Method and system for evaluating cervical cancer image heterogeneity based on spatial autocorrelation

Abstract

The present application relates to the technical field of medical image processing, spatial statistical analysis and artificial intelligence assisted risk assessment, and particularly relates to a cervical cancer image heterogeneity evaluation method and system based on spatial autocorrelation. The method obtains pelvic magnetic resonance images of cervical cancer patients before treatment, and through image preprocessing, cross-sequence registration, tumor region of interest acquisition, spatial autocorrelation parameter calculation, three-dimensional image topology subregion division and tumor habitat heterogeneity score calculation, generates complete remission probability, concurrent chemoradiotherapy failure risk score and risk stratification results, and can further output progression-free survival risk stratification information when having follow-up data, which is used for computer-aided risk assessment before concurrent chemoradiotherapy and multidisciplinary consultation reference.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, spatial statistical analysis, artificial intelligence-assisted risk assessment, and clinical decision support. Specifically, it relates to a method and system for assessing the heterogeneity of cervical cancer images based on spatial autocorrelation, and more particularly to a computer-aided risk assessment method and system for cervical cancer before concurrent chemoradiotherapy based on spatial autocorrelation parameters such as pre-treatment magnetic resonance imaging and local Moran's index. Background Technology

[0002] CCRT is an important standard treatment for LAC, but treatment response and long-term prognosis vary significantly among patients. Even patients at the same International Federation of Gynecology and Obstetrics (FIGO) stage may experience different treatment outcomes due to differences in tumor cell density, necrosis, hypoxia, blood supply, and radiosensitivity. Traditional clinicopathological variables can reflect some of the disease burden, but they are insufficient to fully characterize the spatial heterogeneity within the tumor.

[0003] Pre-treatment MRI can reveal information such as morphology, signal intensity, enhancement, and diffusion restriction of cervical tumors. Existing radiomics or deep learning methods often rely on average features of the entire tumor, texture features of the entire tumor, or unconstrained clustering, which tends to weaken the spatial relationships between voxels and their neighbors. While some image habitat methods can divide subregions, they typically rely mainly on intensity or clustering results, failing to adequately utilize local spatial autocorrelation.

[0004] The local Moran index describes the spatial autocorrelation between a voxel and its neighboring voxels, and is used to identify high-value clusters, low-value clusters, and transitional regions with alternating high and low values. Introducing it into pre-treatment MRI can generate three-dimensional image topological subregions (also known as image habitats) with spatial topological meaning within the region of interest (ROI) of the tumor, and further calculate the Hidden Hierarchical Structure (HHS) to quantify the degree of fragmentation and spatial heterogeneity within the tumor.

[0005] Therefore, it is necessary to propose a data processing technique based on conventional pre-treatment MRI that can output interpretable, verifiable, and scalable cervical cancer CCRT risk assessment information without requiring positron emission tomography (PET), immunohistochemistry (IHC), or omics testing as necessary inputs. Summary of the Invention

[0006] The core problem this invention aims to solve is: how to utilize routine pelvic MRI before treatment in patients with LAC (Laboratory-Invasive Collateral Catheteritis) to quantify the spatial heterogeneity within the tumor using a non-invasive and repeatable data processing method before the onset of CCRT (Coronary Tumor Therapy), and based on this, generate CR probability, CCRT failure risk scores, and PFS risk stratification information. Specifically, this includes:

[0007] 1. How to obtain the entire tumor region of interest (ROI) in pre-treatment MRI and ensure that multiple sequence images correspond in the same spatial coordinate system.

[0008] 2. How to define neighborhood topology and spatial weights in three-dimensional voxel space to make local Moran index calculation feasible and repeatable.

[0009] 3. How to divide the topological subregions of three-dimensional images into high-high, low-low, high-low, and low-high based on local spatial autocorrelation.

[0010] 4. How to convert the fragmentation degree of 3D image topological subregions into quantifiable HHS.

[0011] 5. How to integrate HHS with optional clinical variables to output interpretable and calibrable CR probability and CCRT failure risk scores.

