Multi-modal magnetic resonance aneurysm occlusion assessment method and system

Abstract

The invention relates to the field of medical image processing, and discloses a multi-mode magnetic resonance aneurysm occlusion assessment method and system. The method comprises the following steps: acquiring scanning configuration and a multi-modal magnetic resonance sequence, and generating fused image data through cross-modal registration, artifact suppression and denoising processing; image segmentation, three-dimensional structure reconstruction and topology correction are carried out, and a three-dimensional anatomy and blood flow model is obtained; then, through gridding, index normalization and time sequence alignment processing, individual feature vectors are extracted; and finally, obtaining a blocking mechanism score and probability based on multi-model fusion reasoning, completing three-dimensional visual rendering, and backfilling key parameters to a configuration structure to form closed-loop optimization. According to the method, multi-modal data are effectively integrated, the evaluation accuracy is improved, and self-adaptive optimization and result traceability of the system are realized through parameter backfilling.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing, and in particular to a multi-modal magnetic resonance aneurysm occlusion evaluation method and system. BACKGROUND

[0002] Intracranial aneurysms have high mortality and disability rates, and their treatment is a major challenge in the field of neurosurgery. The emergence of flow diverters (FDs) has revolutionized the treatment strategy for intracranial aneurysms. Instead of traditional aneurysm cavity filling, FDs promote aneurysm occlusion through a "blood flow reconstruction" mechanism. The therapeutic effect of FDs is mainly achieved through three key physiological processes: blood flow blockage, aneurysm cavity thrombosis, and endothelialization of the aneurysm neck.

[0003] Despite the increasing use of FDs, there are still significant bottlenecks in the clinical evaluation of their post-treatment effects, mainly including:

[0004] 1. Single evaluation dimension: Traditional time-of-flight MRA (TOF-MRA) or contrast-enhanced MRA (CE-MRA) mainly provides morphological information on "patency or not", which is easily disturbed by FD metal artifacts, and cannot reflect blood flow dynamics, which is the core factor driving occlusion.

[0005] 2. Lack of mechanism interpretation: Conventional MRI cannot distinguish between blood flow stagnation, thrombus at different stages (such as acute stage thrombus rich in water and old stage thrombus that has been organized), and complete endothelial coverage. The lack of this mechanism information makes it impossible for doctors to determine which stage the occlusion process is in, whether it is stable or at risk of recanalization.

[0006] 3. Subjectivity and lag: The evaluation relies on physician visual inspection, which is highly subjective. Moreover, it can only be confirmed after occlusion occurs, and cannot quantitatively predict the final occlusion outcome in the early postoperative period, missing the best opportunity to guide individualized clinical intervention.

[0007] 4. Lack of predictability: Existing methods are limited to describing the status quo and cannot predict the final occlusion outcome based on quantitative changes in blood flow dynamics in the early postoperative period, thus failing to provide prospective guidance for clinical intervention decisions such as whether additional treatment is needed.

[0008] In the field of medical image processing, the existing scheme of multi-modal magnetic resonance aneurysm occlusion evaluation generally processes based on a single imaging sequence or a simple registration method, which has limitations such as insufficient registration accuracy, insufficient artifact suppression, and large segmentation error. Existing methods rely on manual or semi-automatic operation, which is prone to inconsistent evaluation and poor repeatability in multi-device and multi-protocol environments, making it difficult to meet the stable implementation of aneurysm occlusion evaluation. For the acquisition and calibration configuration, the joint processing of 3D HR-VWI sequence and 4D-Flow sequence, the existing technology generally has common shortcomings such as cross-modal alignment difficulty, inconsistent index calibration, and closed-loop learning absence, which makes it difficult to form a consistent process of acquisition, alignment, segmentation, judgment, control, and record in the application scenario of aneurysm occlusion evaluation, resulting in unreliable evaluation results. SUMMARY

[0009] To solve the above technical problems, the present application provides a multi-modal magnetic resonance aneurysm occlusion evaluation method, comprising:

[0010] Acquire the acquisition and calibration configuration, 3D HR-VWI sequence and 4D-Flow sequence, perform calibration mapping processing based on the scan protocol identification and coordinate system definition in the acquisition and calibration configuration, perform cross-modal registration by using mutual information-based metrics for coarse registration and multi-resolution strategy-driven non-rigid deformation for fine registration, perform artifact suppression and denoising processing by using frequency domain direction filtering and spatial direction consistency evaluation joint determination stripe main direction and execution along the main direction, and generate a cross-modal registration fusion voxel stack structure;

[0011] Perform training set consistency processing based on intensity normalization strategy alignment of historical training data intensity statistical summary, perform deep segmentation of multi-scale feature extraction and cross-layer feature splicing by using an encoder-decoder network, perform topology modification and surface reconstruction processing by skeleton extraction centerline and branch pruning and hole repair, and generate a three-dimensional anatomic and blood flow structure set;

[0012] Perform meshing and wall mapping processing based on centerline stake point generated radial geodesic ring, perform index normalization calibration based on scan protocol identification and coil layout selection scale conversion scheme and offset correction, perform feature splicing and timing alignment processing by rearranging entries on a unified time grid and deriving period mean and period fluctuation, and obtain individual feature vectors;

[0013] Perform trained model inference using a neural network branch with multi-channel input, integrated learning branch, and timing sub-model fusion reasoning, generate three-dimensional mapping and rendering of the standard viewing path according to the centerline stake point and aneurysm neck boundary line, and perform parameter backfilling and version registration processing of model version, fusion strategy, and scale conversion card backfilling, and generate an acquisition and calibration configuration structure.

[0014] Further, the trained model inference comprises:

[0015] Read the version number of the individual feature vector and the parameter snapshot number, load the corresponding trained model list; the trained model list is composed of a multi-channel input neural network branch, an ensemble learning branch and a sequential data based time series sub-model, which work together according to the preset fusion strategy in the inference stage.

[0016] Further, the trained model inference further comprises:

[0017] Split into continuous feature sub-vectors, discrete encoding sub-vectors and time series derived sub-vectors, and send them into neural network branches, ensemble learning branches and time series sub-models respectively; the inference scheduler starts each branch inference according to the order defined in the model list.

[0018] Further, the trained model inference further comprises:

[0019] Call the fusion strategy to map the three-way intermediate results to a unified mechanism score space; the fusion strategy gives adaptive weighting according to branch confidence and slot weight markers, generating blood flow blocking mechanism score, thrombosis mechanism score and endothelialization mechanism score.

[0020] Further, the three-dimensional mapping and rendering comprises:

[0021] Generate a standard browsing path and lay out key perspective nodes on the path; in the multi-channel superposition stage, map the mechanism score contribution heat map to the wall and intraluminal sampling layer, and attach the complete occlusion probability and state level in the form of numerical badges and text superposition on the corresponding area.

[0022] Further, the process of three-dimensional mapping and rendering further comprises:

[0023] If there are multiple versions of the same object, retrieve the historical scene from the version registration index, replay the historical view under the same camera path and rendering parameters, and arrange and compare them in the form of a time axis to form an intuitive expression of time series transformation.

[0024] Further, the parameter backfilling and version registration comprise:

[0025] Enter the parameter backfilling process; parameter backfilling refers to extracting key parameters, decision thresholds and scale conversion cards from the report object during the running process, writing them into the traceable slot of the acquisition and calibration configuration, and forming a closed loop dependency.

[0026] Further, the process of parameter backfilling and version registration further comprises:

[0027] Enter the version registration process; version registration refers to creating a version record in the data registration module for this inspection, and the record includes object identification, inspection time, model version, parameter snapshot number, explainability entry hash, scene replay handle, and report storage location.

[0028] Further, the parameter backfilling and version registration process also includes:

[0029] Construct an updated acquisition and calibration configuration; add or update policy entries, scale conversion cards, rule set patches, and model version anchors from this session.

[0030] Further, a multi-modal magnetic resonance aneurysm occlusion evaluation system applied to the method of any one of the above, comprising:

[0031] A data acquisition and calibration module for acquiring acquisition and calibration configurations, 3D HR-VWI sequences, and 4D-Flow sequences, and performing calibration mapping processing to generate original imaging and velocity fields;

[0032] A cross-modality registration and fusion module for extracting anatomic reference frames and velocity voxels from the original imaging and velocity fields, performing cross-modality registration, and performing artifact suppression and denoising processing on the cross-modality registration voxel set to generate a cross-modality registration fusion voxel stack;

[0033] A training set uniformization module for performing training set uniformization processing on the cross-modality registration fusion voxel stack to obtain segmentation input voxels;

[0034] A deep segmentation and topology correction module for extracting parent artery voxels, aneurysm capsule voxels, flow diversion device voxels, and residual blood flow voxels from the segmentation input voxels, performing deep segmentation, and performing topology correction and surface reconstruction on the structure segmentation label set to generate a three-dimensional anatomic and blood flow structure set;

[0035] A gridding and index calculation module for performing gridding and wall mapping processing on the three-dimensional anatomic and blood flow structure set to obtain an index calculation input set, and extracting thrombus volume fraction, average wall shear stress, blood flow stagnation index, and aneurysm neck reinforcement thickness from the index calculation input set, performing normalization calibration, and generating an index set;

[0036] A feature vector generation module for performing feature splicing and time series alignment on the index set to generate individual feature vectors;

[0037] A model inference module for obtaining the individual feature vectors, performing trained model inference, and obtaining mechanism scores and probability results;

[0038] A visualization and updating module is configured to extract occlusion state levels and complete occlusion probabilities from the mechanism score and probability results, perform three-dimensional mapping and rendering, generate three-dimensional visualization and follow-up comparison reports, perform parameter backfilling and version registration on the three-dimensional visualization and follow-up comparison reports, and generate acquisition and calibration configurations.

