Glioma radiotherapy effect prediction method based on habitat analysis radiomics

Abstract

The invention relates to a glioma radiotherapy effect prediction method based on habitat analysis imageomics, and belongs to the technical field of medical image artificial intelligence and tumor radiotherapy, and the method comprises the steps: collecting a multi-mode MRI image, a radiotherapy plan computed tomography CT image, a dose image and a local recurrence label before postoperative radiotherapy, and converting the images into a standard format; sequentially performing craniocerebral boning processing, N4 bias field correction, isotropic resampling, gray level normalization processing and image registration on the acquired image to obtain a preprocessed image; based on the planned target volume PTV, dividing the interior of the PTV into three types of physiological structure habitat areas and three types of physiological habitat areas; image omics features of the multi-mode MRI images are extracted in the PTV and all habitat areas, and an optimal feature subset is screened out; taking the screened optimal feature subset as an input, taking the local recurrence label as an output label, and selecting an optimal model; and generating an output result by using the optimal model.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging artificial intelligence and tumor radiotherapy, and in particular to a method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics. Background Art

[0002] Glioma is the most common intracranial malignancy in adults. Despite surgical resection and subsequent combined chemoradiotherapy, most patients experience recurrence within the irradiated area. Spatial heterogeneity within tumor tissue is a key factor influencing the efficacy of radiotherapy. Traditional radiotherapy plans generally use the planned target volume (PTV) as a whole for dose coverage, ignoring the histological differences between microregions within the target volume. Furthermore, most existing radiomics studies have only extracted overall PTV characteristics at the macroscopic level, failing to deeply characterize regional diversity.

[0003] Furthermore, many artificial intelligence (AI) models have been questioned in clinical practice due to their "black box" nature. Their lack of clear, interpretable outputs makes it difficult for physicians to make individualized treatment decisions based on model results. Therefore, there is an urgent need for a technical approach that can fully exploit the heterogeneity within target regions while also being interpretable and generalizable. Summary of the Invention

[0004] To address the above technical issues, the present invention provides a method for predicting the efficacy of glioma radiotherapy based on habitat analysis radiomics. By preprocessing and segmenting preoperative images, the PTV (Peripheral Video Volume) is divided into multiple biologically significant subregions. High-dimensional radiomics features are extracted from each habitat. A recurrence risk prediction model is constructed using a machine learning algorithm. This model's decision logic is then output using an interpretable algorithm to achieve personalized radiotherapy efficacy assessment.

[0005] The present invention provides the following technical solutions:

[0006] A method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics, comprising the following steps:

[0007] Data acquisition: Multimodal MRI images before postoperative radiotherapy, radiotherapy planning computed tomography (CT) images, dose images, and local recurrence annotations were collected and converted into a standard format.

[0008] Image preprocessing: The collected MRI images, radiotherapy planning CT images, and dose images are sequentially subjected to brain bone removal, N4 bias field correction, isotropic resampling, grayscale normalization, and image registration to obtain the preprocessed MRI images, radiotherapy planning CT images, and dose images.

[0009] Region of Interest (ROI) definition and habitat division: Based on the planning target volume (PTV), the ROI is defined in the preprocessed image and dose information of the ROI is extracted. Simultaneously, based on the voxel-level grayscale values of the preprocessed MRI image, the planning target volume (PTV) is divided into three types of physiological structural habitat regions and three types of physiological habitat regions, and corresponding habitat masks are generated.

[0010] Feature extraction and selection: Extract radiomic features from multimodal MRI images within the planned target volume (PTV) and each habitat area; select the optimal feature subset based on the radiomic features;

[0011] Model construction and training: The optimal feature subset is selected as input, local recurrence annotation is used as output label, the model performance of the validation set is tested, and the best model is selected;

[0012] Interpretability Analysis: Perform interpretability analysis on the best model, including generating feature importance SHAP summary plots and individual-level SHAP force plots.

