A pancreatic image processing method and system fusing double architecture and interactive interpretation

Abstract

This invention relates to the fields of medical image processing and artificial intelligence, and discloses a pancreatic image processing method and system integrating dual architecture and interactive interpretation. The method includes: reading pre-stored 3D medical images, registration masks, and structured attribute data; performing resampling, grayscale limiting, and discretization according to model version-related parameters; extracting radiomics features through the original image, spatial filtering, and multi-scale transformation pathways; obtaining target features through normalization, correlation-redundancy constraints, and sparse mapping determined and solidified by training samples; fusing the target features and structured attributes and inputting them into two classification branches with different mapping structures; calculating classification reference values, branch differences, feature contributions, and interaction strength; generating and outputting an interpretable data package containing model and preprocessing verification information. This invention is applicable to computer-aided analysis of 3D medical images of the pancreatic region.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence, specifically to a method, system, electronic device, and computer-readable storage medium for performing standardization, multi-path quantitative feature extraction, training domain-closed feature mapping, dual-architecture classification, and interactive interpretation on pre-acquired three-dimensional pancreatic region medical images. Background Technology

[0002] Enhanced computed tomography (CT) images can record the intensity, morphology, and spatial texture information of the pancreatic region in three-dimensional voxel format. Radiomics processing typically calculates features such as first-order statistics, shape, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size region matrix, adjacent gray-level tonal difference matrix, and gray-level dependency matrix within the region of interest, and then maps these high-dimensional features into numerical outputs using a statistical learning model. This type of technique can be used for computer-aided quantitative analysis of existing images.

[0003] Existing solutions have publicly disclosed technical approaches such as training machine learning models using pancreatic CT radiomics features, selecting a single optimal model using multiple medical images and different candidate algorithms, and characterizing transfer-related categories using enhanced CT radiomics scoring. However, these existing technologies still have room for further improvement. First, different slice thicknesses, voxel spacing, and tube current conditions may cause non-targeted fluctuations in the same texture definition. If preprocessing parameters are not fixed together with the model version, the feature distribution in the application stage may deviate from that in the training stage. Second, if validation data is involved in normalization, sorting, or parameter tuning during high-dimensional feature selection, data leakage may occur; if only a single statistical test or a single sparse regression is used, an imbalance between relevance and redundancy may occur. Third, many solutions only output a numerical value of a so-called optimal model, which cannot explain the difference in utilization of the same input by linear discriminant relations and nonlinear interaction relations. Fourth, giving feature importance alone cannot indicate the non-additive interaction between paired features, nor is it convenient to verify whether model differences come from specific attribute combinations.

[0004] Therefore, a technical solution is needed that is closely integrated with 3D medical image processing. Based on unified standardization and target feature mapping, it enables two classification branches with different inductive biases to process the same fused feature vector, while outputting branch difference, local and global contribution information and interaction matrix. Furthermore, it ensures that preprocessing parameters, feature order and model version remain verifiable, thereby improving the stability, traceability and review convenience of the data processing process. Summary of the Invention

[0005] The technical problem that this invention aims to solve is: how to obtain a fusion feature expression with consistent training domain and controlled dimensionality when there are differences in the acquisition parameters of three-dimensional medical images and the candidate feature dimension is high, and how to provide verifiable cross-model differences, feature contributions and pairwise interaction information while generating classification reference values.

[0006] To address the aforementioned technical problems, this invention provides a three-dimensional medical image processing method that integrates dual architecture and interactive interpretation. This method is executed by a computer device on pre-stored data and sequentially includes data reading, image normalization, multi-pathway radiomics feature extraction, target feature mapping with a closed training domain, structured attribute fusion, parallel computation of two classification branches with different mapping structures, feature contribution and interaction strength calculation, and interpretable data packet output.

[0007] During the image standardization phase, the 3D voxel matrix is ​​resampled to a preset isotropic voxel spacing, and voxel values ​​are subjected to fixed-range limiting and discretization using a fixed-box width. Standardization parameters are bound to the model version for reuse during application phases. For discrete masks, nearest-neighbor interpolation can be used; for continuous grayscale images, linear, B-spline, or other interpolation methods that maintain grayscale continuity can be used.