[0039] The key innovations of the present application include:

[0040] (1) Based on the scanning protocol identification and coordinate system definition in the acquisition and calibration configuration, a calibration mapping process is performed, a mutual information-based metric is used for coarse registration, and a multi-resolution strategy is used to drive non-rigid deformation for fine registration to perform cross-modality registration, and a frequency domain direction filter and spatial direction consistency evaluation are used to jointly determine the main direction of the stripes and perform suppression along the main direction to perform artifact suppression and denoising processing, generating a cross-modality registration fusion voxel stack structure.

[0041] (2) Perform training set uniformization processing based on intensity statistical summary of historical training data, use an encoder-decoder network for multi-scale feature extraction and cross-layer feature splicing for deep segmentation, perform topology modification and surface reconstruction processing by skeleton extraction, centerline pruning and hole repair, to generate a three-dimensional anatomical and blood flow structure set.

[0042] (3) Perform trained model inference using a multi-channel input neural network branch, an ensemble learning branch, and a time series sub-model fusion inference, generate a standard viewing path based on the centerline stake points and the aneurysm neck boundary line, and perform parameter backfilling and version registration processing by backfilling the model version, fusion strategy, and scale conversion card, to generate an acquisition and calibration configuration structure.

[0043] The main benefits of the present application are as follows:

[0044] (1) Through the cross-modality registration and artifact suppression processing, the alignment accuracy and noise suppression ability of multi-modal data in a unified coordinate are improved, and the evaluation errors caused by inaccurate registration and artifact interference in the prior art are overcome.

[0045] (2) With the help of the training set uniformization, deep segmentation, and topology modification means, accurate segmentation and three-dimensional reconstruction of the parent artery, aneurysm sac, blood flow guiding device, and residual blood flow are achieved, solving the problem of coarse segmentation and topology inconsistency in existing methods.

[0046] (3) Using the parameter backfilling and version registration mechanism, a closed-loop update of the acquisition and calibration configuration is formed, enhancing the adaptability and traceability of the system for cross-device and cross-protocol evaluation, and making up for the lack of adaptive optimization in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1A whole flowchart of a multi-modal magnetic resonance aneurysm occlusion evaluation method provided by an embodiment of the present application is shown.

[0048] Figure 2 A flowchart of a multi-modal magnetic resonance aneurysm occlusion evaluation method provided by an embodiment of the present application is shown.

[0049] Figure 3 A multi-modal MRI evaluation schematic diagram provided by an embodiment of the present application is shown.

[0050] Figure 4 A whole architecture diagram of a multi-modal magnetic resonance aneurysm occlusion evaluation system according to an embodiment of the present application is shown.

[0051] Figure 5 A structure block diagram of a multi-modal magnetic resonance aneurysm occlusion evaluation system provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0052] Figure 1 A whole flowchart of a multi-modal magnetic resonance aneurysm occlusion evaluation method provided by an embodiment of the present application is shown in Figure 1 The method starts from a multi-modal magnetic resonance imaging scan at a specific follow-up time point after a patient's blood flow guiding device implantation, so as to obtain original data containing a 3D high-resolution tube wall imaging sequence and a four-dimensional blood flow magnetic resonance imaging sequence. Then, image preprocessing and multi-modal fusion are performed on the multi-modal magnetic resonance imaging sequence to obtain a comprehensive image data set. Subsequently, data analysis is performed on the four-dimensional blood flow magnetic resonance imaging sequence to quantify hemodynamic parameters, and data analysis is performed on the three-dimensional high-resolution tube wall imaging sequence to segment key structures and quantify morphological parameters. Then, multi-dimensional fusion is performed on the hemodynamic parameters and morphological parameters to construct an individualized parameter feature matrix. The individualized parameter feature matrix is input into a pre-trained machine learning prediction model to obtain evaluation and prediction results of aneurysm occlusion state, thereby assisting clinical decision-making. Through the above-mentioned process, the method realizes comprehensive and quantitative evaluation of the aneurysm occlusion process after blood flow guiding device treatment.

[0053] Embodiment I: with reference to Figure 2 A flowchart of a multi-modal magnetic resonance aneurysm occlusion evaluation method provided by an embodiment of the present application is shown. The flowchart can at least include steps S100-S400:

[0054] S100, acquire the acquisition and calibration configuration, 3D HR-VWI sequence and 4D-Flow sequence, perform calibration mapping processing based on the scan protocol identification and coordinate system definition in the acquisition and calibration configuration, perform cross-modality registration by using mutual information-based metrics for coarse registration and multi-resolution strategy-driven non-rigid deformation for fine registration, perform artifact suppression and denoising processing by using joint determination of stripe main direction based on frequency domain direction filtering and spatial direction consistency evaluation and performing suppression along the main direction, and generate a cross-modality registration fusion voxel stack structure;

[0055] S200, perform training set consistency processing based on intensity normalization strategy alignment of historical training data intensity statistical summary, perform deep segmentation by using an encoder-decoder network for multi-scale feature extraction and cross-layer feature splicing, perform topology modification and surface reconstruction processing by extracting a center line through skeletonization and performing branch pruning and hole repair, and generate a three-dimensional anatomical and blood flow structure set;

[0056] S300, perform meshing and wall mapping processing based on generating radial geodesic line annuli based on center line stake points, perform index normalization calibration based on selecting a scale conversion scheme based on the scan protocol identification and coil layout and performing offset correction, perform feature splicing and time sequence alignment processing by rearranging entries on a unified time grid and deriving period mean and period fluctuation, and obtain an individual feature vector;

[0057] S400, perform trained model reasoning by using a neural network branch with multi-channel input, ensemble learning branch and temporal sub-model fusion reasoning, generate three-dimensional mapping and rendering of a standard viewing path based on the center line stake points and tumor neck boundary line, perform parameter backfilling and version registration processing of model version, fusion strategy and scale conversion card backfilling, and generate an acquisition and calibration configuration structure.

[0058] Step S100 at least includes steps S110-S130:

[0059] S110, acquire the acquisition and calibration configuration, 3D HR-VWI sequence and 4D-Flow sequence, perform calibration mapping processing, and obtain an original imaging and velocity field link;

[0060] The input source is the acquisition and calibration configuration of the pre-sequence S430 backfill, together with the 3D HR-VWI sequence and 4D-Flow sequence generated by the magnetic resonance console. The acquisition and calibration configuration is a structured parameter set used to drive data acquisition and record mapping in the process of an examination, including scan protocol identification, coil model and channel layout, bed and body position record, coordinate system definition, velocity encoding range, cardiac trigger and respiratory motion gating strategy, timestamp generation rule, DICOM tag mapping table, patient anonymization rule and sequence archiving path. The 3D HR-VWI sequence is a high-resolution vessel wall imaging (HR-VWI) voxel sequence that highlights the wall level and tumor neck transition; the 4D-Flow sequence is a time-resolved three-dimensional phase-contrast blood flow magnetic resonance (4D-Flow) voxel sequence that records multi-phase velocity phase data and amplitude reference. Specifically, the system reads the scan protocol identification and coordinate system definition from the acquisition and calibration configuration, calls the DICOM (Digital Imaging and Communications in Medicine) read module to access the two types of sequences, combines the frame sequence according to the timestamp generation rule, aggregates the channel data according to the coil channel layout, processes the sensitive fields according to the anonymization rule, and writes them into the temporary data area to form the voxel flow to be calibrated. Further, the system performs gain equalization according to the coil model and channel layout, loads the magnetic field gradient nonlinearity correction parameters according to the coordinate system definition, restores the patient's posture description in the scan coordinates according to the bed and body position record, and loads the phase velocity scale according to the velocity encoding range. The above actions are all recorded in the log module with operation time, data batch and parameter snapshot, and when encountering abnormalities such as missing frames, timestamp jumps, and missing tags, the system triggers pre-loading check, temporarily stores the abnormal batch and writes it into the recovery queue; if there is a cardiac trigger or respiratory motion gating strategy, the system obtains the trigger sequence and gating curve from the device log, inserts the time axis alignment process, and completes the mapping between the trigger point, gating window and imaging frame. Subsequently, the system enters the calibration mapping process: geometric distortion correction and intensity normalization are performed on the HR-VWI voxel sequence, and each frame is reconstructed to the scan coordinates; phase bias correction, phase unwrapping, background phase removal and amplitude reference fusion are performed on the 4D-Flow voxel sequence to restore the velocity direction structure of each phase and construct the multi-phase velocity voxel. The coordinate system definition in the acquisition and calibration configuration is uniformly used for the two types of sequences to establish a common index and complete the same examination number, same patient and same time loading aggregation.After processing, the system outputs a field named raw imaging & velocity field, where the raw imaging part contains the HR-VWI voxel sequence that is geometrically corrected and intensity normalized, the velocity field part contains the multi-phase velocity voxel that is phase processed and amplitude reference fused, and the acquisition and calibration configuration snapshots are attached in the subject metadata; the raw imaging & velocity field will be directly called by the raw imaging & velocity field of S120 as the input for anatomical reference frame construction and velocity voxel extraction, while serving as one of the underlying data sources for subsequent S200 deep segmentation and 3D reconstruction, forming a sequential transmission link from S110 to S120.