[0013] The present invention has the following beneficial effects:

[0014] Introducing MRI habitat analysis into the internal structure division of radiotherapy target areas can identify differences in metabolic activity, cell density, etc. within a local range, thereby improving the accuracy of tumor heterogeneity description; extracting multidimensional imaging genomics features based on habitat sub-regions can improve the information density and biological relevance of feature expression, which is conducive to enhancing model prediction performance; through a multi-stage feature screening mechanism, the dimensionality curse is effectively reduced, and the stability and generalization ability of the model are improved; integrating the SHAP algorithm to output the model interpretability SHAP summary diagram and SHAP force diagram, realizing the transition from "black box" to "white box", making the model prediction logic intuitive and transparent, and facilitating doctors' trust and adoption. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments of the present invention. Obviously, the following drawings are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 A flowchart of the method of the present invention;

[0017] Figure 2 Schematic diagram of habitat division;

[0018] Figure 3 The ROC curve diagram of the model of the present invention in the training set and the validation set;

[0019] Figure 4This is the SHAP summary graph;

[0020] Figure 5 Strive for SHAP. DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other. To achieve the above-mentioned objectives, the present invention adopts the following technical solutions.

[0022] The present invention provides a method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics, such as Figure 1 As shown, the following steps are included:

[0023] Data Acquisition: Multimodal MRI images were collected before postoperative radiotherapy, including enhanced T1 isotropic voxel imaging (e-THRIVE), fluid-attenuated inversion recovery (FLAIR), and apparent diffusion coefficient (ADC) sequences; radiotherapy planning computed tomography (CT) images; dose images; and patient follow-up information, including local recurrence annotations. These images were converted to a standard format and desensitized. Patients with local recurrence annotations underwent follow-up MRI assessment within 12 to 24 weeks after radiotherapy. Efficacy was assessed according to the Response Evaluation Criteria in Neuro-Oncology, version 2.0 (RANO 2.0) criteria, including complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). Patients were considered to have relapsed if the target lesion showed persistent progression on two consecutive MRI scans and required a change in treatment regimen. Otherwise, the patient was considered to have relapsed.

[0024] Image preprocessing: MRI images, radiotherapy planning computed tomography (CT) images, and dose images were skull-stripped using the Oxford University Functional Magnetic Resonance Imaging Center (FSL) software library to obtain skull-stripped images. The skull-stripped images were then subjected to N4 bias field correction using the Advanced Normalization Toolset (ANTs) to obtain bias field corrected images. All images were subjected to 1 mm 3Isotropic resampling was performed to obtain isotropically resampled images. The Nyul histogram normalization method (NyulNormalizer) was used to normalize the grayscale of the radiotherapy planning computed tomography (CT) images and MRI images in the isotropically resampled images. The dose images were registered to the radiotherapy planning CT images, and the radiotherapy planning CT images, fluid-attenuated inversion recovery (FLAIR) sequences, and apparent diffusion coefficient (ADC) maps were co-registered to enhanced T1 isotropic voxel imaging (e-THRIVE).

[0025] Region of interest (ROI) definition: Based on the planning target volume (PTV), the region of interest (ROI) is defined in the pre-processed radiotherapy planning computed tomography (CT) images, MRI images, and dose images. The minimum, mean, and maximum doses of the ROI area are automatically extracted for clinical feature analysis. At the same time, multiple sub-regions are constructed based on the heterogeneity of the tumor and surrounding tissue for feature analysis.

[0026] Based on the voxel-level grayscale values of e-THRIVE, FLAIR, and ADC images, the k-means unsupervised clustering algorithm was applied to classify each voxel within the planning target volume (PTV) into three types of physiological structural habitat regions. The k-means unsupervised clustering algorithm takes as input a multidimensional vector composed of the voxel-level grayscale values of the e-THRIVE, FLAIR, and ADC images. It first automatically divides the region into K habitat regions with similar characteristics by minimizing the squared error between the voxel features and the cluster center. Voxel assignment is achieved using the following formula:

[0027] ;

[0028] in, : The multidimensional feature vector of the i-th voxel (such as the gray value combination of different MRI images), : The feature vector of the j-th cluster center (i.e., the mean of the voxel features in this class), : The cluster category index to which the i-th voxel belongs, : represents the Euclidean distance (L2 norm), j∈{1,2,...,K}: cluster number set, K is the preset number of clusters, and the feature center of each class is updated according to the following formula, iterating until convergence:

[0029] ;

[0030] : The set of voxels currently belonging to cluster j, : The number of voxels in cluster j.

