Pancreatic mesenteric area perivascular inflammation imaging-based method and system for preoperative evaluation

Abstract

The present application relates to the technical field of medical image analysis, and particularly relates to a pancreatic mesangial zone perivascular inflammation imageomics identification method and system for preoperative evaluation. The method comprises: acquiring multi-phase enhanced abdominal images, performing data quality control, denoising registration, intensity normalization and voxel resampling processing to generate standardized image structure; performing pancreatic mesangial zone blood vessel candidate region positioning, blood vessel skeleton extraction and centerline tracking, and constructing a ring-shaped region of interest structure; performing blood vessel and surrounding tissue segmentation and mask generation, extracting traditional imageomics features and coupling features reflecting the interaction between blood vessels and surrounding tissues, and generating an effective feature set; through domain correction, identification model training and parameter solidification, an inference report containing inflammation identification labels and explainability mapping is generated. The present application realizes automatic and accurate quantitative identification of perivascular inflammation, effectively improving the objectivity and accuracy of preoperative evaluation.

Description

Technical Field

[0001] This invention relates to the field of medical imaging analysis technology, and in particular to a radiomics identification method and system for perivascular inflammation in the pancreatic mesangial region for preoperative assessment. Background Technology

[0002] In the field of medical image analysis, existing solutions for identifying perivascular inflammation in the pancreatic mesangial region for preoperative assessment typically rely on manual interpretation and experience based on multi-phase enhanced abdominal images, or on locating candidate regions and extracting features within localized areas to provide a judgment result. These solutions suffer from limitations such as inconsistent standardized image structure formation processes, missing or incomplete acquisition parameter recording tables and preprocessing configuration structures, and difficulties in stably constructing ring-shaped regions of interest. Existing methods largely depend on the operator's subjective definition and threshold adjustment of the perivascular area. The segmentation of blood vessels and surrounding tissues, as well as mask generation, are easily affected by imaging noise, registration errors, and voxel resampling differences. Furthermore, variations in equipment and scanning conditions reflected in the acquisition parameter recording table can lead to inconsistencies in the boundaries between vascular and surrounding tissue masks, and difficulties in verifying mask quality markings. These limitations make it difficult to achieve stable implementation of feature matrix construction and effective feature set structure generation. For the joint processing of radiomics feature extraction and coupled feature construction, existing technologies generally lack a consistent definition of the interface neighborhood zone between the vascular mask and the surrounding tissue mask, as well as unified management of the feature naming dictionary. The association between the feature missing marker table, the feature extraction parameter table, and the preprocessing log structure is incomplete, leading to difficulties in aligning feature fields when processing data from the same source in different batches. Regarding feature stability screening and feature selection operations, existing technologies generally lack structured records of consistency statistics and screening rule records and selection rule records under controlled perturbation conditions, making it difficult to form a verifiable list of features for model input. For the continuous chain of domain correction processing, recognition model training, model parameter solidification, case reasoning, and interpretability mapping generation, existing technologies exhibit inconsistencies in the association registration of alignment mapping records, model version structures, and reasoning report structures. Updates to the preprocessing configuration structure lack traceability support, making it difficult to form a consistent process of collection, alignment, judgment, recording, and updating in the preoperative assessment workflow, thus affecting the verification of the reasoning report structure and cross-batch comparison. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a radiomics-based method for identifying perivascular inflammation in the pancreatic mesangial region for preoperative assessment, comprising:

[0004] Acquire multi-phase enhanced abdominal images, clinical information fields, and preprocessing configuration structures; perform data quality control and acquisition parameter recording, noise reduction and registration with intensity normalization, voxel resampling processing, and generate standardized image structures.

[0005] Based on standardized image structures, the process of locating candidate regions of pancreatic mesangial vessels, extracting the vascular skeleton and tracing the centerline, and constructing a ring region of interest (ROI) is performed to obtain a ring ROI structure. The ring ROI structure includes a ring shell voxel set pointer, a ring ROI index table pointer, a cross-sectional normal sequence pointer, an inner and outer radius parameter table pointer, a sampling step size parameter table pointer, and a vascular centerline structure pointer.

[0006] Based on the ring-shaped region of interest structure, the system performs segmentation and mask generation of blood vessels and surrounding tissues, extraction of radiomics features and coupling feature construction to generate an effective feature set structure.

[0007] Based on the effective feature set structure, domain correction processing, recognition model training and model parameter solidification processing are performed to generate an inference report structure.

[0008] Furthermore, the process of performing data quality control and recording collected parameters also includes:

[0009] Data quality control processing includes missing segment checking, artifact marking, and sequence integrity verification. Missing segment checking identifies gaps in consecutive slices by comparing the slice order field in the image sequence index table with the actual loaded slice count, and writes the missing range field into the data quality control marking table. Artifact marking involves performing a coarse screening of each phase slice before intensity anomaly suppression and marking texture anomalies such as stripes, metallic artifacts, and motion blur, and writing the artifact type field into the data quality control marking table. Sequence integrity verification generates a sequence integrity verification result field by checking the phase identifier set against the combined phase set specified in the preprocessing configuration structure. The acquisition parameter recording processing includes recording the device model, slice thickness, reconstruction settings, and contrast agent injection settings. The device model comes from image metadata, the slice thickness from the slice spacing field, the reconstruction settings from the reconstruction kernel and matrix field, and the contrast agent injection settings from the injection rate and dosage field or the in-hospital examination form field. When a field is missing, a missing marker field is written into the acquisition parameter recording table and simultaneously written into the preprocessing log structure.

[0010] Furthermore, the voxel resampling process also includes:

[0011] The voxel resampling process includes reading the target voxel spacing field and the target grid size field, selecting the resampling interpolation method and generating a resampling parameter record table, writing the interpolation method field, the target voxel spacing field, and the target grid size field into the resampling parameter record table, and simultaneously writing them into the preprocessing log structure, while generating a voxel grid description table; generating a standardized image structure, which includes a standardized image data pointer, a standardized image index table, a voxel grid description table pointer, a normalized parameter record table pointer, a resampling parameter record table pointer, and a preprocessing log structure pointer field.

[0012] Furthermore, the process of locating candidate vascular regions in the pancreatic mesangial area also includes:

[0013] The process for locating candidate vascular regions in the pancreatic mesangial area involves first extracting abdominal cavity boundaries and spinal reference plane fields from standardized image structures. Then, a three-dimensional spatial bounding box is constructed based on the relative positional constraints of the pancreatic mesangial area within the body cavity, constraining the area's extent. These constraints include a longitudinal range near the root of the mesentery, an anterior-posterior range near the posterior aspect of the pancreas, and exclude areas outside the body cavity such as air zones and high-density cortical bone. Further, temporal images are extracted from the standardized image structures, and post-enhanced temporal combinations are selected as candidate sources based on a standardized image index table. Finally, tubular structure response calculations are performed for each temporal phase, with the tubular structure response being specific to slender, connected structures. The multi-scale filtering output of the morphology is implemented by calculating the local orientation consistency and boundary gradient stability in the neighborhood of different scales to form a response volume. Then, threshold segmentation is performed on the response volume to generate a candidate voxel set. After the candidate voxel set is generated, connected component analysis is performed under the structural constraints of the pancreatic mesangial region to output a candidate connected component table. The candidate connected component table contains fields such as connected component identifier, number of voxels, principal axis direction, average enhancement intensity, and bifurcation complexity. At the same time, a candidate region quality label table is generated. If the candidate connected component table has too many or too few connected components at any time phase, the localization rules in the preprocessing configuration structure are read and written into the localization rule record table, and then the rule switching process is executed.

[0014] Furthermore, the process of vascular skeleton extraction and centerline tracing also includes:

[0015] The vascular skeleton extraction process includes hole filling and fracture bridging of the candidate voxel set. Hole filling is performed by identifying voxel cavities through local connectivity detection and filling them in their neighborhoods. Fracture bridging is completed by pairing short-distance endpoints and performing morphological closure operations. Subsequently, the completed voxel set is refined to generate a vascular skeleton voxel set. The refinement process involves iteratively stripping boundary voxels while maintaining the topology of endpoints and bifurcation points. A skeleton connectivity graph is constructed for the vascular skeleton voxel set, which consists of a set of nodes and a set of edges. The centerline tracing process includes selecting the main trunk candidate connectivity region based on the main axis direction field in the candidate connectivity region table, selecting the entry endpoint within the main trunk candidate connectivity region, and performing path traversal to generate a centerline coordinate sequence. The path traversal adopts a stepwise expansion strategy that includes the connectivity graph. At the same time, each bifurcation point is selected and recorded according to the bifurcation point number table to form a centerline quality mark table. The centerline quality mark table includes path continuity, bifurcation stability, and loop detection mark fields. When the centerline quality mark table shows insufficient path continuity, the fracture degree field of the candidate region quality mark table is called to trigger a secondary bridging process.

[0016] Furthermore, the process of constructing a ring-shaped region of interest also includes:

[0017] The process of constructing the annular region of interest (ROI) involves extracting the centerline coordinate sequence from the vessel centerline structure and reparameterizing the centerline. Centerline reparameterization generates a sampling point sequence by uniformly sampling along the arc length. The sampling step size is taken from the preprocessing configuration structure and written into the sampling step size parameter table. At each sampling point, a cross-sectional normal sequence is constructed. The cross-sectional normal sequence generates a local tangential direction by the direction difference between adjacent sampling points and combines it with the direction cosine field of the voxel mesh description table for coordinate consistency. Then, an inner and outer radius parameter table is generated according to the inner and outer radius rules in the preprocessing configuration structure. The inner and outer radius rules include two types: a fixed radius scheme and a segmented radius scheme. The fixed radius scheme uses the same inner and outer radius fields for all sampling points. The segmented radius scheme assigns different radius fields to different segments based on the local vessel scale estimation field of the connected domain where the centerline is located. An annular cross-sectional voxel set is generated at each sampling point and accumulated along the centerline to form an annular shell voxel set. At the same time, an index is established between the annular shell voxel set and the centerline coordinate sequence to generate an annular ROI index table.