[0008] In the multi-path feature extraction stage, the computational domain is defined by a region of interest mask registered with the image space, and at least the original image path, the 3D Gaussian-Laplacian filter path, and the 3D discrete wavelet transform path are set. The 3D discrete wavelet transform performs low-pass or high-pass filtering along the three spatial axes respectively, and the filtering type used for each spatial axis is identified by L and H respectively. Each path outputs an imageomics feature vector with a unified field name, transform parameters, and version identifier.

[0009] In the target feature mapping stage, only training samples are used to determine the mean, standard deviation, mutual information ranking, sparsity coefficients, and number of features. First, a maximum correlation minimum redundancy algorithm is used to retain candidate features that are correlated with the class label and have low redundancy. Then, a model with L1 regularization is used to obtain features with non-zero coefficients, and the target dimension is determined based on changes in discriminant strength, sample size, or a preset upper limit. Validation samples only utilize pre-defined transformation parameters and feature indices.

[0010] During the fusion and dual-architecture computation phase, target image features and structured attribute data are encoded and concatenated into a single fused feature vector, which is then fed into a first classification branch and a second classification branch with different mapping structures. The first classification branch can employ a linear mapping model, while the second classification branch can employ a nonlinear mapping model; in one implementation, they employ a linear discriminant model and a multilayer perceptron, respectively. The system can select one branch as the principal numerical output according to a predetermined verification rule, while retaining the output of the other branch and the difference between the two.

[0011] During the interpretation phase, the feature contributions of at least one classification branch can be calculated using additive attribution or perturbation importance methods, and pairwise feature interactions can be calculated using bivariate response decomposition, interaction attribution, or variance decomposition methods. In one implementation, the Shapley additive interpretation values ​​of the two branches are calculated using the same training background samples or a representative subset thereof, and the Friedman H statistic is obtained based on the difference between the bivariate partial dependency function of the second classification branch and the corresponding two univariate partial dependency functions, thereby forming a local contribution vector, a global contribution vector, contribution ranking, interaction matrix, and branch difference.

[0012] The present invention also provides a pancreatic image processing system, an electronic device, and a non-transitory computer-readable storage medium corresponding to the above method.

[0013] Compared to approaches that rely solely on a single imageomics score or retain only one model from multiple candidate models, this invention offers the following advantages: First, unified resampling, amplitude limiting, and discretization parameters help reduce feature fluctuations caused by different acquisition settings. Second, the two-level mapping of maximum correlation minimum redundancy and sparse coefficients within a closed training domain helps reduce the risk of data leakage and redundant computation. Third, the co-originating dual branches can simultaneously present linear and nonlinear mapping results under the same input conditions. Fourth, encapsulating cross-branch differences, Shapley contributions, and the Friedman H interaction matrix helps locate the features and attribute combinations corresponding to output differences. Fifth, the model, parameter, and feature order verification values ​​facilitate version traceability during application. These effects may be influenced by image quality, sample composition, and model training conditions, and should not be construed as an absolute guarantee of performance across all application scenarios. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall functional modules of the pancreatic image processing system of the present invention;

[0015] Figure 2 This is a schematic diagram of the overall process of the three-dimensional medical image processing method of the present invention;

[0016] Figure 3 This is a schematic diagram of the image standardization and multi-pathway radiomics feature extraction process of the present invention;

[0017] Figure 4 This is a schematic diagram of the target feature mapping and model solidification process with closed training domain of the present invention;

[0018] Figure 5 This is a schematic diagram of the dual-architecture parallel classification and joint interpretation output process of the present invention;

[0019] Figure 6 This is a schematic diagram of the processor-centric layered hardware structure of the electronic device of the present invention.

[0020] Figure labeling: 10—Data receiving module; 20—Image preprocessing module; 30—Region of interest management module; 40—Feature extraction module; 41—Original image path; 42—Spatial filtering path; 43—Multi-scale transformation path; 50—Feature mapping module; 51—Normalization unit; 52—Correlation-redundancy constraint mapping unit; 53—Sparse feature mapping unit; 60—Dual-architecture classification module; 61—First classification branch; 62—Second classification branch; 70—Interpretation module; 71—Contribution calculation unit; 72—Interaction strength calculation unit; 73—Branch difference calculation unit; 80—Output module; 90—Electronic device; 91—Processor; 92—Memory; 93—Medical image data interface; 94—Display interface; 95—Network interface; 96—Communication interconnection. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments are used to illustrate the technical concept and are not intended to unduly limit the scope of protection. Unless otherwise stated, the threshold, sample size, software version, and model structure can be adaptively adjusted while maintaining the functional relationships between the modules. Example 1: System Overall Structure