[0031] The three types of physiological structure habitat areas include:

[0032] The first type of structural habitat area (center point: 121,76): enhanced tissue habitat, characterized by high e-THRIVE value and low FLAIR value;

[0033] The second type of structural habitat area (center points: 36, 48): solid low-enhancement habitat, characterized by low FLAIR and e-THRIVE values;

[0034] The third type of structural habitat area (center point: 56,129): inactive tissue habitat, characterized by high FLAIR value and low e-THRIVE value;

[0035] And formed three types of physiological habitat areas:

[0036] The first type of physiological habitat area (center point: 73,128): tumor area with high blood supply and medium cell density, high e-THRIVE value and medium ADC value;

[0037] The second type of physiological habitat area (center points: 47, 52): tumor area with low blood supply and high cell density, low e-THRIVE value and low ADC value;

[0038] The third type of physiological habitat area (center point: 96,48): low blood supply, low cell density tumor area, low e-THRIVE value, high ADC value. And generate the corresponding habitat mask map. High and low relative to the center point, see Figure 2 .

[0039] Feature extraction: The pyradiomics toolkit was used to extract 2153 radiomic features from multimodal MRI images within the planning target volume (PTV) and each habitat subregion. These features included shape features, first-order features, texture features (such as the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and gray-level size-zone matrix (GLSZM), as well as higher-order features after wavelet or LoG filtering.

[0040] Feature selection: Using a multi-stage feature dimensionality reduction method:

[0041] (1) Normalization: Normalize the radiomic features extracted from different modalities of habitat-based images (e.g., e-THRIVE, FLAIR, ADC) to eliminate the effects of dimensional differences and inconsistent numerical ranges, thereby improving the stability and accuracy of subsequent analyses. Z-score standardizes the normalized radiomic features of each modality's MRI images. (2) Balance the number of positive and negative samples in the training set. Sampling Synthetic Minority Oversampling Technique (SMOTE) balances the number of recurrent and non-recurrent patients in the training set. (3) Univariate analysis: Perform a separate statistical test on each candidate radiomic feature to screen out features that are significantly associated with efficacy prediction (recurrence, non-recurrence), reduce the dimension of the feature space, and reduce redundancy. Use the t-test or Mann-Whitney U test to screen features with p < 0.1. t is the test statistic, reflecting the size of the difference between groups, and p is the significance level. The screening condition is usually p < 0.1, which indicates that the feature is significantly correlated with the outcome variable. t-test (Student's t-test): used for continuous variables that conform to normal distribution and have homogeneous variance, to test whether the difference between the means of the two groups is significant. Mann-Whitney U test: non-parametric test, used for continuous or ordinal variables that do not conform to normal distribution, to compare the difference in medians between the two groups. (4) Further screening of the multimodal MRI imaging features after normalization and univariate analysis screening, and optimization of the input feature set. Including: combining the recursive feature elimination (RFE) algorithm, through iterative training of classification models (such as support vector machine SVM and random forest RF), gradually eliminating the features with the smallest contribution from the features screened in steps (1) (2) (3), and finally obtaining the optimal feature subset. (5) Perform correlation test on the features screened in steps (1) (2) (3) (4) to avoid redundant information caused by high correlation between features and improve the generalization ability and interpretability of the model. The correlation test is specifically: calculating the correlation coefficient (Pearson correlation coefficient) between the two features to measure the degree of linear correlation. The formula is:

[0042] ;

[0043] in, is the covariance, , is the standard deviation. The feature pairs with correlation coefficient less than 0.9 are retained, i.e. , to ensure complementarity between the selected features and reduce multicollinearity.

[0044] Model training module: A binary classification machine learning model (such as support vector machines (SVMs) and random forests (RFs)) was constructed, using the input as feature pairs and the patient's recurrence or non-recurrence status as the output label. The training set was used for fitting, and the validation set was used to evaluate model performance (such as area under the curve (AUC), sensitivity, specificity, and accuracy). A total of 56 models were constructed (7 ROIs (PTVs, 3 structural habitats, and 3 physiological habitats) × 4 MRI models for extracting image features (e-THRIVE, FLAIR, ADC, and e-THRIVE+FLAIR+ADC) × 2 machine learning algorithms (SVM and RF)). The model with the best performance was selected for subsequent interpretability analysis.

[0045] Interpretability analysis module: Shapley additive interpretation method (SHAP) is used to explain the output of the best performing model, including: (1) feature importance SHAP summary plot, which reveals the core features that dominate the prediction; (2) individual level SHAP force plot, which shows the prediction composition of the model for a single case;

[0046] Output module: Integrates predicted probability, individual recurrence risk classification (high / low risk), key eigenvalues, SHAP summary graph, and SHAP force graph to generate a clinician-readable radiotherapy efficacy decision-making support report.