[0018] Furthermore, the process of segmenting blood vessels and surrounding tissues and generating a mask also includes:

[0019] The segmentation process for blood vessels and surrounding tissues involves extracting a ring-shaped shell voxel set and a ring-shaped region of interest (ROI) index table from the ring-shaped ROI structure. Based on the sampling point and cross-sectional identifiers in the ring-shaped ROI index table, corresponding voxel blocks from the standardized image structure are retrieved to form a segmentation input block sequence. Simultaneously, a segmentation path record table is generated. Three types of paths—threshold segmentation, region growing, and learning model inference—are used to separate the blood vessel region from the surrounding tissue region. The threshold segmentation path uses the enhancement intensity field and candidate region index field of the standardized image structure to set an intensity threshold and generate an initial blood vessel voxel set. Then, boundary contraction and expansion corrections are performed along the normal direction based on the cross-sectional normal sequence. The region growing path selects seed points at the endpoints of the blood vessel mask candidates. The model is further expanded based on the connected component similarity criterion, which consists of an intensity difference threshold, a gradient consistency threshold, and a morphological continuity threshold. The learning model inference path inputs the segmented input block sequence into the trained segmentation model and outputs a probability map. Then, according to the probability threshold and connectivity constraints, a set of blood vessel voxels and a set of surrounding tissue voxels are generated. The mask generation process includes writing the segmented voxel set into a binary label carrier to generate a blood vessel mask and a mask for the surrounding tissue, and keeping the index consistent with the voxel mesh description table. At the same time, a mask quality label table is generated, which records the fields of break mark, hole mark, and out-of-bounds mark. When a break mark or out-of-bounds mark is triggered, a backtracking mark is written into the segmentation path record table and the segmentation path is switched.

[0020] Furthermore, the process of radiomics feature extraction and coupled feature construction also includes:

[0021] The radiomics feature extraction process includes extracting intensity statistical features from the standardized image structure according to the vascular mask range. These intensity statistical features include mean, quantile, and kurtosis fields. Texture features are extracted from the vascular mask and surrounding tissue masks. Texture features are generated through grayscale discretization and neighborhood co-occurrence statistics. The binning rules and neighborhood definition rules of the grayscale discretization are written into the feature extraction parameter table. Morphological features are extracted, including volume, surface area, thinness, and bifurcation complexity fields obtained from the surface reconstruction of the mask voxel set. The coupled feature construction process includes establishing an interface neighborhood band on the outer wall of the vascular mask. This interface neighborhood band is formed by the vascular mask. The set of dilated voxels is composed of the set of voxels that intersect with the set of voxels of the surrounding tissue mask. The thickness of the interface neighborhood band is constrained by the inner and outer radius parameter table and the sampling step size parameter table. A coupling feature table is constructed, which contains three types of fields: texture difference index, gradient co-occurrence index, and interface morphology index. The texture difference index is generated by the difference between the texture statistics inside the vascular mask and the texture statistics of the interface neighborhood band. The gradient co-occurrence index is generated by the joint statistics of the gradient distribution in the normal direction of the vascular mask boundary and the gradient distribution in the interface neighborhood band. The interface morphology index is generated by the curvature distribution of the vascular mask boundary and the contact area distribution of the surrounding tissue mask.

[0022] Furthermore, the effective feature set structure includes:

[0023] The effective feature set structure includes an input feature list, a stability score table, a screening rule record table, a selection rule record table, a feature screening log structure, and an associated feature matrix index.

[0024] Furthermore, a radiomics identification system for perivascular inflammation in the pancreatic mesangial region for preoperative assessment, applied to any of the methods described above, includes:

[0025] The standardized image acquisition and preprocessing unit is used to perform data quality control and acquisition parameter recording, noise reduction and registration and intensity normalization, voxel resampling processing, and generate standardized image structures.

[0026] The unit for vascular centerline localization and circular region of interest construction is used to perform vascular candidate region localization, vascular skeleton extraction and centerline tracing, and circular region of interest construction to obtain the circular region of interest structure.

[0027] The segmentation and mask generation unit is used to perform segmentation of blood vessels and surrounding tissues and generate blood vessel masks and surrounding tissue masks to form a segmentation mask structure;

[0028] The radiomics feature extraction and coupled feature construction unit is used to extract intensity features, texture features, and morphological features to generate a radiomics feature table.

[0029] The feature stability screening and feature selection unit is used to read the resampling parameter record table, normalization parameter record table, registration parameter record table and acquisition parameter record table to generate a perturbation configuration set;

[0030] The domain correction model training, inference, and version management unit is used to perform domain correction processing to obtain the domain correction feature structure and generate an alignment mapping record table.

[0031] The key innovations of this invention include:

[0032] (1) Based on the standardized image structure, perform the localization of candidate regions of pancreatic mesangial vessels, extraction of vascular skeleton and tracking of center line, and complete the construction of the ring region of interest to form a ring region of interest structure, so that subsequent segmentation of blood vessels and surrounding tissues and mask generation can be carried out within a limited space.

[0033] (2) Starting from the annular region of interest structure, perform segmentation and mask generation of blood vessels and surrounding tissues, and carry out radiomics feature extraction and coupled feature construction under the constraints of the blood vessel mask and the surrounding tissue mask, merge to form a feature matrix and encapsulate it into the feature basis of the effective feature set structure.

[0034] (3) Perform feature stability screening and feature selection operations around the feature matrix to generate an effective feature set structure, and perform domain correction processing, identification model training and model parameter solidification, case reasoning and interpretability mapping generation operations under the constraints of the effective feature set structure, clinical information fields and collection parameter record table, output the reasoning report structure and update the preprocessing configuration structure.

[0035] The following are its main beneficial effects:

[0036] (1) For the localization of candidate blood vessel regions and construction of circular regions of interest in the pancreatic mesangial region, the circular region of interest structure establishes a consistent spatial reference relationship between the candidate blood vessel regions, blood vessel skeleton and centerline tracking results in the standardized image structure and the subsequent processing objects, so that the segmentation of blood vessels and surrounding tissues and mask generation can obtain clear processing boundaries, reduce the difference in range limitation due to manual image reading, and enable the circular region of interest structure to have a verifiable calling path under the same preprocessing configuration structure constraints.

[0037] (2) Regarding the extraction of radiomics features and the construction of coupled features, the radiomics features and coupled features formed by the vascular mask and the surrounding tissue mask are jointly entered into the feature matrix, so that the vascular region and the surrounding tissue region form a consistent feature organization mode under the same standardized image structure and voxel resampling processing results. This alleviates the alignment difficulties caused by the incomplete representation and inconsistent feature naming in the existing scheme that only relies on a single region feature, and makes it easier to maintain consistent field references within the effective feature set structure.

[0038] (3) For the feature stability screening and feature selection operation and subsequent inference link, the effective feature set structure forms the input feature set under the constraint of the collection parameter record table, and forms a continuous connection with the domain correction processing, recognition model training and model parameter solidification, case reasoning and interpretability mapping generation operation, so that the training and reasoning process under different collection parameter record table conditions maintains traceability in the update of the reasoning report structure and preprocessing configuration structure, and alleviates the difficulty of verification caused by feature field drift and unclear version reference when processing across batches in the prior art. Attached Figure Description

[0039] Figure 1 A schematic flowchart of a radiomics identification method for perivascular inflammation in the pancreatic mesangial region for preoperative assessment provided in an embodiment of this application;

[0040] Figure 2 A structural block diagram of a radiomics identification system for perivascular inflammation in the pancreatic mesangial region for preoperative assessment, provided in an embodiment of this application. Detailed Implementation

[0041] Example 1: Refer to Figure 1 This is a schematic flowchart of a radiomics identification method for perivascular inflammation in the pancreatic mesangial region for preoperative assessment provided by an embodiment of the present invention. The flowchart may include at least steps S100-S400:

[0042] S100: Acquire multi-phase enhanced abdominal images, clinical information fields, and preprocessing configuration structures; perform data quality control and acquisition parameter recording, noise reduction and registration, intensity normalization, and voxel resampling to generate standardized image structures.

[0043] S200: Based on standardized image structures, perform candidate region localization of pancreatic mesangial vessels, extraction of vascular skeleton and centerline tracking, and construction of circular regions of interest to obtain the circular region of interest structure.

[0044] S300, based on the ring region of interest structure, performs segmentation and mask generation of blood vessels and surrounding tissues, extraction of radiomics features and coupled feature construction processing to generate an effective feature set structure;

[0045] S400: Based on the effective feature set structure, perform domain correction processing, recognition model training and model parameter solidification processing, and generate an inference report structure.

[0046] S100: Acquire multi-phase enhanced abdominal images, clinical information fields, and preprocessing configuration structures; perform data quality control and acquisition parameter recording, noise reduction and registration, intensity normalization, and voxel resampling to generate standardized image structures.

[0047] The aforementioned radiomics identification method and system for perivascular inflammation in the pancreatic mesangial region for preoperative assessment of pancreaticoduodenectomy (PD) refers to PD. The input sources for this step include three categories: multi-phase enhanced abdominal images, clinical information fields, and preprocessing configuration structures. Multi-phase enhanced abdominal images refer to abdominal tomographic image sequences obtained from the same subject at different time phases before and after enhancement. Clinical information fields refer to a set of structured fields related to the perioperative assessment of the subject. The preprocessing configuration structure refers to a set of rules and parameters required for data quality control, noise reduction and registration, intensity normalization, and voxel resampling. The preprocessing configuration structure for this step comes from the configuration center within the same system or from a rewritten version of the preprocessing configuration structure updated in S400. During the first run, the preprocessing configuration structure is loaded from a built-in template and a version number record field is generated, ensuring that subsequent main steps have a consistent configuration reference. After reading the above input, this step first establishes an image sequence index table for multi-phase enhanced abdominal images. The image sequence index table includes fields such as sequence identifier, phase identifier, slice order, spatial resolution information, and timestamp information. The clinical information fields are then mapped to a unified namespace according to the field dictionary to generate a clinical information field mapping table, which can be used by S400 in the subsequent domain correction and model version management stages.