[0012] 6. How to ensure the system's clinical accessibility and biological rationality without relying on PET, IHC, or omics testing as necessary inputs.

[0013] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0014] A method for assessing cervical cancer imaging heterogeneity based on spatial autocorrelation includes:

[0015] Input data acquisition: Acquire pelvic magnetic resonance imaging data of cervical cancer patients before concurrent chemoradiotherapy;

[0016] Image preprocessing: The pelvic magnetic resonance imaging data is de-identified, spatially aligned, biased field corrected, intensity standardized, and resampled. When multiple sequences of images are included, cross-sequence registration is performed to obtain standardized image data.

[0017] Tumor region of interest acquisition: The tumor region of interest is acquired from the standardized image data;

[0018] Spatial autocorrelation parameter calculation: Within the tumor region of interest, spatial autocorrelation parameters are calculated based on the three-dimensional neighborhood topology and spatial weights of voxels, and a spatial autocorrelation parameter map is generated;

[0019] Three-dimensional image topological subregion division: Based on the spatial autocorrelation parameters, voxel self-standardized image feature values, and neighborhood weighted feature values, the tumor region of interest is divided into different three-dimensional image topological subregions;

[0020] Heterogeneity score calculation: The tumor habitat heterogeneity score is calculated based on the number of connected domains, the volume of the largest connected domain, the proportion of the main connected domain, and / or the proportion of the sub-region volume of each 3D image topological sub-region.

[0021] Risk assessment output: Based on the tumor habitat heterogeneity score, generate the probability of complete remission, the risk score of concurrent chemoradiotherapy failure, and the risk stratification results;

[0022] Quality control and report generation: Perform quality control on image quality, registration quality, tumor region of interest integrity, missing inputs, and model output anomalies, and generate editable evaluation reports.

[0023] Furthermore, the pelvic magnetic resonance imaging data includes one or more of T2-weighted imaging, enhanced T1-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient maps;

[0024] The spatial autocorrelation parameters include the local Moran index, the local Geary C statistic, the Getis-Ord Gi* hotspot statistic, or equivalent statistics that can characterize the local spatial correlation between the target voxel and its neighboring voxels.

[0025] Preferably, the spatial autocorrelation parameter is the local Moran index.

[0026] The spatial weights are binary weights, row-normalized weights, or distance decay weights, used to represent the correlation strength between the target voxel and its neighboring voxels, and the spatial weight matrix is ​​composed of the spatial weights between each voxel pair.

[0027] Furthermore, the three-dimensional neighborhood topology is determined using one of the following methods:

[0028] Method 1: Determining using a three-dimensional 6-neighborhood;

[0029] Method 2: Use a three-dimensional 18-neighborhood for determination;

[0030] Method 3: Use a three-dimensional 26-neighborhood for determination;

[0031] Method 4: Determine the distance according to the preset Euclidean distance threshold, Manhattan distance threshold, or Chebyshev distance threshold.

[0032] Furthermore, when the spatial autocorrelation parameter is the local Moran index, based on the directional relationship of the local Moran index, the three-dimensional image topological sub-regions include high-high clustering regions, low-low clustering regions, high-low transition regions, and low-high transition regions;

[0033] High-high clustering areas indicate that voxels with high image feature values ​​are adjacent to voxels with high image feature values; low-low clustering areas indicate that voxels with low image feature values ​​are adjacent to voxels with low image feature values; high-low transition areas indicate that the voxel itself has high image feature values ​​while the neighborhood weighted image feature values ​​are low; low-high transition areas indicate that the voxel itself has low image feature values ​​while the neighborhood weighted image feature values ​​are high.

[0034] Furthermore, when dividing the three-dimensional image topological subregions, voxels that reach the preset statistical significance threshold or the preset confidence threshold are assigned to the corresponding three-dimensional image topological subregion; voxels that do not reach the preset statistical significance threshold or the preset confidence threshold are retained as non-significant regions, or assigned to adjacent three-dimensional image topological subregions according to spatial connectivity.

