Intelligent prediction method and system for relationship between skull base meningioma and adjacent blood vessels

Abstract

The application discloses an intelligent prediction method and system for the relationship between skull base meningioma and adjacent blood vessels, and relates to the technical field of intelligent auxiliary diagnosis. The method comprises the following steps: acquiring magnetic resonance multi-sequence images of the same subject, performing three-linear interpolation resampling, bias field correction, rigid registration and contrast normalization to obtain preprocessed images; adopting threshold segmentation combined with semi-automatic delineation to generate a three-dimensional binary mask, and labeling intratumoral cystic necrosis area, peritumoral edema area and tumor parenchymal enhancement area to construct a tumor peripheral structure mask; extracting clinical radiological sign features, tumor quantitative indexes and spatial image features to form a joint feature set, performing correlation analysis and least absolute shrinkage and selection operator regression to screen core fusion features; constructing a blood vessel relationship three-classification training set, adjusting the category proportion, inputting the classification model and outputting a separate type, adhesion type or wrapping type label and feature contribution explanation, and using the same for preoperative risk assessment and surgical planning support.

Description

Technical Field

[0001] This application relates to the field of intelligent assisted diagnostic technology, specifically to an intelligent prediction method and system for the relationship between skull base meningiomas and adjacent blood vessels. Background Technology

[0002] Current techniques typically rely on multimodal imaging such as CTA, MRA, DSA, and MRI in the preoperative assessment and surgical planning of skull base meningiomas. Clinicians interpret the images to assess the spatial proximity of the tumor to key vascular structures such as the internal carotid artery, vertebrobasilar artery system and its branches, and venous sinuses, as well as the extent of entrapment or compression displacement. They then combine this information with their experience to determine the approach, resection extent, and vascular protection strategies. However, due to the complex anatomy of the skull base region, significant bony obstruction and artifact interference, substantial individual differences in vascular course, and diverse tumor growth patterns and invasion methods, it is difficult to form stable and repeatable quantitative conclusions about the degree of vascular invasion or entrapment simply by relying on visual inspection and experience. This leads to insufficient consistency in judgment among different doctors for the same case, and preoperative risk assessment results are difficult to standardize and reuse.

[0003] Current technologies typically describe whether blood vessels are involved using imaging reports or image review records, or use a small number of qualitative grading conclusions to assist in decision-making. However, the relationship between tumors and blood vessels in clinical decision-making often involves multi-dimensional information such as distance, contact area ratio, circumferential angle, degree of stenosis, branch involvement level, and blood supply pathway. Current procedures lack a mechanism to structurally express and uniformly quantify the above-mentioned relationship characteristics, resulting in scattered, difficult-to-trace, and difficult-to-compare location data for key high-risk blood vessel segments. Furthermore, it is difficult to form a risk stratification output for surgical decision-making, thereby affecting the reliability of approach selection and vascular protection plans, and increasing the risk of complications such as intraoperative vascular injury and massive bleeding.

[0004] Existing technologies typically focus on single tasks such as tumor segmentation, vascular segmentation, or image classification when introducing artificial intelligence-assisted analysis in order to improve recognition efficiency. However, these methods often fail to output interpretable relationship prediction results for the core clinical question of the relationship between tumors and blood vessels. Furthermore, they have limited adaptability to fluctuations in multimodal image quality, differences in scanning sequences, and heterogeneity of data between centers, making it difficult to directly support the objective assessment and consistency determination of vascular relationships in the surgical planning of skull base meningiomas.

[0005] Therefore, it is necessary to provide an intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels, which can automatically extract spatial features related to the relationship with blood vessels from multimodal images and form quantifiable and traceable prediction results for subsequent surgical risk assessment and individualized decision support. Summary of the Invention

[0006] The purpose of this application is to provide an intelligent prediction method and system for the relationship between skull base meningiomas and adjacent blood vessels, in order to address the shortcomings of the prior art.

[0007] To achieve the above objectives, this application provides the following technical solution: A smart prediction method for the relationship between skull base meningiomas and adjacent blood vessels includes: Step S101: Acquire multiple magnetic resonance imaging images of the same subject, and perform trilinear interpolation resampling, bias field correction, rigid registration and contrast normalization on the multiple sequence images to obtain spatially aligned preprocessed images. Step S102: Based on the preprocessed image, a three-dimensional binary mask is generated by threshold segmentation combined with semi-automatic delineation. The intratumoral cystic necrosis area, peritumoral edema area and tumor parenchymal enhancement area are marked in the three-dimensional binary mask. The tumor body mask is expanded outward and inward to construct the tumor surrounding structure mask. Step S103: Based on the three-dimensional binary mask and the mask of the structure zone around the tumor, extract clinical radiological features, tumor quantitative analysis indicators and spatial radiomics features, and merge them to obtain a joint feature set; Step S104: Perform correlation analysis, minimum absolute contraction, and selection operator regression sequentially on the joint feature set to screen core fusion features. Specifically, compare the correlation statistic with the significance threshold, compare the correlation coefficient with the strength threshold to remove irrelevant features, and compare the regression coefficient with the zero threshold to remove features with a coefficient of zero. Step S105: Construct a three-class training set for vascular relationships based on core fusion features, perform minority class oversampling and class weight correction on the category proportion, and then input it into the preset classification model to obtain the predicted label of the relationship between skull base meningioma and adjacent blood vessels, and interpret the feature contribution of the predicted label output.

[0008] Furthermore, methods for generating three-dimensional binary masks include: A spatially aligned preprocessed image is acquired, and a candidate field of view including skull base meningioma is selected within the preprocessed image. Threshold segmentation is performed on the preprocessed image within the candidate field of view, and the gray-level histogram distribution of the candidate field of view is statistically analyzed to generate a candidate threshold set. The number of connected components and the number of connected component voxels corresponding to the candidate thresholds are compared with the preset connectivity constraint thresholds, and the candidate thresholds that meet the comparison conditions are selected as the segmentation thresholds and the initial binary result is output. Using the initial binary result as a seed constraint, semi-automatic delineation is performed on the preprocessed image. Interactive control points are added along the boundary of skull base meningioma in the 3D view and contour evolution update is performed to obtain a 3D binary mask.

[0009] Furthermore, methods for annotating intratumoral cystic necrosis areas, peritumoral edema areas, and tumor parenchymal enhancement areas within a three-dimensional binary mask include: Based on a three-dimensional binary mask, the gray-level enhancement component within the mask is extracted from longitudinally relaxed contrast agent enhanced three-dimensional sequence images. The gray-level enhancement component is compared with a preset enhancement threshold. Voxels with gray-level enhancement components not lower than the preset enhancement threshold are marked as tumor parenchymal enhancement areas. Based on a three-dimensional binary mask, low grayscale components within the mask are extracted from the transversely relaxed weighted sequence image. The low grayscale components are compared with a preset low grayscale threshold. The set of voxels whose low grayscale components are not higher than the preset low grayscale threshold and whose number of connected voxels is not lower than the preset connected voxel threshold is labeled as the intratumoral cystic necrosis area. Starting from the boundary voxels of the three-dimensional binary mask, the search proceeds outward. High gray-scale abnormal components outside the boundary are extracted from the transverse relaxation weighted inversion recovery sequence image with suppressed liquid signal. The high gray-scale abnormal components are compared with the preset edema threshold. Voxels with high gray-scale abnormal components that are not lower than the preset edema threshold and are spatially adjacent to the outer boundary of the three-dimensional binary mask are marked as peritumoral edema areas.

[0010] Furthermore, methods for constructing a masking structure around a tumor by expanding outwards and inwards from the tumor body include: Based on a multi-label 3D mask, the cystic necrosis area within the tumor and the enhanced area of ​​the tumor parenchyma are merged to obtain a tumor body mask. 3D connected component analysis is performed on the tumor body mask to retain the connected component with the largest number of voxels. An outward expansion operation is performed on the tumor body mask, iteratively expanding the boundary at a preset number of expansion steps per voxel to obtain an outward expansion mask. An inward expansion operation is performed on the tumor body mask, iteratively shrinking the boundary at a preset number of inward expansion steps per voxel to obtain an inward expansion mask. The difference between the outward expansion mask and the inward expansion mask is used to generate a mask of the surrounding structure zone.

