Multi-mode magnetic resonance stent restenosis prediction method and system

Abstract

The invention relates to the field of medical health informatics and clinical decision support, in particular to a multi-mode magnetic resonance stent restenosis prediction method and system. The method comprises the steps of obtaining a multi-modal magnetic resonance image and structured clinical data, and performing preprocessing to generate an alignment index structure; performing segmentation and feature extraction on the image to obtain static and dynamic feature sets; performing normalization and combined coding on the features and the clinical indexes to generate individual feature vectors, performing reasoning by using a pre-training model, constructing evidence tracing mapping capable of being traced back to original data, and generating a risk probability and evidence tracing packet; and finally, performing multi-layer visual rendering, structured report typesetting and parameter adaptive adjustment to generate a visual report including image comparison, a trend curve and the like. According to the method, the multi-modal data and the clinical information are integrated, and through the interpretable prediction model and automatic report generation, the accuracy of stent restenosis prediction and the clinical decision support efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical health informatics and clinical decision support, and particularly relates to a multi-modal magnetic resonance stent restenosis prediction method and system. BACKGROUND

[0002] Stent implantation for intracranial atherosclerotic stenosis (ICAS) is an important means of preventing stroke, but in-stent restenosis (ISR) is a major long-term complication, with an incidence of 10-30%, which seriously threatens the prognosis of patients. ISR is a gradual pathophysiological process, and its core mechanism is in-stent neoatherosclerosis (ISNA), which is characterized by smooth muscle cell proliferation, lipid deposition, foam cell formation, and even new plaque hemorrhage, ultimately leading to lumen restenosis.

[0003] Current clinical management strategies for ISR have significant lag:

[0004] (1) Dependence on late lumen assessment: Current clinical practice mainly relies on follow-up imaging (such as DSA, CTA, MRA) to "diagnose" significant lumen stenosis (usually defined as a lumen stenosis rate of ≥50%) that has already occurred. This is a "post-confirmation" mode, and the patient is already at high risk of stroke at this time, missing the best intervention window.

[0005] (2) Ignoring early process monitoring: The development of ISNA is a dynamic process, much earlier than significant lumen stenosis. Conventional imaging mainly focuses on the "lumen" as the end result, and cannot effectively assess the active pathological processes of the vessel wall (such as inflammation, neovascularization, and intimal proliferation rate), which are key early biomarkers for predicting restenosis.

[0006] (3) Lack of effective prediction tools: Although high-resolution magnetic resonance wall imaging (HR-MRI) can excellently display the structure and composition of the vessel wall, existing technology is only used to describe the current situation (such as finding plaques and assessing wall remodeling). How to extract quantitative features from baseline or early follow-up HR-MRI images and build a model that can individualize the prediction of future restenosis risk is still a technical gap in the current field.

[0007] In addition, in the field of medical health informatics and clinical decision support, the existing scheme for multi-modal magnetic resonance stent restenosis prediction is usually based on a data processing flow combining single-time-point image reading and empirical rules, including image preprocessing, coarse registration, region delineation, and conventional report output, etc. There are limitations such as unstable cross-time-point registration alignment, disconnection between candidate region positioning and segmentation and statistics link, and non-standard clinical initial form field fusion. Existing methods rely on manual judgment and file transfer between independent subsystems, which is prone to static feature set missing and dynamic feature set inconsistency in postoperative follow-up management and long-term evaluation scenarios, making it difficult to meet the stable implementation of individual feature vector construction and prediction reasoning. For the joint processing of multi-modal magnetic resonance and clinical initial forms, the existing technology generally has shortcomings in synchronization under the reference of alignment index structure and evidence mask, in the determination of evidence trace mapping and threshold mapping, and in the control of visual layer rendering, template matching and parameter landing, making it difficult to form a continuous process of acquisition, alignment, segmentation quantization, individual feature vector construction, prediction reasoning, report layout and parameter update in the application scenario of stent restenosis prediction, resulting in the deficiencies of risk probability and evidence trace package being non-recoverable and report configuration structure and processing link being separated. SUMMARY

[0008] To solve the above technical problems, the present application provides a multi-modal magnetic resonance stent restenosis prediction method, comprising:

[0009] Obtain multi-modal magnetic resonance containing high-resolution blood vessel wall imaging, contrast-enhanced blood vessel wall imaging and susceptibility-weighted imaging, and structured clinical initial forms covering lipid metabolism, glucose metabolism, blood pressure history, smoking history and recent medication records, perform standardization and artifact suppression, registration alignment and alignment index generation processing based on report configuration structure, and generate alignment index structure;

[0010] Perform candidate region positioning and segmentation processing, statistics and denoising, cross-time-point alignment and change calculation processing, obtain segmentation results and evidence mask, static feature set and dynamic feature set;

[0011] Perform normalization including segmental quantile mapping and robust limiting, and combined encoding of one-hot encoding and ordered binning encoding, generate individual feature vector; perform prediction model reasoning based on the model artifacts trained and verified offline, and construct trace mapping table from output binning back to input feature segment and source set and coordinates, generate risk probability and evidence trace package;

[0012] Perform visualization layer rendering including cross-section pairing and voxel mapping, report layout based on page template and content block structure, template matching and threshold adjustment and parameter landing processing according to task type and department habit, construct image contrast and reinforcement annotation, trend curve, visual report draft and report configuration structure.

[0013] Further, the candidate region positioning and segmentation processing, statistics and denoising, cross-time point alignment and change calculation processing are performed to obtain segmentation results and evidence masks, static feature sets and dynamic feature sets, specifically including:

[0014] The evidence mask is morphologically refined to generate a boundary vector curve; local noise suppression and transparency gradient are performed on the segmentation boundary neighborhood; image contrast views are generated on the same section according to the alignment index structure, and magnification windows are automatically inserted in the stent neighborhood and upstream and downstream paragraphs.

[0015] Further, the normalization and combined encoding processing including piecewise quantile mapping and robust clipping and the ordered binning encoding and combined encoding processing are performed to generate individual feature vectors, specifically including:

[0016] Normalization refers to implementing scale unification and robust clipping on numerical value fields, scale unification is completed by piecewise quantile mapping, and robust clipping is performed by truncating extreme quantile outside items and retaining truncation markers; the combined strategy of one-hot encoding and ordered binning encoding is used for category fields.

[0017] Further, the prediction model inference is performed based on the model artifacts trained and verified offline, and a traceability mapping table from output grading back to input feature segments and source sets and coordinates is constructed, specifically including:

[0018] The traceability mapping table is constructed to record the mapping relationship from output grading to input feature segments to source sets and coordinates, wherein the source sets include static feature sets, dynamic feature sets and clinical indicator sets, and the coordinates are given by the alignment index structure and the evidence mask.