[0035] Furthermore, the tumor habitat heterogeneity score is calculated based on the number of connected domains, the volume of the largest connected domain, the proportion of the main connected domain, and / or the proportion of the sub-region volume of each three-dimensional image topological sub-region, and normalized to a preset numerical range; the more connected domains there are and / or the lower the proportion of the largest connected domain, the higher the degree of spatial fragmentation, and the higher the tumor habitat heterogeneity score.

[0036] Furthermore, the risk assessment output also includes fusing the tumor habitat heterogeneity score with one or more clinicopathological variables selected from FIGO stage, lymph node status, maximum tumor diameter, squamous cell carcinoma antigen level, and pathological type, and inputting them into the prediction model to generate the complete remission probability and concurrent chemoradiotherapy failure risk score.

[0037] Furthermore, the predictive models used to generate risk assessment outputs employ logistic regression models, nomogram models, random forest models, gradient boosting tree models, support vector machine models, or calibrated neural network models.

[0038] Furthermore, the cervical cancer imaging heterogeneity assessment method based on spatial autocorrelation also includes a prognostic risk stratification step: when follow-up data is available, based on at least one of the complete remission probability, the concurrent chemoradiotherapy failure risk score, the total score of the nomogram, or the comprehensive risk score output by the prediction model, and according to the risk threshold determined by the training set, progression-free survival risk stratification information is generated and output.

[0039] An assessment system for evaluating cervical cancer imaging heterogeneity based on spatial autocorrelation includes:

[0040] Data receiving module: used to receive and verify pelvic magnetic resonance imaging data and optional clinicopathological variables of cervical cancer patients before treatment;

[0041] Image preprocessing module: used to perform de-identification, spatial orientation unification, bias field correction, intensity normalization, and resampling on the pelvic magnetic resonance imaging data, and to perform cross-sequence registration when multiple sequence images are included;

[0042] Tumor Region of Interest Acquisition Module: Used to acquire tumor regions of interest, supporting manual delineation by physicians, automatic segmentation, and manual review;

[0043] Spatial feature calculation module: used to calculate spatial autocorrelation parameters and generate a spatial autocorrelation parameter map based on the three-dimensional neighborhood topology and spatial weights of voxels within the tumor region of interest;

[0044] Three-dimensional image topology sub-region division module: used to divide the three-dimensional image topology sub-regions based on the spatial autocorrelation parameters, voxel self-standardized image feature values ​​and neighborhood weighted feature values;

[0045] Heterogeneity scoring module: used to calculate tumor habitat heterogeneity score based on the number of connected domains, the largest connected domain volume, the proportion of the main connected domain and / or the proportion of the sub-region volume of each 3D image topological sub-region;

[0046] Risk assessment module: used to generate complete remission probability, concurrent chemoradiotherapy failure risk score, risk stratification results and optional progression-free survival risk stratification information based on the tumor habitat heterogeneity score or by further integrating clinicopathological variables.

[0047] The quality control and reporting module is used to check image quality, cross-sequence registration error, tumor region of interest integrity, missing input variables, abnormal subregion volume, and abnormal model output, and to issue a prompt for review when abnormalities are found. The quality control and reporting module is also used to generate an editable evaluation report, which includes the source of the input image, the source of the tumor region of interest, the results of the three-dimensional image topological subregion division, the tumor habitat heterogeneity score, the probability of complete remission, the risk score of concurrent chemoradiotherapy failure, the risk level, the quality control status, and a statement for non-diagnostic purposes.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. Enables the generation of non-invasive pre-treatment risk assessment information. The system uses routine pre-treatment MRI as the core input and does not require PET, IHC, or omics testing as necessary inputs.