[0011] Furthermore, methods for extracting clinical radiological features, quantitative tumor analysis indicators, and spatial radiomics features based on three-dimensional binary masks and masks of structures surrounding the tumor include: A three-dimensional binary mask, along with masks for the intratumoral cystic necrosis area, peritumoral edema area, tumor parenchymal enhancement area, and surrounding structural bands, was obtained and mapped to the corresponding voxel coordinates of the preprocessed image. Clinical radiological features were extracted based on the three-dimensional binary mask and the intratumoral cystic necrosis area, peritumoral edema area, and tumor parenchymal enhancement area. These features included tumor location zoning, tumor boundary clarity grading, tumor morphology classification, peritumoral edema grading, intratumoral cystic necrosis grading, and tumor parenchymal enhancement pattern grading. Quantitative tumor analysis indicators were extracted based on the three-dimensional binary mask, including tumor volume, maximum tumor diameter, tumor surface area, tumor compactness, tumor aspect ratio, and the volume percentages of the intratumoral cystic necrosis area, peritumoral edema area, and tumor parenchymal enhancement area. Spatial radiomics features were extracted based on the three-dimensional binary mask and the surrounding structural band mask, including first-order grayscale statistical features, shape features, and texture-related features.

[0012] Furthermore, methods for merging to obtain a joint feature set include: Clinical radiological features, quantitative tumor analysis indicators, and spatial radiomics features are written into the same feature container according to the subject identifier, and each feature is assigned a unique feature name and feature source label. Missing value processing and scaling are performed on the feature container. Missing value processing includes comparing the missing value ratio with a preset missing value ratio threshold. Features with a missing value ratio higher than the preset missing value ratio threshold are removed from the feature container, and features with a missing value ratio not higher than the preset missing value ratio threshold are filled with the median. Scale scaling includes normalizing the mean and standard deviation of continuous numerical features and performing ordered encoding on graded features. The clinical radiological features, quantitative tumor analysis indicators, and spatial radiomics features that have undergone missing value processing and scaling are merged according to a preset splicing order to output a joint feature set.

[0013] Furthermore, the method of sequentially performing correlation analysis, minimum absolute shrinkage, and selection operator regression on the joint feature set to screen core fusion features includes: Based on feature type, the joint feature set is divided into count-type feature subset, measurement-type feature subset, and spatial image omics feature subset. Standard regularization preprocessing is performed on the joint feature set to map each feature value to a uniform scale. Correlation analysis is then performed on the preprocessed joint feature set to obtain the correlation statistic, correlation coefficient, and probability value for each feature. The probability value is compared with a significance threshold, and the correlation coefficient is compared with an intensity threshold. Irrelevant features are removed to obtain the retained feature set. Minimum absolute shrinkage and selection operator regression are performed on the retained feature set to output a regression coefficient vector. The regression coefficients are compared with a zero threshold to remove features with zero regression coefficients, thus obtaining the core fused features.

[0014] Furthermore, methods for eliminating irrelevant features by comparing correlation statistics with significance thresholds and correlation coefficients with strength thresholds include: For count-type feature subsets, chi-square test statistics and correlation strength coefficients are calculated; for econometric feature subsets, rank correlation coefficients are calculated; and for spatial image omics feature subsets, linear correlation coefficients are calculated. Correlation statistics and correlation coefficients are obtained, and probability values ​​associated with the correlation statistics are read. These probability values ​​are compared with a significance threshold; features with probability values ​​not lower than the significance threshold are marked as irrelevant candidate features. The absolute values ​​of correlation coefficients are compared with a strength threshold; features with absolute values ​​of correlation coefficients not higher than the strength threshold are marked as weakly correlated candidate features. Irrelevant and weakly correlated candidate features are merged to form a removal set. The removal set is then deleted from the joint feature set to obtain the correlation-selected retention set.

[0015] Furthermore, methods for comparing regression coefficients to a zero threshold to eliminate features with a coefficient of zero include: Obtain the correlation screening and retention set and construct the regression input matrix. The regression input matrix is ​​aligned with the target task label according to the subject identifier. Perform regression fitting with a penalty term on the regression input matrix using minimum absolute shrinkage and selection operator regression. Traverse the candidate set of regularization parameters under the preset fold cross division, select the regularization parameter value with the smallest corresponding loss value and output the regression coefficient vector. Compare each regression coefficient in the regression coefficient vector with the zero threshold. Features whose regression coefficients fall into the zero threshold are marked as discarded features, and features whose regression coefficients deviate from the zero threshold are marked as retained features. Output the core fusion feature composed of the retained features.

[0016] A smart prediction system for the vascular relationship of skull base meningiomas, which implements a method for intelligently predicting the relationship between skull base meningiomas and adjacent blood vessels, the system includes: The image preprocessing module is used to acquire multiple magnetic resonance imaging sequences of the same subject and perform trilinear interpolation resampling, bias field correction, rigid registration and contrast normalization on the multiple sequence images to obtain spatially aligned preprocessed images. The mask generation and annotation module generates a three-dimensional binary mask based on the preprocessed image using threshold segmentation combined with semi-automatic delineation. The module then annotates the intratumoral cystic necrosis area, peritumoral edema area, and tumor parenchymal enhancement area within the three-dimensional binary mask. The tumor body mask expands outward and inward to construct a mask for the surrounding structures of the tumor. The extraction and fusion module extracts clinical radiological features, tumor quantitative analysis indicators, and spatial radiomics features based on a three-dimensional binary mask and a mask of the structure zone surrounding the tumor, and merges them to obtain a joint feature set. The feature selection module sequentially performs correlation analysis, minimum absolute contraction, and selection operator regression on the joint feature set to select core fusion features. Specifically, it compares the correlation statistic with the significance threshold, compares the correlation coefficient with the intensity threshold to eliminate irrelevant features, and compares the regression coefficient with the zero threshold to eliminate features with a coefficient of zero. The prediction and interpretation module constructs a three-class classification training set for vascular relationships based on core fusion features, performs minority class oversampling and class weight correction on the category proportion, and then inputs it into the preset classification model to obtain the predicted label of the relationship between skull base meningioma and adjacent blood vessels, and interprets the feature contribution of the predicted label output.

[0017] The technical effects and advantages of the intelligent prediction method and system for the relationship between skull base meningiomas and adjacent blood vessels provided in this application are as follows: By performing trilinear interpolation resampling, bias field correction, rigid registration, and contrast normalization on multiple magnetic resonance imaging (MRI) sequences of the same subject, spatially aligned preprocessed images are obtained. Threshold segmentation combined with semi-automatic delineation is then used to generate a three-dimensional binary mask on the preprocessed images. Within this mask, intratumoral cystic necrosis areas, peritumoral edema areas, and tumor-enhanced areas are labeled. Furthermore, a mask for the surrounding structures is constructed by expanding outwards and inwards from the tumor mask. This transforms the original multiple sequence images into structured, partitioned data with unified coordinates and scale. This reduces the impact of resolution differences, intensity drift, and pose differences between different sequences on subsequent analysis, allowing the main tumor region, enhancement region, necrosis region, edema region, and peritumoral structure zone to form reusable regional constraints within the same coordinate system. This improves the repeatability and consistency of the preoperative assessment process, providing an automated data preparation foundation to address the issues of reliance on experience and insufficient consistency in conclusions during manual image interpretation in the previous techniques.

[0018] Based on a three-dimensional binary mask and a mask of the surrounding structures of the tumor, clinical radiological features, quantitative tumor analysis indicators, and spatial radiomics features are extracted and merged to obtain a joint feature set. Correlation analysis, minimum absolute contraction, and selection operator regression are then performed on the joint feature set to screen core fusion features. Specifically, correlation statistics are compared with significance thresholds, correlation coefficients are compared with intensity thresholds to eliminate irrelevant features, and regression coefficients are compared with a zero threshold to eliminate features with zero coefficients. This transforms multi-dimensional descriptions of signs and spatial statistical information into a computable set of core fusion features. Information related to distance, contact area ratio, and degree of circumference required for assessing the relationship between the tumor and blood vessels is incorporated into a unified feature space through the characteristic expression of tumor zoning and the surrounding structures. Threshold comparison and regression sparsification reduce the interference of redundant and weakly correlated features on classification, enabling the output conclusions to have quantitative basis and structured expression capabilities. This provides a feasible data expression and screening path to address the lack of a structured and unified quantitative mechanism and the difficulty in forming risk stratification outputs in the background technology.