[0019] Further, the process of constructing the traceability mapping table further includes:

[0020] First, according to the position table, the feature domain and key name are traced back, and then the original set item is located through the key name and task number in the domain; if the key name belongs to the dynamic feature domain, the cross-time point pairing relationship is further traced back, and the mask version and section position are associated; if the key name belongs to the static feature domain, the class label and time point in the segmentation result and evidence mask are associated; if the key name belongs to the clinical indicator domain, the time window and source marker of the clinical item are associated.

[0021] Further, the process of visualizing layer rendering including cross-section pairing and voxel mapping further includes:

[0022] The cross-section pairing relationship and voxel mapping relationship are established between the baseline period and follow-up period image groups according to the alignment index structure; the layer superposition order is set in the rendering engine, the basic layer is the original image after resampling, the evidence layer is the evidence mask and the segmentation boundary, and the indicator layer is the annotation element pointed by the traceability package at the spatial position.

[0023] Further, the process of report layout based on page template and content block structure further comprises:

[0024] Under the constraints of page template and content block structure, the image contrast view, evidence annotation layer and trend curve are laid out according to the column and partition, and the textual explanation, grading prompt and reference tag are filled according to the field mapping table; the page template and content block structure have version labels under the task number, which are used for unified page margin, column title, caption style and font size level.

[0025] Further, the process of template matching and threshold adjustment and parameter landing according to task type and department habit further comprises:

[0026] Template matching refers to selecting the optimal template in the template registry according to task type, department habit and number of maps; threshold adjustment refers to synchronous correction of grading text, legend threshold and annotation color scale according to risk probability and grading threshold reference; parameter landing refers to writing final layer rendering parameters, format parameters and threshold parameters into the configuration warehouse, and generating version labels and access control entries.

[0027] Further, the report configuration structure is used as the calling source of pre-processing parameter configuration at the cross-main step level, and is used for parameterization of sequence resampling, intensity standardization, artifact suppression and registration options when a new task is started; at the same time, the structure is used for parameter initialization of the next format and rendering process inside the main step, forming a parameter sedimentation path for task internal loop.

[0028] Further, a multi-modal magnetic resonance stent restenosis prediction system applied to the method of any one of the above, comprising:

[0029] The registration alignment and alignment index generation module is used for standardization, artifact suppression, registration alignment and alignment index generation from the baseline and follow-up multi-modal magnetic resonance and clinical initial table, outputs the alignment index structure, and outputs the call for the segmentation quantization and static and dynamic feature construction module;

[0030] The segmentation quantization and static and dynamic feature construction module is used for candidate region positioning and segmentation processing, statistics and denoising, cross-time point alignment and change calculation under the reference of the alignment index structure, outputs the segmentation result and evidence mask, static feature set and dynamic feature set, and outputs the call for the individual feature vector construction and prediction reasoning module;

[0031] The individual feature vector construction and prediction inference module is used for performing arrangement, normalization and coding on the static feature set, the dynamic feature set and the clinical index set, sending the individual feature vector into the inference pipeline and performing signal bit masking, domain weight scheduling, model calculation and grading threshold mapping, outputting the risk probability and the evidence traceability package, and outputting for calling of the evidence summary report rendering and configuration updating module;

[0032] The evidence summary report rendering and configuration updating module is used for completing visual layer rendering, report layout, template matching, threshold adjustment and parameter landing under the reference of the alignment index structure, the segmentation result and the evidence mask, outputting image contrast, strengthened annotation and trend curve, visual report draft and report configuration structure, and the report configuration structure is called by the registration alignment and alignment index generation module.

[0033] The key innovation points of the application include:

[0034] (1) The alignment index structure is formed in the registration alignment and alignment index generation, and is directly called by the segmentation quantization and static and dynamic feature construction, and the field organization and reference link are through the candidate region positioning and segmentation processing, statistics, denoising, cross-time point alignment and change calculation.

[0035] (2) In the individual feature vector construction and prediction inference, the clinical index arrangement, normalization, coding, prediction model inference and evidence traceability mapping are connected in series, the risk probability and the evidence traceability package are output, and the evidence summary report rendering and configuration updating are directly called.

[0036] (3) In the evidence summary report rendering and configuration updating, the visual layer rendering, report layout, template matching, threshold adjustment and parameter landing form a unified process, and the report configuration structure is output back to the registration alignment and alignment index generation as a parameter source.

[0037] The main beneficial effects are as follows:

[0038] (1) Facing the data link of multi-modal magnetic resonance and clinical primary table, cross-time point consistent alignment is completed, the static feature set and the dynamic feature set are directly referenced in the individual feature vector construction, and it is suitable for multi-time point image and clinical joint processing scene.

[0039] (2) Around the link of individual feature vector to risk probability and evidence traceability package, direct reference is realized in the evidence summary report rendering and report layout, and the alignment index structure and the evidence mask maintain consistent mapping relationship, which is suitable for multi-modal magnetic resonance and clinical primary table joint determination scene.

[0040] (3) For the link from rendering the visualization layer to the report configuration structure, complete the consistent application of template matching, threshold adjustment and parameter disking, and return the report configuration structure to the registration and alignment and alignment index generation, which is suitable for parameter management scenarios in continuous processing flow. Attached Figure Description

[0041] Figure 1 This is an overall flowchart of a multimodal magnetic resonance stent restenosis prediction method provided in an embodiment of this application.

[0042] Figure 2 This is a flowchart illustrating a multimodal magnetic resonance stent restenosis prediction method provided in an embodiment of this application.

[0043] Figure 3 A complete flowchart illustrating the construction and application of a machine learning prediction model to new patients, as provided in this application embodiment.

[0044] Figure 4 This is an overall architecture diagram of a multimodal magnetic resonance stent restenosis prediction system provided in an embodiment of this application.

[0045] Figure 5 This is a structural block diagram of a multimodal magnetic resonance stent restenosis prediction system provided in an embodiment of this application. Detailed Implementation

[0046] The overall flowchart of this method is as follows: Figure 1 As shown, the process begins with patient enrollment, followed by sequential execution of multi-timepoint multimodal MRI data acquisition, image preprocessing and spatiotemporal registration, feature extraction, construction of individualized multi-dimensional feature vectors, input of the feature vectors into a pre-trained machine learning prediction model, model output of individualized prediction results, and finally generation of a visualized prediction report and clinical decision recommendations. This process systematically demonstrates the entire process from data acquisition to risk prediction and report generation, with each step tightly connected through data flow, achieving end-to-end individualized prediction of restenosis risk after intracranial artery stenting. Specifically, the method includes the following steps:

[0047] S1: Multimodal MRI data acquisition at multiple time points: Multimodal MRI sequences of the target vessel were acquired at the postoperative baseline period (T0, e.g., within 1 week postoperatively) and at least one early follow-up period (T1, e.g., 3-6 months postoperatively), including:

[0048] Baseline scan: to obtain the baseline status after stent implantation.