[0022] like Figure 1 As shown, the pancreatic image processing system includes a data receiving module 10, an image preprocessing module 20, a region of interest management module 30, a feature extraction module 40, a feature mapping module 50, a dual-architecture classification module 60, an interpretation module 70, and an output module 80. These modules can be deployed on the same workstation or distributed according to a client-server architecture.

[0023] The data receiving module 10 reads stored 3D medical images, masks, and structured attribute data via a DICOM interface, file interface, or database interface. During reading, it can verify image sequence identifiers, voxel spacing, layer number, grayscale units, spatial orientation matrix, mask size, and other fields required for the model. If the data is incomplete or the spatial matrix is ​​inconsistent, it can return to a pending review status without performing subsequent classification mapping.

[0024] The image preprocessing module 20 performs resampling, limiting, and discretization according to the configuration file bound to the model version. The region of interest management module 30 transforms the mask to the normalized image coordinate system and checks whether the mask is empty, whether it exceeds the image range, and whether the number of connected primes reaches the preset lower limit.

[0025] The feature extraction module 40 includes an original image path 41, a spatial filtering path 42, and a multi-scale transformation path 43. The feature mapping module 50 includes a normalization unit 51, a correlation-redundancy constraint mapping unit 52, and a sparse feature mapping unit 53. The dual-architecture classification module 60 includes a first classification branch 61 and a second classification branch 62, which employ different mapping structures; in one embodiment, the first classification branch 61 is a linear discriminant branch, and the second classification branch 62 is a multilayer perceptron branch. The interpretation module 70 includes a contribution calculation unit 71, an interaction strength calculation unit 72, and a branch difference calculation unit 73. The output module 80 is used to save JSON, XML, DICOM structured reports, or database records, and can also generate contribution bar charts and interaction matrix charts for manual review.

[0026] like Figure 2 As shown, this method can be executed according to the following processing stages: Step S101, read the pre-stored 3D medical image, registration mask, and structured attribute data; Step S103, perform voxel resampling, grayscale limiting, and grayscale discretization according to the preprocessing configuration associated with the model version; Step S105, extract features within the mask-defined area through the original image, spatial filtering, and multi-scale transformation pathways; Step S107, call the parameters determined and fixed only by the training samples to perform normalized feature transformation; Step S109, sequentially perform correlation-redundancy constraint mapping and sparse feature mapping; Step S111, encode and concatenate the target image features and structured attributes; Step S113, input the same fused feature vector into the first classification branch and the second classification branch respectively; Step S115, calculate the branch feature contribution, pairwise feature interaction information, and classification reference value difference; Step S117, generate an interpretable data package and output it to a storage device or display device. The above step numbers are used to identify the processing stages; while maintaining the input-output dependency, some processing can be executed in parallel or in combination. Example 2: 3D Enhanced Image Standardization

[0027] This embodiment reads portal venous phase enhanced CT scan data of the pancreatic region that has already been acquired and stored. Exemplary acquisition parameters are tube voltage 120 kV, tube current 50-413 mAs, and original slice thickness 0.625-5 mm. The acquisition process is not part of the computer image processing method of this invention; data generated by other scanning devices and verified by configuration can also be used as input.

[0028] The image preprocessing module 20 first converts the stored pixel values ​​into CT values ​​according to the DICOM grayscale mapping parameters. Then, it uses trilinear interpolation to resample the continuous grayscale image to isotropic voxels of 1 mm × 1 mm × 1 mm, and uses nearest neighbor interpolation to resample the discrete mask. Subsequently, the voxel values ​​are limited to -1000 HU to 2000 HU, and discretized with a fixed bin width of 25 HU. A symmetric grayscale co-occurrence matrix is ​​enabled during texture calculation. This forms a normalized voxel matrix V and a binary mask M registered with it.