[0019] A three-class classification training set for vascular relationships is constructed using core fusion features. Minority class oversampling and class weight correction are applied to the class proportions. This data is then input into a pre-defined classification model to obtain predicted labels for the relationship between skull base meningiomas and adjacent blood vessels. The contribution of the predicted labels to the output features is explained, resulting in automated classification outputs for dissociated, adherent, and encapsulated types, along with their contributing feature ranking. The benefits are: adjusting class proportions reduces the impact of class imbalance on training bias, leading to more balanced three-class label outputs; and interpreting feature contributions correlates the prediction results with core fusion features, creating interpretable outputs for clinical decision-making, reducing situations where conclusions are given without supporting evidence. This addresses the problems of low efficiency, difficulty in automation, and lack of interpretable relationship prediction results in existing manual assessments, providing objective auxiliary results that can be used for preoperative risk assessment and surgical planning. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a schematic diagram of the intelligent prediction method for the relationship between skull base meningioma and adjacent blood vessels in this application; Figure 2 Flowchart for the development of the classification model in this application; Figure 3 The modeling features and weighting graphs for the fusion model of radiological features and spatial image omics features in this application; Figure 4 This is the ROC curve of the fusion model based on radiological features and spatial image omics features in this application; Figure 5 This is a confusion matrix diagram of the fusion model based on radiological features and spatial image omics features in this application; Figure 6 Radar chart showing the performance comparison of the radiological sign model, spatial image omics model, and fusion model in this application; Figure 7 This is a global map showing the distribution of sapride features in the classification of tumor-adjacent vessel relationships in this application – a dissociated type. Figure 8 Individualized interpretation of tumor-adjacent vessel separation samples in this application - Shapley feature waterfall plot and intensity plot; Figure 9 This is a global map showing the distribution of Shapley characteristics for the tumor-adhesive type in the classification of tumor-adhesive relationships in this application. Figure 10 Individualized interpretation of tumor-adhesive-to-adjacent-blood-vessel samples in this application - Shapley feature waterfall plot and intensity plot; Figure 11 This is a global map showing the distribution of Shapley features for the tumor-adjacent vessel relationship classification in this application; Figure 12 Individualized interpretation of tumor-and-adjacent-vascular-encapsulated samples in this application - Shapley feature waterfall plot and intensity map; Figure 13 This is a schematic diagram of an intelligent prediction system for vascular relationships in skull base meningiomas according to this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] It should be noted that when a component is said to be "fixed to" another component, it can be directly attached to the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component.

[0024] Example 1 Reference Figure 1 The method in this embodiment is executed according to steps S101 to S105; refer to Figure 2 Show the flowchart of the classification model development and design; refer to Figure 3 The fusion model of radiological features and spatial image omics characteristics is presented, along with its modeling features and weighting graphs, referencing... Figure 4 The receiver operating characteristic (ROC) curve of the fusion model is shown, with reference to... Figure 5 This displays the classification results of the multilayer perceptron training set; refer to Figure 6 This section presents a performance comparison of the radiological features model, the spatial image omics model, and the fusion model; (Refer to...) Figure 7 and Figure 8 This section presents the global results of the dissociative Shapley feature distribution and the individualized interpretation results of the dissociative samples, respectively; (Refer to...) Figure 9 and Figure 10 This section presents the global results of the Shapley feature distribution for adhesive-type samples and the individualized interpretation results for adhesive-type samples; (Refer to...) Figure 11 and Figure 12The global results of the distribution of enclosed Shapley features and the individualized interpretation results of enclosed samples are presented respectively. This embodiment provides an intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels, including: Step S101: Acquire multiple magnetic resonance imaging (MRI) sequences of the same subject, and perform trilinear interpolation resampling, bias field correction, rigid registration, and contrast normalization on the multiple sequences to obtain spatially aligned preprocessed images.

[0025] In one embodiment, the method for obtaining a spatially aligned preprocessed image includes: Multiple magnetic resonance imaging (MRI) sequences were acquired from the same subject within the same examination cycle. These sequences included longitudinal relaxation-weighted contrast-enhanced 3D sequences, transverse relaxation-weighted sequences, and transverse relaxation-weighted inversion recovery sequences with liquid signal suppression. The multiple sequences were read into the processing environment, and the corresponding voxel spacing, image origin, and orientation matrix were read. Voxel spacing characterizes the spatial distance between adjacent voxel centers, while the image origin and orientation matrix characterize the position and orientation of the multiple sequences in a unified coordinate system, thus providing a unified coordinate conversion basis for subsequent resampling and registration. For example, if the voxel spacing in the longitudinal relaxation-weighted contrast-enhanced 3D sequences was found to be smaller in the transverse and longitudinal directions than the slice-thickness direction voxel spacing in the transverse relaxation-weighted sequences, this difference indicates resolution inconsistency among the multiple sequences.

[0026] Trilinear interpolation resampling is performed on multiple sequence images to unify the voxel spacing to an isotropic voxel spacing. The isotropic voxel spacing is determined by a preset voxel spacing threshold, which is used to limit the resampling output resolution. Trilinear interpolation resampling includes mapping the output voxel center coordinates to the input voxel coordinate system, locating the eight adjacent input voxels that enclose the coordinates, calculating weights based on the relative distances of the coordinates in the three axes, and performing a weighted summation of the gray values ​​of the eight input voxels to generate the output voxel gray value, so that different sequences can express the same anatomical location on the same voxel grid. For example, the transverse relaxation weighted sequence image has a large layer thickness. After trilinear interpolation resampling, denser voxel sampling points are generated in the layer thickness direction to align the voxel grid of the longitudinal relaxation weighted contrast agent enhanced 3D sequence image.

[0027] Bias field correction is performed on each sequence of images obtained by trilinear interpolation resampling. The bias field correction uses a low-frequency field estimation and iterative update mechanism to suppress the intensity component of slow spatial drift. The low-frequency field is used to characterize the slowly changing gray background caused by magnetic field inhomogeneity. The bias field correction includes estimating the low-frequency field in the image domain and compensating the original image so that the gray difference of the same tissue at different spatial locations converges to within the preset gray fluctuation threshold, thereby reducing the sensitivity of subsequent segmentation threshold setting to spatial drift. For example, the gray of the same brain tissue on the left and right sides of the image has a continuous increasing trend. After performing bias field correction, the increasing trend is suppressed and the gray distribution of the tissue tends to be concentrated.

[0028] Using longitudinally relaxed contrast-enhanced 3D sequence images as reference sequences, rigid registration was performed on transversely relaxed sequence images and liquid-suppressed transversely relaxed inversion recovery sequence images to obtain registered images spatially consistent with the reference sequence. Rigid registration employed a rigid transformation model including three-axis translation and three-axis rotation, which was used to describe the head position differences of the same subject in different sequence acquisitions. Rigid registration involved calculating the similarity metric between the reference sequence and the sequence to be registered under rigid transformation and optimizing the search to obtain the rigid transformation parameter that reached the extreme value of the similarity metric. Subsequently, the rigid transformation parameter was used to perform resampling mapping on the sequence to be registered to obtain the registered image, so that the tumor boundary and blood vessel course of different sequences fell into the same voxel coordinate position. For example, the liquid-suppressed transversely relaxed inversion recovery sequence image had a slight rotational deviation relative to the reference sequence. After rigid registration, the tumor contour corresponded to the same coordinate position in both sequences.

[0029] Contrast normalization is performed on each sequence image obtained through rigid registration. The contrast normalization adopts mean and standard deviation normalization based on the region of interest (ROI). The ROI is consistent with the ROI used in the subsequent threshold segmentation combined with semi-automatic delineation step. The ROI includes the tumor region and the peritumoral edema region, which are the coverage areas of the subsequent 3D binary mask. Contrast normalization includes statistically analyzing the mean and dispersion parameters of gray levels in the ROI, and mapping the gray levels of multiple sequence images to a uniform relative scale based on the statistical parameters. This ensures that the gray level scale differences introduced by different scanning conditions converge to within a preset scale deviation threshold, thereby improving the consistency of cross-sequence feature extraction. For example, if the gray level range of the tumor parenchyma of the same subject varies across different sequences, the contrast normalization compresses the variation to a uniform scale, facilitating subsequent segmentation operations under the same threshold principle.