[0049] (1) 3D high-resolution vascular wall imaging (HR-VWI): including T1WI plain scan and enhanced sequence, used to assess the morphology, thickness and basic components of the vessel wall.

[0050] (2) Contrast-enhanced vessel wall imaging (CE-VWI): Delayed scan after gadolinium contrast injection, used to quantify wall enhancement (a surrogate marker for vascular inflammation and neovascularization).

[0051] (3) Susceptibility-weighted imaging (SWI): Used to detect intrastent or adjacent wall hemorrhage (IPH).

[0052] S2: Image preprocessing and registration:

[0053] Preprocess all sequence images at T0 and T1 time points (standardization, denoising, metal artifact reduction).

[0054] Use rigid or non-rigid registration algorithms to perform high-precision spatiotemporal registration of images at different time points, ensuring pixel-level comparison of the same anatomical location, which is crucial for calculating dynamic change parameters.

[0055] S3: Feature extraction and quantification:

[0056] (1) Static features (single time point): Extract on images at T1 time point:

[0057] Wall enhancement features: On CE-VWI, measure the normalized enhancement rate of the wall at the proximal, middle, and distal stent segments (e.g., compared to the signal of the sternocleidomastoid muscle).

[0058] Intima features: Measure the maximum wall thickness and wall area.

[0059] Composition features: Identify and quantify low signal lesions on SWI (suspected IPH).

[0060] (2) Dynamic features (cross-time point changes): Based on registered T0 and T1 images, calculate:

[0061] Wall proliferation dynamics: Wall area growth rate, maximum wall thickness growth rate.

[0062] Inflammation activity changes: Rate of change in wall enhancement degree.

[0063] (3) Clinical features: Integrate patient clinical data, such as low-density lipoprotein (LDL-C) levels, history of diabetes, history of smoking, etc.

[0064] S4: Individualized risk prediction modeling:

[0065] Fuse the extracted static, dynamic, and clinical features to construct a multi-dimensional feature vector representing the overall state of the patient.

[0066] Input this feature vector into a pre-trained machine learning prediction model.

[0067] The model output result is a prediction risk probability value (for example: 30% risk) or a risk level (for example: low, medium, high) of the patient in the future M months (for example: 24 months) after the operation to have a significant hemodynamic restenosis (defined as a lumen stenosis rate of ≥50%).

[0068] In an embodiment, with reference to Figure 2 is a flowchart of a multi-modal magnetic resonance stent restenosis prediction method provided by an embodiment of the application, which can at least include steps S100-S400:

[0069] S100, acquire multi-modal magnetic resonance containing high-resolution blood vessel wall imaging, contrast-enhanced blood vessel wall imaging and susceptibility-weighted imaging and structured clinical initial table covering lipid metabolism, sugar metabolism, blood pressure history, smoking history and recent medication records, perform standardization and artifact suppression, registration alignment and alignment index generation processing based on report configuration structure, generate alignment index structure;

[0070] S200, perform candidate region positioning and segmentation processing, statistics and denoising, cross-time point alignment and change calculation processing, obtain segmentation results and evidence mask, static feature set and dynamic feature set;

[0071] S300, perform normalization including segmentation quantile mapping and robust limiting, and combined encoding processing of one-hot encoding and ordered binning encoding, generate individual feature vector; perform prediction model reasoning based on the model artifacts trained and verified offline, and construct a traceability mapping table from the output subgrade back to the input feature segments and the source set and coordinates, generate risk probability and evidence traceability package;

[0072] S400, perform visualization layer rendering including cross-section pairing and voxel mapping, report layout based on page template and content block structure, template matching and threshold adjustment and parameter landing processing according to task type and department habit, construct image comparison and reinforcement annotation, trend curve, visual report draft and report configuration structure.

[0073] S100, acquire multi-modal magnetic resonance and clinical initial table, perform standardization and artifact suppression and registration alignment and alignment index generation processing, generate alignment index structure;

[0074] Specifically, the registration alignment and the alignment index generation refer to the report configuration structure formed by the previous operation, which contains threshold, template, rendering parameter and preprocessing parameter configuration. The system obtains original entries from the baseline period and follow-up period multi-modal magnetic resonance and clinical initial form, the multi-modal magnetic resonance in this embodiment includes high-resolution vessel wall imaging (HR-VWI), contrast-enhanced vessel wall imaging (CE-VWI) and susceptibility-weighted imaging (SWI), and the clinical initial form is a set of structured clinical records initially entered, covering lipid metabolism, glucose metabolism, blood pressure history, smoking history and recent medication records. After reading the preprocessing parameter configuration, the system completes uniform resampling, intensity standardization and spatial direction correction for each sequence, performs phase inconsistency correction, stripe suppression and inter-frame steady-state reconstruction for metal-related artifacts and motion artifacts, forms a multi-channel image block with sequence label, time point label and channel label; performs field legality check and coding mapping on the clinical initial form, generates an initial form segment with key name specification and consistent coding. The above image block and initial form segment are aggregated into a preprocessing data package under the same task number, the preprocessing data package is a set of images and tables after sequence normalization and artifact suppression, and is recorded into a preprocessing log, the log item includes source, processing link, alarm code and summary description. If there are situations such as missing sequence, insufficient layer coverage and missing clinical entries, the system registers alarm entries and gives placeholder outline strategies, and still outputs a preprocessing data package with complete structure, providing a stable entry for subsequent actions. The product is marked as an output field name preprocessing data package, and is passed to the preprocessing data package field call of the subsequent step in the same task context, while providing traceable sources for the threshold adjustment and template strategy rewrite of S400 at the cross-main step level.