[0050] 2. Enhance the ability to characterize tumor spatial heterogeneity. Spatial autocorrelation parameters such as the local Moran index explicitly utilize the spatial correlation between voxels and neighboring voxels to identify high-high clustering areas, low-low clustering areas, high-low transition areas, and low-high transition areas. This allows for a more comprehensive characterization and quantification of the spatial heterogeneity and topological fragmentation characteristics within the tumor compared to the average features of the entire tumor, the texture features of the entire tumor, or unconstrained clustering.

[0051] 3. Provides quantifiable HHS. HHS incorporates the number of connected components in the topological subregions of 3D imagery, the maximum connected component volume, and the volume weight into a unified calculation.

[0052] 4. Improve the performance of CR risk assessment. In confirmatory studies, the model combining HHS with FIGO staging and lymph node status had AUCs of 0.83, 0.81 and 0.80 on three external validation sets, which were superior to the traditional clinical variable model.

[0053] 5. It has prognostic stratification value. The risk score or the total score of the nomogram can be used to generate PFS risk stratification information.

[0054] 6. Enhanced clinical interpretability. The system can simultaneously output three-dimensional image topology sub-region maps, HHS (Heat, Health, and Safety) and risk levels, facilitating MDT (Multidisciplinary Team) review of risk sources.

[0055] 7. Facilitates multi-center deployment. Image preprocessing, resampling, intensity normalization, and registration steps help reduce the impact of differences in scanner, field strength, and acquisition protocol on the results. Attached Figure Description

[0056] The invention will now be further described with reference to the accompanying drawings:

[0057] Figure 1 This is a flowchart of the spatial autocorrelation-based cervical cancer image heterogeneity assessment method described in this invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this invention. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0059] This invention is primarily used for computer-aided risk assessment of patients with locally advanced cervical cancer (LACC) before concurrent chemoradiotherapy (CCRT). The system outputs a Habitat Heterogeneity Score (HHS), Complete Response (CR) probability, CCRT failure risk score, and Progression-Free Survival (PFS) risk stratification information. These results are used for data discussion and risk stratification reference within a multidisciplinary team (MDT) and do not directly replace physician diagnostic or treatment decisions.

[0060] Commonly used English terms appearing for the first time in the main text should be described in the format of "Chinese name (English full name, abbreviation)"; subsequent abbreviations can be used directly.

[0061]

[0062] I. Overall Concept

[0063] This invention proposes a method for assessing cervical cancer risk based on pre-treatment MRI spatial heterogeneity characterization. For example... Figure 1 As shown, the system takes CCRT anterior pelvic MRI as the core input, and after image preprocessing, cross-sequence registration, tumor region of interest acquisition, spatial autocorrelation parameter calculation, three-dimensional image topological subregion division, HHS calculation, prediction model evaluation and risk stratification output, it generates CR probability and CCRT failure risk score, and can further generate PFS risk stratification information when follow-up data is available.

[0064] The core process of this invention is as follows: acquiring MRI image data, identifying the region of interest (ROI) of the tumor, calculating spatial autocorrelation parameters within this region based on three-dimensional spatial topological relationships, preferably calculating the local Moran's index, dividing the three-dimensional image topological subregions according to the local spatial autocorrelation relationships, calculating the hematocrit (HHS) based on the degree of fragmentation of the connected domains in each subregion, and generating risk assessment information accordingly. The MRI sequence, clinical variables, spatial autocorrelation parameters, and prediction model type can all be adjusted according to specific implementation conditions without affecting the aforementioned core technical concept.

[0065] II. The steps of the method for assessing the heterogeneity of cervical cancer images based on spatial autocorrelation are shown in Table 1.

[0066] Table 1

[0067]

[0068] III. Spatial autocorrelation parameters, 3D image topological subregions and HHS;

[0069] In one implementation, after resampling the MRI to a unified voxel grid, the voxel positions within the tumor region of interest are represented in three-dimensional coordinates. A three-dimensional neighborhood topology is used to determine the set of neighboring voxels for the target voxel; this can be a 6-neighborhood, 18-neighborhood, or 26-neighborhood, or it can be determined based on a preset Euclidean distance threshold, Manhattan distance threshold, or Chebyshev distance threshold. Spatial weights are used to characterize the correlation strength between the target voxel and its neighboring voxels; these can be binary weights, row-normalized weights, or distance-attenuated weights. A spatial weight matrix is ​​formed from the spatial weights between each voxel pair. The aforementioned neighborhood topology, weight form, and weight matrix construction method can be equivalently replaced based on image resolution, voxel spacing, and model calibration results.