[0047] Specific embodiments are provided below:

[0048] 1. Research Subjects and Data Sources

[0049] A total of 145 patients with gliomas confirmed by histopathology were included in this study. Inclusion criteria included: (1) prior brain radiotherapy; (2) prior brain surgical resection; and (3) image quality meeting the requirements for radiomics analysis. Exclusion criteria included: (1) concurrent other brain diseases; (2) incomplete radiotherapy course; and (3) inability to assess radiotherapy response.

[0050] Patients were divided into two cohorts based on the time of diagnosis. The training cohort (Cohort 1) consisted of 102 patients diagnosed between January 2018 and December 2021, of whom 42 experienced tumor recurrence and 60 remained relapse-free. The validation cohort (Cohort 2) consisted of 43 patients diagnosed between January 2022 and May 2024, of whom 16 experienced recurrence and 27 remained relapse-free.

[0051] 2. Radiotherapy Plan and Target Volume Delineation

[0052] All patients received 6MV X-ray volumetric modulated rotational radiotherapy (VMAT) delivered by an Elekta Axesse (4D) medical accelerator. Target volumes were defined as follows: GTV (Gross Tumor Volume): visible glioma solid tumor; GTVtb (Gross Tumor Volume tumor bed): the area of the visible tumor bed after surgery; CTV (Clinical Target Volume): the area defined by a 1.0-2.5 cm extension of the GTV; and PTV (Planning Target Volume): the area defined by a 3 mm extension of the CTV. The prescribed dose ranged from 40 to 70 Gy, with target volumes manually delineated by experienced radiation oncologists according to Radiation Therapy Oncology Group (RTOG) criteria. For WHO grade III-IV patients and some WHO grade II patients, oral temozolomide (75 mg / m² / day) was administered concurrently with chemotherapy during radiotherapy.

[0053] 3. Definition of Tumor Recurrence

[0054] All patients underwent follow-up MRI assessment within 12 to 24 weeks after the completion of radiotherapy. Efficacy was assessed according to the RANO 2.0 criteria, including complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). Local recurrence was defined as persistent progression of the target lesion on two consecutive MRI scans, necessitating a change in treatment regimen.

[0055] IV. Clinical Variable Collection and Statistical Analysis

[0056] Clinical information, including patient gender, age, KPS score, tumor type, grade, chemotherapy use, IDH mutation status, MGMT promoter methylation status, Ki-67 index, and PTV dose parameters (minimum dose, maximum dose, and mean dose), was extracted from the electronic medical record system. Univariate logistic regression analysis was performed to identify clinical variables with predictive value (p < 0.1). Subsequently, multivariate logistic regression (forward selection) was performed to identify significant variables (p < 0.05). Given the clinical significance of age, age was included as a predictor in both the training and validation sets.

[0057] 5. Image acquisition and export include:

[0058] (1) MRI images: e-THRIVE, FLAIR, and ADC were retrieved from the hospital's PACS system; the contrast agent used was 0.1 mmol / kg Gd-DTPA.

[0059] (2) Radiotherapy planning computed tomography (CT) images and dose images: BrillianceCT Big Bore scans were performed 1-3 days before radiotherapy with parameters of 140 kVp, 223 mA, slice thickness 3-5 mm, pixel 0.642-1.1 mm, matrix 512×512, and number of slices 100-180. The dose distribution was generated by the MONACA planning system.

[0060] (3) All images were exported in DICOM format and converted to NIfTI using AFNI dcm2niix.

[0061] 6. Image preprocessing includes:

[0062] (1) Skull removal: The brain extraction tool of the Oxford University Functional Magnetic Resonance Imaging Center Software Library (FSL) was used to perform skull stripping on e-THRIVE, FLAIR, ADC, radiotherapy planning computed tomography (CT) images, and dose images, preserving the brain parenchyma area.

[0063] (2) Non-uniform field correction: The N4 BiasField Correction algorithm of the Advanced Normalization Toolset (ANTs) is used to eliminate the intensity offset caused by magnetic field inhomogeneity.

[0064] (3) Resampling: All images were normalized to a voxel size of 1 mm × 1 mm × 1 mm using the Advanced Normalization Toolset (ANTs) linear interpolation method.

