturbation representation is extracted. By integrating the aforementioned voxel abnormality representation, brain region structural representation, and brain connectivity perturbation representation, a trans-temporal change representation is generated.

5. The method according to claim 4, characterized in that, The construction of a brain region evidence field based on the voxel abnormality representation, brain region structural representation, brain connectivity perturbation representation, and trans-temporal change representation includes: Using the brain region anchors as aggregation benchmarks, the voxel abnormality representations are mapped to the corresponding brain regions, and evidence credibility weights are generated based on the quality labels. The brain region structural representations, brain connectivity perturbation representations, and trans-temporal change representations are correlated according to spatial adjacency, connectivity propagation, and temporal evolution to form multi-scale evidence units; Based on the aforementioned evidence credibility weights, uncertainty propagation and conflict constraint labeling are applied to the multi-scale evidence units to generate a brain region evidence field that includes evidence strength, evidence source, and evolutionary direction.

6. The method according to claim 1, characterized in that, The generation of evidential cueing context based on the brain region evidence field, including spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains, includes: Evidence role analysis is performed on the evidence units in the brain region evidence field to determine the brain region location, image source, temporal state and conflict attributes corresponding to each evidence unit; Based on the spatial adjacency of brain regions, cross-modal support relationships, cross-temporal evolution relationships, and conflict constraint relationships, the evidence units are chained together to generate a multi-dimensional evidence chain; Based on the multidimensional evidence chain and evidence credibility weight, an evidence-based prompting context with limited reasoning scope, evidence citation rules, and conflict retention rules is generated.

7. The method according to claim 6, characterized in that, The process involves inputting the evidence-based context into a large model for hierarchical reasoning, generating initial analysis results that include brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships, including: The evidence-based prompt context is input into the large model, and evidence node retrieval and reasoning task decomposition are performed on the multidimensional evidence chain based on the limited reasoning scope. According to the evidence citation rules, the large model is driven to generate corresponding candidate conclusions according to the brain region layer, abnormal region layer and temporal evolution layer, and a citation mapping between candidate conclusions and evidence units is established simultaneously. Based on the conflict retention rule, the candidate conclusions are subjected to consistency constraints and uncertainty marking, generating initial analysis results that include brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships.

8. The method according to claim 1, characterized in that, The step of performing evidence chain backtracking verification and conflict constraint verification on the initial analysis results, and generating a correction prompt context based on the abnormal reference identifiers obtained from the verification, includes: The conclusion objects, abnormality types, change directions, and evidence citation relationships in the initial analysis results are analyzed, and the analysis results are back-matched to the corresponding evidence chains in the brain region evidence field. Based on the backtracking matching results, abnormal citation markers are generated for cases where conclusions are missing, evidence points are deviating, evolution directions are inconsistent, and conflict constraints are not preserved. The abnormal reference identifier, the corresponding evidence chain, and the conflict constraint relationship are associated and encapsulated to generate a correction prompt context for constraint correction.

9. The method according to claim 8, characterized in that, The step of inputting the correction prompt context into the large model for constraint correction and generating target analysis results for brain imaging data analysis includes: Input the correction prompt context into the large model, and determine the conclusion to be corrected and the corresponding correction constraints based on the abnormal reference identifier; Based on the corresponding evidence chain and conflict constraint relationship, the conclusion to be corrected is supplemented with evidence, reconstructed with citations, and remarked with uncertainty to generate a correction analysis result. The consistency between the corrected analysis results and the brain region evidence field is verified, and target analysis results for brain imaging data analysis are generated after the verification is passed.

10. A brain imaging data analysis system incorporating a large model, characterized in that, include: The acquisition module is used to acquire multi-temporal brain image data and corresponding acquisition parameters of the target object, perform quality screening, deformation registration, intensity domain correction and individual brain region segmentation on the multi-temporal brain image data, and generate individual brain region image data bound with spatial coordinates, time markers and quality markers. The extraction module inputs the individual brain region image data into a three-dimensional image coding network and a brain region topology coding network to extract voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and transtemporal change representations, and constructs a brain region evidence field based on the voxel abnormality representations, brain region structural representations, brain connectivity perturbation representations, and transtemporal change representations. The reasoning module generates an evidential cueing context containing spatial evidence chains, modal evidence chains, temporal evolution chains, and conflict constraint chains based on the brain region evidence field. The evidential cueing context is then input into the large model for hierarchical reasoning to generate initial analysis results containing brain region analysis conclusions, abnormal region conclusions, trend conclusions, and evidence citation relationships. The generation module performs evidence chain backtracking verification and conflict constraint verification on the initial analysis results, generates a correction prompt context based on the abnormal reference identifier obtained from the verification, and inputs the correction prompt context into the large model for constraint correction, thereby generating the target analysis results for brain imaging data analysis.

Images