[0061] S120, extracting anatomical reference frame and velocity voxel from raw imaging & velocity field, performing cross-modality registration, generating cross-modality registered voxel set, and

[0062] The input source is the raw imaging and velocity field output by S110, where the raw imaging part undertakes the construction of the anatomical reference frame, and the velocity field part undertakes the extraction of velocity voxels. The anatomical reference frame is a reference framework describing the intracranial anatomical structure in a unified three-dimensional coordinate, which is composed of anatomical reference points, direction bases, scale parameters and spatial masks; the anatomical reference points are derived from the bone or membrane markers that can be stably identified in the HR-VWI sequence, the direction bases follow the head-foot, front-back and left-right standards, the scale parameters record the voxel spacing and slice thickness in each direction, and the spatial masks limit the calculation range of intracranial blood vessels and surrounding tissues. Velocity voxels are a set of three-dimensional voxels extracted from the velocity field part, containing velocity phase information and amplitude reference, and are organized according to the phase index. Specifically, the system first performs anatomical landmark detection in the raw imaging part, uses edge and texture composite operators to search for stable points such as the saddle and petrous apex, and constructs a set of anatomical reference points; then restores the standard orientation of the HR-VWI sequence in the scanning coordinate through rigid alignment, generates the direction base, records the voxel spacing and slice thickness to form the scale parameter; and then generates the spatial mask through a joint strategy based on intensity threshold and morphology, excludes extracranial tissues and bone high signals, and obtains the limited area. The velocity voxel extraction process loads the phase and amplitude voxels from the velocity field part by phase, performs background phase removal residual review and amplitude reference clipping for each phase, eliminates voxels with insufficient reliability due to low amplitude, clips the candidate velocity voxel set according to the limited area, and maps the candidate velocity voxel set to the anatomical reference frame coordinate framework. After the anatomical reference frame and velocity voxels are constructed, the system enters the cross-modality registration: stage one is coarse registration, which establishes an alignment relationship between the raw imaging and the amplitude reference using a mutual information-based metric, and outputs a rigid transformation; stage two is fine registration, which uses a multi-resolution strategy to drive non-rigid deformation, constrains the deformation to change slowly in the vicinity of the vessel wall, and maintains smooth transition in the brain parenchyma area; stage three is phase synchronization, which aligns the velocity voxels of each phase with the time index of the anatomical reference frame according to the trigger points and the gating window recorded by S110, and forms a spatiotemporally consistent mapping. During the registration process, the system records the registration metric curve, the deformation field energy and the key point deviation, and if the registration metric is below the threshold or the key point deviation is too large, the rollback strategy is triggered to restart the parameter grid in the coarse registration stage; if some phases are missing triggers or have abnormal gating windows, the time index is supplemented by adjacent interpolation in the phase synchronization stage and a marker is recorded for reference by subsequent modules. After registration is completed, the system resamples the raw imaging part and the velocity voxels in the unified coordinate and the unified time index, constructs a cross-modality registered voxel set containing the registered HR-VWI voxels, the velocity voxels of each phase and the corresponding anatomical reference frame description, and binds the acquisition and calibration configuration snapshot number from S110, and writes it into the sequence metadata.

[0063] Figure 3 A multi-modality MRI evaluation schematic diagram provided for an embodiment of the present application is shown in FIG. 1. Figure 3As shown, it intuitively demonstrates the ability of the method to assess key features of aneurysm occlusion. Figure 3 (a) Based on 4D-Flow MRI data, the residual blood flow status within the aneurysm is displayed in the form of color flow streamlines, reflecting the effectiveness of blood flow blockage mechanism. Figure 3 (b) Based on 3D HR-VWI data, the endothelialization condition at the aneurysm neck is displayed by arrows indicating the enhanced signal, and the thrombus area within the aneurysm cavity is marked by asterisks. This schematic diagram demonstrates that through the cross-modality registration and deep segmentation process, hemodynamic and tubular morphological information can be effectively fused and visualized, providing intuitive basis for subsequent individualized assessment.

[0064] This output field is named cross-modality registration voxel set, which will be directly called by the cross-modality registration voxel set of S130 as the input for artifact suppression and denoising processing, while preparing the aligned voxel base for S200 deep segmentation and three-dimensional reconstruction, and connecting the subsequent index calculation and reasoning link.

[0065] S130, artifact suppression and denoising processing on the cross-modality registration voxel set, generating a cross-modality registration fusion voxel stack structure link;

[0066] The input is the cross-modality registered voxel set output by S120, which contains the registered HR-VWI voxels, the phase-wise velocity voxels, and the anatomical reference frame description. The artifact suppression and denoising process is a composite process constructed for susceptibility, metal streak, chemical shift, motion ghosting, and background noise, and outputs the cross-modality registered fusion voxel stack; this voxel stack converges the HR-VWI voxels and the velocity voxels after suppression and denoising under unified coordinates and time index, and retains the necessary metadata for subsequent segmentation and reconstruction. Specifically, the system first performs susceptibility artifact direction estimation and streak detection on the registered HR-VWI voxels, uses frequency domain directional filtering and spatial directional consistency evaluation to jointly determine the main direction of the streak, and then performs suppression along the main direction to suppress the streak energy below the background threshold; for chemical shift related mismatch, the system corrects the boundary drift through frequency domain resampling and adjacent band fusion; for motion ghosting, the system calculates the displacement field between adjacent slices within the anatomical reference frame limited area, implements inter-slice elastic alignment, and repairs the misplacement. For the velocity voxels, the system performs phase residue review phase by phase, performs background model regression for residual phase bias in the resting area, and then implements temporal and spatial denoising according to the anatomical reference frame limited area: in the time axis direction, an adaptive window smoothing strategy is used to suppress high-frequency jitter, and in the spatial direction, a block matching non-local average strategy is used to fuse similar block information and reduce shot noise; amplitude reference is involved in confidence weighting, and low-amplitude voxels are given less weight to reduce velocity estimation jitter. To avoid the alignment relationship between cross-modalities being affected by processing side effects, the system calls a registration check operator after each processing sub-stage to review the spatial consistency of the intensity and velocity structure in the neighborhood of the anatomical reference point; once the deviation exceeds the threshold, the current sub-stage parameters are traced back, the processing intensity is reduced, and the process is executed again. An exception recording mechanism runs throughout the process, and if a local strong susceptibility area is detected that causes wall information to be missing, the system marks the area as a low-confidence area and records the confidence marker in the metadata for reference by the subsequent segmentation stage. After the above processing, the system performs fusion and assembly of the HR-VWI voxels and the phase-wise velocity voxels: organizes them into a multi-layer stack structure in the unified coordinates, records the source, processing sequence, and time point in the inter-layer index, and hangs the anatomical reference frame description, registration metric summary, and processing parameter summary on the object metadata layer to form the cross-modality registered fusion voxel stack. The output field is named cross-modality registered fusion voxel stack, which will be directly called by the cross-modality registered fusion voxel stack of S210 as the only data source for segmentation input voxel construction, while providing a stable foundation for subsequent three-dimensional reconstruction of S200, and running through the entire link of S300 index calculation and S400 inference rendering. The technical effect of this step is summarized as follows: by performing directional suppression, temporal and spatial denoising, and fusion and assembly on the cross-modality registered voxel set, a cross-modality registered fusion voxel stack with clear structure and temporal and spatial consistency is formed, meeting the input stability requirements of subsequent segmentation, reconstruction, and index calculation.

[0067] Step S200 includes at least steps S210-S230:

[0068] S210, a cross-modal registration fusion voxel stack is acquired, training set unification processing is performed, and a segmentation input voxel link is obtained;