[0048] In the data quality control and acquisition parameter recording stage, the system performs data quality control processing on the multi-phase enhanced abdominal images. This data quality control refers to a screening process focused on sequence integrity, image usability, and metadata consistency, encompassing three processing links: missing segment detection, artifact marking, and sequence integrity verification. Missing segment detection identifies gaps in consecutive slices by comparing the slice order field in the image sequence index table with the actual loaded slice count, and writes the missing range field into the data quality control marking table. Artifact marking involves a coarse screening of each phase slice before intensity anomaly suppression, combined with texture anomalies such as stripes, metallic artifacts, and motion blur, and marks them with an artifact type field into the data quality control marking table. Sequence integrity verification generates a sequence integrity verification result field by checking the phase identifier set against the phase set specified in the preprocessing configuration structure. In parallel, the system performs acquisition parameter recording processing. The acquisition parameter recording table is a structured record carrier, containing fields such as equipment model, slice thickness, reconstruction settings, and contrast agent injection settings. The equipment model comes from image metadata, slice thickness from the slice spacing field, reconstruction settings from the reconstruction kernel and matrix field, and contrast agent injection settings from the injection rate and dosage field or the in-hospital examination form field. When any of these fields are missing, the system writes a missing marker field into the acquisition parameter recording table and simultaneously writes it into the preprocessing log structure. The preprocessing log structure records the triggering conditions, anomaly types, handling actions, timestamps, and associated sequence identifiers for each step of this process, used for subsequent S400 operation auditing and version write-back records. After completing this step, this step produces an intermediate product with the output field named "Original Image Parameter Structure." This original image parameter structure aggregates multi-phase enhanced abdominal image pointers, multi-temporal metadata, data quality control marker table pointers, and acquisition parameter recording table pointers, and serves as the input source for denoising registration and intensity normalization in subsequent processing steps.

[0049] In the denoising and registration stage, the system imports multi-phase enhanced abdominal images from the original image parameter structure and, in conjunction with the data quality control label table, applies a skipping or weighting strategy to segments that fail the test, followed by denoising processing. This denoising process involves suppressing image intensity noise and local anomalies. First, a spatial filtering strategy is selected based on the preprocessing configuration structure, and a denoising parameter record table is generated. The spatial filtering strategy can be neighborhood-based smoothing or edge-preserving filtering. Both strategies are parameterized using fields such as window size, iteration count, and edge threshold. The denoising parameter record table is written into the preprocessing log structure along with a sequence identifier. After denoising, the system proceeds to registration, which involves aligning multi-phase sequences in a spatial coordinate system. First, a reference phase is selected and a reference phase identifier field is generated. Then, spatial alignment and interpolation resampling are performed on the remaining phases. Spatial alignment includes two levels: rigid alignment and non-rigid alignment. Rigid alignment eliminates overall displacement and rotation differences, while non-rigid alignment compensates for local deformation differences. Interpolation resampling restores the image to a unified grid after alignment transformation. If a registration failure occurs during the alignment process, the system writes the failure sequence identifier into the registration failure list and records the failure reason field in the preprocessing log structure. Simultaneously, a rollback strategy is triggered, which includes switching the reference phase or switching the alignment level. The output field of this step is named "Registered Image Sequence," which serves as the input source for subsequent intensity normalization and voxel resampling. The image sequence index table is also transmitted along with this sequence, allowing S200 to call the corresponding image by phase index when performing pancreatic mesangial vascular candidate region localization.

[0050] In the intensity normalization stage, the system performs intensity normalization processing on the registered image sequence. Intensity normalization refers to the process of mapping the intensity distribution of images under different scanning batches and reconstruction settings to a unified intensity domain. During implementation, a normalization grouping key is generated by grouping the sequences. The normalization grouping key includes fields such as equipment model, slice thickness, and reconstruction settings. Based on this, a normalization strategy is selected and a normalization parameter record table is generated. The normalization strategy includes two types of links: intensity range mapping and histogram matching records. Intensity range mapping completes normalization by setting an input intensity cutoff interval and mapping it to a target interval. Histogram matching records complete normalization by establishing a reference histogram and aligning the histograms of each sequence. Both strategies have strategy type, threshold, and reference sequence identifier fields written into the normalization parameter record table. During intensity normalization, the system performs elimination or interpolation rules on voxels whose quality is marked as unacceptable by a mask. The quality mask is generated by the data quality control stage and passed in with the registered image sequence to constrain the sampling range of normalization statistics. After intensity normalization, voxel resampling is performed. Voxel resampling refers to the process of unifying the image voxel spacing and voxel grid to the target grid specified in the preprocessing configuration structure. During implementation, the target voxel spacing field and the target grid size field are read, a resampling interpolation method is selected, and a resampling parameter record table is generated. The resampling parameter record table is written with the interpolation method field, the target voxel spacing field, and the target grid size field, and is simultaneously written to the preprocessing log structure. At the same time, a voxel grid description table is generated. The voxel grid description table contains fields such as original grid information, target grid information, coordinate origin information, and direction cosine information, providing a unified grid reference for subsequent cross-sequence alignment by the S300 in the image omics feature extraction stage.

[0051] In the standardized image structure generation stage, the system encapsulates the registered image sequence after intensity normalization and voxel resampling into a standardized image structure. This standardized image structure includes fields such as a standardized image data pointer, a standardized image index table, a voxel grid description table pointer, a normalization parameter record table pointer, a resampling parameter record table pointer, and a preprocessing log structure pointer. The standardized image index table is generated by expanding the image sequence index table, supplementing it with fields such as a normalization grouping key, a registration reference phase identifier, and a resampling target grid identifier. For abnormal paths, if a registration failure list exists, the system writes a degradation flag field to the standardized image index table and retains the index entry for the failed sequence, allowing subsequent steps to skip or replace it based on the index entry. After encapsulation, this step inputs the product with the output field name "Standardized Image Structure" into the standardized image structure of S200, which is then used by S200 to perform pancreatic mesangial region vascular candidate region localization, vascular skeleton extraction and centerline tracking, and circular region of interest construction. At the same time, the output field name "Voxel Mesh Description Table" is input into the voxel mesh description table of S300, so that S300 uses a consistent mesh when performing vascular and surrounding tissue segmentation and mask generation, radiomics feature extraction and coupled feature construction. The acquisition parameter record table, normalization parameter record table, resampling parameter record table, and preprocessing log structure are input as traceable record fields into the domain correction processing and model version management link of S400, serving as the reference source when generating the group key structure, alignment mapping record table, and version number record table, thus forming a cross-main step connection relationship from S100 to S200, S300, and S400.

[0052] In summary, the technical effects of this step are as follows: By performing data quality control, denoising and registration, intensity normalization, and voxel resampling on multi-phase enhanced abdominal images to form a standardized image structure, a unified input format across sequences is achieved. At the same time, the acquisition parameters and preprocessing parameters are recorded and can be used in subsequent steps.

[0053] S200: Based on standardized image structures, perform candidate region localization of pancreatic mesangial vessels, extraction of vascular skeleton and centerline tracking, and construction of circular regions of interest to obtain the circular region of interest structure.

[0054] Specifically, in the aforementioned radiomics identification method and system for preoperative assessment of pancreatic mesangial perivascular inflammation, this step receives the input product from the S100 output field named "Standardized Image Structure," and establishes the calculation coordinate reference for this step by combining the standardized image index table and voxel grid description table within the standardized image structure. The pancreatic mesangial vascular candidate region localization in this step refers to the screening of candidate voxels, merging of connected components, and quality constraints on tubular high-enhancement structures within the pancreatic mesangial region within the abdominal enhanced image, thereby forming the input set for subsequent vascular skeleton extraction and centerline tracing. When the pancreatic mesangial region structure is first generated in this step, the abdominal cavity boundary and spinal reference plane fields are extracted from the standardized image structure. Subsequently, a three-dimensional spatial constraint box is constructed based on the relative position constraints of the pancreatic mesangial region within the body cavity. These relative position constraints include a longitudinal range near the root of the mesentery, an anterior-posterior range near the posterior aspect of the pancreas, and an exclusion range avoiding the air area outside the body cavity and the high-density area of ​​the cortical bone. Furthermore, the system extracts images from each temporal phase from the standardized image structure, selects enhanced temporal phase combinations as candidate screening sources based on the standardized image index table, and performs tubular structure response calculation for each temporal phase. The tubular structure response is a multi-scale filtering output for slender connected morphologies. Its implementation involves calculating local directional consistency and boundary gradient stability within different scale neighborhoods to form a response volume, and then performing threshold segmentation on the response volume to generate a candidate voxel set. After the candidate voxel set is generated, the system performs connected component analysis under the structural constraints of the pancreatic mesangial region. The connected component analysis outputs a candidate connected component table, which includes fields such as connected component identifier, voxel count, principal axis direction, average enhancement intensity, and bifurcation complexity. Simultaneously, a candidate region quality label table is generated, recording anomalies such as artifact interference, breakage degree, and spatial boundary violations. If the candidate connected component table exhibits an excessive or insufficient number of connected components at any given time phase, the system reads the localization rules from the preprocessing configuration structure and writes them into the localization rule record table. Subsequently, a rule switching process is executed, which includes three types of actions: threshold self-adaptive adjustment, scale combination adjustment, and time phase combination adjustment. The triggering conditions and actions are recorded in the preprocessing log structure. After candidate screening and connected component merging, this step generates an intermediate product with the output field named "vascular candidate region structure." This structure includes a candidate voxel set pointer, a candidate connected component table pointer, a pancreatic mesangial region range structure pointer, and a candidate region quality label table pointer. This vascular candidate region structure serves as the input source for subsequent vascular skeleton extraction and centerline tracing in this step. Simultaneously, at the field level, the vascular candidate region structure is used for verification against the candidate region index table during subsequent segmentation mask generation in S300, forming a cross-main step connection from S100 to S200 and then to S300.