[0030] The formula for calculating normalized grayscale values ​​is as follows: ; In the formula, I To preprocess the image in voxel coordinates ( The original gray value at (); x is the coordinate component of the voxel in the first coordinate axis direction, y is the coordinate component of the voxel in the second coordinate axis direction, and z is the coordinate component of the voxel in the third coordinate axis direction. The normalized grayscale value is obtained by normalizing the contrast of the original grayscale value. σ is the mean of the original gray values ​​of all voxels within the region of interest; σ is the standard deviation of the original gray values ​​of all voxels within the region of interest; the region of interest includes the tumor region and the peritumoral edema region.

[0031] Through the above steps, a spatially aligned preprocessed image is obtained. The purpose is to simultaneously achieve resolution unification, intensity slow drift suppression, pose difference elimination, and grayscale unification. This ensures that when a three-dimensional binary mask is subsequently generated on the preprocessed image and further labeled with intratumoral cystic necrosis areas, peritumoral edema areas, and tumor parenchymal enhancement areas, the labeling results have consistent spatial placement across multiple sequences. For example, the same voxel coordinate corresponds to the tumor parenchymal enhancement signal in a longitudinally relaxed contrast-enhanced three-dimensional sequence image, and corresponds to the peritumoral edema signal in a transversely relaxed inversion recovery sequence image with suppressed liquid signals. Spatial alignment allows the two types of signals to be used together in the same coordinate system.

[0032] Step S102: Based on the preprocessed image, a three-dimensional binary mask is generated by threshold segmentation combined with semi-automatic delineation. The intratumoral cystic necrosis area, peritumoral edema area and tumor parenchymal enhancement area are marked in the three-dimensional binary mask. The tumor body mask is expanded outward and inward to construct the tumor surrounding structure mask.

[0033] In one embodiment, a method for generating a 3D binary mask based on a preprocessed image using threshold segmentation combined with semi-automatic delineation includes: Acquire spatially aligned preprocessed images and select candidate field of view including skull base meningiomas in the preprocessed images. The candidate field of view is manually selected in the coronal, sagittal, and axial planes to limit the calculation range of threshold segmentation to reduce the interference of bony structures and cerebrospinal fluid signals on segmentation; for example, a rectangular area covering the tumor and a ring of soft tissue around it is selected in the axial plane.

[0034] Threshold segmentation is performed on the preprocessed image within the candidate field of view. Threshold segmentation includes statistically analyzing the gray-level histogram distribution of the candidate field of view and generating a set of candidate thresholds. The number of connected components and the number of connected voxels corresponding to each candidate threshold are compared with a preset connectivity constraint threshold. Candidate thresholds with the number of connected components falling within a preset range and the maximum number of connected voxels not less than a preset voxel count threshold are selected as the segmentation thresholds, and the initial binary result is output. For example, when scanning from low to high candidate thresholds, a candidate threshold is selected where the maximum connected component covers the main body of the tumor and the number of tiny artifact connected components does not exceed the upper limit.

[0035] Using the initial binary result as a seed constraint, semi-automatic delineation is performed on the preprocessed image. Semi-automatic delineation includes adding interactive control points along the tumor boundary in the 3D view and performing contour evolution update. The contour evolution update process uses both the boundary gradient intensity and the region consistency cost as evolution drivers, so that the evolving contour stops at the strong gradient boundary and fills in the region with high region consistency, resulting in a 3D binary mask. For example, when the boundary gradient is weak near the venous sinus of the tumor, control points are added to limit the evolution contour from going out of bounds. Tumor-related voxels in the preprocessed image are separated from the background tissue and a computable 3D binary mask is formed, providing region constraints for partition labeling and subsequent feature extraction.

[0036] Methods for annotating intratumoral cystic necrosis areas, peritumoral edema areas, and tumor parenchymal enhancement areas within a three-dimensional binary mask include: Based on a three-dimensional binary mask, the gray-level enhancement component within the mask is extracted from the longitudinally relaxed contrast-enhanced three-dimensional sequence image, and the enhancement component is compared with a preset enhancement threshold. Voxels with enhancement components not lower than the enhancement threshold are marked as tumor parenchymal enhancement areas. For example, if a patchy continuous high gray-level enhancement area appears within the mask, the voxel of the area is marked as a tumor parenchymal enhancement area.

[0037] Based on a three-dimensional binary mask, low grayscale components within the mask are extracted from the laterally relaxed weighted sequence image. The low grayscale components are then compared with a preset low grayscale threshold. A set of voxels whose low grayscale components are not higher than the low grayscale threshold and whose number of connected voxels is not lower than the preset connected voxel threshold is labeled as an intratumoral cystic necrosis region. For example, if a continuous low grayscale region appears in the center of the tumor and remains connected in multiple sections, the continuous region is labeled as an intratumoral cystic necrosis region.

[0038] Starting from the boundary voxels of the 3D binary mask, the search proceeds outward to extract high-grayscale abnormal components outside the boundary in the transverse relaxation-weighted inversion recovery sequence image with suppressed liquid signals. These abnormal components are then compared with a preset edema threshold. Voxels whose abnormal components are not lower than the edema threshold and are spatially adjacent to the outer boundary of the 3D binary mask are labeled as peritumoral edema areas. For example, a band-shaped high-grayscale area around the tumor boundary is labeled as a peritumoral edema area. The labeling process uses a multi-label voxel encoding method to write the intratumoral cystic necrosis area, peritumoral edema area, and tumor parenchymal enhancement area into the same 3D mask container for subsequent unified retrieval and statistical analysis. The aim is to divide the interior of the 3D binary mask into functional zones with different pathological meanings, allowing subsequent quantitative indicators and spatial radiomics features to be calculated separately under the constraints of these zones. This transforms the overall tumor into a structured expression of enhancement, necrosis, and edema. For example, the ratio of cystic necrosis to edema volume can form comparable zonal quantitative results across different cases.

[0039] In one embodiment, a method for constructing a masking band of structures surrounding a tumor by expanding outward and inward from a tumor body mask includes: Based on a multi-label 3D mask, the intratumoral cystic necrosis area and the tumor parenchyma enhancement area are merged to obtain a tumor body mask that only includes the tumor body. Then, 3D connected component analysis is performed on the tumor body mask to retain the connected component with the largest number of voxels to exclude small artifact connected components. For example, a small number of isolated points are scattered around the outer periphery of the mask. The connected component analysis retains the main connected components and removes the isolated points.

[0040] An outward expansion operation is performed on the tumor mask. The outward expansion operation iterates the boundary in units of voxels according to a preset number of outward expansion steps to obtain the outward expansion mask.

[0041] An inward expansion operation is performed on the tumor mask. The inward expansion operation iterates and shrinks the boundary in units of voxels according to a preset number of inward expansion steps to obtain the inward expansion mask. A structural band mask surrounding the tumor is generated by the difference between an external expansion mask and an internal expansion mask. This structural band mask covers a thin band-like region inside and outside the tumor boundary, used to express information about the peritumoral microenvironment and adjacent structures. For example, the external expansion step takes 5 voxels and the internal expansion step takes 1 voxel, resulting in a band-like set of voxels surrounding the tumor boundary. The aim is to construct a surrounding structural band closely adjacent to the tumor boundary without changing the definition of tumor zonation, so that subsequent spatial radiomics features can be extracted within a peripheral band of uniform thickness. This allows spatial information related to vascular relationships, such as the number of peritumoral vessels and the eccentricity of peritumoral morphology, to be included in feature calculation. For example, the structural band mask covers the local vascular course area outside the tumor edge, facilitating the counting of the total number of peritumoral vessels and their use for vascular relationship prediction.

[0042] Step S103: Based on the three-dimensional binary mask and the mask of the structure surrounding the tumor, extract clinical radiological features, tumor quantitative analysis indicators and spatial radiomics features, and merge them to obtain a joint feature set.

[0043] In one embodiment, a method for extracting clinical radiological features, quantitative tumor analysis indicators, and spatial radiomics features based on a three-dimensional binary mask and a mask of structures surrounding the tumor includes: Obtain a three-dimensional binary mask, an intratumoral cystic necrosis area, a peritumoral edema area, a tumor parenchymal enhancement area, and a mask of the surrounding structures of the tumor. Map the mask back to the voxel coordinates of the preprocessed image output in step S101 for performing feature statistics in the same coordinate system. For example, the tumor parenchymal enhancement area falls into a continuous set of high grayscale voxels in a longitudinally relaxed contrast-enhanced three-dimensional sequence image.