[0029] In an alternative implementation, the grayscale limiting range, voxel spacing, and bin width can be determined from training data for different imaging modalities, but the application phase should read a parameter file consistent with the current model version. The parameter file should at least record the modality, phase, interpolation method, target voxel spacing, limiting upper and lower bounds, discretization rules, texture matrix settings, and software version. Example 3: Region of Interest and Multipath Feature Extraction

[0030] Regions of interest (ROI) masks can be obtained through manual delineation, semi-automatic region growing, or a trained segmentation network. In one reproduction embodiment, two observers independently delineate portions of the samples and cross-check them; for samples with boundary differences, a third observer forms the final mask. The system only reads the generated mask file and does not perform any operations on the human body.

[0031] like Figure 3 As shown, the normalized voxel matrix V and the mask M are input into the feature extraction module 40. The original image path 41 calculates first-order statistics, 3D shape, and texture features within the voxel set defined by M; the spatial filtering path 42 performs 3D Gaussian-Laplacian filtering using example scale parameters, and calculates the filtered first-order statistics and texture features; the multi-scale transformation path 43 performs low-pass or high-pass filtering along the first, second, and third spatial axes respectively, obtaining eight frequency sub-bands: LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH, and calculates first-order statistics and texture features in each frequency sub-band. Here, L represents low-pass filtering along the corresponding spatial axis, and H represents high-pass filtering along the corresponding spatial axis; the LLL frequency sub-band performs low-pass filtering along all three spatial axes, and the HLL frequency sub-band performs high-pass filtering along the first spatial axis and low-pass filtering along the second and third spatial axes.

[0032] In a reproducible implementation, a feature definition compatible with PyRadiomics 3.0.1 was used, resulting in 957 candidate features, including 120 original features, 93 Gaussian-Laplace filter features, and 744 wavelet transform features. Texture features can include GLCM, GLRLM, GLSZM, NGTDM, and GLDM categories. The system records the path, filter parameters, feature category, mask identifier, and implementation version for each field for subsequent verification. Example 4: Target Feature Mapping in a Closed Training Domain

[0033] like Figure 4 As shown, the offline training data is stratified according to a predetermined random seed. The example data includes 340 data objects, of which 190 are first-labeled and 150 are second-labeled; they are divided into a training set of 238 objects and an independent validation set of 102 objects in a 7:3 ratio. The labels are formed from the data already prepared before training and are only used as identifiers for offline supervised learning.

[0034] Sixty data objects were extracted from the training set. Masks were repeatedly generated by two independent mask makers, and the intra-group correlation coefficients of candidate features were calculated, retaining features with an ICC greater than 0.75. Subsequently, the mean of each feature was calculated using only the training set. and standard deviation and in accordance with Perform Z-score transformation. The validation set directly uses the values ​​from the training set. and .

[0035] The maximum relevance minimum redundancy mapping unit 52 ranks candidate features using the MID criterion based on mutual information. In one implementation, the top 10% of features are retained, i.e., 96 features are retained out of 957 candidate features. Subsequently, the sparse coefficient mapping unit 53 uses logistic regression with L1 regularization, setting 50 candidate regularization parameters on a logarithmic scale from 10⁻⁴ to 10², and uses five-fold hierarchical cross-validation, a liblinear solver, a maximum number of iterations of 1000, and a random seed of 42 to determine the target parameters, while retaining features with non-zero coefficients.

[0036] In this example, sparse mapping initially yielded 8 features; further, the first 5 features were retained according to feature importance order and a pre-defined simplification rule. Compared to the 8-feature version, the area under the curve for the 5-feature version changed by approximately 0.0037, and the difference did not reach the pre-defined significance threshold in this example. The final five target image features are: the run-length variance feature of the grayscale run-length matrix after a pre-defined scale 3D Gaussian-Laplace filter (…). ), the first-order robust mean absolute deviation characteristics of the LLL frequency subband in three-dimensional discrete wavelet decomposition ( The gray-level emphasis feature of the gray-level dependence matrix of the LLL frequency sub-band () ), Busyness characteristics of adjacent gray-level tone difference matrix of HLL frequency subbands in three-dimensional discrete wavelet decomposition ( ), and the busyness feature of the adjacent grayscale tone difference matrix after preset scale three-dimensional Gaussian-Laplace filtering ( LLL indicates that low-pass filtering is performed along all three spatial axes, while HLL indicates that high-pass filtering is performed along the first spatial axis and low-pass filtering is performed along the second and third spatial axes. The median inter-observer ICC for the selected features is approximately 0.870, ranging from approximately 0.850 to 0.883.