[0055] Further, this step proceeds to the vascular skeleton extraction and centerline tracing processing chain. Vascular skeleton extraction refers to transforming the candidate voxel set within the vascular candidate region structure into a thin line representation that maintains the connected topology, thus forming the graph structure basis for centerline tracing. The implementation process involves first performing hole filling and break bridging processing on the candidate voxel set. Hole filling identifies voxel voids through local connectivity detection and performs neighborhood completion. Break bridging is completed through short-distance endpoint pairing and morphological closure operations. Subsequently, the completed voxel set undergoes thinning processing. This thinning process iteratively strips boundary voxels while maintaining the topology of endpoints and bifurcation points, generating a vascular skeleton voxel set. The system constructs a skeleton connected graph for the vascular skeleton voxel set. The skeleton connected graph consists of a node set and an edge set, where the node set is formed by endpoints and bifurcation points, and the edge set is formed by the connectivity relationships between adjacent skeleton voxels. To support the subsequent construction of circular regions of interest, this step involves numbering the bifurcation points. The bifurcation point numbering table records fields such as bifurcation point identifier, number of connected branches, branch direction vector, and adjacent edge list. Noise branch cleanup is performed on the skeleton connected graph, eliminating short pseudo-branches based on branch length and intensity consistency thresholds. The cleanup rules are then written into the centerline tracing parameter table. Next, centerline tracing is performed, which involves generating ordered paths and outputting coordinate sequences on the skeleton connected graph. This is achieved by first selecting candidate trunk connected regions based on the principal axis direction field in the candidate connected region table, then selecting entry endpoints within these candidate trunk connected regions and performing path traversal. Path traversal uses a progressive expansion strategy based on the connected graph to generate the centerline coordinate sequence. Simultaneously, for each bifurcation point, branch selection and recording are performed according to the bifurcation point numbering table, forming a centerline quality marker table. The centerline quality marker table includes fields such as path continuity, bifurcation stability, and loop detection markers. Loop detection identifies repeated node visits and triggers a loop-breaking strategy. The loop-breaking strategy adds loop edges to a shielding list and writes them into the centerline tracing parameter table. When the centerline quality marker table indicates insufficient path continuity, this step triggers a secondary bridging process by calling the breakage degree field of the candidate region quality marker table. This secondary bridging process re-executes breakage bridging and repeats centerline tracing without altering the structural constraints of the pancreatic mesangial region. Upon completion of this step, an intermediate product named "vascular centerline structure" is generated. This structure includes pointers to the centerline coordinate sequence, vascular skeleton voxel sets, bifurcation point numbers, centerline tracing parameters, and centerline quality markers. Within this step, it serves as the input source for constructing the annular region of interest. Simultaneously, the centerline coordinate sequence and bifurcation point numbers are used at the field level for cross-sectional normal generation and index consistency verification during S300 feature extraction, forming a consistent reference basis across steps.

[0056] Further, this step proceeds to the annular region of interest (ROI) construction processing link. The annular ROI construction refers to the process of generating a three-dimensional annular shell sampling region around the vessel centerline structure and outputting a structured index. The annular ROI is used to constrain the spatial range of subsequent vessel and surrounding tissue segmentation and radiomics feature extraction. The implementation process involves first extracting the centerline coordinate sequence from the vessel centerline structure and then reparameterizing the centerline. Centerline reparameterization generates a sampling point sequence by uniformly sampling according to arc length. The sampling step size is taken from the preprocessing configuration structure and written into the sampling step size parameter table. Subsequently, the system constructs a cross-sectional normal sequence at each sampling point. The cross-sectional normal sequence generates a local tangential direction through the direction difference between adjacent sampling points and combines it with the direction cosine field of the voxel mesh description table for coordinate consistency, forming a stable cross-sectional normal definition. Furthermore, the system generates an inner and outer radius parameter table based on the inner and outer radius rules in the preprocessing configuration structure. These rules include two types: a fixed radius scheme and a segmented radius scheme. The fixed radius scheme uses the same inner and outer radius fields for all sampling points. The segmented radius scheme assigns different radius fields to different segments based on the local vessel scale estimation field of the connected domain where the centerline is located. The local vessel scale estimation is obtained through statistics of voxel density and principal axis cross-sectional width in the candidate connected domain table. Based on this, the system generates an annular cross-sectional voxel set at each sampling point and accumulates it along the centerline to form an annular shell voxel set. Simultaneously, it establishes an index association between the annular shell voxel set and the centerline coordinate sequence to generate an annular region of interest index table. This annular region of interest index table includes fields such as sampling point identifier, cross-sectional identifier, corresponding voxel set pointer, inner and outer radius fields, and cross-sectional normal field. If the annular shell voxel set crosses the body cavity boundary or overlaps with the cortical bone region at certain sampling points, this step triggers local clipping processing based on the spatial boundary crossing field of the candidate region quality marker table. Local clipping removes the out-of-bounds voxels and writes a clipping marker field into the annular region of interest index table. Simultaneously, the clipping rules are written into the localization rule record table and the preprocessing log structure. After the above construction is completed, this step generates the final product with the output field named "annular region of interest structure." The annular region of interest structure includes pointers to the annular shell voxel set, the annular region of interest index table, the cross-sectional normal sequence, the inner and outer radius parameter tables, the sampling step size parameter table, and the vessel centerline structure. This annular region of interest structure is then input into the annular region of interest structure of S300 for S300 to perform operations such as vessel and surrounding tissue segmentation and mask generation, radiomics feature extraction, and coupled feature construction. Meanwhile, the centerline tracking parameter table, positioning rule record table, inner and outer radius parameter table, and sampling step size parameter table generated in this step are entered into the S400 version write-back link as parameter record output fields and participate in the preprocessing configuration structure update, so that subsequent batches can reuse the running rules of this step under the same configuration framework.

[0057] In summary, the technical effects of this step are as follows: by locating candidate regions of pancreatic mesangial vessels, extracting the vascular skeleton, and tracing the centerline on a standardized image structure, and constructing a circular region of interest structure, subsequent segmentation and radiomics feature extraction have clear spatial constraints. At the same time, key rules and parameters are recorded and can be used for subsequent version management.

[0058] S300, based on the ring region of interest structure, performs segmentation and mask generation of blood vessels and surrounding tissues, extraction of radiomics features and coupled feature construction processing to generate an effective feature set structure;

[0059] Specifically, in the aforementioned radiomics identification method and system for pancreatic mesangial perivascular inflammation in preoperative PD assessment, this step receives the input product from the S200 output field named "ring region of interest structure" and reads the standardized image structure and voxel grid description table formed by S100 as the coordinate reference for the same source image, thereby enabling the segmentation, feature extraction, and screening in this step to operate under a unified grid. The segmentation of blood vessels and surrounding tissues in this step refers to the process of separating the vascular region and surrounding tissue region from the standardized image structure within the spatial range defined by the ring region of interest structure and forming a voxel-level labeled set. The surrounding tissue is defined as a set of adipose tissue and soft tissue voxels within the vicinity of the outer wall of the blood vessel, and its spatial boundaries are constrained by the inner and outer radius parameter table and the cross-sectional normal sequence. The mask generation refers to the process of writing the segmented voxel set into a binary label carrier and maintaining consistency with the index of the voxel grid description table. The mask includes two types: vascular masks and surrounding tissue masks. During runtime, the system first extracts the annular shell voxel set and the annular region of interest index table from the annular region of interest structure. Then, it retrieves the corresponding voxel blocks from the standardized image structure according to the sampling point identifier and cross-section identifier in the annular region of interest index table, forming a segmentation input block sequence. Simultaneously, the index range field of this segmentation input block sequence is written into the segmentation path record table. Subsequently, segmentation processing is performed, which includes three paths: threshold segmentation, region growing, and learning model inference. All three paths are parameterized and registered through the segmentation parameter record table. The threshold segmentation path sets an intensity threshold based on the enhancement intensity field and candidate region index field of the standardized image structure and generates an initial set of blood vessel voxels. Then, it performs boundary contraction and expansion corrections along the normal direction according to the cross-section normal sequence, thereby forming blood vessel mask candidates. The region growing path selects seed points at the endpoints of the blood vessel mask candidates and expands them according to the connected component similarity criterion. The connected component similarity criterion consists of an intensity difference threshold, a gradient consistency threshold, and a morphological continuity threshold, and is written into the segmentation parameter record table. The learning model inference path inputs the segmentation input block sequence into the trained segmentation model and outputs a probability map. Then, it generates a set of blood vessel voxels and a set of voxels from the surrounding tissue according to the probability threshold and connectivity constraints. In case of anomalies, the system makes a judgment based on the mask quality mark table, which records fields such as break mark, hole mark, and boundary cross mark. The break mark is generated by counting the number of connected segments in the vascular mask, the hole mark is generated by voxel hole detection, and the boundary cross mark is generated by the boundary detection of the annular shell voxel set. When a break mark or boundary cross mark is triggered, the system writes a rollback mark to the segmentation path record table and switches the segmentation path. After the rollback is completed, the handling action is written to the preprocessing log structure and the original path record is retained.After segmentation and mask generation are completed, the system generates an intermediate product with the output field name "segmentation mask structure". The segmentation mask structure aggregates fields such as vascular mask, surrounding tissue mask, segmentation path record table, segmentation parameter record table, and mask quality label table. The segmentation mask structure is used as the input source of the segmentation mask structure of S320. At the same time, the vascular mask and the surrounding tissue mask are used as the input sources of the vascular mask and the surrounding tissue mask of S320, so that the subsequent radiomics feature extraction and coupled feature construction can obtain a clear spatial object.