[0065] (4) Intensity normalization: Except for the dose map, all images were normalized to the grayscale range of 0-255 using the Nyul histogram normalization method (NyulNormalizer).

[0066] (5) Image registration: First, the dose image is rigidly / elasticly registered to the corresponding radiotherapy plan computed tomography (CT) image; then, the radiotherapy plan CT, FLAIR, and ADC are registered to the e-THRIVE image using the nonlinear symmetric normalization (SyN) algorithm; finally, the planning target volume (PTV) outlined on the radiotherapy plan CT image is used as a mask to crop all modalities, and the minimum, average, and maximum doses of the planning target volume (PTV) are calculated for subsequent clinical variable analysis.

[0067] 7. Planning target volume (PTV) habitat sub-region segmentation, including:

[0068] (1) Input modality: pre-processed e-THRIVE, FLAIR, and ADC data.

[0069] (2) Clustering algorithm: K-means is used, with the number of clusters k=3, and the voxel grayscale similarity is measured by squared Euclidean distance; there is no prior threshold.

[0070] (3) Habitat definition:

[0071] Structural habitat (based on e-THRIVE+FLAIR):

[0072] Type 1 structural habitat area: e-THRIVE high / FLAIR low → enhanced tissue habitat;

[0073] The second type of structural habitat area: e-THRIVE low / FLAIR low → solid low enhanced habitat;

[0074] The third type of structural habitat area: e-THRIVE low / FLAIR high → inactive tissue habitat.

[0075] Physiological habitat (based on e-THRIVE+ADC):

[0076] Type 1 physiological habitat: e-THRIVE high / ADC medium → tumor area with high blood supply and medium cell density;

[0077] The second type of physiological habitat area: low e-THRIVE / low ADC → tumor area with low blood supply and high cell density;

[0078] The third type of physiological habitat area: e-THRIVE low / ADC high → tumor area with low blood supply and low cell density.

[0079] Output: Each voxel obtains a double label of structure and physiology, and obtains six complementary habitat sub-regions; the habitat is shown in the attached figure. Figure 2 .

[0080] 8. Radiomics feature extraction and selection, including:

[0081] In this implementation, global filtering is first performed on the entire clustered PTV and its habitat subregions, including wavelet transform (8 directions), LoG σ = 1-5 mm Laplace-Gaussian filter, square transform (Square, Squareroot, Logarithm, Exponential, Gradient, 2D local binary pattern (LBP-2D) and 3D local binary pattern (3D LBP) to enrich the texture expression ability of the original image. Then, the radiomics feature extraction toolkit PyRadiomics (grayscale bin width bin) is used to extract the grayscale bin width of the PTV. Width=25, grayscale bin width specifies the width of each grayscale bin when the image grayscale is rebinned (resampled). The larger the value, the fewer bins the image grayscale range is divided into, which is used to control the stability and computational complexity of the feature. A total of 2153 imaging features were extracted from images of three modalities: e-THRIVE, FLAIR, and ADC, including 18 first-order features, 14 shape features, and 68 texture features (22 gray-level co-occurrence matrices (GLCMs), 16 gray-level run-length matrices (GLRLMs), and 100% texture features). ), 16 grayscale size region matrices (GLSZM), 14 grayscale dependency matrices (GLDM) and 5 neighboring grayscale difference matrices (NGTDM)) are used to quantify the grayscale distribution, structural complexity and texture pattern inside the tumor tissue, and the first-order features and texture features are filtered by wavelet or LoG to generate high-order features. Feature selection adopts a four-step process: ① Z-score normalization of each ROI feature, and use SMOTE to handle category imbalance in the training set; ② Features with p>0.1 are eliminated by t-test or Mann-Whitney U test; ③ 3 significant features are screened out by combining recursive feature elimination (RFE-SVM / RFE-RF); ④ Spearman correlation coefficient is calculated. If r>0.9 and it has similar biological significance, one item is randomly retained, otherwise all are retained. The remaining features are used to construct a radiotherapy recurrence prediction model, and a pure clinical model and a radiomics-clinical joint model are also established at the same time;

[0082] Model building, including:

[0083] Random forest (RF) and radial basis kernel support vector machine (SVM) classifiers were selected. Hyperparameters were optimized using a 5-fold crossover grid search, aiming to maximize the area under the curve (AUC). Model performance was evaluated using receiver operating characteristic (ROC) curves, area under the curve (95% confidence interval (CI), sensitivity, specificity, and precision), and calibration curves were used to verify consistency across training and validation cohorts. The optimal model was further evaluated using the SHAP method to calculate the marginal contribution of each feature, generating interpretable outputs to guide clinical radiotherapy decisions.