[0075] Further, the system extracts anatomical landmarks and stent-related features from the pre-processed data. The anatomical landmarks in this embodiment indicate spatial reference points and lines that can be stably positioned, including the inflection points of the skull base bony structures, the bifurcations of the major intracranial arteries, the boundaries of the sinuses, and the texture patterns of the sulci; the stent-related features indicate signal patterns such as the high signal of the stent end cap, the stent body axis, the local lumen bright-dark contrast transition, and the neighborhood susceptibility mutation. The system first runs multi-scale filtering and morphological operations under the HR-VWI channel to generate a candidate centerline network and a bifurcation candidate point set; then constructs an end cap candidate region based on the pre- and post-enhancement intensity difference under the CE-VWI channel; and constructs a neighborhood suppression weight map based on the susceptibility mutation under the SWI channel. After multi-channel candidates are subjected to cross-channel consistency voting and spatial connectivity constraints, an initial anatomical landmark table is generated, and the stent-related features are jointly entered into the registration solving process. In the rigid registration stage, a global pose search and a fine stepping strategy are adopted to complete coarse positioning and fine positioning in two rounds. Coarse positioning is completed under the bony window, and fine positioning is completed under the blood vessel wall window; if the convergence speed of the objective function is too slow or the stepping number triggers the upper limit, the system records a registration warning entry and triggers a regional limited retry. In the non-rigid registration stage, a control grid deformation field estimation and a smoothing regularization term are adopted. The control grid is set to a higher resolution in the arterial direction area and a lower deformation strength in the stent neighborhood combined with the suppression weight map to avoid metal-related abnormalities from being pulled. During the deformation estimation process, the system sets the registration anchor weight for the anatomical landmarks and the stent-related features, and optimizes them layer by layer from coarse to fine through a multi-resolution pyramid. After two-stage solving, the system outputs the spatio-temporally aligned sequence group and the deformation parameter stub, which together constitute the registration aligned data under the task number and are written into the registration quality summary, anchor residual summary, and abnormal segment list. The product is labeled as the output field name registration aligned data and is passed to the registration aligned data field call of the subsequent steps in the same task context.

[0076] The system performs alignment index generation and version label writing on the registration alignment data. Alignment index generation refers to mapping each pixel coordinate of the follow-up period to the corresponding pixel coordinate of the baseline period and forming a key-value mapping table and a coordinate system label in this embodiment. The mapping key is composed of three parts: time point, channel, and voxel position. The coordinate system label records the resampling stack information and direction convention. The system first constructs forward and inverse mapping candidates according to the deformation parameter stub, then updates the mapping value by running a local resampling kernel in the neighborhood of anatomical landmarks and stent-related features, and records the interpolation mode label and local confidence level grading. For difficult registration areas, the system reverts to the neighborhood consistency rule, gives an alternative mapping through the surrounding multi-point consistency strategy, and writes an abnormality label in the index table. The version label writing records the data version number, model version summary, processing timestamp, and configuration summary value. All labels and index primary keys are written together in the index header, and the index body saves the pixel mapping entries and local confidence level grading. The alignment index structure in this embodiment is composed of an index header and an index body. The index header describes the version, task number, and coordinate system, and the index body stores the pixel-level mapping and interpolation label. The system writes the alignment index structure to the task warehouse and registers the access control entry and the traceability path, which facilitates the subsequent consumer to quickly retrieve and review according to the task number. The product is marked as the output field name alignment index structure, which is passed to the alignment index structure field call of the subsequent main step S200 in the same task context, used as a spatial reference for candidate region positioning and segmentation processing, and provides a coordinate mapping entry for the visualization layer rendering of the main step S400, constituting a data alignment base across main steps.

[0077] S200, performing candidate region positioning and segmentation processing, statistics and denoising, cross-time point alignment and change calculation processing, obtaining segmentation results and evidence masks, static feature sets and dynamic feature sets;

[0078] Specifically, the segmentation and quantification are constructed with the alignment index structure output by the previous main step as a spatial reference, which records the pixel-level mapping key, coordinate system marker, interpolation label, and local confidence score. The system synchronously accesses the baseline image group and the follow-up image group as the reference bottom map for candidate region positioning and the target input for segmentation and quantification. The candidate region in this embodiment indicates a three-dimensional connected body set located in the stent neighborhood and a certain length range upstream and downstream, which meets the continuous lumen direction and stable wall texture; the stent neighborhood is derived from the stent-related feature labeling in the previous step, including the end cap estimated position, axial direction, and neighborhood suppression weight map. The system first generates multi-section projections according to the manifold direction under the mapping relationship given by the alignment index structure, and then performs multi-scale filtering and morphological opening and closing operations on each section to obtain the blood vessel centerline seed and the candidate edge of the tube wall; then the candidate edge and the centerline are subjected to connected domain screening and length constraint to remove isolated short fragments and interference regions adjacent to bony high signal. In terms of abnormal scene processing, if the local confidence score is lower than the threshold, the system records an alarm entry and triggers the neighborhood expansion strategy to recover the missing candidate by means of the surrounding section consistency rule. After completing the candidate region positioning, the system runs the encoder-decoder segmentation network and the boundary refinement process based on the graph model on the multi-channel image group, the former produces a coarse segmentation voxel map, and the latter performs boundary rebound on the coarse segmentation according to the smoothness constraint and edge gradient intensity. The output consists of the lumen region and the tube wall region, and the evidence mask is generated synchronously; the evidence mask in this embodiment indicates the three-tuple mapping of the segmentation region pixel set and the corresponding source layer, time point, and category label, which is used for tracing and visualizing superimposition. After the above processing is completed, the system forms the segmentation result and the evidence mask, which are recorded as the output field names segmentation result and evidence mask. This field enters the input position segmentation result and evidence mask in the next step in the task context, and at the same time, provides a mask resource entry for the visualization layer rendering of the main step S400 at the cross-main-step level.

[0079] Further, the enhancement ratio, thickness and area are extracted from the segmentation result and evidence mask, and statistics and denoising are completed to generate a static feature set. The enhancement ratio in this embodiment indicates the intensity ratio metric in the tube wall region before and after contrast enhancement, the thickness indicates the shortest normal distance between the inner and outer boundaries, and the area indicates the projected area of the tube wall on a given cross-section. The system constructs an intensity sampling channel within the pixel set defined by the evidence mask, obtains multi-channel intensity using a resampling kernel and interpolation label consistent interpolation method; then performs paired sampling and intensity difference before and after enhancement, and then groups and statistics according to channel weight and time point label to obtain cross-section level and paragraph level enhancement ratio summary. In the thickness and area calculation link, the system uses the centerline orthogonal cross-section stack to carry out boundary point set reconstruction, and carries out sub-pixel level boundary tracking and topological repair for the inner and outer boundaries of the tube wall respectively. In the face of holes and narrow gaps, the boundary is repaired through neighborhood projection and local smoothing strategy, and then the thickness distribution is generated according to the normal distance field from the boundary to the centerline; the area comes from the discrete grid accumulation of the region surrounded by the boundary point set, and the numerical stability is guaranteed by the boundary direction consistency constraint and small spot rejection strategy. In the statistics and denoising link, the system replaces the outliers with the neighborhood median, performs robust quantile aggregation on the outlying paragraphs across the cross-sections, and writes the data quality summary into the log, marking the source channel and mask version number. After processing, the system generates a static feature set containing three types of fields: enhancement ratio, thickness and area, and registers it as the output field name static feature set in the task context. This field enters the input position static feature set in the next step, and at the same time, it provides an image phenotype basic field for the individual feature vector construction of the main step S300 at the cross-main step level.