[0037] The aforementioned target image features can be directly constructed into feature sub-vectors according to the fixed order, or weighted aggregation can be performed using coefficients determined by training samples. In a reproducible implementation, the radiomics score is calculated according to the following formula: In the formula, f1 to f5 represent the five features after Z-score transformation based on the training set parameters. This formula and coefficients are merely parameters for this example and do not constitute a limitation on coefficients obtained from other training samples or on the direct fusion method. The model file must at least solidify the target feature fields, their order, missing value rules, and... , Sparsity coefficients, training timestamps, and version verification values. Example 5: Structured Attribute Fusion and Dual-Architecture Computation

[0038] Structured attribute data may include at least one of the following: age-related numerical fields, laboratory test numerical fields, pre-stored image grading fields, and image location fields. In one embodiment, the structured attribute data includes age values, logarithmically transformed carcinoembryonic antigen (CEA) values, and pre-stored lymph node image grading. Continuous fields are transformed using Z-scores based on the training set mean and standard deviation; grading fields may use one-hot encoding based on a single grade. The target image features are concatenated with the above-mentioned structured attribute codes in a fixed order to form a fused feature vector z.

[0039] The first classification branch 61 can employ a linear discriminant model, a generalized linear model, or a linear kernel support vector machine. In this embodiment, linear discriminant analysis is used. Based on the intra-class scatter matrix SW and the inter-class scatter matrix SB of the training samples, the projection direction that increases the ratio of inter-class scatter to intra-class scatter is obtained, and the projected value is calibrated and mapped to the first classification reference value pL. This branch can provide a clearer linear feature utilization structure.

[0040] The second classification branch 62 can employ a multilayer perceptron, a nonlinear kernel support vector machine, or a tree ensemble model. In this embodiment, a lightweight multilayer perceptron consisting of an input layer, one or more fully connected hidden layers, a ReLU or hyperbolic tangent activation layer, a regularization layer, and a binary classification output layer is used to map the same fused feature vector z to a second classification reference value pN. The model structure and parameters can be determined through parameter search and hierarchical cross-validation applied only to training samples. Both branches can employ class-balanced weights.

[0041] In one training implementation, in-domain hyperparameter tuning is performed on 26 candidate models covering linear models, discriminant analysis, support vector machines, tree models, ensemble models, Bayesian models, and neural networks. The area under the curve, Brier score, and calibration results are then comprehensively compared. Ultimately, two branches with different mapping structures—linear discriminant analysis and multilayer perceptron—are retained, instead of saving only a single so-called optimal model. The system can mark one branch as the primary reference value according to preset rules, while always saving pL, pN, and their difference. The selected model structure, training parameters, and the order of fused features are fixed together with the model version. Example 6: Joint Interpretation and Interpretable Data Packets

[0042] like Figure 5 As shown, the contribution calculation unit 71 can use an additive attribution algorithm or a perturbation importance algorithm to calculate the feature contribution of the classification branch. In this embodiment, the training set or a representative subset thereof is used as the common background samples to calculate the Shapley additive interpretation values ​​of the first classification branch 61 and the second classification branch 62, respectively. For input z, the branch output can be expressed as... Where φ0 is the expected background output. This represents the local contribution of the j-th feature component. The global contribution can be measured using the sample set. The mean is represented by .

[0043] The interaction strength calculation unit 72 can obtain pairwise feature interaction information using bivariate response decomposition, interaction attribution, or variance decomposition. In this embodiment, a bivariate partial dependency function is calculated for the second classification branch 62; for features j and k, the difference between the centered bivariate partial dependency and the two univariate partial dependencies is taken as a non-additive interaction term, and the Friedman H statistic is obtained from the square root of the ratio of the sum of squares of this interaction term to the sum of squares of the bivariate partial dependencies. The grid resolution and numerical integration method can be set according to the training data range. The closer the H value is to 0, the weaker the non-additive interaction observed within the current sample range.

[0044] In one example The H-value for age is approximately 0.236. The H-value for age is approximately 0.205; other exemplary interactions include The carcinoembryonic antigen (CEA) value is approximately 0.148, and its... Approximately 0.143, and the carcinoembryonic antigen (CEA) value and second-level hierarchical coding value are approximately 0.119. These values ​​are used to illustrate how the interaction matrix was generated and should not be interpreted as fixed thresholds for other datasets.