[0069] The input source is the cross-modality registration fused voxel stack output by S130; the voxel stack contains high-resolution vessel wall imaging (HR-VWI) voxels and multi-phase velocity voxels that have been assembled under a unified coordinate and time index, and is hung with an anatomical reference frame description, a registration metric summary and a processing parameter summary in the subject metadata. Specifically, the system first reads the coordinate definition, voxel spacing, slice thickness, time index and confidence label of the voxel stack, and links the scan protocol identification and coil channel layout in the calibration configuration to establish the preprocessing context of this batch; then enters the voxel resampling stage, and according to the anisotropic resolution produced by different sequence sources, the system performs voxel resampling in the anatomical reference frame coordinates to unify the grid in the spatial dimension of HR-VWI voxels and velocity voxels; in the time dimension, according to the gating curve and trigger point, the multi-phase frames of the velocity voxels are mapped to the unified time index, and the traces of frame loss, jump and interpolation are recorded to form a traceable time alignment track. Further, the system constructs a training set uniformization processing strategy; the training set uniformization processing is a set of data normalization, cropping, slicing and weight configuration processes for the segmentation inference stage, and the core actions include intensity normalization strategy alignment, candidate region cropping strategy alignment, block-level slicing strategy alignment and class weight configuration alignment. In the intensity normalization strategy alignment link, the system reads the intensity statistical summary of the historical training data, performs intensity normalization on the HR-VWI voxels according to the gray distribution mapping hyperparameter, and synchronously updates the metadata; in the candidate region cropping strategy alignment link, the system generates candidate regions in the limited area of the anatomical reference frame according to the blood vessel centerline neighborhood, instrument metal trace neighborhood and velocity high-confidence corridor, merges the overlapping candidate regions according to the priority to form candidate cubes for slicing; in the block-level slicing strategy alignment link, the system slices the candidate cubes according to the fixed edge length, and records the block index, original position, overlap ratio and time frame number; in the class weight configuration alignment link, the system generates the loss weight mapping and sampling tendency mapping of this batch based on the class proportion of the historical training data, and records them in the metadata layer. In the face of abnormal scenarios, if there is a large range of low confidence label in the candidate region, the system reduces the sampling frequency of the region and writes a candidate region weight reduction entry in the log; if there is a long period of time index vacancy, the system generates a compensation block using an adjacent frame expansion window and marks a time interpolation marker for reference by the subsequent segmentation stage. Subsequently, the system enters the multi-channel assembly; the multi-channel assembly refers to stacking HR-VWI voxels and velocity voxels in block-level coordinates by channel, while adding the spatial mask and confidence map of the anatomical reference frame as auxiliary channels; in this link, the system completes the block-level intensity standardization, time frame selection and channel order registration.After the above processing, the system output field is named Segment Input Voxel. This output is presented at the object level as a set of multi-channel voxels with block-level index and time frame index, and is bound to the parameter snapshot of the training set consistency processing strategy. The segment input voxel is directly called for depth segmentation in the subsequent S220, and also provides spatial anchor points for the 3D reconstruction in S200 to be stitched with the segmentation label, forming a sequential transfer from S210 to S220, and is consistent with the meshing and wall mapping input for the subsequent S300 index calculation.

[0070] S220. Extract the aneurysm-carrying artery voxel, aneurysm cyst voxel, blood flow guiding device voxel and residual blood voxel from the segmented input voxel, perform deep segmentation, and generate a structural segmentation label set.

[0071] The input source is the segmentation input voxels output by S210; this input consists of multi-channel block-level voxels, including HR-VWI voxels, velocity voxels, spatial masks, and confidence maps, along with block indexes, original locations, and time frame numbers. Specifically, the system first performs block-level quality gating on the segmentation input voxels; quality gating is a lightweight screening process before blocks enter network inference, including determining the block content proportion, metal artifact residue, and velocity confidence threshold; when the effective tissue proportion within a block is too low, the metal stripe residue exceeds the threshold, or the velocity confidence is too low, the block is moved to a low-priority queue and skipped or downgraded in this round of inference. Subsequently, the system initiates the deep segmentation inference subsystem. This subsystem is an encoder-decoder network with weights frozen during training. It receives multi-channel block-level voxels at input and generates class probability volumes of the same size as the input blocks at output. This subsystem distinguishes between blood vessel walls, tumor sacs, and instruments through multi-scale feature extraction, cross-layer feature concatenation, and spatial attention gating. It also provides residual blood flow candidates after temporal convergence of velocity voxel channels. During network inference, the system simultaneously records the model version, parameter checksum, inference time, hierarchical feature summary, and peak position of the probability volume, writing these to the inference log for subsequent version registration by the S430. Further, the system performs intra-block post-processing on the class probability volumes. Post-processing includes confidence threshold pruning, small connected component removal, hole filling, and boundary smoothing. Considering the overlap of block slices, the system constructs a block overlap weight map during the block synthesis stage, merges the probability volumes of overlapping regions through weighted fusion, and then performs class consistency verification in global coordinates. For residual blood flow candidates, the system performs multi-frame consistency screening at the time frame level to eliminate occasional candidates generated by isolated frames. For blood flow guidance device candidates, the system performs metal neighborhood re-determination in the joint space of HR-VWI voxels and velocity voxels to avoid misclassifying metal artifacts as device surfaces. Subsequently, the system generates labels for four types of entities; the labels are voxel annotations with the same coordinates as the original image, and the categories correspond to aneurysm-bearing artery voxels, aneurysm sac voxels, blood flow guidance device voxels, and residual blood flow voxels, respectively. During the generation process, the system clips the label boundaries according to the anatomical reference system limit area and records a summary such as the number of voxels of each category, the number of connected branches, and the boundary length in the object metadata. If cross-class overlap occurs, the system triggers a conflict resolution strategy; the conflict resolution strategy determines the final assignment based on the relationship between neighborhood priority and probability peak, and writes the conflict points in the conflict map for use in the topology correction stage of S230.After the above reasoning and post-processing, the system output field is named Structure Segmentation Label Set. This output is a set of voxel annotations for four types of entities in global coordinates, with block synthesis records, conflict maps and time consistency markers. The Structure Segmentation Label Set is directly called in S230 for topology correction and surface reconstruction, and serves as the geometric source for meshing and wall mapping in S300. At the same time, it provides object-level visual elements for 3D mapping in S400.

[0072] S230. Perform topological correction and surface reconstruction on the structural segmentation label set to generate three-dimensional anatomical and blood flow structure set structural links.

[0073] The input source is the structural segmentation label set output by S220. This input provides voxel annotations for the tumor-bearing artery voxel, aneurysm sac voxel, blood flow guiding device voxel, and residual blood flow voxel in global coordinates, along with a conflict map, temporal consistency markers, and category summaries. Specifically, the system first enters the topology correction stage. Topology correction is used to restore reasonable anatomical connectivity within the label body, and the actions include centerline extraction, branch pruning, hole repair, and cross-class boundary cleanup. During centerline extraction, the system performs skeletonization within the tumor-bearing artery voxel to obtain a centerline with an approximate single-pixel width. Subsequently, path tracing is performed based on the anatomical reference system direction basis, pruning short pseudo-branches with large curvature abrupt changes to form a continuous vascular trunk. In the neighborhood of the aneurysm sac, the system detects the boundary gradient of the neck transition zone, restoring the contact zone between the aneurysm sac and the tumor-bearing artery, making the transition zone coherent. The branch pruning stage addresses burr branches caused by segmentation probability residuals by pruning according to the connected volume size, endpoint degree, and neighborhood consistency threshold. The hole repair stage searches for closed cavities in tumor cysts and instrument tag bodies, repairing isolated cavities through morphological expansion-contraction sequences. The cross-class boundary purification stage re-evaluates intersecting voxels based on the conflict map, updating their affiliations according to neighborhood dominance and temporal consistency markers to obtain non-overlapping segmented volumes. After topology correction, the system enters the surface reconstruction stage. Surface reconstruction generates closed, renderable, and computable geometric surfaces for various tag bodies within a unified coordinate system and establishes a bidirectional mapping with the voxel domain. For the aneurysm-bearing artery voxels and aneurysm sac voxels, the system calculates isosurfaces within the voxel domain to form a triangular mesh. Subsequently, a mesh smoothing and shape-preserving strategy at the boundaries is implemented to avoid excessive contraction in the aneurysm neck region. For the blood flow guiding device voxels, the system employs a stricter mesh density in the metal neighborhood to restore the texture of the slender stent and the overall cylindrical outline. For residual blood flow voxels, the system generates a dedicated surface for semi-transparent volume rendering while retaining the corresponding index with the velocity voxels, facilitating the subsequent overlay of flow visual elements in 3D space. After surface reconstruction, the system constructs a structure-level index; this index is a lookup structure connecting the geometric and voxel domains, with entries including object identifiers, mesh patch indexes, voxel coordinate ranges, centerline stake points, and neck boundary lines. This index supports the meshing and wall mapping operations of the S300 and provides an object lookup channel for the 3D mapping and rendering of the S400. In case of anomalies, if self-intersection or flipping is detected during the surface reconstruction stage, the system invokes the self-intersection detection and normal alignment process to repair erroneous patches. If a ruptured vascular trunk is detected, the system reverts to the topology correction stage, performs local expansion-contraction and refinement at both ends of the rupture, and then reconstructs again. Through the above processing, the system assembles a set of three-dimensional anatomical and blood flow structures at the object level. This set of anatomical and blood flow structures is a persistent data object, carrying the surface of the aneurysm-bearing artery, the surface of the aneurysm sac, the surface of the blood flow guiding device, and the residual blood flow surface, along with a structural-level index, temporal consistency markers, and a snapshot of the reconstruction parameters.The output field, named "3D Anatomy and Blood Flow Structure Set," will be directly used as input for meshing and wall mapping in S310. Simultaneously, the 3D Anatomy and Blood Flow Structure Set guides wall mesh construction and index calculation in S300, and drives 3D mapping and rendering in S400, forming a stable geometric base throughout subsequent modules. In summary, this step achieves the following technical results: by repairing connectivity and generating geometric surfaces on the structural segmentation tag set, a topologically sound, geometrically continuous, and bidirectionally mappable 3D anatomical and blood flow structure set is obtained, satisfying the object integrity requirements of subsequent meshing, wall mapping, and 3D visual rendering.

[0074] Step S300 includes at least steps S310-S330:

[0075] S310. Obtain the three-dimensional anatomical and blood flow structure set, perform meshing and wall mapping processing, and obtain the index calculation input set.