[0084] In this implementation, to address the difficulty of interpreting "black box" machine learning models, the SHAP method, based on cooperative game theory, is used to analyze the interpretability of prediction models. The specific approach is as follows: First, the difference between the output of each candidate model when a feature is included and when it is not included in the prediction is calculated to obtain the marginal contribution of that feature to a single prediction. The SHAP values of multiple predictions are then aggregated to determine the importance of each feature to the overall model output.

[0085] After feature screening and five-fold cross-validation, the optimal model was selected as the support vector machine (SVM) radiomics model based on the first type of physiological habitat area of e-THRIVE. This model consists of three features:

[0086] ① First-order features of the image after exponential transformation: exponential_firstorder_Kurtosis of the support vector machine (SVM) model based on the first-class physiological habitat area of e-THRIVE;

[0087] ② Grayscale size area matrix features after three-dimensional Gaussian logarithmic filtering (σ=5.0 mm): small area emphasis log-sigma-5-0-mm-3D_glszm_SmallAreaEmphasis;

[0088] ③The maximum two-dimensional column diameter (shape feature) of the original image: original_shape_Maximum2DDiameterColumn.

[0089] like Figure 3 As shown in the data, it achieved AUC = 0.703 (95% CI 0.569-0.836) and AUC = 0.670 (95% CI 0.623-0.717) in the training cohort and validation cohort, respectively, and its performance was better than that of other radiomic group joint models, so it was selected as the object of interpretability analysis.

[0090] like Figure 4 As shown in the SHAP summary diagram, "original_shape_Maximum2DDiameterColumn (MDC) of the original image's two-dimensional maximum column diameter (shape feature)" is the most weighted recurrence predictor, and an increase in its value will significantly improve the model output; on the contrary, when the values of the other two features decrease, the risk of recurrence will also increase.

[0091] like Figure 5As shown, SHAP attempts to use two sample patients as examples: Patient A's overall SHAP value is 0.52, which is greater than the baseline value. The model determines that they have relapsed. This is mainly due to the "grayscale size area matrix feature after three-dimensional Gaussian logarithmic filtering (σ=5.0 mm) - small area emphasis log-sigma-5mm-glszm_SmallAreaEmphasis" feature (value -1.199) that has a positive boosting effect. Patient B's SHAP value is 0.48, which is less than the baseline value. The model determines that they have not relapsed. The same feature (value 1.683) has a negative inhibitory effect. Through the above visualization, this implementation not only improves the transparency and clinical acceptability of the model, but also provides an intuitive basis for optimizing radiotherapy strategies.

[0092] The present invention is further described below with reference to specific embodiments, but is not limited to the following contents:

[0093] Example 1

[0094] A 62-year-old male patient with postoperative glioma had pathological findings of glioblastoma and a KPS score >80. He underwent a multimodal MRI scan before radiotherapy in January 2023 and was subsequently scheduled for postoperative precision targeted radiotherapy. After image preprocessing, habitat region division, feature extraction, and feature screening, the three radiomics standardized values based on the first type of physiological habitat were:

[0095] 1.exponential_firstorder_Kurtosis:-0.638

[0096] 2.log sigma 5 0 mm 3D_glszm_SmallAreaEmphasis: -1.199

[0097] 3.original_shape_Maximum2DDiameterColumn: -0.864

[0098] The prediction value was calculated by substituting the formula of the Physiological_Habitat1 (e-THRIVE)_radiomic SVM model of the present invention: According to the Physiological_Habitat1 (e-THRIVE)_radiomic SVM model, when the prediction value is 0.861, the probability of the prediction being effective is greater than 0.5, and the output result is recurrence.

[0099] SHAP was used to calculate the feature marginal contribution of each prediction and draw a SHAP summary plot; original_shape_Maximum2DDiameterColumn was the most important positive indicator, and a high value significantly increased the risk of recurrence; log sigma5 0 mm 3D_glszm_SmallAreaEmphasis was the second most important indicator, and a low value significantly increased the risk of recurrence;

[0100] According to SHAP interpretation, the patient's prediction results were as follows: the total SHAPForce plot value of 0.52 was greater than the baseline value, and recurrence was predicted; this was mainly due to the low value of log sigma 5 0 mm 3D_glszm_SmallAreaEmphasis (-1.199), which pushed the patient towards recurrence.