[0080] Again, cross-time point alignment and change calculations are performed on the static feature set to generate a dynamic feature set. In this embodiment, cross-time point alignment follows the pixel mapping relationship given by the alignment index structure. First, same-site pairing is completed at the cross-section level, and then sequence pairing along the centerline is completed at the paragraph level. During alignment, the system uses a neighborhood majority consensus strategy to supplement pairing for low-confidence regions. When encountering strong artifact segments in the scaffold neighborhood, local weight decay and cross-section jump rules are triggered to avoid the spread of incorrect pairings. Change calculations establish time difference channels and growth channels for the two core fields of thickness and enhancement rate. The time difference channel characterizes the change magnitude from the baseline period to the follow-up period, while the growth channel characterizes the growth trend per unit time. The area field calculates the difference and ratio on the aligned cross-section and summarizes at the paragraph level. Regarding abnormal scenario handling, if a cross-section has missing test markers or inconsistent mask versions, the system reverts to the most recent version of the same source mask according to the log backtracking strategy, and then generates a complete sequence through a local interpolation strategy. In this embodiment, the dynamic feature set includes three types of fields: thickness growth rate, enhancement change rate, and area change. It retains pairing relationships and mask version references to facilitate subsequent evidence tracing and mapping. After the above processing is completed, the system outputs a dynamic feature set. This field enters the input position dynamic feature set in the main step S300 within the task context and participates in the construction of individual feature vectors together with the clinical indicator set generated in the main step S300. Simultaneously, this set is read and rendered by the visualization layer in the main step S400 to generate trend curves and annotation text.

[0081] S300: Perform clinical indicator organization and normalization coding, predictive model inference and evidence tracing mapping to generate individual feature vectors, risk probabilities and evidence tracing packages;

[0082] Subsequently, the system performs risk prediction based on machine learning, such as Figure 3 As shown:

[0083] Figure 3 This application provides a complete flowchart for constructing and applying a machine learning prediction model to new patients. The process mainly includes two stages: a model training stage and a prediction application stage. In the training stage, the system trains the model using machine learning algorithms based on a historical patient database containing early multimodal MRI features and clear clinical outcomes (such as whether stent restenosis occurred). This results in optimized model parameters that can be used for prediction, i.e., the "trained prediction model." In the prediction application stage, the system inputs the new patient's multidimensional feature vector (constructed from the static and dynamic feature sets and clinical indicator sets extracted in step S200) into the trained prediction model. The model then calculates and outputs a personalized predicted risk probability or risk level for that new patient.

[0084] Specifically, the individual feature vector construction and the prediction inference receive the static feature set and the dynamic feature set formed by the main step S200 in the task context, where the static feature set contains the reinforcement rate, thickness, area, etc. fields statistically obtained within the evidence mask defined range, and the dynamic feature set contains the thickness growth rate, reinforcement change rate and area change, etc. fields from the baseline period to the follow-up period. While reading the two sets, the system extracts the original entries corresponding to the task number from the registered clinical initial form copy. The clinical initial form is a structured clinical record set containing lipid metabolism, glucose metabolism, blood pressure history, smoking history and recent medication record fields. The clinical indicator processing first triggers field alignment under the condition that the task number and the subject identifier match successfully. Field alignment in this embodiment refers to merging synonymous fields into a unified key name and completing numerical unit conversion according to the preset unit table; for time-related entries, establish the latest visit time window and the test time window, and the time window uses different spans on different entries, and marks the time limit state on the super window entry. Subsequently, the system performs missing value strategy and conflict entry processing. The missing value strategy preferentially uses homologous time sequence completion and adjacent entry completion, and records as explicit missing when it cannot be completed; the conflict entry processing follows the two-level rules of source priority and time priority, source priority means preferentially using traceable source when electronic medical record (EMR) and manual input coexist, and time priority means using the latest entry when multiple entries coexist from the same source. In order to ensure the consistency of key name and caliber, the system caliber checks each key value with the dictionary snapshot in the task context, and writes the checking result and abnormal alarm into the log. After the processing is completed, the set containing the above structured fields, time window markers and missing markers is formed, which is recorded as the output field name clinical indicator set in the task context, and enters the input location clinical indicator set in the next step in the same task, while providing readable text source for the report layout of the main step S400.

[0085] Further, fields are extracted from the static feature set, the dynamic feature set, and the clinical indicator set, normalized and encoded to generate individual feature vectors. Normalization in this embodiment refers to implementing scale unification and robust clipping on numerical fields, scale unification is completed by piecewise quantile mapping, and robust clipping is performed by truncating extreme quantile outliers and retaining truncation markers; for category fields, a combination strategy of One-Hot Encoding and ordered binned encoding is adopted, One-Hot Encoding is used for unordered categories, and ordered binned encoding is used for hierarchical items, both of which generate fixed-position sparse or dense segments under the encoding table version number. Subsequently, the system performs field splicing and position solidification, field splicing follows the feature domain priority order, splicing the static feature domain first, then the dynamic feature domain, and finally the clinical indicator domain; position solidification in this embodiment refers to mapping each key name to a fixed index position and writing it into a position table, the position table carries a version number and a dictionary digest. For missing markers and abnormal markers, the system configures corresponding position signal bits in the vector, the signal bits are used to prompt the masking logic in the subsequent inference stage. In terms of abnormal scenario processing, if the field missing rate exceeds the threshold or the encoding table version does not match, the system records an alarm and triggers a fallback encoding table or a resampling strategy; if the time bases of static and dynamic features are inconsistent, the system re-coalesces at the paragraph level according to the alignment index structure and synchronously corrects the corresponding positions. After the above processing is completed, the system generates a fixed-length encoding field sequence in the task context, the sequence is arranged in a fixed order and has a position table reference and a dictionary version reference, which is called an individual feature vector in this embodiment; this product is recorded as the output field name individual feature vector, and enters the input position individual feature vector of the next step in the same task, while providing readable feature names and value segments for the visual annotation text module of step S400 at the cross-main step level.