[0070] In a preferred embodiment, the spatial autocorrelation parameter is the local Moran index. Let the standardized image feature value of the i-th voxel be x_i, the mean within the ROI be μ, the variance be m2, and w_ij be the spatial weight between the i-th voxel and its neighboring voxel j. Then, the local Moran index I_i of the i-th voxel can be summarized as: I_i = ((x_i - μ) / m2) × Σ_j w_ij(x_j - μ). Where I_i represents the local Moran index of the i-th voxel, Σ_j represents the summation of voxels j within the neighborhood of the i-th voxel, and x_j represents the standardized image feature value of the neighboring voxel j. This calculation is based on the spatial topology of the three-dimensional tumor ROI and performs spatial weighting processing on the voxel image signal and its neighboring image signals. In other equivalent embodiments, the spatial autocorrelation parameter can also be replaced by the local Geary C statistic, the Getis-Ord Gi* hotspot statistic, or other statistics that can characterize the local spatial correlation between the target voxel and its neighboring voxels.

[0071] When the spatial autocorrelation parameter is the local Moran's index, four types of 3D image topological subregions can be formed based on the directional relationship between the voxel's own standardized image feature values, neighborhood weighted feature values, and the local Moran's index: high-high clustering region, low-low clustering region, high-low transition region, and low-high transition region. Voxels that reach the preset statistical significance threshold or the preset confidence threshold are preferentially assigned to the corresponding subregion; voxels that do not reach the preset statistical significance threshold or the preset confidence threshold can be retained as non-significant regions, or assigned to adjacent subregions according to spatial connectivity.

[0072] HHS (Heterochromatic Hierarchy Structure) is used to quantify the fragmentation degree of topological subregions in 3D imaging. Its core logic is: given a given total volume of a subregion, the more connected components there are, and / or the lower the proportion of the largest connected component to the total volume of that subregion, the higher the degree of spatial fragmentation, and the higher the HHS. HHS can be calculated comprehensively based on the number of connected components, the volume of the largest connected component, the proportion of the main connected component, and / or the proportion of the subregion volume, and normalized to a preset numerical range. Higher values ​​indicate stronger spatial heterogeneity within the tumor, typically corresponding to a lower CR (complete resection) probability and a higher risk of CCRT (tumor-associated tumor resection) failure.

[0073] IV. Prediction Model and Output Results;

[0074] In one implementation, the complete response (CR) determined 3 to 6 months after CCRT according to the Response Evaluation Criteria in Solid Tumors version 1.1 (RECIST 1.1) is used as the training label. The model may use only HHS, or it may further incorporate conventional clinical variables such as FIGO stage, lymph node status, maximum tumor diameter, squamous cell carcinoma antigen, and pathological type.

[0075] Predictive models can employ logistic regression, nomograms, random forests, gradient boosting trees, support vector machines, or calibrated neural networks. A preferred implementation involves inputting HHS and clinical variables into a multivariate logistic regression model, outputting the CR probability; the CCRT failure risk score can be obtained by subtracting the CR probability from 1, or the model can directly output the failure probability.

[0076] The system output should include at least HHS, CR probability, CCRT failure risk score, low / medium / high risk level, contribution of key variables, and quality control tips. When follow-up validation data is available, PFS risk stratification information can be further output.

[0077] V. The module composition of the cervical cancer image heterogeneity assessment system based on spatial autocorrelation is shown in Table 2;

[0078] Table 2

[0079]

[0080] VI. System Output Interface and Report Presentation;

[0081] The system can be configured with an image display area, a tumor region of interest display area, a 3D image topological subregion display area, a HHS and risk score display area, a quality control prompt area, and a report generation area. The interface layout is not the focus of this invention; its technical function is to visualize the output results of the risk assessment module, HHS calculation module, and quality control module.