[0045] Branch Difference Calculation Unit 73 Output Furthermore, it can calculate the rank correlation coefficient of the global contribution ranking of the two branches, the angle between the local contribution vectors, or the weighted difference index. Interpretable data packets include at least pL, pN, and The data includes local contribution information, global contribution information, contribution ranking, pairwise feature interaction information, preprocessing configuration, target feature order, model version, and timestamp. When using the Friedman H statistic, the pairwise feature interaction information can be represented as an H statistic matrix. The application can use this to locate the source of feature differences between the two models for the same input. Example 7 Verification

[0046] On the independent validation set of 102 data objects, the area under the curve for the linear discriminant branch is approximately 0.828, with a 95% confidence interval of approximately 0.735–0.897 and a Brier score of approximately 0.168; the area under the curve for the multilayer perceptron branch is approximately 0.822, with a 95% confidence interval of approximately 0.732–0.895 and a Brier score of approximately 0.172. The average precision examples for the precision-recall curves are approximately 0.875 and 0.860, respectively.

[0047] In exploratory stratification based on age, sex, hierarchical coding, and pancreatic region location, the area under the curve (AUC) for the multilayer perceptron branch ranged from approximately 0.66 to 0.77, while the stratification of the linear discriminant branch showed relatively larger fluctuations. These results indicate that even when the overall discriminative power of the two branches is similar, they may still utilize different attributes in different ways. Therefore, simultaneously preserving branch outputs, contribution vectors, and interaction matrices is valuable for verification.

[0048] The performance data described above is only used to illustrate the implementation results of the technical process on example data. Since sample sources, equipment platforms, mask quality, and data distribution may differ, this invention does not consider specific performance values ​​as essential technical features, nor does it guarantee the same results on other data. Example 8: Alternative Implementation

[0049] like Figure 6As shown, the electronic device 90 adopts a processor-centric hierarchical structure, including a processor 91, a memory 92, a medical image data interface 93, a display interface 94, a network interface 95, and a communication interconnect 96. The medical image data interface 93 and the display interface 94 establish input and output connections with the processor 91, respectively. The processor 91 accesses the memory 92 and the network interface 95 via the communication interconnect 96. The processor 91 can be a central processing unit, a graphics processing unit, a neural network processor, or a combination thereof; the memory 92 is used to store preprocessing configurations, feature dictionaries, normalization parameters, model parameters, and interpreters.

[0050] After obtaining pre-stored data from the medical image data interface 93, the processor 91 sequentially performs image standardization, feature extraction, target feature mapping, dual-branch calculation, and interpretation calculation. Parallel task queues can be established for different frequency sub-bands and different filtering paths; the output is presented through the display interface 94, or written by the processor 91 to the image archiving communication system or database via the communication interconnect 96 and the network interface 95.

[0051] While maintaining pre-acquired 3D digital images, closed feature maps within the training domain, parallel output from a homogeneous dual-architecture model, and joint encapsulation of contribution and interaction strength, the input can be replaced with CT scans, MRI scans, positron emission tomography (PET) scans of other body parts, or combinations thereof, and the preprocessing and model parameters are redefined based on the training data. The first classification branch can employ a linear discriminant model, a generalized linear model, or a linear kernel support vector machine, while the second classification branch can employ a multilayer perceptron, a nonlinear kernel support vector machine, or a tree ensemble model. The two branches maintain different mapping structures and process the same fused feature vector.

[0052] Those skilled in the art can make equivalent substitutions for module deployment, processing order, model structure, file format, and threshold range without departing from the technical concept of this invention; contribution calculation can employ additive attribution or perturbation importance, and interaction calculation can employ bivariate response decomposition, interaction attribution, or variance decomposition. The specific scope of protection is determined by the claims.