[0086] The individual feature vector is subjected to prediction model inference and evidence trace mapping to generate a risk probability and evidence trace package. Prediction model inference in this embodiment refers to loading a model artifact that has been trained and validated offline in a controlled runtime environment, with the model artifact accompanying a model version summary, training data summary, and threshold configuration; after successful loading, the system feeds the individual feature vector together with the position table and dictionary version reference into the inference pipeline through the vector assembler, which performs signal bit masking and domain weight scheduling in the pre-processing stage, model calculation in the core stage, and binning strategy and threshold mapping in the post-processing stage. To support interpretable output, the system also constructs a trace mapping table in parallel, which records the mapping relationship from output binning to input feature segments to source sets and coordinates in this embodiment, where the source sets include static feature sets, dynamic feature sets, and clinical indicator sets, and the coordinates are given by alignment index structure and evidence mask. During the mapping generation process, the system first traces back to the feature domain and key name according to the position table, and then locates the original set entry through the key name and task number within the domain; if the key name belongs to the dynamic feature domain, it continues to trace back to the cross-time point pairing relationship, and associates the mask version and section position; if the key name belongs to the static feature domain, it associates the class label and time point in the segmentation result and evidence mask; if the key name belongs to the clinical indicator domain, it associates the time window and source marker of the clinical entry. In terms of abnormal scenario processing, if the model version and position table version are inconsistent, the system triggers the version compatibility adaptation module to perform key name mapping; if the trace mapping has a hanging entry, the system logs the hanging reason in the log and adds a replacement note in the mapping table. After inference and mapping are completed, the system generates a set containing numerical binning, binning threshold reference, and trace mapping table, which is referred to as a risk probability and evidence trace package in this embodiment; this product is recorded as the output field name risk probability and evidence trace package, and enters the input position risk probability and evidence trace package of main step S400 at the main step level, which is used for subsequent visualization layer rendering and report layout, while the access control entry and version label are registered in the task warehouse.

[0087] S400, visualization layer rendering and report layout, template matching and threshold adjustment, and parameter disk processing are performed to construct image contrast and reinforcement annotation, trend curve, visualization report draft, and report configuration structure.

[0088] Specifically, the evidence summary report renders and configures the risk probability and evidence provenance package from the output of the main step S300 as core inputs in the task context, and links the alignment index structure generated by the main step S100 and the segmentation result and evidence mask generated by the main step S200 as visual reference under the same task number. The risk probability and evidence provenance package indicates numerical grading, grading threshold reference, and mapping table from the output entry back to the static feature set, dynamic feature set, and clinical indicator set in this embodiment; the alignment index structure is a pixel-to-pixel key-value mapping and coordinate system marker; the segmentation result and evidence mask are voxel sets and class label mappings of lumen and wall regions. The system first registers the task number and rendering queue entry in the rendering scheduler, and then establishes cross-sectional pairing and voxel mapping between the baseline and follow-up image groups according to the alignment index structure; sets the layer overlay order in the rendering engine, the base layer is the resampled original image, the evidence layer is the evidence mask and segmentation boundary, and the indicator layer is the labeled element at the spatial position pointed by the provenance package. Specifically, the system performs morphological refinement on the evidence mask to generate boundary vector curves for high-contrast overlay; performs local noise suppression and transparency gradient on the segmentation boundary neighborhood to avoid interface sawtooth; performs color scale mapping and label text generation on the risk-related fields, and the label text includes field name, time point label, and grading marker. Further, the system generates left-right side-by-side or top-bottom contrast image pair views on the same cross section according to the alignment index structure, and automatically inserts a zoom window in the stent neighborhood and upstream and downstream paragraphs; for time evolution information, the system finds the paragraph-level pairing sequence from the corresponding key name in the provenance package from the dynamic feature set, constructs the thickness and reinforcement-related trend curve, and hangs the positioning link from the segmentation result at the curve node for jumping back to the corresponding cross section. In terms of abnormal scene processing, if a certain evidence entry is suspended during mapping playback, the system generates a placeholder marker in the layer and records the reason in the log; if the alignment index structure has insufficient confidence in a local area, the system adopts a neighborhood majority consensus strategy to generate an alternative mapping and mark a prompt. After the above processing is completed, the system obtains three types of visualization products, namely image contrast, reinforcement annotation, and trend curve, which together form a set of visualization layers and are recorded as output field name image contrast, reinforcement annotation, and trend curve in the task context, and enter the next step input image contrast, reinforcement annotation, and trend curve in the same task, while providing a usable spatial overlay view entry for the backtracking review of the main step S200 at the cross-main step level.

[0089] Further, key fields are extracted from image contrast, enhanced annotation and trend curve to perform report layout and field filling, and generate a visual report draft. Report layout in this embodiment refers to layout of image contrast view, evidence annotation layer and trend curve according to column and partition under the constraint of page template and content block structure, and filling of textual description, grading prompt and reference label according to field mapping table. Page template and content block structure have version label under task number, which is used for unified page margin, column title, figure caption style and font size level. The system first reads the template registry, selects the matching template according to the task type and the subject category, and establishes the binding relationship between the content block and the data field. Then, the image contrast is embedded into the image comparison partition, the enhanced annotation details are embedded into the evidence annotation partition, the trend curve is embedded into the time evolution partition, and the grading text is written in the data summary partition according to the threshold reference in the trace package. During field filling, the system performs numerical format standardization, unit homogenization and Chinese key name mapping, and inserts footnote entries when there are missing markers or alternative instructions; when the number of graph layers is inconsistent with the reserved slot of the template, the system uses priority selection and column overflow strategy to complete layout rollback, and records overflow information in the layout log. After the layout engine completes pagination, header and footer generation and directory index construction, a layout document object with embedded layer reference and field backlink is generated; this object is called visual report draft in this embodiment, which includes page template and content block structure, image contrast reference, enhanced annotation reference and trend curve reference. After the above processing is completed, the system registers the product as the output field name visual report draft in the task context, and enters the input location visual report draft in the next step in the same task, while providing readable field name and symbol mapping reference for the individual feature vector construction module of the main step S300 in the cross-main step level, for field alignment in subsequent version iteration.