[0082] The report generation area can generate editable data analysis paragraphs, including the source of the input image, the source of the tumor region of interest, the results of the three-dimensional image topological subregion division, HHS, CR probability, CCRT failure risk score, PFS risk stratification information, quality control status, and a statement for non-diagnostic purposes. VII. Specific Implementation Examples;

[0084] Example 1: MRI-based spatial heterogeneity risk assessment;

[0085] Pre-treatment pelvic MRI of LAC patients was acquired. After preprocessing and acquisition of the entire tumor region of interest, spatial autocorrelation parameters were calculated within this region. The local Moran's index was preferentially calculated, and a three-dimensional image topology sub-region map was generated. Subsequently, the HHS was calculated and input into a trained and calibrated risk assessment model, which outputs the CR probability and CCRT failure risk score.

[0086] Example 2: Risk stratification of HHS combined with clinical variables;

[0087] The HHS, FIGO stage, and lymph node status are input into a multivariate logistic regression model to obtain the CR probability and CCRT failure risk score. The system can determine the cutoff value based on the training concentrated index, preset sensitivity or specificity requirements, and classify patients into low-risk, intermediate-risk, and high-risk groups, or at least into low-risk and high-risk groups.

[0088] Example 3: Quality control, follow-up validation, and biological explanation;

[0089] Before outputting risk stratification, the system can check for MRI artifacts, cross-sequence registration quality, abnormal tumor region of interest volume, excessively small subregion volume, and missing key inputs. If significant abnormalities are found, the system prompts for manual review. When long-term follow-up data is available, risk scores can be used for PFS stratification; when transcriptome data is available, Gene Set Enrichment Analysis (GSEA) can be used to explore pathway differences in hypoxia, DNA replication, and extracellular matrix. Follow-up and omics data are only used for validation or interpretation and are not required inputs for clinical deployment.

[0090] Example 4: Multicenter experimental verification;

[0091] In a set of confirmatory studies, the CR outcome 3 to 6 months after CCRT was used as the primary assessment label. HHS was included as an independent factor associated with CR in the multivariate model; the nomogram model constructed from HHS, FIGO stage, and lymph node status achieved area under the receiver operating characteristic curve (AUC) of 0.83, 0.81, and 0.80 in three external validation sets, respectively, which was superior to the control model containing only traditional clinical variables. The above results are used to illustrate the feasibility of the technical approach of this invention and do not limit the scope of protection under different datasets, different thresholds, or different model forms.

[0092] In confirmatory studies, when HHS was considered an independent relevant factor, the odds ratio (OR) was 0.18, and the 95% confidence interval (CI) ranged from 0.08 to 0.39. Since the model uses CR as the positive outcome, an OR less than 1 indicates that a higher HHS level corresponds to a lower probability of CR and a higher risk of CCRT failure. Low nomogram scores were associated with poor PFS, with a hazard ratio (HR) of 3.85 and a 95% CI ranging from 1.82 to 8.15. These values ​​are for illustrative purposes only and do not limit the scope of protection.

[0093] VIII. Boundary Description

[0094] 1. This invention is positioned as a computer-aided image data processing and risk assessment information generation tool based on pre-treatment MRI. The output results are used for risk stratification before concurrent chemoradiotherapy, multidisciplinary consultation and discussion, and data review and reference. It does not replace doctors in making diagnoses, staging, treatment plan selection or treatment instructions.

[0095] 2. The core necessary input of this invention is pre-treatment MRI image data; the clinical variable is preferred fusion information, and PET, IHC, transcriptome or other omics data are only used as optional research verification materials.

[0096] 3. The CR probability, CCRT failure risk score, and PFS risk stratification information should be interpreted comprehensively in conjunction with the patient's clinical manifestations, pathological results, treatment accessibility, and physician judgment, and should not be used as automatic treatment recommendations.