Claims

1. A three-dimensional medical image processing method integrating dual architecture and interactive interpretation, characterized in that, The method, executed by a computer device on pre-acquired and stored three-dimensional medical image data, includes: Read the three-dimensional medical image data, the region of interest mask spatially registered with the three-dimensional medical image data, and the corresponding structured attribute data; Obtain the preprocessing configuration associated with the target model version, and perform spatial normalization and grayscale normalization on the three-dimensional medical image data and the region of interest mask according to the preprocessing configuration to obtain the normalized voxel matrix and registration mask; Using the set of voxels defined by the registration mask as the computational domain, quantitative features are extracted through the original image path, spatial filtering path, and multi-scale transformation path to form an image omics feature vector; Based on the target feature mapping configuration constructed and solidified using only training samples, normalization transformation, correlation-redundancy constraint mapping, and sparse feature mapping are sequentially performed on the image omics feature vector to obtain the target image features; The target image features and the structured attribute data are encoded and fused into the same fused feature vector, and the fused feature vector is input into a first classification branch and a second classification branch with different mapping structures to obtain a first classification reference value and a second classification reference value. Calculate feature contribution information and pairwise feature interaction information based on the input and output of at least one classification branch, and determine branch difference information based on the first classification reference value and the second classification reference value; Generate an interpretable data packet including the first classification reference value, the second classification reference value, the feature contribution information, the pairwise feature interaction information, the branch difference information, and the model and preprocessing verification information, and output the interpretable data packet to a storage device or a display device.

2. The three-dimensional medical image processing method according to claim 1, characterized in that, The preprocessing configuration includes target voxel spacing, grayscale limiting boundary, grayscale discretization rules, texture matrix settings, and version identifier; The target voxel spacing, the grayscale limiting boundary, and the grayscale discretization rule are determined based on the image acquisition information of the training samples or selected from a preset candidate configuration, and are solidified together with the model parameters of the corresponding classification branch. The spatial normalization includes processing the three-dimensional medical image data using an interpolation method that preserves grayscale continuity, and processing the region of interest mask using a discrete value-preserving interpolation method.

3. The three-dimensional medical image processing method according to claim 1, characterized in that, The original image path is used to extract first-order statistical features, three-dimensional shape features, and texture features from the normalized voxel matrix; The spatial filtering path is used to perform three-dimensional Gaussian-Laplace filtering or equivalent scale spatial transformation at one or more preset scales, and extract first-order statistical features and texture features from the transformation results; The multi-scale transformation path is used to perform low-pass filtering or high-pass filtering along the first, second and third spatial axes respectively, forming eight frequency sub-bands of three-dimensional discrete wavelet decomposition, and extracting first-order statistical features and texture features from each frequency sub-band. Where L represents low-pass filtering along the corresponding spatial axis, H represents high-pass filtering along the corresponding spatial axis, LLL frequency sub-band is the frequency sub-band that performs low-pass filtering along all three spatial axes, and HLL frequency sub-band is the frequency sub-band that performs high-pass filtering along the first spatial axis and low-pass filtering along the second and third spatial axes. The texture features include at least one of the following: gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level size region matrix features, adjacent gray-level tone difference matrix features, and gray-level dependency matrix features.

4. The three-dimensional medical image processing method according to claim 1, characterized in that, The target feature mapping configuration is constructed through the following offline training process: normalization parameters for each candidate feature are determined using only training samples; Determine a stable feature set based on feature consistency under repeated masking, mask perturbation, or resampling perturbation; A correlation-redundancy constraint mapping is generated based on the correlation between candidate features and training labels, as well as the redundancy between candidate features. Perform a coefficient mapping with sparse constraints on the candidate features preserved by the correlation-redundancy constraint mapping to obtain non-zero coefficient features; The target feature dimension is determined based on at least one of the following: performance variation, training sample size, and preset upper limit of dimension. The normalization parameters, the stable feature set, the feature arrangement order, the correlation-redundancy constraint mapping, the sparse feature mapping, and the target feature dimension are all solidified together, and the verification sample only calls the solidified target feature mapping configuration.

5. The three-dimensional medical image processing method according to claim 4, characterized in that, The target image features include: grayscale run-length matrix run-length variance features obtained by three-dimensional Gaussian-Laplace filtering at a preset scale; First-order robust mean absolute deviation characteristics of LLL frequency subbands in three-dimensional discrete wavelet decomposition; The gray-level dependence matrix of the LLL frequency sub-band emphasizes low gray-level features; Busyness characteristics of adjacent gray-level tone difference matrix of HLL frequency subband in three-dimensional discrete wavelet decomposition; And the busyness characteristics of the adjacent gray-level tone difference matrix obtained by three-dimensional Gaussian-Laplace filtering with a preset scale; The LLL frequency sub-band performs low-pass filtering along all three spatial axes, and the HLL frequency sub-band performs high-pass filtering along the first spatial axis and low-pass filtering along the second and third spatial axes. The target image features are directly constructed into feature sub-vectors according to the solidification order, or weighted aggregation is performed using coefficients determined by training samples and solidified together with the model version.