[0044] Clinical radiological features are extracted based on three-dimensional binary masks and regional masks. These features include tumor location zoning, tumor boundary clarity grading, tumor morphology classification, peritumoral edema grading, intratumoral cystic degeneration and necrosis grading, and tumor parenchymal enhancement pattern grading. Grading is achieved through threshold comparison and count comparison. Specifically, the ratio of the number of voxels in the peritumoral edema area to the number of voxels in the three-dimensional binary mask is compared with a preset edema grading threshold; the ratio of the number of voxels in the intratumoral cystic degeneration and necrosis area to the number of voxels in the three-dimensional binary mask is compared with a preset necrosis grading threshold; and the ratio of the number of voxels in the tumor parenchymal enhancement area to the number of voxels in the three-dimensional binary mask is compared with a preset enhancement grading threshold. This converts the image reading description into a computable label. For example, if the proportion of voxels in the peritumoral edema area falls into the second grading level according to the edema grading threshold, the output edema grading is set to the second level.

[0045] Based on a three-dimensional binary mask, quantitative tumor analysis indicators are extracted. These indicators include tumor volume, maximum tumor diameter, tumor surface area, tumor compactness, tumor aspect ratio, and the proportions of cystic necrosis, peritumoral edema, and solid enhancement within the tumor. Tumor volume is obtained by multiplying the number of voxels in the three-dimensional binary mask by their volume. The maximum tumor diameter is obtained by the maximum Euclidean distance between pairs of boundary points in the three-dimensional binary mask. The surface area is obtained by triangulating the three-dimensional binary mask into a mesh and calculating the mesh area. For example, if the number of voxels in the three-dimensional binary mask is 10,000 and the voxel volume is 1 cubic millimeter, the tumor volume output will be 10,000 cubic millimeters.

[0046] Spatial radiomics features are extracted based on a 3D binary mask and a mask of the surrounding structures. These features include first-order gray-level statistical features, shape features, and texture-related features. Texture-related features are obtained by constructing gray-level co-occurrence statistics, gray-level run statistics, gray-level region size statistics, and neighborhood gray-level difference statistics, and are calculated separately under the constraints of the 3D binary mask and the mask of the surrounding structures, so that the information inside the tumor and the information of the surrounding structures are quantified simultaneously. For example, the contrast index of the gray-level co-occurrence statistics within the mask of the surrounding structures is used to characterize the degree of texture change in the surrounding tissue. By fusing the tumor information from the 3D binary mask and the surrounding information from the mask of the surrounding structures, the interpretable description of clinical radiological features and the spatial statistical information of the spatial radiomics features are transformed into numerical inputs that can be used for classification learning.

[0047] In one implementation, the method for merging to obtain a joint feature set includes: Clinical radiological features, tumor quantitative analysis indicators, and spatial radiomics features are written into the same feature container according to the subject identifier, and each feature is assigned a unique feature name and feature source marker. The feature source marker includes tumor origin and peritumoral structural zone origin to ensure that features with the same name can be distinguished under different sources; for example, the same texture-related feature is formed into two independent features by tumor origin and peritumoral structural zone origin respectively.

[0048] Missing value processing and scaling are performed on the feature container. Missing value processing includes comparing the missing value ratio with a preset missing value ratio threshold. Features with a missing value ratio higher than the threshold are removed from the feature container, while features with a missing value ratio not higher than the threshold are filled with the median. Scaling scaling includes normalizing the mean and standard deviation of continuous numerical features and performing ordered encoding on hierarchical features to make features of different dimensions fall into a unified numerical range. For example, tumor volume and surface area are normalized and mapped to similar numerical ranges, and hierarchical features are encoded as incremental integers in order of rank.

[0049] The processed clinical radiological features, tumor quantitative analysis indicators, and spatial radiomics features are merged according to a preset splicing order to output a joint feature set. The joint feature set is then associated with subject identifiers for use in correlation analysis and regression screening steps. For example, the joint feature set includes several radiological grading features, several quantitative analysis indicators, and several radiomics features of the tumor and its surrounding structures. The purpose of merging is to incorporate clinically relevant radiographic information and radiomics information including spatial statistics into the same feature space, providing a unified input interface for the joint prediction of the relationship between skull base meningiomas and adjacent blood vessels.

[0050] Step S104: Perform correlation analysis, minimum absolute contraction, and selection operator regression sequentially on the joint feature set to screen core fusion features. Specifically, compare the correlation statistic with the significance threshold and the correlation coefficient with the strength threshold to remove irrelevant features. Compare the regression coefficient with the zero threshold to remove features with a coefficient of zero.

[0051] In one implementation, a method for sequentially performing correlation analysis, minimum absolute shrinkage, and selection operator regression to screen core fusion features on a joint feature set includes: A joint feature set is obtained and divided into a count feature subset, a measurement feature subset, and a spatial radiomics feature subset based on the feature type. This division is used to ensure that the calculation paths for different types of features corresponding to different correlation statistics are consistent. For example, cortical vascular embedding, presence or absence of deep cerebral edema, cerebral edema crossing the midline, and cortical gyri embedding are classified into the count feature subset; the maximum tumor diameter, tumor lobulation index, and tumor necrosis ratio are classified into the measurement feature subset; and the total number of peritumoral vessels and the peritumoral region eccentricity are classified into the spatial radiomics feature subset.

[0052] Standard regularization preprocessing is performed on the joint feature set to map each feature value to a uniform scale, thereby reducing the impact of dimensional differences on correlation analysis, minimum absolute shrinkage, and selection operator regression screening; for example, the maximum tumor diameter and the total number of peritumoral vessels form a comparable numerical range on a uniform scale.

[0053] Correlation analysis is performed on the joint feature set that has been preprocessed by standard regularization to obtain the correlation statistics, correlation coefficient and probability value corresponding to each feature, and then the threshold comparison and elimination process is entered. The feature set retained by correlation analysis is subjected to minimum absolute shrinkage and selection operator regression, and the regression coefficient vector is output. Features with zero coefficients are removed according to the zero threshold to obtain the core fusion features, which are used as input for subsequent classification model training. For example, the initial fusion feature set is formed by splicing clinical radiology features and spatial radiomics features. The core fusion feature set is obtained by screening through correlation analysis and minimum absolute shrinkage and selection operator regression. Under the condition that the joint feature set includes features from multiple sources and of multiple types, irrelevant and redundant features are first removed at the statistical level by correlation analysis, and then features are compressed at the regression coefficient level by minimum absolute shrinkage and selection operator regression to obtain the core fusion features, so that the input dimension of the subsequent model converges and the output of the interpretable feature set is maintained.

[0054] ; In the formula, The number of samples; The number of features that enter the regression screening stage; For the first The target task value for each sample; For the regression constant term; This is a vector of regression coefficients; For the first The regression coefficients corresponding to the features; For the first The sample at the th Values ​​on the feature; Here, is the regularization parameter used to control the intensity of the penalty; when When the comparison with the zero threshold falls into the zero threshold, the first... Features that are determined to have a coefficient of zero are removed from the core fusion features.

[0055] In one implementation, a method for comparing a correlation statistic with a significance threshold and a correlation coefficient with a strength threshold to remove irrelevant features includes: For count-type feature subsets, chi-square test statistics and correlation strength coefficients are calculated; for econometric feature subsets, rank correlation coefficients are calculated; and for spatial image omics feature subsets, linear correlation coefficients are calculated, resulting in correlation statistics and correlation coefficients. The correlation statistics are used to characterize the degree of association between features and target tasks, while the correlation coefficients are used to characterize the direction and strength of the association.

[0056] ; In this formula, This is the chi-square test statistic; For the first Line number The actual observation frequency of the column; For the first Line number The theoretical expected frequency of the column; This is the index of the value category of the count-type feature. A category index for the three-category label of vascular relationships; The number of categories for the count-type feature; The number of categories for the three-category label of the vascular relationship; The corresponding probability value is used for comparison with the significance threshold.

[0057] ; In this formula, Here is the Cramer correlation coefficient; The number of samples; The smaller of the number of count-type feature categories and the number of three-class label categories for vascular relationships; Used for comparison with intensity threshold.