[0097] 4. Different medical institutions can recalibrate the resampling scale, neighborhood weights, statistical thresholds, HHS calculation details, and risk cutoff values ​​according to scanning protocols, data scale, and regulatory requirements.

[0098] 5. The three-dimensional image topological subregion (also known as image habitat) referred to in this invention is a computer-aided image subregion obtained based on the spatial autocorrelation of MRI, and is not equivalent to the fixed histological structure that can be directly observed in pathological sections.

[0099] 6. This invention can be extended to other scenarios of efficacy risk assessment for solid tumors receiving radiotherapy, chemotherapy, or combined radiotherapy and chemotherapy, but this invention is mainly applied to cervical cancer CCRT risk stratification.

[0100] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be included within the scope of protection of the present invention. Furthermore, all content not described in detail in this specification is prior art known to those skilled in the art.

Claims

1. A method for assessing cervical cancer image heterogeneity based on spatial autocorrelation, characterized in that, include: Input data acquisition: Acquire pelvic magnetic resonance imaging data of cervical cancer patients before concurrent chemoradiotherapy; Image preprocessing: The pelvic magnetic resonance imaging data is de-identified, spatially aligned, biased field corrected, intensity standardized, and resampled. When multiple sequences of images are included, cross-sequence registration is performed to obtain standardized image data. Tumor region of interest acquisition: The tumor region of interest is acquired from the standardized image data; Spatial autocorrelation parameter calculation: Within the tumor region of interest, spatial autocorrelation parameters are calculated based on the three-dimensional neighborhood topology and spatial weights of voxels, and a spatial autocorrelation parameter map is generated; Three-dimensional image topological subregion division: Based on the spatial autocorrelation parameters, voxel self-standardized image feature values, and neighborhood weighted feature values, the tumor region of interest is divided into different three-dimensional image topological subregions; Heterogeneity score calculation: The tumor habitat heterogeneity score is calculated based on the number of connected domains, the volume of the largest connected domain, the proportion of the main connected domain, and / or the proportion of the sub-region volume of each 3D image topological sub-region. Risk assessment output: Based on the tumor habitat heterogeneity score, generate the probability of complete remission, the risk score of concurrent chemoradiotherapy failure, and the risk stratification results; Quality control and report generation: Perform quality control on image quality, registration quality, tumor region of interest integrity, missing inputs, and model output anomalies, and generate editable evaluation reports.

2. The method according to claim 1, characterized in that: The pelvic magnetic resonance imaging data includes one or more of the following: T2-weighted imaging, enhanced T1-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient maps. The spatial autocorrelation parameters include the local Moran index, the local Geary C statistic, the Getis-Ord Gi* hotspot statistic, or equivalent statistics that can characterize the local spatial correlation between the target voxel and its neighboring voxels. The spatial weights are binary weights, row-normalized weights, or distance decay weights, used to represent the correlation strength between the target voxel and its neighboring voxels, and the spatial weight matrix is ​​composed of the spatial weights between each voxel pair.

3. The method according to claim 1, characterized in that, The three-dimensional neighborhood topology is determined using one of the following methods: Method 1: Determining using a three-dimensional 6-neighborhood; Method 2: Use a three-dimensional 18-neighborhood for determination; Method 3: Use a three-dimensional 26-neighborhood for determination; Method 4: Determine the distance according to the preset Euclidean distance threshold, Manhattan distance threshold, or Chebyshev distance threshold.

4. The method according to claim 2, characterized in that: When the spatial autocorrelation parameter is the local Moran index, based on the directional relationship of the local Moran index, the three-dimensional image topological subregion includes high-high clustering region, low-low clustering region, high-low transition region and low-high transition region; High-high clustering areas indicate that voxels with high image feature values ​​are adjacent to voxels with high image feature values; low-low clustering areas indicate that voxels with low image feature values ​​are adjacent to voxels with low image feature values; high-low transition areas indicate that the voxel itself has high image feature values ​​while the neighborhood weighted image feature values ​​are low; low-high transition areas indicate that the voxel itself has low image feature values ​​while the neighborhood weighted image feature values ​​are high.