[0090] Again, template matching and threshold adjustment and parameter landing of the visual report draft generate the report configuration structure. Template matching in this embodiment refers to selecting the optimal template in the template registry according to the task type, department habit and number of map layers, and making fine adjustments to the content block order and size if necessary; threshold adjustment in this embodiment refers to reference according to risk probability and grading threshold, and synchronous correction of grading text, legend threshold and annotation color scale within the same report draft range; parameter landing in this embodiment refers to writing the final layer rendering parameters, format parameters and threshold parameters into the configuration warehouse, and generating version labels and access control entries. The system first checks the integrity of the visual report draft, and if it finds that the layer reference is hanging or the field is missing rate exceeds the threshold, it triggers the rollback strategy, and preferentially rolls back to the last version of the template or the last version of the threshold parameters; for cross-page layers, the system applies partition rearrangement and annotation merging strategy to avoid disconnection between graphics and text. After the above matching and correction are completed, the system generates a structure containing threshold, template and rendering parameter set, which is called report configuration structure in the task context, and writes the task number, model version summary, template version label and timestamp in the configuration warehouse. The report configuration structure, as the output field name of this main step, is registered and used by step S100 as the calling source of the pre-processing parameter configuration at the cross-main step level, so that the sequence resampling, intensity standardization, artifact suppression and registration options are parameterized when a new task starts; at the same time, this structure is used for parameter initialization of the next version of the format and rendering process within this main step, forming a parameter sedimentation path for task internal loop.

[0091] Embodiment two: Figure 4 The overall architecture diagram of a multi-modal magnetic resonance stent restenosis prediction system provided in the embodiments of the present application is shown in FIG. 1. Figure 4 As shown in the figure, the system comprises:

[0092] A data management module for centralized storage and management of multi-modal magnetic resonance images (including HR-VWI, CE-VWI, SWI sequences) and structured clinical initial form data from the baseline period and the follow-up period;

[0093] An image analysis engine as a core processing unit, which further comprises: a registration unit responsible for standardization, artifact suppression and high-precision spatio-temporal registration of multi-modal sequences at different time points, generating an alignment index structure; a segmentation and labeling unit for performing positioning and segmentation of candidate regions such as lumen and wall on the registered images, and generating evidence masks; a feature extraction unit based on segmentation results and evidence masks, performing statistics, denoising and cross-time point change calculation to extract static feature sets and dynamic feature sets;

[0094] a prediction core module, which is built-in machine learning model artifacts trained and validated via large-scale cohort data, for receiving fused feature vectors extracted by the image analysis engine, performing prediction model inference and evidence provenance mapping, and generating evidence provenance packages containing risk probability and interpretable justifications;

[0095] a visualization report interface, for rendering visualization layers, report layout and template matching based on the output of the prediction core module, and finally generating visualization clinical reports integrated with image comparison, reinforced annotation, trend curve and personalized suggestions.

[0096] Embodiment Three: Figure 5 shows a structural block diagram of a multi-modal magnetic resonance stent restenosis prediction system according to an embodiment of the present application. As shown in Figure 5 , the structure can include:

[0097] a registration alignment and alignment index generation module 01, for standardized, artifact suppression, registration alignment and alignment index generation from baseline and follow-up multi-modal magnetic resonance and clinical initial forms, outputting an alignment index structure, outputting for segmentation and static and dynamic feature construction module calling; specifically, receiving original entries from multi-modal magnetic resonance and clinical initial forms, completing standardization and artifact suppression according to preset parameters, forming image and entry sets that can be registered and aligned; in the registration alignment stage, the corresponding relationship is established according to the time point and sequence label, the consistent processing of position and time is completed, and the processing state and parameter summary are recorded; when the alignment index is generated, the mapping entries and version labels of the processed image entries are established, forming structured entries that can be accessed and called, as the output of the alignment index structure; the alignment index structure is transmitted to the segmentation and static and dynamic feature construction module as a spatial and temporal reference, while retaining the record information for subsequent rendering and format calling.

[0098] The segmentation and quantization and static and dynamic feature construction module 02 is configured to complete candidate region positioning and segmentation processing, statistics and denoising, cross-time point alignment and change calculation under the reference of the alignment index structure, output segmentation results and evidence masks, static feature sets and dynamic feature sets, and output for calling by the individual feature vector construction and prediction reasoning module; specifically, the alignment index structure and image entries from the previous processing are received, candidate region positioning and segmentation processing are completed according to the reference relationship, lumen region and pipe wall region entries are obtained, and evidence masks are generated; in the statistics and denoising stage, the enhancement rate, thickness and area fields are extracted around the segmentation results, entry-level aggregation and abnormal marking processing are completed, and a stable field entry set is formed; in the cross-time point alignment and change calculation stage, homologous pairing and paragraph pairing are completed according to the alignment index structure, thickness growth rate, enhancement change rate and area change entries are generated, and a dynamic feature set is formed; after the processing is completed, the segmentation results and evidence masks, the static feature set and the dynamic feature set are output, and are transmitted to the individual feature vector construction and prediction reasoning module for calling, and state information is registered for subsequent layout and parameter updating.

[0099] The individual feature vector construction and prediction reasoning module 03 is configured to perform arrangement, normalization and coding on the static feature set, the dynamic feature set and the clinical indicator set, send the individual feature vector into the reasoning pipeline, and perform signal bit masking, domain weight scheduling, model calculation and grading threshold mapping, output risk probability and evidence traceability package, and output for calling by the evidence summary report rendering and configuration updating module; specifically, the static feature set, the dynamic feature set and the clinical indicator set are received, field arrangement and coding mapping are completed, and a fixed-order individual feature vector is generated; signal bit masking and domain weight scheduling are performed in the pre-processing stage of the reasoning pipeline, model calculation is performed in the core stage, grading threshold mapping is performed in the post-processing stage, and grading entries are formed; a mapping relationship from the grading entries back to the feature segments and then to the source set and the spatial entries is constructed synchronously, and the risk probability and the evidence traceability package are combined to form; the risk probability and the evidence traceability package are transmitted to the evidence summary report rendering and configuration updating module as the input objects of visualization and layout, and version and state entries are registered in the process record.

[0100] The evidence summary report rendering and configuration updating module 04 is used to complete the rendering of visual layers, the layout of reports, the matching of templates, the adjustment of thresholds and the landing of parameters under the reference of the alignment index structure, the segmentation results and the evidence mask, and output the image comparison, the strengthened annotation and the trend curve, the draft of the visual report and the report configuration structure, and the report configuration structure is called by the registration alignment and the alignment index generation module. Specifically, the risk probability and the evidence traceability package are received, and the alignment index structure, the segmentation results and the evidence mask are combined, the rendering of visual layers and the layout of reports are completed, the image comparison, the strengthened annotation and the trend curve are generated and assembled into the draft of the visual report; in the template matching and threshold adjustment stage, the format field and the grading threshold are corrected synchronously, the parameter landing is completed, and the template, the threshold and the rendering parameter item are registered to form the report configuration structure; the image comparison, the strengthened annotation, the trend curve and the draft of the visual report are used as the input object of the issued report, and the report configuration structure is returned to the registration alignment and the alignment index generation module as the parameter source, and the cyclic updating relationship of the processing configuration and the data item is maintained.