[0060] Further, this step performs radiomics feature extraction and coupled feature construction based on the segmented mask structure. Radiomics feature extraction refers to the process of extracting verifiable numerical descriptions of intensity, texture, and morphology from a standardized image structure according to the mask range and forming a structured feature table. Coupled feature construction refers to the process of constructing cross-regional combined features reflecting the boundary interface and neighborhood differences, based on the relationship between the vascular mask and the surrounding tissue mask, and forming a structured feature table. During runtime, the system first performs a consistency check on the coordinate indices of the vascular mask and the surrounding tissue mask based on the voxel grid description table. If an offset is found between the mask index and the standardized image index table, an offset marker is written to the feature missing marker table, triggering the mask remapping process. The mask remapping process reconstructs the mask index based on the direction cosine and coordinate origin fields of the voxel grid description table, and then writes the remapping action into the feature extraction parameter table. The system then proceeds to radiomics feature extraction. It extracts intensity statistical features from the standardized image structure according to the vascular mask range. These features include mean, quantile, and kurtosis fields, which are written into the radiomics feature table. In parallel, the system extracts texture features from the vascular mask and surrounding tissue masks. Texture features are generated through grayscale discretization and neighborhood co-occurrence statistics. The binning rules and neighborhood definition rules of the grayscale discretization are written into the feature extraction parameter table. The system continues to extract morphological features. Morphological features are obtained based on surface reconstruction of the mask voxel set, yielding fields such as volume, surface area, thinness, and bifurcation complexity, which are also written into the radiomics feature table. Next, the system enters the coupled feature construction stage. It first establishes an interface neighborhood band on the outer wall of the vascular mask. This band consists of the expanded voxel set of the vascular mask and the intersection voxel set of the surrounding tissue mask. The thickness of the interface neighborhood band is constrained by the inner and outer radius parameter tables and the sampling step size parameter table, and is written into the feature extraction parameter table. Subsequently, a coupled feature table is constructed, containing three types of fields: texture difference index, gradient co-occurrence index, and interface morphology index. The texture difference index is generated by the difference between the texture statistics within the vascular mask and the texture statistics of the interface neighborhood band. The gradient co-occurrence index is generated by the joint statistics of the gradient distribution in the normal direction of the vascular mask boundary and the gradient distribution in the interface neighborhood band. The interface morphology index is generated by the curvature distribution of the vascular mask boundary and the contact area distribution of the surrounding tissue mask. For abnormal cases, the system writes a missing type field into the feature missing marker table. The missing type field includes categories such as empty mask, insufficient voxel count, grayscale binning failure, and texture matrix degradation, and writes the missing feature name into the missing mapping segment of the feature naming dictionary. The feature naming dictionary refers to a dictionary structure that uniformly names, identifies versions, and identifies the source object for all feature fields. The source object identifier is limited to three categories: vascular mask, surrounding tissue mask, and interface neighborhood band. The version identifier is bound to the feature extraction parameter table.After completing radiomics feature extraction and coupled feature construction, the system merges the radiomics feature table and the coupled feature table according to the sample index to generate an intermediate product with the output field named feature matrix. This feature matrix serves as the input source for the feature matrix in S330, and the feature naming dictionary serves as the input source for the feature naming dictionary in S420. This ensures consistent naming and traceable referencing of the input feature fields during subsequent model training and model parameter solidification. This image processing link falls within the image analysis scope defined by G06T7 / 00. The feature extraction parameter table and the feature missing marker table, as record carriers of the image analysis process, are synchronously written into the preprocessing log structure and retrieved by the S400 version write-back link.

[0061] Further, this step performs feature stability screening and feature selection on the feature matrix and generates an effective feature set structure. Feature stability screening refers to the process of quantitatively evaluating the repeatability of feature fields under controlled perturbation and batch repetition conditions, and accordingly removing or retaining markers. Feature selection refers to the process of filtering for correlation, redundancy, and missing values ​​based on stability screening and forming an input feature list. During runtime, the system first reads the resampling parameter record table, normalization parameter record table, registration parameter record table, and acquisition parameter record table to generate a perturbation configuration set. The perturbation configuration set includes three types of perturbation entries: voxel grid perturbation, intensity bin perturbation, and registration reference phase switching. The system also registers the triggering conditions and amplitude fields for each type of perturbation in the screening rule record table. Subsequently, the system recalculates or remaps the feature matrix for each perturbation entry, obtaining a multi-version feature matrix sequence. It then calculates consistency statistics for the same feature field to generate a stability score table, which includes fields such as field identifier, perturbation entry identifier, score value, and anomaly marker. When a break or out-of-bounds marker appears in the mask quality marker table, the system writes a masking marker to the stability score table and removes the sample from the statistical set of the corresponding perturbation entry, simultaneously recording the removal action in the feature screening log structure. Next, the system enters the outlier removal rule record. It performs outlier detection on the feature matrix, generating an outlier list based on quantile truncation and neighborhood consistency checks. The truncation and check rules are registered in the screening rule record table, and the outlier list is processed and written to the feature screening log structure. After stability screening, the system performs feature selection processing on the feature fields. Feature selection first performs correlation filtering, which calculates the correlation between feature fields and deletes highly correlated fields according to threshold rules; these threshold rules are written to the selection rule record table. Next, redundancy removal is performed, which determines the retention strategy for duplicate descriptive fields within the same source object based on the source object identifier and type identifier in the feature naming dictionary, and this determination is written to the selection rule record table. Finally, missing threshold filtering is performed, which calculates the missing proportion based on the feature missing marker table and deletes fields exceeding the threshold, while simultaneously writing the missing handling record to the feature screening log structure. After feature selection, the system generates an input feature list, which consists of a set of feature field identifiers and their mapping entries in the feature naming dictionary. This input feature list serves as the input feature list source for the S420.Finally, the system encapsulates the stability scoring table, screening rule record table, selection rule record table, and input feature list with the screened feature matrix to generate an output field named "effective feature set structure". The effective feature set structure serves as the input source for the effective feature set structure of S410, and the stability scoring table serves as the input source for the stability scoring table of S410, which is then used by S400 to perform domain correction processing, recognition model training and model parameter solidification, case reasoning and interpretability mapping generation operations. Furthermore, the feature screening log structure serves as the associated input source for the training log structure and the running audit log structure of S430, thus providing a consistent screening rule reference when S400 updates the preprocessing configuration structure.

[0062] In summary, the technical effects of this step are as follows: By forming a segmentation mask structure under the constraint of a ring-shaped region of interest structure and completing radiomics feature extraction, coupled feature construction and stability screening, an effective feature set structure that can be used for subsequent domain correction and recognition model training is obtained, and key parameters and rules are recorded for subsequent version management links to call.

[0063] S400: Based on the effective feature set structure, perform domain correction processing, identify model training and model parameter solidification processing, and generate an inference report structure;

[0064] Specifically, in the aforementioned radiomics identification method and system for perivascular inflammation in the pancreatic mesangial region for preoperative assessment of PD, this step receives the input product from the output field named "Effective Feature Set Structure" from S300, and receives the clinical information field and acquisition parameter record table formed from S100 as batch-related inputs. Simultaneously, it reads the preprocessing log structure and feature naming dictionary formed from S100 for version consistency verification. The "Effective Feature Set Structure" in this step refers to the encapsulated structure of the feature set obtained after feature stability screening and feature selection, including an input feature list, stability scoring table, screening rule record table, selection rule record table, feature screening log structure, and associated feature matrix index. The "Clinical Information Field" refers to a structured set of fields surrounding the basic information of the examinee, preoperative assessment fields, and imaging examination information fields. The "Acquisition Parameter Record Table" refers to a structured record carrier of parameters such as the model of the imaging acquisition device, slice thickness, reconstruction settings, and enhancement injection settings. At the start of operation, the system first performs a consistency check on the input data. Specifically, it aligns the input feature list with the feature naming dictionary, writes missing fields into a feature missing marker table, and records them in the operation audit log structure. It also performs an association check between the acquisition parameter record table and the subject identifier in the clinical information field, writing mismatched records into the anomaly marker segment of the training data index table and recording them in the operation audit log structure. Finally, it compares the screening rule record table within the effective feature set structure with the version number record field of the preprocessing configuration structure, writing inconsistent version records into the version number record table and adding them to the version write-back candidate queue. After completing the consistency check, the system binds the effective feature set structure, clinical information field, and acquisition parameter record table together to generate a grouping key structure. This grouping key structure consists of a device model field, slice thickness field, reconstruction setting field, and phase identifier field, and serves as the basis for grouping and mapping indexes in subsequent domain correction processing.

[0065] Further, this step performs domain correction processing. Domain correction refers to the process of constructing a distribution alignment map and applying it to feature fields to address the feature distribution differences introduced by different device models and different combinations of acquisition parameters. During runtime, the system retrieves the feature matrix from the effective feature set structure and divides it into multiple feature subsets according to the grouping key structure. Simultaneously, it writes the grouping key identifier and sample index range for each group into the alignment map record table. The system then performs distribution statistical extraction on each feature subset. Distribution statistical extraction includes three types of fieldized results: central statistics, discrete statistics, and quantile statistics of the feature fields, and writes them into the domain correction parameter record table as the basis for mapping generation. When the sample size of a group is insufficient to trigger a threshold condition, the system records an insufficient flag in the domain correction log structure and merges the group into the adjacent grouping key according to the merging rule. The merging rule is written into the merging segment of the alignment map record table. Subsequently, the system generates an alignment mapping record table based on the domain correction parameter record table. This alignment mapping record table includes a grouping key identifier, feature field identifier, mapping type identifier, mapping parameter field, and applicable scope field. During generation, a verification flag is written to the mapping parameter field. The verification flag comes from boundary constraint checks and outlier rollback checks on the mapped feature subset. When an outlier rollback occurs, the system writes a rollback flag to the domain correction log structure and switches the corresponding mapping type identifier to an alternative mapping type identifier. After mapping generation is complete, the system applies the alignment mapping record table to the feature matrix to form a domain-corrected feature structure. This domain-corrected feature structure includes pointers to the domain-corrected feature matrix, grouping key structure, alignment mapping record table, domain correction parameter record table, and domain correction log structure. Within this step, the domain-corrected feature structure serves as the input source for recognition model training and case reasoning. Simultaneously, the alignment mapping record table and domain correction parameter record table are written to the preprocessing log structure for version write-back. The image analysis and feature alignment operation in this step falls within the image analysis processing scope defined by G06T7 / 00, and related records form a traceable link through the domain correction log structure and the operation audit log structure.