[0076] The input source is a 3D anatomical and blood flow structure set output from S230. This set includes the surface of the aneurysm-bearing artery, the aneurysm sac, the blood flow guiding device, and the residual blood flow surface, along with a structural-level index, temporal consistency markers, and reconstruction parameter snapshots. The structural-level index records object identifiers, mesh patch indexes, voxel coordinate ranges, centerline stakes, and neck boundary lines. Specifically, the system first loads objects in a unified coordinate system and then uses the structural-level index to restore the geometric dependencies and spatial adjacencies between objects. Addressing the differences in sampling density between different objects, the system performs a mesh quality pre-check on all surfaces, detecting overlapping, flipped faces, self-intersections, and acute-angled patches. Each detected surface is repaired, with repair strategies including local resampling, normal unification, boundary stitching, and angle reallocation. Processing records are written to object metadata, and abnormal segments are simultaneously written to the log. Furthermore, the system generates radial geodesic rings based on the centerline stake points to limit the local resampling range in the aneurysm sac region and prevent geometric collapse in the vicinity of the aneurysm neck. On the surface of the blood flow guiding device, the system establishes anisotropic sampling rules based on the stent texture scale, using a finer resolution along the texture direction and a moderate resolution along the circumferential direction to maintain the geometric recognizability of the overall cylindrical outline and local metallic texture. After completing quality pre-check and repair, the system enters the meshing stage; meshing in this step refers to converting the surface object into a closed, manifold triangular mesh that meets quality constraints, while simultaneously constructing the intraluminal sampling layer. In this stage, a highly regular triangular mesh is first generated on the surface of the aneurysm-bearing artery and the aneurysm sac, and then constraint point arrays are set on the boundary line of the aneurysm neck transition zone to control the boundary morphology of the transition zone; subsequently, an intraluminal sampling layer is generated near the residual blood flow surface. The sampling layer consists of multiple radial sampling lines emanating from the centerline and several ring surfaces, used to attach velocity samples. A bidirectional index relationship exists between the intracavitary sampling layer and the surface mesh. The index entries include surface vertex numbers, corresponding voxel coordinates, corresponding time frame numbers, and confidence weights, with the weights derived from the confidence markers in stage S130. Subsequently, the system performs wall mapping processing. In this step, wall mapping refers to mapping the velocity samples and residual blood flow distribution in the voxel domain to wall mesh vertices and mesh patches according to anatomical reference coordinates, generating a set of neighboring layer samples. The system first retrieves the corresponding entries between the residual blood flow surface and the intracavitary sampling layer, then reads velocity samples frame by frame according to the time frame number, completing the coordinate mapping from voxels to the mesh. The mapping action establishes a distance threshold in the proximity direction of the surface normal. Sampling points within the threshold generate neighboring layer samples through distance and confidence weights. Neighboring layer samples and mesh vertices form a one-to-many relationship, which is structured and stored in the object metadata. If a neighboring region has missing samples or gating anomalies, the system introduces a sliding window in the temporal neighborhood to collect substitute samples from adjacent frames and adds time interpolation markers to the entries.For the surface of the blood flow guide device, the system additionally generates a device neighborhood mapping entry to preserve the relative position and contact zone line between the device and the vessel wall. This entry consists of the device surface vertex, the nearest vessel wall vertex, and the local normal, and is subsequently used for index screening and geometric constraints. After the wall mapping is completed, the system generates a wall mesh, a set of neighboring layer samples, a projection of the centerline to the wall, a projection of the aneurysm boundary line to the wall, and a device neighborhood mapping entry at the object layer. All entries are linked to a unified time index and carry a quality scoring summary and a list of abnormal segments. After the above processing, the system obtains a structurally complete, coordinate-unified, time-labeled geometric-physical hybrid input object at the data level, which serves as the direct carrier for subsequent index calculations. In this step, this object is recorded as the output field name index calculation input set, and the subsequent consumption location is specified as the index calculation input set of S320 in the object metadata. At the same time, a lightweight rendering snapshot is registered in the visualization channel for reference by the S400 rendering process.

[0077] S320. Extract thrombus volume fraction, average wall shear stress, blood flow stagnation index and aneurysm neck enhancement thickness from the indicator calculation input set, perform normalization calibration, and generate indicator set step;

[0078] The input source is the index calculation input set output by S310. The index calculation input set includes wall grid, neighboring layer sample set, centerline to wall projection, neck boundary line to wall projection, and instrument neighborhood mapping entries, and carries a unified time index and quality score summary. Specifically, the system first enters the thrombus volume fraction extraction process; the thrombus volume fraction is defined as the ratio of the volume of voxels in a coagulated or poorly flowing state inside the cyst to the total volume of the cyst. The determination is based on two types of information: one is the range of solidified tissue indicated by low signal or high signal bands of wall adhesion in the HR-VWI sequence, and the other is the long-term low-velocity or resting characteristics of neighboring layer samples in the time domain. The system constructs a voxel mask under the surface envelope of the tumor cyst, reads samples from neighboring layers frame by frame using the time index, and marks long-term low-velocity segments as resting voxels. Simultaneously, it searches the voxel domain for the depth distribution of high-signal bands adhering to the wall and marks solidified voxels. The two types of voxels are merged after connectivity verification to generate thrombus candidates. If near-field perturbations of metal, indicated by device neighborhood mapping entries, exist in the vicinity of the tumor neck, the system reduces the resting determination weight of that region using a neighborhood exclusion strategy to prevent misclassification. After obtaining the thrombus candidates, the system outputs two statistics in the voxel domain: candidate volume and total tumor cyst volume, forming the original value of the thrombus volume fraction, and registers the candidate boundary and time coverage summary in the object metadata. Subsequently, the system proceeds to the average wall shear stress extraction process; the average wall shear stress is defined as the mean of the tangential force intensity generated by the near-wall velocity gradient on the wall mesh over the time dimension. The process retrieves neighboring layer samples for each grid vertex within the wall mapping entry, extracting them layer by layer according to the distance in the normal direction to generate a near-wall velocity profile. After local smoothing, the profile is mapped onto grid patches to obtain the tangential force intensity sequence for each patch. The sequences are statistically analyzed under a unified time index to form an average value. For the device neighborhood, the system classifies patches into contact or non-contact zones based on the device neighborhood mapping entry. Contact zones use an independent statistical channel, recording the contact zone area and sequence fluctuation for subsequent model sensitivity learning of device-wall interactions. Next, the system proceeds to the blood flow stagnation index extraction process. The blood flow stagnation index is defined as a combined indicator of the duration and coverage of neighboring layer samples in a low-velocity state over time. The system reads the velocity time series of neighboring layer samples for each grid patch, constructs a set of low-velocity interval segments, and calculates the duration, frequency, and coverage of these segments. If interpolation markers exist in the time index, the system adds a weighted marker to the interpolated segments within the segment set and retains the interpolation influence factor in the indicator entry to ensure that subsequent models can identify differences in data sources. Subsequently, the system extracts the enhancement thickness of the aneurysm neck; the enhancement thickness of the aneurysm neck is defined as the thickness estimate of the wall signal intensity rise zone along the aneurysm neck boundary line, reflecting the enhancement range of the wall layer in the neck region.The system acquires normal intensity profiles along the aneurysm neck boundary, uses segmented thresholding and morphological peak-valley localization methods to determine the outer edge of the high-signal ring, and then maps the outer edge point set back to the wall grid, using statistical normal distances to obtain thickness estimates. For boundary segments covered by devices, the system enables contact zone marking of the device neighborhood mapping entries and switches to an alternative intensity channel within the neighborhood to avoid false high-signal recordings caused by metal. All four types of indicators are initially extracted in the grid or voxel domain and then undergo normalization calibration. Normalization calibration refers to scaling, offset correction, and distribution alignment of indicator values ​​across devices, sequences, and patients. The correction factor is derived from the acquisition and calibration configuration snapshots and the distribution summary accumulated during the training phase. The system first selects the corresponding scaling scheme based on the scanning protocol identifier, coil layout, and voxel spacing, and then performs offset correction based on the distribution summary of historical similar data. If an indicator exhibits an abnormal distribution in the current batch, the system records the abnormal summary and adds a weight reduction flag to the entry for that indicator. After the above processing, the four types of indicators form a set of entries with a consistent scale. Each entry includes a statistical value, confidence weight, time coverage, and anomaly marker. The set of entries is assembled into a structured record at the object layer. The record items are organized according to object identifier, region identifier, and time index, and maintain a traceable relationship with the geometric object. After processing, the system registers this structured record as an output field name indicator set, which is subsequently consumed as the indicator set of S330. At the same time, a lightweight projection is generated in the rendering channel, which can be called by the 3D mapping process of S400 for overlaying numerical annotations in the 3D visual environment.