[0058] Read the probability values ​​associated with the correlation statistics and compare the probability values ​​with the significance threshold. Features with probability values ​​not lower than the significance threshold are marked as irrelevant candidate features. The absolute value of the correlation coefficient is compared with the strength threshold. Features whose absolute value of the correlation coefficient is not higher than the strength threshold are marked as weakly correlated candidate features. Irrelevant candidate features and weakly correlated candidate features are merged to form a removal set. The removal set is deleted from the joint feature set to obtain the correlation screening retention set. The correlation screening retention set is written into the input port of the regression step of the minimum absolute shrinkage and selection operator. For example, when the absolute value of the correlation coefficient corresponding to the maximum diameter of the tumor is higher than the strength threshold, it is included in the correlation screening retention set. The purpose is to remove irrelevant features in the joint feature set in advance by comparing the probability value threshold and the correlation coefficient strength threshold, so that the feature set entering the regression screening stage has less redundancy and reduces the risk of collinear input.

[0059] In one implementation, a method for comparing regression coefficients to a zero threshold to eliminate features with a coefficient of zero includes: A correlation-filtered retention set is obtained and a regression input matrix is ​​constructed, which is aligned with the target task label according to the subject identifier. A minimum absolute contraction and selection operator regression is used to perform a penalized regression fit on the regression input matrix, and the candidate set of regularization parameters is traversed under a preset fold cross-partition. The regularization parameter value that minimizes the comparison between the loss value and the loss threshold is selected, and the corresponding regression coefficient vector is output. Each regression coefficient in the regression coefficient vector is compared with a zero threshold. Features whose regression coefficients fall below the zero threshold are marked as discarded features, and features whose regression coefficients deviate from the zero threshold are marked as retained features. The core fusion features composed of the retained features are output and used for subsequent vascular relationship classification model training. The purpose is to further compress the correlation-filtered retention set into core fusion features, so that the retained features form a sparse expression at the regression coefficient level, thereby providing a more convergent set of input features for subsequent model training. For example, when the regression coefficient corresponding to the total number of peritumoral vessels deviates from the zero threshold, it is included in the core fusion features.

[0060] Step S105: Construct a three-class training set for vascular relationships based on core fusion features, perform minority class oversampling and class weight correction on the category proportion, and then input it into the preset classification model to obtain the predicted label of the relationship between skull base meningioma and adjacent blood vessels, and interpret the feature contribution of the predicted label output.

[0061] In one implementation, a method for constructing a three-class classification training set for vascular relationships based on core fusion features includes: The core fusion features are acquired and organized into a feature vector record set according to the subject's identification. Each record in the feature vector record set corresponds to one imaging evaluation result of a subject. Next, the vascular relationship annotation results associated with the subject's identification are obtained. The vascular relationship annotation results include three types of labels: separation type, adhesion type, and encapsulation type. Then, the feature vector record set and the three types of labels are matched one by one according to the subject's identification to form a three-class classification training set of vascular relationships, which includes input features and output labels. For example, the core fusion feature vector of a subject is bound to the encapsulation type label to form a training sample. The purpose is to form a structured sample set of core fusion features and three-class classification labels of vascular relationships, so that the preset classification model has training input and training target that can be directly read. For example, each sample in the training set includes both tumor origin features and peritumoral structure band origin features to distinguish different vascular relationship labels.

[0062] In one implementation, the method for performing minority class oversampling and class weight correction on class proportions includes: The training set for the three-class classification of vascular relationships first calculates the percentage of samples in each category: separation, adhesion, and encapsulation. Then, it compares this percentage with a preset category percentage threshold, labeling categories with percentages not exceeding the threshold as low-percentage categories. The sample set corresponding to these low-percentage categories is then processed using minority class oversampling. Minority class oversampling involves repeatedly sampling and amplifying or feature space interpolation to increase the number of amplified low-percentage category samples, ensuring the number falls within the target deviation threshold. Finally, based on the sample percentages before amplification, category weights are calculated for each category, and these weights are compared with a weight upper limit threshold. The comparison process involves truncating weights above the upper weight threshold to obtain the corrected category weights. For example, if a category has a sample percentage of 0.10 and is not higher than the preset category percentage threshold, minority class oversampling is performed to increase the number of category samples to the target percentage range, and the category weights are increased to within the upper weight threshold. The aim is to mitigate the impact of category percentage bias on the training process by simultaneously increasing the number of samples and adjusting the loss weights, so that the classification model can obtain a more balanced source of category gradients when learning the three types of vascular relationships. For example, when the number of encapsulated samples is too small, the amplification process and weight correction give the classification model more training opportunities for encapsulated samples.

[0063] In one implementation, the method for interpreting the contribution of the predicted label output features to the classification model to obtain a predicted label of the relationship between the skull base meningioma and adjacent blood vessels includes: A classification model is selected and written into a three-class classification training set for vascular relationships. The preset classification model includes a tree model classifier or a support vector classifier that supports multi-class output. Then, the training parameters are configured using minority class oversampling and class weight correction results. Iterative training is performed on the classification model and the model parameters are fixed. Next, the core fusion features of the subjects to be predicted are obtained and input into the classification model. The output class scores for separation, adhesion, and encapsulation are used, and the class corresponding to the highest score is selected as the predicted label for the relationship between skull base meningioma and adjacent blood vessels. Furthermore, the feature contribution explanation is output for the predicted label. The feature contribution explanation includes performing perturbation replacement on each core fusion feature and calculating the change in class score. The change in score is compared with the contribution threshold, and a list of features with contributions higher than the contribution threshold and the contribution ranking are output.

[0064] Specifically, classification model training methods include: The core fusion features are obtained, and each subject's core fusion features are organized into a feature vector record according to the subject's identifier. At the same time, the three-class label of vascular relationship associated with the subject's identifier is obtained. The three-class label of vascular relationship includes separation type, adhesion type and encapsulation type. The feature vector record and the three-class label of vascular relationship are matched one-to-one to form a three-class training set of vascular relationship. For example, the feature vector record of a subject is bound to the encapsulation type label to form a training sample.

[0065] For the three-class classification training set of vascular relationships, the sample proportion of each category is statistically analyzed. The sample proportion is compared with the preset category proportion threshold. Categories with a sample proportion not higher than the preset category proportion threshold are marked as low proportion categories. Sample amplification processing is performed on low proportion categories. The sample amplification processing includes repeated extraction amplification or feature space interpolation synthesis amplification, so that the amplified category proportion converges to the target proportion range. At the same time, the category weight is calculated based on the category sample proportion before amplification processing, and the category weight is compared with the upper limit threshold of weight. When the category weight is higher than the upper limit threshold of weight, it is truncated to the upper limit threshold of weight, and the category weight correction result is obtained. For example, when the proportion of encapsulated samples is not higher than the preset category proportion threshold, encapsulated samples are amplified and the encapsulated category weight is increased but not exceeding the upper limit threshold of weight.

[0066] A candidate classification algorithm set is established and cross-split training is performed. The candidate classification algorithm set includes logistic regression classifier, support vector classifier, gradient boosting tree classifier, lightweight gradient boosting tree classifier, and multilayer perceptron classifier. For each candidate classifier, multi-fold cross-training and evaluation are performed under the same data splitting rules. In each fold training, sample amplification processing and class weight correction results are loaded simultaneously to keep the training input of different candidate classifiers consistent with the class imbalance processing. For example, the three-class classification training set of blood vessel relationship is divided into five subsets. One subset is selected in turn as the evaluation subset, and the remaining subsets are used as training subsets.

[0067] The optimal classifier is selected and the model parameters are fixed based on the cross-evaluation results. The cross-evaluation results include consistency coefficient, area under the equilibrium curve, area under the curve for multi-class one-to-many, accuracy, recall, and comprehensive evaluation score. The optimal classifier is determined based on the comprehensive evaluation score and the preset selection rules, and the parameter configuration corresponding to the optimal classifier in the cross-training is fixed as the final training parameters. For example, when the multilayer perceptron classifier ranks first in the comprehensive evaluation score ranking, the multilayer perceptron classifier is used as the final model and its parameter configuration is fixed.