[0066] Further, this step performs recognition model training and model parameter solidification, and executes case reasoning and interpretability mapping generation within the same version management chain. The recognition model training refers to the process of constructing a classifier structure and generating a model parameter package based on the domain correction feature structure and clinical information fields. The model parameter solidification refers to the process of encapsulating the trained parameters, input feature list, domain correction mapping version number, and preprocessing parameter version number into a reusable version entity. During runtime, the system first divides the domain correction feature structure into a training subset and a validation subset according to the training data index table and records the division rules to the training log structure. The division rules include a timestamp field, a grouping key coverage field, and a sample index range field. The system then extracts the structured variables involved in modeling from the clinical information fields and performs field normalization and missing value imputation. The imputation rules are written into the field processing section of the training log structure. Subsequently, the system calls the input feature list to select columns for the domain correction feature structure, generates a training feature tensor, and inputs it into the recognition model training process. In this invention, the recognition model is limited to a classifier structure from features to inflammation recognition labels, including an algorithm type field, a hyperparameter table field, a threshold table field, a loss function selection field, and a stopping condition field. These fields are written to the model review record table and the training log structure. During training, the system writes a weight update summary field and records a random seed identifier field and a grouping key overlay field for each iteration, for subsequent experimental reproduction. When training is interrupted or data is abnormal, the system writes the interruption reason field to the training log structure and triggers the breakpoint recovery process. The breakpoint recovery process writes the pointer to the most recent model parameter package temporary file to the model parameter solidification candidate queue. After training is complete, the system generates a model parameter package and a threshold table, which includes decision threshold fields and calibration label fields for various labels. Subsequently, the system performs model parameter solidification processing. It encapsulates the model parameter package, threshold table, input feature list, alignment mapping record table version number, and preprocessing parameter version number into a model version structure. The version number record table contains the version number, generation timestamp, applicable grouping key range, and dependency record fields. These dependency record fields are linked to pointers to the normalization parameter record table, resampling parameter record table, centerline tracking parameter table, feature extraction parameter table, screening rule record table, and selection rule record table, thus providing the model version structure with the dependency information required for end-to-end operation. Within this step, the model version structure serves as the input source for case inference and is registered as a searchable version entity in the system storage layer. During the case inference phase, the system calls the aforementioned model version structure for the case to be evaluated. First, it searches for a matching alignment mapping record table according to the grouping key structure and performs homology domain correction on the feature vector of the case to be evaluated. Then, it inputs the corrected feature vector into the recognition model to generate inflammation recognition labels and confidence fields. The system also writes the grouping key, feature missing marker, and threshold table version number of the inference process into the runtime audit log structure.Subsequently, an interpretability mapping generation process is performed. This interpretability mapping refers to the process of mapping the contribution relationship of input features to the inference output back to the annular region of interest (ROI) structure index space and forming a gridded contribution representation. This is achieved by determining the feature source object identifier based on a feature naming dictionary, aggregating contribution values ​​to the voxel set indexes corresponding to the vascular mask, surrounding tissue mask, and interface neighborhood band according to the source object identifier, and generating an interpretability mapping index table by combining the annular ROI index table and the voxel grid description table. The interpretability mapping index table includes a mapping grid identifier, voxel index range, contribution value encoding segment, associated mask index segment, and version number field. This interpretability mapping index table is then written into the inference report structure. The inference report structure is a structured report carrier containing an inflammation identification label, confidence field, interpretability mapping index, version number record field, and runtime audit log index field. After generating the inference report structure, the system stores it and registers the retrieval key. Simultaneously, the index information of the inference report structure is written back to the version write-back record segment of the preprocessing configuration structure, forming a traceable entry point for subsequent batch runs. Finally, this step performs an update process on the preprocessing configuration structure. The update process includes writing the version number of the alignment mapping record table, the version number of the threshold table, the version number of the input feature list, the coverage range of the grouping key, and the exception rollback rules into the preprocessing configuration structure. The updated preprocessing configuration structure is then sent back as the output field name preprocessing configuration structure for S100 to call, thus forming a closed-loop connection across the main steps.

[0067] In summary, the technical effects of this step are as follows: by performing domain correction on the effective feature set structure, training and solidifying the recognition model, and completing the generation of case reasoning and interpretability mapping, a reasoning report structure is formed and key version information is written back to the preprocessing configuration structure, so that subsequent operations have a consistent version reference and recording link.

[0068] Example 2: Figure 2 A structural block diagram of a radiomics identification system for perivascular inflammation in the pancreatic mesangial region, oriented towards preoperative assessment, is shown according to an embodiment of the present invention. Figure 2 As shown, the structure may include:

[0069] The standardized image acquisition and preprocessing unit 01 is used to acquire multi-phase enhanced abdominal images, clinical information fields, and preprocessing configuration structures. It performs data quality control and acquisition parameter recording, denoising registration and intensity normalization, voxel resampling, and generates a standardized image structure, as well as an acquisition parameter record table and a preprocessing log structure. Specifically, the standardized image acquisition and preprocessing unit receives sequence data and metadata of multi-phase enhanced abdominal images, receives clinical information fields associated with the same subject, and receives a preprocessing configuration structure as an operational constraint. The preprocessing configuration structure includes data quality control rules, an acquisition parameter record field set, denoising configuration items, registration configuration items, intensity normalization configuration items, voxel resampling configuration items, and a version number record field. After loading multi-phase enhanced abdominal images, the standardized image acquisition and preprocessing unit generates an image sequence index and performs temporal consistency and slice continuity checks. Missing segments, artifact interference, and incomplete sequences are written to the data quality control marker area and simultaneously recorded in the preprocessing log structure. During the acquisition parameter recording process, the standardized image acquisition and preprocessing unit extracts information such as device model, slice thickness, reconstruction settings, and enhancement injection settings from image metadata and clinical information fields and writes them to the acquisition parameter record table. Missing items are marked with missing markers and the corresponding sequence identifiers are recorded in the preprocessing log structure. Denoising, registration, and intensity normalization are performed under the constraints of the preprocessing configuration structure. Denoising performs noise suppression on each temporal image and records denoising parameters. Registration performs alignment on multi-temporal images and records reference temporal identifiers and registration treatment records. Intensity normalization maps the image intensity domain according to the grouping key and records normalization parameters. Voxel resampling reads the target voxel spacing and target grid size, performs interpolation resampling, and records resampling parameters. Simultaneously, a voxel grid description table is generated and associated with the image sequence index. The standardized image acquisition and preprocessing unit encapsulates the above-mentioned processing into a standardized image structure. The standardized image structure includes a standardized image data pointer, a standardized image index table, and a voxel grid description table index. The standardized image structure is provided to the vessel centerline localization and annular region of interest construction unit as its input object. At the same time, the acquisition parameter record table and the preprocessing log structure are provided to the domain correction model training inference and version management unit as its input objects. The preprocessing configuration structure and the preprocessing log structure maintain a consistent association with the version number record field and are available for subsequent write-back calls.

[0070] The vessel centerline localization and annular region of interest (ROI) construction unit 02 is connected to the standardized image acquisition and preprocessing unit. It is used to perform pancreatic mesangial vessel candidate region localization, vessel skeleton extraction and centerline tracking, and annular ROI construction based on the standardized image structure, obtaining an annular ROI structure and generating an annular ROI index table, cross-sectional normal sequence, inner and outer radius parameter table, and sampling step size parameter table. Specifically, the vessel centerline localization and annular ROI construction unit receives the standardized image structure output from the standardized image acquisition and preprocessing unit, reads the standardized image index table and voxel grid description table within the standardized image structure as spatial index references, retrieves corresponding image data according to the temporal identifier of the standardized image index table, and forms candidate filtering inputs. The localization of candidate vascular regions in the pancreatic mesangial region is performed under the structural constraints of the pancreatic mesangial region. The structure of the pancreatic mesangial region is generated by the vascular centerline localization and the annular region of interest construction unit based on the body cavity boundary, the spinal reference level, and the spatial bounding box, and is associated with the sequence identifier. The candidate region localization calculates the tubular structure response in the neighborhood at different scales and forms a candidate voxel set. Then, connected component merging is performed to generate a candidate connected component table and a candidate region quality label table. The candidate connected component table records the connected component identifier, voxel count, principal axis direction, and enhancement intensity statistics. The candidate region quality label table records artifact interference, breakage degree, and spatial boundary crossing markers. The localization rule handling records are written into the preprocessing log structure. The vascular skeleton extraction and centerline tracing are performed based on the candidate voxel set. The vascular skeleton extraction process involves filling holes and bridging breaks in the candidate voxel set, refining it to generate a vascular skeleton voxel set, and simultaneously constructing a skeleton connectivity graph and a bifurcation point numbering table, as well as generating a centerline quality marker table. The centerline tracing process selects the main path based on the principal axis direction field and forms a centerline coordinate sequence. When loop detection markers appear or path continuity is insufficient, a backtracking process is performed, and the centerline tracing parameter table is recorded. The annular region of interest (ROI) construction process reads the centerline coordinate sequence, performs centerline reparameterization, and generates a sampling point sequence. The sampling point spacing is generated based on the sampling step size parameter table, and the inner and outer radius fields are generated based on the inner and outer radius parameter tables. A cross-sectional normal sequence is generated at each sampling point, subsequently forming an annular shell voxel set and generating an annular ROI index table. The annular ROI index table records the sampling point identifier, cross-sectional identifier, corresponding voxel set index, inner and outer radius fields, and cross-sectional normal field. When spatial boundary violations occur, local clipping is performed, and the clipping marker field is recorded.The vessel centerline localization and annular region of interest construction unit outputs an annular region of interest structure and provides the annular region of interest structure as an input object to the segmentation and mask generation unit. At the same time, the indexes of the annular region of interest index table, cross-sectional normal sequence, inner and outer radius parameter table, sampling step size parameter table and centerline tracking parameter table are provided to the domain correction model training inference and version management unit as version-dependent input objects, so that the indexes remain consistent in the subsequent solidification and write-back process.

[0071] The segmentation and mask generation unit 03, connected to the vessel centerline positioning and annular region of interest (ROI) construction unit, is used to perform segmentation of blood vessels and surrounding tissues within the spatial range defined by the annular ROI structure and generate vessel masks and surrounding tissue masks to form a segmentation mask structure. It also generates a segmentation path record table, a segmentation parameter record table, and a mask quality label table. The segmentation and mask generation unit determines and switches segmentation paths based on the mask quality label table. Specifically, the segmentation and mask generation unit receives the annular ROI structure output by the vessel centerline positioning and annular ROI construction unit, retrieves the standardized image structure formed by the standardized image acquisition and preprocessing unit as the image data source, retrieves the corresponding voxel blocks in the standardized image structure according to the sampling point identifier and cross-section identifier in the annular ROI index table to form a segmentation input block sequence, and writes the index range field of the segmentation input block sequence into the segmentation path record table. The segmentation of blood vessels and surrounding tissues is performed within the spatial range defined by the annular region of interest structure. The segmentation path consists of threshold segmentation, region growing, and learning model inference, and the path identifier and processing parameters are registered in the segmentation parameter record table. Threshold segmentation generates an initial set of blood vessel voxels based on the enhancement intensity field and candidate region index field of the standardized image structure, and performs boundary contraction and expansion correction according to the cross-sectional normal sequence. Region growing selects seed points at the endpoints of the blood vessel mask candidates and expands them according to the connected component similarity criterion. The learning model inference inputs the segmentation input block sequence into the trained segmentation model and outputs a probability map and generates a set of voxels according to the probability threshold and connectivity constraints. The segmentation and mask generation unit writes the vascular voxel set into the vascular mask and the surrounding tissue voxel set into the surrounding tissue mask, forming a segmentation mask structure. The mask quality labeling table generates break markers by counting the number of connected segments in the vascular mask, generates hole markers by detecting voxel holes, and generates boundary crossing markers by detecting the boundaries of the annular shell voxel set. When a break marker or boundary crossing marker is triggered, the segmentation and mask generation unit writes a rollback marker into the segmentation path record table and switches the segmentation path. At the same time, it writes the action into the preprocessing log structure and retains the original path record. The segmentation and mask generation unit provides the segmentation mask structure, vascular mask, and surrounding tissue mask as input objects to the radiomics feature extraction and coupling feature construction unit, and provides the indexes of the segmentation path record table, segmentation parameter record table, and mask quality labeling table to the domain correction model training inference and version management unit as associated input objects for runtime auditing.