[0079] S330, Perform feature splicing and time-series alignment on the indicator set to generate individual feature vector structure;

[0080] The input source is the indicator set output by S320. The indicator set is a structured record, with records organized according to object identifier, region identifier, and time index, covering four categories of items: thrombus volume fraction, mean wall shear stress, blood flow stagnation index, and aneurysm neck enhancement thickness, and carrying confidence weights, time coverage, and anomaly markers. Specifically, the system first performs time-series alignment on all items based on a unified time index; time-series alignment refers to rearranging various items on a fixed time grid, and filling missing positions using window interpolation or step sampling, while retaining the interpolation identity and source label to prevent key statistics from being unseen and rewritten. For time periods with gated anomalies, the system only retains the statistics of stable segments, writes a downweighting marker on abnormal segments, and applies lower weights to related items in subsequent splicing stages. Subsequently, the system proceeds to feature derivation and region aggregation. Feature derivation refers to calculating advanced statistics reflecting morphological, amplitude, and time-varying characteristics from the original indicator entries. Statistical items include time-period average, time-period volatility, peak duration, peak frequency, and regional proportion. All statistics are generated using standard statistical processes without introducing formulas, and include source labels and window lengths. Region aggregation refers to aggregating statistical values ​​from bottom to top along the object hierarchy. The system first completes statistics at the grid patch level, and then aggregates along the patch-region-object path. The region boundary includes three sub-regions: the neck transition zone, the instrument contact zone, and the non-contact zone. Statistics for each of the three sub-regions are retained separately to prevent heterogeneity from being diluted. After completing feature derivation and region aggregation, the system performs feature filtering and consistent coding. Feature filtering compresses redundant or strongly correlated entries, prioritizing those with high correlation to weights during training. Consistent encoding transforms discrete identifiers into fixed-order encoded vectors. Object identifiers, region identifiers, and instrument contact zone markers use multi-hot encoding, while anomaly markers and deweighting markers exist as Boolean bits. Confidence weights and time coverage remain as numerical slots. Anomaly handling is implemented throughout the encoding process; if an entry is completely missing at the object level, the system creates a missing placeholder and attaches a structural missing marker and deweighting factor to it to prevent out-of-bounds access during subsequent inference. Subsequently, the system constructs an individual-level vector layout. This layout consists of a fixed-order set of slots arranged according to four types of indicators, three types of sub-regions, and various statistical derivations. Finally, summary slots for acquisition and calibration configurations, such as scanning protocol identifiers, coil layouts, and voxel spacing, are added to absorb cross-device differences. The system writes the values ​​obtained from the aforementioned encoding, aggregation, and derivation into the corresponding slots, forming a dense array with complete metadata. The array header includes the object identifier, inspection time, version number, and parameter snapshot number, while the array tail includes a complete list of fields and verification information, forming a traceable object. To match the reading specifications of subsequent inference channels, the system also generates a lightweight index. Index entries record the mapping from slot name to offset position, the partition where the weighting flag is located, the abnormal placeholder segment, and the length of the time alignment window.After processing, the system registers the output field name individual feature vectors at the object level and specifies in the metadata that the subsequent consumption location is the individual feature vector of S410; simultaneously, it retains a summary in the rendering pipeline for S400 to attach numerical logos and time profile curves in 3D mapping. The technical effect of this step can be summarized as follows: by completing index rearrangement, statistical derivation, and consistent encoding at a unified time grid and unified geometric level, a structurally stable, traceable, and multi-source information-covering individual feature vector is formed, meeting the strict requirements for input form and order in subsequent inference stages.

[0081] Step S400 includes at least steps S410-S430:

[0082] S410. Obtain individual feature vectors, perform inference on the trained model, and obtain mechanism scoring and probability results.

[0083] The input source is the individual feature vector output by S330. The individual feature vector is a dense array obtained through feature derivation, region aggregation, and uniform encoding under a unified time grid and unified geometric level. The slot order is fixed, and it carries object identifier, region identifier, instrument contact zone marker, confidence weight, time coverage, abnormal occupancy information, and summary slots for acquisition and calibration configuration. Specifically, the system first reads the version number and parameter snapshot number of the individual feature vector during the inference session initialization phase, and loads the corresponding list of trained models. The list of trained models consists of a multi-channel input neural network (NN) branch, an ensemble learning branch, and a temporal sub-model based on sequential data. The three work together in the inference phase according to a preset fusion strategy. Understandably, the neural network branch handles the coupling relationship between high-dimensional continuous slots and multi-hot encoded slots, the ensemble learning branch forms robust boundary judgments in a low-noise, low-dimensional statistical space, and the temporal sub-model sequentially models temporal coverage and segment persistence features. The structure and weights of these three components are frozen during the training phase and are only read, not updated, during the inference phase. The system splits individual feature vectors into continuous feature sub-vectors, discrete encoded sub-vectors, and temporally derived sub-vectors according to the model input description, and feeds them into the neural network branch, ensemble learning branch, and temporal sub-model, respectively. Before entering each branch, the system applies weight decay to the corresponding slots based on the weight reduction marker in the individual feature vectors and records the weight reduction strategy index in the session context to ensure that the source of anomalies is identified. Furthermore, the inference scheduler initiates inference for each branch according to the order defined in the model list; after the neural network branch completes forward propagation, it outputs intermediate representations related to three mechanisms, corresponding to blood flow occlusion-related features, thrombosis-related features, and endothelialization-related features, respectively; the ensemble learning branch refines the boundary sensitive points based on the statistical entries after region aggregation, and outputs the independent confidence scores for the three mechanisms; the temporal sub-model sequentially models low-speed continuous segments, volatility, and temporal coverage, and outputs supplementary confidence scores reflecting temporal behavior. The inference scheduler then enters the fusion stage, calling the fusion strategy to map the three intermediate results to a unified mechanism scoring space; the fusion strategy provides adaptive weighting based on branch confidence and slot weighting labels, generating three numerical values: blood flow occlusion mechanism score, thrombosis mechanism score, and endothelialization mechanism score, and registers the fusion parameter summary in the object metadata layer. To enhance the traceability of inference, the system also generates interpretability entries; the interpretability entries record the main source slots, regional contribution decomposition, and temporal segment contribution decomposition for each mechanism score, and attach rendering prompts for the S420 to overlay local contributions in a 3D visual environment.Subsequently, the system enters the outcome probability generation stage. This stage primarily uses the mechanism score as input, while also incorporating the collection and calibration configuration summary slots from individual feature vectors for out-of-domain calibration. The system outputs the complete occlusion probability and the interim outcome probability through a frozen posterior mapping module. During the training phase, the posterior mapping module obtains probability mapping parameters based on historical follow-up samples; during the inference phase, it only performs numerical mapping and threshold pruning. Throughout the inference session, the system records the model version, weight checksum, input summary hash, inference time, and abnormal interruption information. If the proportion of missing input slots exceeds the session threshold, the system triggers a degradation path, using a simplified fusion strategy to generate the mechanism score solely based on stable slots, and attaches a degradation marker to the output entries, prompting subsequent visualization processes to perform style degradation during rendering. After processing, the system assembles the three mechanism scores, the probability of complete occlusion, the probability of staged outcome, and their interpretability entries into an object-level data structure and registers them as output field names: mechanism score and probability result. The subsequent consumption location is the mechanism score and probability result of S420. At the same time, lightweight curves and contribution heatmap hook handles are generated in the visualization channel for S420 to call when performing 3D mapping and rendering. The version index entry for this inference is reserved in the version registration list of S430, forming a sequential transmission from S410 to S420 and then to S430.

[0084] S420. Extract the occlusion status level and complete occlusion probability from the mechanism score and probability results, perform three-dimensional mapping and rendering, and generate a three-dimensional visualization and follow-up comparison report.

[0085] The input source is the mechanism score and probability result output by S410. The mechanism score and probability result include the blood flow occlusion mechanism score, thrombosis mechanism score, endothelialization mechanism score, complete occlusion probability, and staged outcome probability, and carries interpretability entries and an inference version index. Specifically, the system first enters the state level extraction process; the state level is a hierarchical description of the occlusion state, generated by numerical segmentation and rule combination. The rule combination simultaneously references the three mechanism scores and the complete occlusion probability, and is fine-tuned based on the regional decomposition results of the device contact zone and the aneurysm neck transition zone. The system reads the regional contribution decomposition from the interpretability entries, calculates the local judgment intensity for the aneurysm neck boundary neighborhood and the device contact zone respectively, and forms a hierarchical judgment. When the three mechanism scores contradict each other, the system calls the conflict resolution strategy, prioritizing the one with more stable time coverage, and records the source of conflict and the basis for resolution in the level entry. After the state level is determined, the system constructs a 3D mapping dataset. This dataset uses the 3D anatomical and blood flow structure set output by S230 as its geometric base, and attaches a lightweight projection of the mechanism score contribution heatmap from S410 and the indicator set from S320 for overlay display in a 3D environment. Understandably, 3D mapping refers to the process of spatially locating, color mapping, and controlling the transparency of numerical entries on the surface of a 3D object or an intracavitary sampling layer. The system locates the display position based on the object identifier and mesh patch index, selects different rendering channels based on the region identifier, and uses separate material parameters for the instrument contact zone to avoid conflicts with the vessel wall surface style. Subsequently, the system enters the rendering process, which consists of three parts: scene construction, camera path planning, and multi-channel overlay. In the scene construction phase, the system loads the surface of the aneurysm-bearing artery, the surface of the aneurysm sac, the surface of the blood flow guiding device, and the surface of residual blood flow, restoring the lighting, shadow, and transparency parameters. In the camera path planning phase, the system generates a standard browsing path based on the centerline stakes and the aneurysm neck boundary line, and places key viewpoint nodes on the path. In the multi-channel overlay phase, the system maps the mechanism score contribution heatmap to the wall and intraluminal sampling layers, attaches the probability of complete occlusion and the status level to the corresponding areas in the form of numerical badges and text overlays, and adds lightweight projections of thrombus volume fraction, average wall shear stress, blood flow stagnation index, and aneurysm neck enhancement thickness from S320 to the tool panel for interactive inspection. Furthermore, the system generates follow-up comparison views; if there are multiple versions of the same object that have been checked, the system retrieves historical scenes from the version registration index, replays the historical views under the same camera path and the same rendering parameters, and arranges the comparisons in the form of a timeline to form an intuitive expression of temporal changes; for cross-device or cross-protocol checks, the system calls the scale conversion card in the acquisition and calibration configuration summary slot, applies scale alignment in the rendering layer, and ensures that the color and numerical annotations of different batches are visually consistent.After rendering, the system proceeds to report construction. Report construction integrates scene screenshots, numerical badges, level determinations, and version indexes into a document object, and attaches traceable metadata. The system organizes chapters according to object identifiers and inspection times, embedding key perspective screenshots, enlarged details, and tool panel numerical snapshots in each chapter. Simultaneously, interpretable entries from S410 and summary indicators from S320 are added to the appendix, forming verifiable data endorsement. Anomaly logging is maintained throughout the report construction process; if certain areas have input weight reduction or degradation markers, the system displays prompts at the corresponding locations in the report and lists the scope of impact and handling strategies in the appendix, ensuring readers can reconstruct the process during review. After processing, the system packages the document object, scene data, and version index into an object-layer data structure, registering it as the output field name "3D Visualization and Follow-up Comparison Report," and indicating in the object metadata that the subsequent consumption location is the S430 "3D Visualization and Follow-up Comparison Report." Simultaneously, a scene replay handle is retained in the visualization channel for subsequent review and re-rendering. This output also works in conjunction with the cross-step channels of S200 and S300: when a report is opened for clinical review or research verification, the system can trace back the three-dimensional anatomical and blood flow structure set and indicator set based on the object identifier to ensure consistent citation across steps.