[0068] Based on the final model output, predicted labels and feature contribution explanations are provided. The core fusion feature vector is used to output category scores for separation, adhesion, and encapsulation types. The category with the highest score is selected as the predicted label for the relationship between skull base meningiomas and adjacent vessels. Simultaneously, perturbation replacement is performed on each core fusion feature, and the change in category score is calculated. The score change is compared with a contribution threshold, and a list of features with contributions exceeding the threshold and their ranking are output to explain the basis for the predicted label formation. For example, if the score drop for encapsulation type exceeds the contribution threshold when the total number of peritumoral vessels is replaced with the number of digits in the training set, then the total number of peritumoral vessels is output as a high-contribution feature and ranked first. The aim is to output vascular relationship predicted labels that can be used for clinical interpretation, while providing a set of contributing features driving the predicted labels, making the vascular relationship prediction results traceable. For example, the predicted label is adhesion type, accompanied by boundary morphology features and peritumoral structural texture features, with a contribution ranking output.

[0069] Methods for performing perturbation replacements on each core fusion feature and calculating the change in category score include: Step S1051: Obtain the trained classification model and the target sample set for interpretation. The target sample set is selected from the three-class classification training set, validation set, or test set of vascular relationships, and obtain the core fusion feature vector corresponding one-to-one with the target sample set. At the same time, construct a feature reference value table. The feature reference value table records the reference value of each core fusion feature in the training set according to the feature name. The reference value includes the median reference value and the quantile interval reference value, which are used to provide a fixed replacement value source for perturbation replacement. For example, the number of digits in the training set is recorded as the reference value for the total number of peritumoral vessels.

[0070] Step S1052: For any target sample in the target sample set, input the core fusion feature vector of the target sample into the classification model, output the category score vectors of separation, adhesion, and encapsulation, and record the category score vectors as the baseline score; each item of the category score vector is the confidence score or probability score output by the classification model for the corresponding category, and the score is used for subsequent change calculation; for example, the encapsulation category score is recorded as the baseline score.

[0071] Step S1053: Select each feature in the core fusion feature vector as the feature to be perturbed, and perform perturbation replacement on the feature to be perturbed to generate a perturbed sample vector. The perturbation replacement includes: reading the median reference value of the feature to be perturbed in the feature reference value table, replacing the original value of the feature to be perturbed in the target sample vector with the median reference value to obtain a first perturbed sample vector; and reading the quantile interval reference value of the feature to be perturbed in the feature reference value table, replacing the original value with the lower quantile reference value and the upper quantile reference value to obtain a second perturbed sample vector and a third perturbed sample vector. The quantile points of the lower quantile reference value and the upper quantile reference value are determined by a preset quantile threshold to limit the perturbation amplitude to be within the statistical range of the training set. For example, the maximum diameter of the tumor is replaced with the quantile, lower quantile value and upper quantile value in the training set to generate three perturbed sample vectors.

[0072] Step S1054: Input the perturbation sample vectors into the classification model to obtain the corresponding perturbation category score vectors, and calculate the change in category score for each perturbation sample vector. The calculation of the change in category score includes: for each category, taking the absolute value of the difference between the category score corresponding to the perturbation category score vector and the category score corresponding to the baseline score as the change in category; and in multiple perturbation replacements of the same feature to be perturbed, taking the maximum value or average value of the change to obtain the contribution measure of the feature to be perturbed to the category; at the same time, assigning a direction label to the change, the direction label is determined by the relationship between the perturbation category score and the baseline score, and is used to indicate the direction of improvement or inhibition of the category score by the feature to be perturbed; for example, if the encapsulation score decreases when the total number of peritumoral vessels is replaced to the median, then the direction label is recorded as a decrease in the encapsulation score.

[0073] Step S1055: Repeat steps S1053 and S1054 for all features to be perturbed in the target sample to obtain the feature contribution table of the target sample in the separation, adhesion, and encapsulation types. The feature contribution table records the contribution measure and direction label of each feature to each type using the feature name as the index. The contribution measure is compared with the contribution threshold. Features with contribution measures higher than the contribution threshold are written into the contribution feature list and output in the contribution ranking from high to low to form the interpretation result of the prediction label. For example, when the total number of peritumoral blood vessels, the maximum diameter of the tumor, and the proportion of edema volume are among the top in the encapsulation type contribution ranking, the three features are output as high contribution features of the encapsulation type.

[0074] Using the above method, without introducing additional model structures, the perturbation replacement and category score change calculation are performed on the core fusion features one by one to output the set of contributing features and contribution ranking corresponding to the predicted labels. The purpose is to provide traceable feature basis for the three-class classification results of vascular relationships, so as to facilitate the association of radiological signs, tumor quantitative analysis indicators and spatial radiomics features in the context of clinical image reading.

[0075] Example 2 See Figure 13 As shown, this embodiment provides an intelligent prediction system for the vascular relationship of skull base meningiomas, which implements an intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels. The system includes: The image preprocessing module is used to acquire multiple magnetic resonance imaging sequences of the same subject and perform trilinear interpolation resampling, bias field correction, rigid registration and contrast normalization on the multiple sequence images to obtain spatially aligned preprocessed images. The mask generation and annotation module generates a three-dimensional binary mask based on the preprocessed image using threshold segmentation combined with semi-automatic delineation. The module then annotates the intratumoral cystic necrosis area, peritumoral edema area, and tumor parenchymal enhancement area within the three-dimensional binary mask. The tumor body mask expands outward and inward to construct a mask for the surrounding structures of the tumor. The extraction and fusion module extracts clinical radiological features, tumor quantitative analysis indicators, and spatial radiomics features based on a three-dimensional binary mask and a mask of the structure zone surrounding the tumor, and merges them to obtain a joint feature set. The feature selection module sequentially performs correlation analysis, minimum absolute contraction, and selection operator regression on the joint feature set to select core fusion features. Specifically, it compares the correlation statistic with the significance threshold, compares the correlation coefficient with the intensity threshold to eliminate irrelevant features, and compares the regression coefficient with the zero threshold to eliminate features with a coefficient of zero. The prediction and interpretation module constructs a three-class classification training set for vascular relationships based on core fusion features, performs minority class oversampling and class weight correction on the category proportion, and then inputs it into the preset classification model to obtain the predicted label of the relationship between skull base meningioma and adjacent blood vessels, and interprets the feature contribution of the predicted label output.

[0076] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligently predicting the relationship between skull base meningiomas and adjacent blood vessels, characterized in that, include: Step S101: Acquire multiple magnetic resonance imaging images of the same subject, and perform trilinear interpolation resampling, bias field correction, rigid registration and contrast normalization on the multiple sequence images to obtain spatially aligned preprocessed images. Step S102: Based on the preprocessed image, a three-dimensional binary mask is generated by threshold segmentation combined with semi-automatic delineation. The intratumoral cystic necrosis area, peritumoral edema area and tumor parenchymal enhancement area are marked in the three-dimensional binary mask. The tumor body mask is expanded outward and inward to construct the tumor surrounding structure mask. Step S103: Based on the three-dimensional binary mask and the mask of the structure zone around the tumor, extract clinical radiological features, tumor quantitative analysis indicators and spatial radiomics features, and merge them to obtain a joint feature set; Step S104: Perform correlation analysis, minimum absolute contraction, and selection operator regression sequentially on the joint feature set to screen core fusion features. Specifically, compare the correlation statistic with the significance threshold, compare the correlation coefficient with the strength threshold to remove irrelevant features, and compare the regression coefficient with the zero threshold to remove features with a coefficient of zero. Step S105: Construct a three-class training set for vascular relationships based on core fusion features, perform minority class oversampling and class weight correction on the category proportion, and then input it into the preset classification model to obtain the predicted label of the relationship between skull base meningioma and adjacent blood vessels, and interpret the feature contribution of the predicted label output.

2. The intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels according to claim 1, characterized in that, Methods for generating 3D binary masks include: A spatially aligned preprocessed image is acquired, and a candidate field of view including skull base meningioma is selected within the preprocessed image. Threshold segmentation is performed on the preprocessed image within the candidate field of view, and the gray-level histogram distribution of the candidate field of view is statistically analyzed to generate a candidate threshold set. The number of connected components and the number of connected component voxels corresponding to the candidate thresholds are compared with the preset connectivity constraint thresholds, and the candidate thresholds that meet the comparison conditions are selected as the segmentation thresholds and the initial binary result is output. Using the initial binary result as a seed constraint, semi-automatic delineation is performed on the preprocessed image. Interactive control points are added along the boundary of skull base meningioma in the 3D view and contour evolution update is performed to obtain a 3D binary mask.