5. The method according to claim 4, characterized in that: When dividing a 3D image topological subregion, voxels that reach a preset statistical significance threshold or a preset confidence threshold are assigned to the corresponding 3D image topological subregion; voxels that do not reach the preset statistical significance threshold or the preset confidence threshold are retained as non-significant regions, or assigned to adjacent 3D image topological subregions based on spatial connectivity.

6. The method according to claim 1, characterized in that: The tumor habitat heterogeneity score is calculated based on the number of connected domains, the volume of the largest connected domain, the proportion of the main connected domain, and / or the proportion of the sub-region volume of each three-dimensional image topological sub-region, and is normalized to a preset numerical range. The more connected domains there are, and / or the lower the percentage of the largest connected domain, the higher the degree of spatial fragmentation, and the higher the tumor habitat heterogeneity score.

7. The method according to claim 1, characterized in that: The risk assessment output also includes fusing the tumor habitat heterogeneity score with one or more clinicopathological variables selected from FIGO stage, lymph node status, maximum tumor diameter, squamous cell carcinoma antigen level, and pathological type, and inputting them into the prediction model to generate the complete remission probability and concurrent chemoradiotherapy failure risk score.

8. The method according to claim 1 or 7, characterized in that: The predictive models used to generate risk assessment outputs include logistic regression models, noctilinear plot models, random forest models, gradient boosting tree models, support vector machine models, or calibrated neural network models.

9. The method according to claim 8, characterized in that, It also includes a prognostic risk stratification step: when follow-up data is available, based on at least one of the complete remission probability, the concurrent chemoradiotherapy failure risk score, the total score of the nomogram, or the comprehensive risk score output by the prediction model, and according to the risk threshold determined by the training set, progression-free survival risk stratification information is generated and output.

10. The assessment system for a spatial autocorrelation-based cervical cancer image heterogeneity assessment method according to claim 1, characterized in that, include: Data receiving module: used to receive and verify pelvic magnetic resonance imaging data and optional clinicopathological variables of cervical cancer patients before treatment; Image preprocessing module: used to perform de-identification, spatial orientation unification, bias field correction, intensity normalization, and resampling on the pelvic magnetic resonance imaging data, and to perform cross-sequence registration when multiple sequence images are included; Tumor Region of Interest Acquisition Module: Used to acquire tumor regions of interest, supporting manual delineation by physicians, automatic segmentation, and manual review; Spatial feature calculation module: used to calculate spatial autocorrelation parameters and generate a spatial autocorrelation parameter map based on the three-dimensional neighborhood topology and spatial weights of voxels within the tumor region of interest; Three-dimensional image topology sub-region division module: used to divide the three-dimensional image topology sub-regions based on the spatial autocorrelation parameters, voxel self-standardized image feature values ​​and neighborhood weighted feature values; Heterogeneity scoring module: used to calculate tumor habitat heterogeneity score based on the number of connected domains, the largest connected domain volume, the proportion of the main connected domain and / or the proportion of the sub-region volume of each 3D image topological sub-region; Risk assessment module: used to generate complete remission probability, concurrent chemoradiotherapy failure risk score, risk stratification results and optional progression-free survival risk stratification information based on the tumor habitat heterogeneity score or by further integrating clinicopathological variables. The quality control and reporting module is used to check image quality, cross-sequence registration error, tumor region of interest integrity, missing input variables, abnormal subregion volume, and abnormal model output, and to issue a prompt for review when abnormalities are found. The quality control and reporting module is also used to generate an editable evaluation report, which includes the source of the input image, the source of the tumor region of interest, the results of the three-dimensional image topological subregion division, the tumor habitat heterogeneity score, the probability of complete remission, the risk score of concurrent chemoradiotherapy failure, the risk level, the quality control status, and a statement for non-diagnostic purposes.

Images