[0086] S430. Perform parameter backfilling and version registration for the 3D visualization and follow-up comparison report, and generate the data collection and calibration configuration structure.

[0087] The input source is the 3D visualization and follow-up comparison report output by S420. This report includes scene screenshots, level determinations, numerical badges, rendering parameters, version index, and indicator summaries and interpretability entries in the appendix. Specifically, the system first enters the parameter backfilling process. Parameter backfilling refers to extracting key parameters, judgment thresholds, and scale conversion cards from the report object and writing them into traceable slots in the acquisition and calibration configuration, forming a closed-loop dependency. The system retrieves the model version, fusion strategy, posterior mapping parameters, and rendering style cards referenced in this inference from the report's version index and writes these entries into the corresponding fields according to the structure of the acquisition and calibration configuration. Simultaneously, the system extracts the scale conversion scheme, anomaly weighting strategy, and time interpolation window length of key indicators from the report appendix, updating the data governance entries in the configuration so that the next check can directly reference the updated strategy set during the S110 calibration and mapping processing stage. Understandably, parameter backfilling also includes change feedback from both the device and protocol sides. If the report records alignment cards across devices or protocols, the system adds that card as a new entry to the protocol mapping table of the data acquisition and calibration configuration. If the level determination in the report triggers a conflict resolution strategy, the system writes the resolution criteria into the configured rule set, forming a replayable resolution chain. Subsequently, the system enters the version registration process. Version registration refers to creating a version record for this inspection in the data registration module. The record items include the object identifier, inspection time, model version, parameter snapshot number, interpretability entry hash, scene replay handle, and report storage location. After version registration is completed, the system maintains a chain in the registration table from the current record to the historical record. The chain nodes are arranged in chronological order for easy retrieval in the S420 follow-up comparison view. To ensure operational security and traceability, the system triggers read-only snapshot generation during version registration, summarizing and persisting key intermediate products of the current session, including mechanism scoring and probability result summaries, indicator set summaries, and rendering parameter summaries. If degradation paths or abnormal placeholders exist during operation, the system emphasizes and records them in the snapshot metadata, attaching the scope and source of the anomaly to the registration entry for easy review later. After parameter backfilling and version registration, the system constructs an updated acquisition and calibration configuration. This configuration is structurally consistent with the configuration referenced by S110, including scanning protocol identifier, coil channel layout, coordinate system definition, velocity encoding range, cardiac triggering and respiratory gating strategies, timestamp generation rules, DICOM label mapping table, anonymization rules, and sequence archiving paths. It also adds or updates policy entries, scale conversion cards, rule set patches, and model version anchors from the current session.The system registers this configuration as the output field name for collection and calibration configuration at the object layer, and explicitly specifies the subsequent consumption location as the collection and calibration configuration of S110 in the object metadata, forming a closed loop from S430 back to S110. Furthermore, the system records configuration change audit entries in the security module, including the change time, operator identifier (marked as a system account if in an automated process), change summary, and rollback handle, supporting restoration to any historical version when necessary via the rollback handle. After processing, the system sends a lightweight notification to the rendering channel, containing a summary of the new configuration and version anchors, for subsequent checks to read the default style and alignment card during scene initialization, thus forming a stable parameter inheritance relationship across the main step's execution path. In summary, the technical effect of this step is: by establishing a parameter backfilling and version registration channel between the report object and the configuration object, the execution process is encapsulated as a traceable, replayable, and iterative configuration closed loop, and the output collection and calibration configuration provides a complete and consistent strategy foundation for the next direct S110 call.

[0088] As a specific embodiment of the present invention, a follow-up study of a 50-year-old female patient 6 months after implantation of a blood flow diverter device for a left internal carotid artery posterior communicating segment aneurysm was conducted. First, a head scan was performed on the patient using a 3.0T magnetic resonance imaging (MRI) scanner to acquire raw data of 3D high-resolution vessel wall imaging sequences and four-dimensional blood flow MRI sequences. The 3D high-resolution vessel wall imaging sequence included plain and enhanced scans, and the four-dimensional blood flow MRI sequence used specific velocity encoding values. Subsequently, the imaging data of the multimodal sequences were transmitted to the system. The system's image processing module automatically performed image registration and fusion. Using a pre-trained deep learning segmentation model, the carrier artery, aneurysm contour, and blood flow diverter device location were automatically segmented on the fused image, and a three-dimensional model was generated. The system automatically calculated quantitative parameters, including the total volume of the aneurysm cavity, thrombus volume, aneurysm neck enhancement features, and the average flow velocity, wall shear stress, and blood flow stagnation index obtained from the four-dimensional blood flow data. Subsequently, the system input the extracted quantitative parameters into a pre-trained machine learning prediction model. The model outputs a mechanistic score, including the occlusion status level and sub-scores for blood flow obstruction, thrombosis, and endothelialization, and predicts the probability of complete stable occlusion. Finally, the system generates a comprehensive assessment report and presents a 3D model integrating blood flow, thrombosis, and enhancement information, a predicted score card, and parameter trend graphs on the interface.

[0089] Example 2: Figure 4 This diagram illustrates the overall architecture of a multimodal magnetic resonance aneurysm occlusion assessment system according to an embodiment of the present invention. Figure 4As shown, the system adopts a layered design, mainly comprising a data acquisition layer, a data processing and analysis layer, a support layer, and an intelligent evaluation and output layer from top to bottom. The data acquisition layer acquires multimodal magnetic resonance raw data through a magnetic resonance scanning device. The data processing and analysis layer sequentially performs preprocessing, image segmentation, and feature extraction and quantization operations on the raw data. The support layer provides historical data support and model iteration capabilities for the system. The intelligent evaluation and output layer uses a machine learning model to perform individualized evaluation based on the processed data and generates visualized results. The layers are connected via data flow, forming a complete closed-loop system from data acquisition to intelligent evaluation.

[0090] Example 3: Figure 5 A structural block diagram of a multimodal magnetic resonance aneurysm occlusion assessment system according to an embodiment of the present invention is shown. Figure 5 As shown, the structure may include:

[0091] The data acquisition and calibration module 01 is used to acquire the acquisition and calibration configuration, 3D HR-VWI sequence, and 4D-Flow sequence, and perform calibration mapping processing to generate the original imaging and velocity field. Specifically, it receives the acquisition and calibration configuration backfilled from the visualization and update module, along with the 3D HR-VWI sequence and 4D-Flow sequence generated by the MRI console, and performs calibration mapping processing according to the scanning protocol identifier and coordinate system definition in the acquisition and calibration configuration. This includes reading configuration parameters, processing sequence data, and completing gain equalization to form the original imaging and velocity field. The original imaging and velocity field are then transmitted to the cross-modal registration and fusion module as input.

[0092] The cross-modal registration and fusion module 02 is used to extract the anatomical reference frame and velocity voxels from the original image and velocity field, perform cross-modal registration, and perform artifact suppression and denoising on the cross-modal registration voxel set to generate a cross-modal registration fusion voxel stack. Specifically, it receives the original image and velocity field from the data acquisition and calibration module, extracts the anatomical reference frame from the original image, extracts velocity voxels from the velocity field, performs coarse registration using a mutual information-based metric and fine registration using a multi-resolution strategy to drive non-rigid deformation, and performs artifact suppression by frequency domain directional filtering and spatial direction consistency evaluation on the registered voxel set to generate a cross-modal registration fusion voxel stack. The cross-modal registration fusion voxel stack is then passed to the training set unification module as an input object.

[0093] The training set unification module 03 is used to perform training set unification processing on the cross-modal registration and fusion voxel stack to obtain segmentation input voxels. Specifically, it receives the cross-modal registration and fusion voxel stack from the cross-modal registration and fusion module, performs intensity normalization strategy alignment based on the intensity statistical summary of historical training data, generates candidate regions within the anatomical reference system defined region, completes block-level slicing and multi-channel assembly, and forms segmentation input voxels. The segmentation input voxels are then passed to the depth segmentation and topology correction module as input objects.