6. The three-dimensional medical image processing method according to claim 5, characterized in that, The three-dimensional medical imaging data is CT image data; The target image features are composed of the five features listed in the order of claim 5; The spatial normalization includes resampling the continuous grayscale image to an isotropic voxel of 1 mm × 1 mm × 1 mm using trilinear interpolation, and resampling the region of interest mask using nearest neighbor interpolation. The grayscale normalization includes limiting the voxel values ​​to -1000 HU to 2000 HU and discretizing them according to a fixed bin width of 25 HU; The scale parameter σ of the three-dimensional Gaussian-Laplace filter is 3.0 mm; The five features after Z-score transformation based solely on the mean and standard deviation determined from the training samples are denoted as f1 to f5, and the weighted aggregation of the radiomics score Rad-score satisfies: Rad-score = 0.3612×f1 + 0.2852×f2 - 0.1939×f3 + 0.1697×f4 + 0.1548×f5.

7. The three-dimensional medical image processing method according to claim 6, characterized in that, The first classification branch adopts a linear discriminant analysis model, and the second classification branch adopts a multilayer perceptron model including at least one fully connected hidden layer and a nonlinear activation layer. The model parameters of both classification branches are determined by a model selection process that only applies to training samples, and are solidified together with the preprocessing configuration, the target feature mapping configuration, and the order of fused features. Using training samples or a representative subset thereof as common background samples, the local contribution vector, global contribution vector, and contribution ranking of the two classification branches are obtained by the Shapley additive interpretation algorithm. Based on the centering difference between the bivariate partial dependency function of the second classification branch and the corresponding two univariate partial dependency functions, the pairwise feature interaction information is generated using the Friedman H statistic. The first classification reference value is denoted as pL, and the second classification reference value is denoted as pN. The branch difference information includes Δp = pN - pL. The interpretable data package also includes the Friedman H statistics matrix, model version identifier, preprocessing parameter verification value, target feature order verification value, and generation time identifier.

8. A pancreatic image processing system integrating dual architecture and interactive interpretation, characterized in that, include: The data receiving module is used to read pre-acquired and stored three-dimensional medical image data, a region of interest mask spatially registered with the three-dimensional medical image data, and corresponding structured attribute data; The image preprocessing module is used to call the preprocessing configuration associated with the target model version to generate a normalized voxel matrix and a registration mask; The feature extraction module is used to form radiomics feature vectors by using the set of voxels defined by the registration mask as the computation domain, through the original image path, spatial filtering path and multi-scale transformation path; The feature mapping module is used to call the target feature mapping configuration that is constructed and fixed using only training samples, and to perform normalization transformation, correlation-redundancy constraint mapping and sparse feature mapping on the radiomics feature vector to obtain the target image features. The dual-architecture classification module is used to fuse the target image features and the structured attribute data into the same fused feature vector, and input a first classification branch and a second classification branch with different mapping structures respectively to generate a first classification reference value and a second classification reference value. The interpretation module is used to generate feature contribution information, pairwise feature interaction information, and branch difference information between two classification reference values; The output module is used to generate and output an interpretable data packet that includes the feature contribution information, the pairwise feature interaction information, the branch difference information, and the model and preprocessing verification information.

9. An electronic device comprising a processor, a memory, a medical imaging data interface, and an output interface, wherein the processor is communicatively connected to the memory, the medical imaging data interface, and the output interface, characterized in that, The memory stores a computer program, model parameters, target feature mapping configuration, and preprocessing configuration. When the computer program is executed by the processor, the electronic device performs the three-dimensional medical image processing method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing a computer program, model parameters, target feature mapping configuration, and preprocessing configuration thereon, characterized in that, When the computer program is executed by the processor, it implements the three-dimensional medical image processing method according to any one of claims 1 to 7.

Images