[0072] The radiomics feature extraction and coupling feature construction unit 04, connected to the segmentation and mask generation unit, is used to extract intensity features, texture features, and morphological features from the standardized image structure based on the vascular mask and the surrounding tissue mask to generate a radiomics feature table. It also constructs a coupling feature table by establishing an interface neighborhood band on the outer wall of the vascular mask, merges the features to generate a feature matrix, and generates a feature naming dictionary, a feature extraction parameter table, and a feature missing marker table. Specifically, the radiomics feature extraction and coupling feature construction unit receives the segmentation mask structure output by the segmentation and mask generation unit and reads the standardized image structure and voxel grid description table output by the standardized image acquisition and preprocessing unit as coordinate index references. It first performs a consistency check on the coordinate indices of the vascular mask and the surrounding tissue mask. If an offset is found between the mask index and the standardized image index table, an offset marker is written into the feature missing marker table, and a mask remapping process is triggered. The mask remapping process reconstructs the mask index based on the direction cosine and coordinate origin fields of the voxel grid description table and writes the remapping action into the feature extraction parameter table. When generating the radiomics feature table, the radiomics feature extraction and coupling feature construction unit extracts intensity statistical features according to the vascular mask range and writes them into the radiomics feature table. In parallel, it extracts texture and morphological features from the vascular mask and surrounding tissue masks. Texture features are generated by grayscale discretization and neighborhood co-occurrence statistics, and binning rules and neighborhood definition rules are written into the feature extraction parameter table. Morphological features are generated based on surface reconstruction of the mask voxel set and written into the radiomics feature table. When constructing the coupling feature table, the radiomics feature extraction and coupling feature construction unit establishes an interface neighborhood band on the outer wall of the vascular mask. The interface neighborhood band is composed of the expanded voxel set of the vascular mask and the intersection voxel set of the surrounding tissue mask, and the thickness constraint of the interface neighborhood band is written into the feature extraction parameter table. Subsequently, the coupling feature table is constructed, and the field values ​​of texture difference index, gradient co-occurrence index, and interface morphology index are written. For missing features, the missing type field is written into the feature missing marker table, and the missing feature name is written into the missing mapping segment of the feature naming dictionary. The radiomics feature extraction and coupled feature construction unit merges the radiomics feature table and the coupled feature table according to the sample index to generate a feature matrix and generates a feature naming dictionary. The feature naming dictionary records the unified name of the feature fields, the version identifier and the source object identifier and is bound to the feature extraction parameter table. The radiomics feature extraction and coupled feature construction unit provides the feature matrix, the feature naming dictionary, the feature extraction parameter table and the feature missing label table as input objects to the feature stability screening and feature selection unit, and writes their index information into the preprocessing log structure for use by the domain correction model training inference and version management unit.

[0073] The feature stability screening and feature selection unit 05, connected to the radiomics feature extraction and coupled feature construction unit, is used to read the resampling parameter record table, normalization parameter record table, registration parameter record table, and acquisition parameter record table to generate a perturbation configuration set, calculate consistency statistics on the feature matrix to generate a stability score table, perform correlation filtering, redundancy removal, and missing threshold filtering to generate an input feature list, and encapsulate it to generate an effective feature set structure. The effective feature set structure includes a stability score table, a screening rule record table, a selection rule record table, and a feature screening log structure. Specifically, the feature stability screening and feature selection unit receives the feature matrix output by the radiomics feature extraction and coupled feature construction unit and reads the feature naming dictionary. At the same time, it obtains the resampling parameter record table, normalization parameter record table, and registration parameter record table from the standardized image acquisition and preprocessing unit, obtains the acquisition parameter record table from the standardized image acquisition and preprocessing unit, and completes correlation verification at the input end. The perturbation configuration set consists of voxel grid perturbations, intensity bin perturbations, and registration reference phase switching. Trigger conditions and amplitude fields are registered in the screening rule record table. The feature stability screening and feature selection unit performs recalculation or remapping on the feature matrix for each perturbation entry to generate multiple versions of the feature matrix sequence. It also calculates consistency statistics for the same feature field to generate a stability score table. The stability score table records field identifiers, perturbation entry identifiers, score values, and anomaly markers. When a break or out-of-bounds marker exists in the mask quality marker table, the feature stability screening and feature selection unit writes a masking marker into the stability score table and removes the corresponding sample from the statistical set, simultaneously writing the removal action into the feature screening log structure. The feature stability screening and feature selection unit performs outlier detection on the feature matrix and generates an outlier list. Outlier detection is based on quantile truncation and neighborhood consistency checks, and truncation rules and check rules are registered in the screening rule record table. The handling records of the outlier list are written into the feature screening log structure. After completing the stability scoring, the feature stability screening and feature selection unit performs feature selection processing. The feature selection processing first performs relevance filtering and writes threshold rules into the selection rule record table, then performs redundancy removal and writes retention strategies based on the source object identifier and type identifier in the feature naming dictionary, and finally performs missing threshold filtering and writes deletion records based on the missing feature marker table to calculate the missing proportion. The input feature list consists of a set of feature field identifiers and their mapping items in the feature naming dictionary, and is provided to the domain correction model training inference and version management unit as input field constraints for training and inference.The feature stability screening and feature selection unit encapsulates the stability scoring table, screening rule record table, selection rule record table, input feature list, and screened feature matrix into an effective feature set structure. The effective feature set structure is then provided to the domain correction model training inference and version management unit as an input object. At the same time, the association index between the feature screening log structure and the operation audit log structure is written into the preprocessing log structure to form an index relationship consistent with subsequent write-back actions.

[0074] The Domain Correction Model Training Inference and Version Management Unit 06, connected to the Feature Stability Screening and Feature Selection Unit, performs domain correction processing based on the effective feature set structure, clinical information fields, and acquisition parameter record table to obtain a domain correction feature structure and generate an alignment mapping record table. It then performs recognition model training based on the domain correction feature structure to generate a model parameter package and threshold table, performs model parameter solidification to generate a model version structure, performs case inference on the cases to be evaluated to generate an interpretability mapping index table, constructs an inference report structure, and writes the inference report structure back to the preprocessing configuration structure to complete the preprocessing configuration structure update. Specifically, the Domain Correction Model Training Inference and Version Management Unit receives the effective feature set structure output by the Feature Stability Screening and Feature Selection Unit, and receives the clinical information fields and acquisition parameter record table output by the Standardized Image Acquisition and Preprocessing Unit. Simultaneously, it reads the preprocessing configuration structure and preprocessing log structure for version number record field verification and registration. The domain correction process is executed under the constraint of a grouping key structure, which consists of a device model field, a slice thickness field, a reconstruction setting field, and a phase identifier field. The domain correction model training, inference, and version management unit divides the feature matrix into feature subsets according to the grouping key structure and generates an alignment mapping record table. The alignment mapping record table records the grouping key identifier, feature field identifier, mapping type identifier, mapping parameter field, and applicable scope field. During the mapping generation process, verification and rollback flags are written, and the handling records are written to the domain correction log structure. The domain correction model training, inference, and version management unit applies the alignment mapping record table to the feature matrix to generate a domain correction feature structure and writes the indexes of the domain correction parameter record table and the alignment mapping record table to the preprocessing log structure. The recognition model training is executed under the constraint of a training data index table. The domain correction model training, inference, and version management unit generates a training feature tensor based on the domain correction feature structure and clinical information fields and generates a training log structure. For training interruptions, an interruption reason field is written, triggering the breakpoint recovery process. At the same time, a model parameter package and a threshold table are generated, and their indexes are written to the model review record table and the training log structure. The model parameter solidification encapsulates the model parameter package, threshold table, input feature list, alignment mapping record table version number, and preprocessing parameter version number to generate a model version structure. The version number record table is written with the version number, generation timestamp, applicable grouping key range, and dependency record field. The dependency record field is associated with the indexes of the normalization parameter record table, resampling parameter record table, centerline tracking parameter table, feature extraction parameter table, screening rule record table, and selection rule record table.The case inference process calls the model version structure for the case to be evaluated. First, it searches the alignment mapping record table according to the grouping key structure and performs domain correction on the feature vector of the case to be evaluated. Then, it inputs the recognition model to generate inflammation recognition labels and confidence fields, and writes the inference process into the runtime audit log structure. The interpretability mapping generation, based on the source object identifier of the feature naming dictionary, aggregates the contribution values ​​into the voxel set index of the vascular mask, surrounding tissue mask, and interface neighborhood zone. It then combines the annular region of interest index table and voxel grid description table to generate an interpretability mapping index table and writes it into the inference report structure. The inference report structure includes inflammation recognition labels, confidence fields, interpretability mapping index tables, version number record fields, and runtime audit log index fields. The domain correction model training inference and version management unit writes the inference report structure back to the preprocessing configuration structure and writes it into the version write-back record segment. At the same time, it updates the version number of the alignment mapping record table, the version number of the threshold table, the version number of the input feature list, the grouping key coverage range, and the anomaly fallback rules in the preprocessing configuration structure and outputs them to the standardized image acquisition and preprocessing unit as its preprocessing configuration structure input object, completing the cross-unit closed-loop call relationship.