[0094] The depth segmentation and topology correction module 04 is used to extract the aneurysm-bearing artery voxel, aneurysm sac voxel, blood flow guiding device voxel, and residual blood flow voxel from the segmentation input voxels, perform depth segmentation, and perform topology correction and surface reconstruction on the structural segmentation label set to generate a three-dimensional anatomical and blood flow structure set. Specifically, it receives the segmentation input voxels from the training set unification module, uses an encoder-decoder network to perform depth segmentation to generate a structural segmentation label set, then extracts the centerline through skeletonization and performs branch pruning and hole repair to complete topology correction, and then performs surface reconstruction to generate a three-dimensional anatomical and blood flow structure set; the three-dimensional anatomical and blood flow structure set is then passed to the meshing and index calculation module as input.

[0095] The meshing and index calculation module 05 is used to perform meshing and wall mapping processing on the three-dimensional anatomical and blood flow structure set to obtain an index calculation input set. It then extracts thrombus volume fraction, average wall shear stress, blood flow stagnation index, and aneurysm neck enhancement thickness from the index calculation input set, performs normalization calibration, and generates an index set. Specifically, it receives the three-dimensional anatomical and blood flow structure set from the depth segmentation and topology correction module, generates a radial geodesic ring based on centerline stake points, performs meshing and wall mapping to obtain an index calculation input set, extracts the indices from the index calculation input set, and performs normalization calibration based on the scanning protocol identifier and coil layout, selecting a scale conversion scheme to generate an index set. The index set is then passed to the feature vector generation module as input.

[0096] The feature vector generation module 06 is used to perform feature concatenation and time-series alignment on the indicator set to generate individual feature vectors. Specifically, it receives the indicator set from the gridding and indicator calculation module, rearranges the entries on a unified time grid and derives advanced statistics such as time period mean and time period volatility, completes feature concatenation and time-series alignment, and generates individual feature vectors. The individual feature vectors are then passed to the model inference module as input objects.

[0097] The model inference module 07 is used to obtain the individual feature vector, perform inference on the trained model, and obtain mechanism scores and probability results. Specifically, it receives individual feature vectors from the feature vector generation module, and performs fusion inference using a multi-channel input neural network branch, an ensemble learning branch, and a temporal sub-model to generate blood flow blocking mechanism scores, thrombosis mechanism scores, endothelialization mechanism scores, complete occlusion probability, and staged outcome probability. The mechanism scores and probability results are then passed to the visualization and update module as input.

[0098] The visualization and update module 08 is used to extract the occlusion state level and the probability of complete occlusion from the mechanism score and probability results, perform 3D mapping and rendering, generate a 3D visualization and follow-up comparison report, and perform parameter backfilling and version registration on the 3D visualization and follow-up comparison report to generate a data acquisition and calibration configuration. Specifically, it receives the mechanism score and probability results from the model inference module, extracts the occlusion state level and the probability of complete occlusion, generates a standard browsing path based on the centerline stake points and the neck boundary line, performs 3D mapping and rendering, generates a 3D visualization and follow-up comparison report, then extracts key parameters from the report for parameter backfilling and version registration, and generates a data acquisition and calibration configuration; the data acquisition and calibration configuration is backfilled into the data acquisition and calibration module for subsequent processing.

Claims

1. A multimodal magnetic resonance imaging (MRI) method for assessing aneurysm occlusion, characterized in that, include: Acquire acquisition and calibration configuration, 3D HR-VWI sequence and 4D-Flow sequence. Perform calibration mapping processing based on the scanning protocol identifier and coordinate system definition in the acquisition and calibration configuration. Use mutual information-based metric for coarse registration and multi-resolution strategy to drive non-rigid deformation for fine registration. Perform cross-modal registration. Use frequency domain directional filtering and spatial direction consistency evaluation to jointly determine the main direction of the fringes and perform suppression along the main direction. Perform artifact suppression and denoising processing to generate a cross-modal registration fusion voxel stack structure. The training set is aligned by intensity normalization strategy based on intensity statistical summary of historical training data, and deep segmentation is performed by multi-scale feature extraction and cross-layer feature stitching using encoder-decoder network. The topology correction and surface reconstruction are performed by extracting the center line through skeletonization and pruning branches and repairing holes to generate a three-dimensional anatomical and blood flow structure set. The process involves performing gridding and wall mapping based on the centerline stake points to generate radial geodesic loops, index normalization calibration based on the selection of scale conversion scheme and offset correction based on scanning protocol identifiers and coil layout, and rearranging entries on a unified time grid to derive the time period mean and time period fluctuations through feature splicing and time sequence alignment to obtain individual feature vectors. The process involves performing inference on trained models using multi-channel input neural network branches, ensemble learning branches, and temporal sub-models; generating 3D mapping and rendering of standard browsing paths based on centerline stakes and neck boundary lines; parameter backfilling and version registration processing by backfilling model version, fusion strategy, and scale conversion cards; and generating acquisition and calibration configuration structures.

2. The method according to claim 1, characterized in that, Trained model inference includes: Read the version number and parameter snapshot number of the individual feature vector, and load the corresponding list of trained models. The list of trained models consists of a neural network branch with multi-channel input, an ensemble learning branch, and a temporal sub-model based on sequential data. The three work together in the inference phase according to a preset fusion strategy.

3. The method according to claim 2, characterized in that, Trained model inference also includes: The model is split into continuous feature sub-vectors, discrete encoding sub-vectors, and temporal derived sub-vectors, which are then fed into the neural network branch, ensemble learning branch, and temporal sub-model, respectively. The inference scheduler starts the inference of each branch in the order defined in the model list.

4. The method according to claim 3, characterized in that, Trained model inference also includes: The fusion strategy maps the three intermediate results to a unified mechanism scoring space. The fusion strategy provides adaptive weighting based on branch confidence and slot weighting labeling, generating three values: blood flow blocking mechanism score, thrombosis mechanism score, and endothelialization mechanism score.

5. The method according to claim 1, characterized in that, 3D mapping and rendering include: A standard browsing path is generated, and key viewpoint nodes are set up on the path; in the multi-channel overlay stage, the heat map of mechanism score contribution is mapped to the wall and cavity sampling layers, and the probability of complete occlusion and state level are attached to the corresponding areas in the form of numerical logos and text overlays.

6. The method according to claim 5, characterized in that, The process of 3D mapping and rendering also includes: If multiple versions of the same object are checked, the historical scene is retrieved from the version registration index, the historical view is replayed under the same camera path and the same rendering parameters, and arranged and compared in the form of a timeline to form an intuitive expression of the temporal change.

7. The method according to claim 1, characterized in that, Parameter backfilling and version registration include: Enter the parameter backfilling process; parameter backfilling refers to extracting key parameters, judgment thresholds and scale conversion cards from the report object during the operation and writing them into the traceable slots of the collection and calibration configuration to form a closed-loop dependency.

8. The method according to claim 7, characterized in that, The process of parameter backfilling and version registration also includes: Enter the version registration process; version registration refers to creating a version record for this inspection in the data registration module. The record items include object identifier, inspection time, model version, parameter snapshot number, interpretability entry hash, scene replay handle, and report storage location.

9. The method according to claim 8, characterized in that, The parameter backfilling and version registration process also includes: Construct updated data collection and calibration configurations; add or update policy entries, scale conversion cards, rule set patches, and model version anchors from this session.

10. A multimodal magnetic resonance aneurysm occlusion assessment system, applied to the method of any one of claims 1-9, characterized in that, include: The data acquisition and calibration module is used to acquire acquisition and calibration configurations, 3D HR-VWI sequences and 4D-Flow sequences, and perform calibration mapping processing to generate raw images and velocity fields; The cross-modal registration and fusion module is used to extract the anatomical reference frame and velocity voxels from the original imaging and velocity field, perform cross-modal registration, and perform artifact suppression and denoising on the cross-modal registration voxel set to generate a cross-modal registration fusion voxel stack. The training set unification module is used to perform training set unification processing on the cross-modal registration fusion voxel stack to obtain segmented input voxels; The depth segmentation and topology correction module is used to extract the tumor-bearing artery voxel, aneurysm sac voxel, blood flow guiding device voxel and residual blood flow voxel from the segmentation input voxel, perform depth segmentation, and perform topology correction and surface reconstruction on the structural segmentation label set to generate a three-dimensional anatomical and blood flow structure set. The meshing and index calculation module is used to perform meshing and wall mapping processing on the three-dimensional anatomical and blood flow structure set to obtain the index calculation input set. The module extracts thrombus volume fraction, average wall shear stress, blood flow stagnation index and aneurysm neck enhancement thickness from the index calculation input set, performs normalization calibration, and generates an index set. The feature vector generation module is used to perform feature concatenation and time-series alignment on the indicator set to generate individual feature vectors. The model inference module is used to obtain the individual feature vector, perform inference on the trained model, and obtain mechanism scores and probability results. The visualization and update module is used to extract the occlusion status level and complete occlusion probability from the mechanism scoring and probability results, perform three-dimensional mapping and rendering, generate a three-dimensional visualization and follow-up comparison report, and perform parameter backfilling and version registration on the three-dimensional visualization and follow-up comparison report, and generate collection and calibration configuration.

Images