Claims

1. A method for predicting restenosis in a multimodal magnetic resonance stent, characterized in that, include: We acquire multimodal magnetic resonance imaging (MRI) including high-resolution vascular wall imaging, contrast-enhanced vascular wall imaging, and magnetic susceptibility-weighted imaging, as well as structured clinical preliminary reports covering lipid metabolism, glucose metabolism, blood pressure history, smoking history, and recent medication records. Based on the report configuration structure, we perform standardization and artifact suppression, registration alignment, and alignment index generation to generate an alignment index structure. Candidate region localization and segmentation, statistics and denoising, cross-time point alignment and change calculation are performed to obtain segmentation results, evidence masks, static feature sets and dynamic feature sets; Normalization, which includes segmented quantization and robust amplitude limiting, and combined encoding processing of one-hot encoding and ordered binning encoding are performed to generate individual feature vectors. Based on the model artifacts trained and validated offline, prediction model inference is performed, and a source mapping table that traces back from the output tiers to the input feature fragments and source sets and coordinates is constructed to generate risk probability and evidence source tracing packages. The system performs visualization layer rendering including cross-section pairing and voxel mapping, report layout based on page templates and content block structures, template matching and threshold adjustment based on task type and departmental habits, and parameter disk processing. It also constructs image comparison and enhanced annotation, trend curves, visualization report drafts, and report configuration structures.

2. The method according to claim 1, characterized in that, The process involves candidate region localization and segmentation, statistical analysis and denoising, cross-timepoint alignment and change calculation, resulting in segmentation results, evidence masks, static feature sets, and dynamic feature sets. Specifically, this includes: Morphological refinement of the evidence mask is performed to generate boundary vector curves; local noise reduction and transparency gradient are applied to the segmentation boundary neighborhood; image comparison views are generated on the same cross section based on the alignment index structure, and magnified windows are automatically inserted in the scaffold neighborhood and upstream and downstream segments.

3. The method according to claim 1, characterized in that, The process involves normalization with segmented quantization and robust amplitude limiting, combined with one-hot coding and ordered bin coding, to generate individual feature vectors. Specifically, this includes: Normalization refers to applying scaling uniformity and robust limiting to numerical fields. Scaling uniformity is achieved through segmented quantile mapping, while robust limiting is achieved by truncating entries outside extreme quantiles and retaining truncation markers. For categorical fields, a combination strategy of one-hot encoding and ordered binning encoding is adopted.

4. The method according to claim 1, characterized in that, Predictive model inference is performed based on model artifacts trained and validated offline, and a source mapping table is constructed to trace back from output tiers to input feature fragments and source sets and coordinates, specifically including: Construct a source mapping table to record the mapping relationship from output classification to input feature fragments and then to source set and coordinates. The source set includes static feature set, dynamic feature set and clinical indicator set, and the coordinates are given by alignment index structure and evidence mask.

5. The method according to claim 4, characterized in that, The process of constructing the source mapping table also includes: First, backtrack to the feature domain and key name based on the location table, and then locate the original set item through the key name and task number within the domain; if the key name belongs to the dynamic feature domain, continue backtracking to the cross-time point pairing relationship and associate the mask version and cross-sectional position; if the key name belongs to the static feature domain, associate the segmentation result with the category label and time point in the evidence mask; if the key name belongs to the clinical indicator domain, associate the time window and source mark of the clinical item.

6. The method according to claim 1, characterized in that, The process of rendering visualization layers, which includes cross-section pairing and voxel mapping, also includes: Establish cross-sectional pairing and voxel mapping relationships between baseline and follow-up image groups based on the alignment index structure; set the layer stacking order in the rendering engine, with the base layer being the resampled original image, the evidence layer being the evidence mask and segmentation boundary, and the indicator layer being the annotation elements of the field pointed to by the tracing package at the spatial location.

7. The method according to claim 1, characterized in that, The report formatting process based on page templates and content block structures also includes: Under the constraints of page templates and content block structures, image comparison views, evidence annotation layers, and trend curves are laid out according to columns and sections, and textual descriptions, classification prompts, and reference tags are filled according to field mapping tables. Page templates and content block structures have version tags under task numbers to unify page margins, column titles, figure caption styles, and font sizes.

8. The method according to claim 1, characterized in that, The process of template matching, threshold adjustment, and parameter loading based on task type and departmental habits also includes: Template matching refers to selecting the optimal template in the template registry based on task type, departmental habits, and number of layers; threshold adjustment refers to synchronously correcting the classification text, legend threshold position, and annotation color level based on risk probability and classification threshold reference; parameter disk entry refers to writing the final layer rendering parameters, layout parameters, and threshold parameters into the configuration repository and generating version tags and access control entries.

9. The method according to claim 8, characterized in that, The report configuration structure is used as a source of calls for preprocessing parameter configuration at the cross-main step level, parameterizing sequence resampling, intensity normalization, artifact suppression, and registration options when a new task starts; at the same time, this structure is used within this main step for parameter initialization of the next layout and rendering process, forming a parameter accumulation path in the task loop.

10. A multimodal magnetic resonance stent restenosis prediction system, applied to the method of any one of claims 1-9, characterized in that, include: The registration and alignment and alignment index generation module is used to standardize, suppress artifacts, register and align and generate alignment indexes from baseline and follow-up multimodal magnetic resonance imaging and initial clinical manifestations, and output the alignment index structure for use by the segmentation quantification and static and dynamic feature construction modules. The segmentation quantization and static / dynamic feature construction module is used to complete candidate region localization and segmentation, statistics and denoising, cross-time point alignment and change calculation under the reference of the alignment index structure. It outputs segmentation results and evidence masks, static feature sets and dynamic feature sets, and outputs them for the individual feature vector construction and prediction inference module to call. The individual feature vector construction and prediction inference module is used to organize, normalize and encode the static feature set, dynamic feature set and clinical indicator set, send the individual feature vector into the inference pipeline and perform signal bit masking, domain weight scheduling, model calculation and grading threshold mapping, output risk probability and evidence tracing package, and output for the evidence summary report rendering and configuration update module to call. The Evidence Summary Report Rendering and Configuration Update module is used to complete the rendering of visualization layers, report layout, template matching, threshold adjustment and parameter disking under the alignment index structure, segmentation results and evidence mask reference. It outputs image comparison, enhanced annotation and trend curve, visualization report draft and report configuration structure. The report configuration structure is called by the registration and alignment and alignment index generation modules.

Images