3. The intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels according to claim 2, characterized in that, Methods for annotating intratumoral cystic necrosis areas, peritumoral edema areas, and tumor parenchymal enhancement areas within a three-dimensional binary mask include: Based on a three-dimensional binary mask, the gray-level enhancement component within the mask is extracted from longitudinally relaxed contrast agent enhanced three-dimensional sequence images. The gray-level enhancement component is compared with a preset enhancement threshold. Voxels with gray-level enhancement components not lower than the preset enhancement threshold are marked as tumor parenchymal enhancement areas. Based on a three-dimensional binary mask, low grayscale components within the mask are extracted from the transversely relaxed weighted sequence image. The low grayscale components are compared with a preset low grayscale threshold. The set of voxels whose low grayscale components are not higher than the preset low grayscale threshold and whose number of connected voxels is not lower than the preset connected voxel threshold is labeled as the intratumoral cystic necrosis area. Starting from the boundary voxels of the three-dimensional binary mask, the search proceeds outward. High gray-scale abnormal components outside the boundary are extracted from the transverse relaxation weighted inversion recovery sequence image with suppressed liquid signal. The high gray-scale abnormal components are compared with the preset edema threshold. Voxels with high gray-scale abnormal components that are not lower than the preset edema threshold and are spatially adjacent to the outer boundary of the three-dimensional binary mask are marked as peritumoral edema areas.

4. The intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels according to claim 3, characterized in that, Methods for constructing a membrane around a tumor by expanding outwards and inwards from the tumor body include: Based on a multi-label 3D mask, the cystic necrosis area within the tumor and the enhanced area of ​​the tumor parenchyma are merged to obtain a tumor body mask. 3D connected component analysis is performed on the tumor body mask to retain the connected component with the largest number of voxels. An outward expansion operation is performed on the tumor body mask, iteratively expanding the boundary at a preset number of expansion steps per voxel to obtain an outward expansion mask. An inward expansion operation is performed on the tumor body mask, iteratively shrinking the boundary at a preset number of inward expansion steps per voxel to obtain an inward expansion mask. The difference between the outward expansion mask and the inward expansion mask is used to generate a mask of the surrounding structure zone.

5. The intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels according to claim 4, characterized in that, Methods for extracting clinical radiological features, quantitative tumor indicators, and spatial radiomics features based on three-dimensional binary masks and masks of surrounding structures include: A three-dimensional binary mask, along with masks for the intratumoral cystic necrosis area, peritumoral edema area, tumor parenchymal enhancement area, and surrounding structural bands, was obtained and mapped to the corresponding voxel coordinates of the preprocessed image. Clinical radiological features were extracted based on the three-dimensional binary mask and the intratumoral cystic necrosis area, peritumoral edema area, and tumor parenchymal enhancement area. These features included tumor location zoning, tumor boundary clarity grading, tumor morphology classification, peritumoral edema grading, intratumoral cystic necrosis grading, and tumor parenchymal enhancement pattern grading. Quantitative tumor analysis indicators were extracted based on the three-dimensional binary mask, including tumor volume, maximum tumor diameter, tumor surface area, tumor compactness, tumor aspect ratio, and the volume percentages of the intratumoral cystic necrosis area, peritumoral edema area, and tumor parenchymal enhancement area. Spatial radiomics features were extracted based on the three-dimensional binary mask and the surrounding structural band mask, including first-order grayscale statistical features, shape features, and texture-related features.

6. The intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels according to claim 5, characterized in that, Methods for merging features to obtain a joint feature set include: Clinical radiological features, quantitative tumor analysis indicators, and spatial radiomics features are written into the same feature container according to the subject identifier, and each feature is assigned a unique feature name and feature source label. Missing value processing and scaling are performed on the feature container. Missing value processing includes comparing the missing value ratio with a preset missing value ratio threshold. Features with a missing value ratio higher than the preset missing value ratio threshold are removed from the feature container, and features with a missing value ratio not higher than the preset missing value ratio threshold are filled with the median. Scale scaling includes normalizing the mean and standard deviation of continuous numerical features and performing ordered encoding on graded features. The clinical radiological features, quantitative tumor analysis indicators, and spatial radiomics features that have undergone missing value processing and scaling are merged according to a preset splicing order to output a joint feature set.

7. The intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels according to claim 6, characterized in that, Methods for sequentially performing correlation analysis, minimum absolute shrinkage, and selection operator regression on the joint feature set to screen core fusion features include: Based on feature type, the joint feature set is divided into count-type feature subset, measurement-type feature subset, and spatial image omics feature subset. Standard regularization preprocessing is performed on the joint feature set to map each feature value to a uniform scale. Correlation analysis is then performed on the preprocessed joint feature set to obtain the correlation statistic, correlation coefficient, and probability value for each feature. The probability value is compared with a significance threshold, and the correlation coefficient is compared with an intensity threshold. Irrelevant features are removed to obtain the retained feature set. Minimum absolute shrinkage and selection operator regression are performed on the retained feature set to output a regression coefficient vector. The regression coefficients are compared with a zero threshold to remove features with zero regression coefficients, thus obtaining the core fused features.

8. The intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels according to claim 7, characterized in that, Methods for eliminating irrelevant features by comparing correlation statistics with significance thresholds and correlation coefficients with strength thresholds include: For count-type feature subsets, chi-square test statistics and correlation strength coefficients are calculated; for econometric feature subsets, rank correlation coefficients are calculated; and for spatial image omics feature subsets, linear correlation coefficients are calculated. Correlation statistics and correlation coefficients are obtained, and probability values ​​associated with the correlation statistics are read. These probability values ​​are compared with a significance threshold; features with probability values ​​not lower than the significance threshold are marked as irrelevant candidate features. The absolute values ​​of correlation coefficients are compared with a strength threshold; features with absolute values ​​of correlation coefficients not higher than the strength threshold are marked as weakly correlated candidate features. Irrelevant and weakly correlated candidate features are merged to form a removal set. The removal set is then deleted from the joint feature set to obtain the correlation-selected retention set.

9. The intelligent prediction method for the relationship between skull base meningiomas and adjacent blood vessels according to claim 8, characterized in that, Methods for comparing regression coefficients to a zero threshold to eliminate features with a coefficient of zero include: Obtain the correlation screening and retention set and construct the regression input matrix. The regression input matrix is ​​aligned with the target task label according to the subject identifier. Perform regression fitting with a penalty term on the regression input matrix using minimum absolute shrinkage and selection operator regression. Traverse the candidate set of regularization parameters under the preset fold cross division, select the regularization parameter value with the smallest corresponding loss value and output the regression coefficient vector. Compare each regression coefficient in the regression coefficient vector with the zero threshold. Features whose regression coefficients fall into the zero threshold are marked as discarded features, and features whose regression coefficients deviate from the zero threshold are marked as retained features. Output the core fusion feature composed of the retained features.

10. An intelligent prediction system for the relationship between a skull base meningioma and adjacent blood vessels, used to execute the intelligent prediction method for the relationship between a skull base meningioma and adjacent blood vessels as described in any one of claims 1-9, characterized in that the system... include: The image preprocessing module is used to acquire multiple magnetic resonance imaging sequences of the same subject and perform trilinear interpolation resampling, bias field correction, rigid registration and contrast normalization on the multiple sequence images to obtain spatially aligned preprocessed images. The mask generation and annotation module generates a three-dimensional binary mask based on the preprocessed image using threshold segmentation combined with semi-automatic delineation. The module then annotates the intratumoral cystic necrosis area, peritumoral edema area, and tumor parenchymal enhancement area within the three-dimensional binary mask. The tumor body mask expands outward and inward to construct a mask for the surrounding structures of the tumor. The extraction and fusion module extracts clinical radiological features, tumor quantitative analysis indicators, and spatial radiomics features based on a three-dimensional binary mask and a mask of the structure zone surrounding the tumor, and merges them to obtain a joint feature set. The feature selection module sequentially performs correlation analysis, minimum absolute contraction, and selection operator regression on the joint feature set to select core fusion features. Specifically, it compares the correlation statistic with the significance threshold, compares the correlation coefficient with the intensity threshold to eliminate irrelevant features, and compares the regression coefficient with the zero threshold to eliminate features with a coefficient of zero. The prediction and interpretation module constructs a three-class classification training set for vascular relationships based on core fusion features, performs minority class oversampling and class weight correction on the category proportion, and then inputs it into the preset classification model to obtain the predicted label of the relationship between skull base meningioma and adjacent blood vessels, and interprets the feature contribution of the predicted label output.

Images