Claims

1. A radiomics method for identifying perivascular inflammation in the pancreatic mesangial region for preoperative assessment, characterized in that, include: Acquire multi-phase enhanced abdominal images, clinical information fields, and preprocessing configuration structures; perform data quality control and acquisition parameter recording, noise reduction and registration with intensity normalization, voxel resampling processing, and generate standardized image structures. Based on standardized image structures, the process of locating candidate regions of pancreatic mesangial vessels, extracting the vascular skeleton and tracing the centerline, and constructing a ring region of interest (ROI) is performed to obtain a ring ROI structure. The ring ROI structure includes a ring shell voxel set pointer, a ring ROI index table pointer, a cross-sectional normal sequence pointer, an inner and outer radius parameter table pointer, a sampling step size parameter table pointer, and a vascular centerline structure pointer. Based on the ring-shaped region of interest structure, the system performs segmentation and mask generation of blood vessels and surrounding tissues, extraction of radiomics features and coupling feature construction to generate an effective feature set structure. Based on the effective feature set structure, domain correction processing, recognition model training, and model parameter solidification processing are performed to generate an inference report structure. The process of locating candidate regions of pancreatic mesangial vessels further includes: First, extracting the abdominal cavity boundary and spinal reference plane fields from standardized image structures. Then, constructing a three-dimensional spatial bounding box as the pancreatic mesangial region extent structure based on the relative position constraints of the pancreatic mesangial region within the body cavity. These relative position constraints include a longitudinal range near the root of the mesentery, an anterior-posterior range near the posterior aspect of the pancreas, and an exclusion range avoiding air areas outside the body cavity and high-density areas of the cortical bone. Further, extracting images from each temporal phase from the standardized image structures, selecting post-enhanced temporal phase combinations as candidate screening sources based on a standardized image index table, and performing tubular structure response calculations for each temporal phase. The structural response is a multi-scale filtered output for slender connected morphologies. Its implementation involves calculating local directional consistency and boundary gradient stability within neighborhoods at different scales to form a response volume. Threshold segmentation is then performed on the response volume to generate a candidate voxel set. After the candidate voxel set is generated, connected component analysis is performed under structural constraints within the pancreatic mesangial region to output a candidate connected component table. This table includes fields for connected component identifier, voxel count, principal axis direction, average enhancement intensity, and bifurcation complexity. A candidate region quality label table is also generated. If the candidate connected component table triggers an excessive or insufficient number of connected components at any given time, the positioning rules in the preprocessed configuration structure are read and written into the positioning rule record table, followed by a rule switching process. The process of vascular skeleton extraction and centerline tracing also includes: vascular skeleton extraction processing involves performing hole filling and break bridging on the candidate voxel set. Hole filling identifies voxel holes through local connectivity detection and performs neighborhood completion. Break bridging is completed by pairing short-distance endpoints and performing morphological closure operations. Subsequently, the completed voxel set is refined to generate a vascular skeleton voxel set. The refinement process iteratively strips boundary voxels while maintaining the topology of endpoints and bifurcation points, and constructs a skeleton connectivity graph for the vascular skeleton voxel set. The skeleton connectivity graph consists of a set of nodes and a set of edges. The centerline... The tracing process involves selecting the main candidate connected domain based on the main axis direction field in the candidate connected domain table, then selecting the entry endpoint within the main candidate connected domain and performing path traversal to generate the centerline coordinate sequence. The path traversal adopts a progressive expansion strategy that includes the connected graph. At the same time, for each bifurcation point, branches are selected and recorded according to the bifurcation point number table to form a centerline quality mark table. The centerline quality mark table includes path continuity, bifurcation stability, and loop detection mark fields. When the centerline quality mark table shows insufficient path continuity, the breakage degree field of the candidate region quality mark table is called to trigger the secondary bridging process. The process of constructing the annular region of interest (ROI) also includes: extracting the centerline coordinate sequence from the vessel centerline structure and reparameterizing the centerline; generating a sampling point sequence by uniformly sampling according to the arc length, with the sampling step size taken from the preprocessing configuration structure and written into the sampling step size parameter table; constructing a cross-sectional normal sequence at each sampling point; generating a local tangential direction by the direction difference between adjacent sampling points and combining it with the direction cosine field of the voxel mesh description table for coordinate consistency; and generating an inner and outer radius parameter table according to the inner and outer radius rules in the preprocessing configuration structure. The inner and outer radius rules include two types: a fixed radius scheme and a segmented radius scheme. The fixed radius scheme uses the same inner and outer radius fields for all sampling points, while the segmented radius scheme assigns different radius fields to different segments based on the local vessel scale estimation field of the connected domain where the centerline is located. A set of annular cross-sectional voxels is generated at each sampling point and accumulated along the centerline to form an annular shell voxel set. At the same time, an index is established between the annular shell voxel set and the centerline coordinate sequence to generate an annular ROI index table. The process of segmenting blood vessels and surrounding tissues and generating a mask also includes: Blood vessel and surrounding tissue segmentation processing involves extracting a ring-shaped shell voxel set and a ring-shaped region of interest index table from the ring-shaped region of interest structure. Based on the sampling point identifiers and cross-sectional identifiers in the ring-shaped region of interest index table, corresponding voxel blocks from the standardized image structure are retrieved to form a segmentation input block sequence. Simultaneously, a segmentation path record table is generated. Three types of paths—threshold segmentation, region growing, and learning model inference—are used to separate the blood vessel region from the surrounding tissue region. The threshold segmentation path includes setting an intensity threshold using the enhancement intensity field and candidate region index field of the standardized image structure and generating an initial blood vessel voxel set. Then, boundary contraction and expansion corrections are performed along the normal direction based on the cross-sectional normal sequence. The region growing path is used in the blood vessel masking process. The selected endpoints are seed points, and expansion is performed based on the connected component similarity criterion. The connected component similarity criterion consists of an intensity difference threshold, a gradient consistency threshold, and a morphological continuity threshold. The learning model inference path inputs the segmented input block sequence into the trained segmentation model and outputs a probability map. Then, a set of blood vessel voxels and a set of surrounding tissue voxels are generated according to the probability threshold and connectivity constraints. The mask generation process includes writing the segmented voxel set into a binary label carrier to generate a blood vessel mask and a mask for the surrounding tissue, keeping the index consistent with the voxel mesh description table. At the same time, a mask quality label table is generated, which records the fields of break mark, hole mark, and out-of-bounds mark. When a break mark or out-of-bounds mark is triggered, a backtracking mark is written into the segmentation path record table and the segmentation path is switched. The process of radiomics feature extraction and coupled feature construction also includes: radiomics feature extraction includes extracting intensity statistical features from the standardized image structure according to the vascular mask range. The intensity statistical features include mean, quantile, and kurtosis fields. Texture features are extracted from the vascular mask and surrounding tissue masks. Texture features are generated through gray-level discretization and neighborhood co-occurrence statistics. The binning rules and neighborhood definition rules of gray-level discretization are written into the feature extraction parameter table. Morphological features are extracted, including volume, surface area, thinness, and bifurcation complexity fields obtained from the surface reconstruction of the mask voxel set. The coupled feature construction process includes establishing an interface neighborhood on the outer wall of the vascular mask. The interface neighborhood band is composed of the set of dilated voxels of the vascular mask and the set of voxels of the intersection of the surrounding tissue mask. The thickness of the interface neighborhood band is constrained by the inner and outer radius parameter table and the sampling step size parameter table, and a coupling feature table is constructed. The coupling feature table contains three types of fields: texture difference index, gradient co-occurrence index, and interface morphology index. The texture difference index is generated by the difference between the texture statistics inside the vascular mask and the texture statistics of the interface neighborhood band. The gradient co-occurrence index is generated by the joint statistics of the gradient distribution in the normal direction of the vascular mask boundary and the gradient distribution in the interface neighborhood band. The interface morphology index is generated by the curvature distribution of the vascular mask boundary and the contact area distribution of the surrounding tissue mask. The effective feature set structure includes: an input feature list, a stability score table, a screening rule record table, a selection rule record table, a feature screening log structure, and an associated feature matrix index.

2. The method according to claim 1, characterized in that, The process of performing data quality control and recording acquisition parameters also includes: Data quality control processing includes missing segment checking, artifact marking, and sequence integrity verification. Missing segment checking identifies gaps in consecutive slices by comparing the slice order field in the image sequence index table with the actual loaded slice count, and writes the missing range field into the data quality control marking table. Artifact marking involves performing a coarse screening of each phase slice before intensity anomaly suppression and marking texture anomalies such as stripes, metallic artifacts, and motion blur, and writing the artifact type field into the data quality control marking table. Sequence integrity verification generates a sequence integrity verification result field by checking the phase identifier set against the phase set specified in the preprocessing configuration structure. The acquisition parameter recording processing includes recording the device model, slice thickness, reconstruction settings, and contrast agent injection settings. The device model comes from image metadata, the slice thickness from the slice spacing field, the reconstruction settings from the reconstruction kernel and matrix field, and the contrast agent injection settings from the injection rate and dosage field or the in-hospital examination form field. When a field is missing, a missing marker field is written into the acquisition parameter recording table and simultaneously written into the preprocessing log structure.

3. The method according to claim 1, characterized in that, The voxel resampling process also includes: The voxel resampling process includes reading the target voxel spacing field and the target grid size field, selecting the resampling interpolation method and generating a resampling parameter record table, writing the interpolation method field, the target voxel spacing field, and the target grid size field into the resampling parameter record table, and simultaneously writing them into the preprocessing log structure, while generating a voxel grid description table; generating a standardized image structure, which includes a standardized image data pointer, a standardized image index table, a voxel grid description table pointer, a normalized parameter record table pointer, a resampling parameter record table pointer, and a preprocessing log structure pointer field.

4. A radiomics identification system for perivascular inflammation in the pancreatic mesangial region for preoperative assessment, applied to the method described in any one of claims 1-3, characterized in that, include: The standardized image acquisition and preprocessing unit is used to perform data quality control and acquisition parameter recording, noise reduction and registration and intensity normalization, voxel resampling processing, and generate standardized image structures. The unit for vascular centerline localization and circular region of interest construction is used to perform vascular candidate region localization, vascular skeleton extraction and centerline tracing, and circular region of interest construction to obtain the circular region of interest structure. The segmentation and mask generation unit is used to perform segmentation of blood vessels and surrounding tissues and generate blood vessel masks and surrounding tissue masks to form a segmentation mask structure; The radiomics feature extraction and coupled feature construction unit is used to extract intensity features, texture features, and morphological features to generate a radiomics feature table. The feature stability screening and feature selection unit is used to read the resampling parameter record table, normalization parameter record table, registration parameter record table and acquisition parameter record table to generate a perturbation configuration set; The domain correction model training, inference, and version management unit is used to perform domain correction processing to obtain the domain correction feature structure and generate an alignment mapping record table.

Images