[0101] The patient's follow-up skull enhanced MRI 3 months later showed that the brain lesions had become larger and progressed, and the actual clinical outcome was completely consistent with the prediction conclusion of the model of the present invention.

Claims

1. A method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics, characterized in that: The following steps are involved: Data acquisition: Multimodal MRI images before postoperative radiotherapy, radiotherapy planning computed tomography (CT) images, dose images, and local recurrence annotations were collected and converted into a standard format. Image preprocessing: The collected MRI images, radiotherapy planning CT images, and dose images are sequentially subjected to brain bone removal, N4 bias field correction, isotropic resampling, grayscale normalization, and image registration to obtain the preprocessed MRI images, radiotherapy planning CT images, and dose images. Region of Interest (ROI) definition and habitat division: Based on the planning target volume (PTV), the ROI is defined in the preprocessed image and dose information of the ROI is extracted. Simultaneously, based on the voxel-level grayscale values of the preprocessed MRI image, the planning target volume (PTV) is divided into three types of physiological structural habitat regions and three types of physiological habitat regions, and corresponding habitat masks are generated. Feature extraction and selection: Extract radiomic features from multimodal MRI images within the planned target volume (PTV) and each habitat area; select the optimal feature subset based on the radiomic features; Model construction and training: The optimal feature subset is selected as input, local recurrence annotation is used as output label, the model performance of the validation set is tested, and the best model is selected; Interpretability Analysis: Perform interpretability analysis on the best model, including generating feature importance SHAP summary plots and individual-level SHAP force plots.

2. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 1, characterized in that: The multimodal MRI images include enhanced T1 isotropic voxel imaging e-THRIVE, fluid attenuated inversion recovery sequence FLAIR, and apparent diffusion coefficient map ADC.

3. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 2, characterized in that: The image preprocessing is as follows: performing brain bone removal on MRI images, radiotherapy planning CT images, and dose images to obtain skull-stripped images; performing N4 bias field correction on the skull-stripped images to obtain bias field-corrected images; performing isotropic resampling on the obtained images to obtain isotropic resampled images; performing grayscale normalization on the radiotherapy planning CT images and MRI images in the isotropic resampled images; aligning the dose image to the radiotherapy planning CT images, and jointly aligning the radiotherapy planning CT images and the fluid-attenuated inversion recovery sequence (FLAIR) and the apparent diffusion coefficient map (ADC) to the enhanced T1 isotropic voxel imaging (e-THRIVE).

4. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 3, characterized in that: Based on the planned target volume (PTV), the region of interest (ROI) is defined in the pre-processed radiotherapy planning computed tomography (CT) images, MRI images, and dose images, and the minimum, average, and maximum doses of the ROI area are extracted.

5. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 3, characterized in that: Based on the voxel-level grayscale values of enhanced T1 isotropic voxel imaging (e-THRIVE), fluid-attenuated inversion recovery (FLAIR), and apparent diffusion coefficient (ADC), the k-means unsupervised clustering algorithm was used to classify each voxel within the planned target volume (PTV) into three types of physiological structure habitats.

6. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 5, characterized in that: The k-means unsupervised clustering algorithm is specifically as follows: the input is a multidimensional vector composed of voxel-level grayscale values of enhanced T1 isotropic voxel imaging e-THRIVE, fluid-attenuated inversion recovery sequence FLAIR, and apparent diffusion coefficient map ADC. First, by minimizing the squared error of the distance between voxel features and cluster centers, the region is divided into K habitat areas with similar characteristics to achieve voxel allocation, and the feature center of each class is updated, and the algorithm is iterated until convergence.

7. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 5, characterized in that: The three types of physiological structure habitat areas include: The first type of structural habitat area: enhanced tissue habitat, characterized by high e-THRIVE value and low FLAIR value; The second type of structural habitat area: solid low-enhancement habitat, characterized by low FLAIR and e-THRIVE values; The third type of structural habitat area: inactive tissue habitat, characterized by high FLAIR value and low e-THRIVE value.

8. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 7, characterized in that: According to the three types of physiological structure habitat areas, three types of physiological habitat areas are formed: Type I physiological habitat: tumor area with high blood supply and medium cell density, high e-THRIVE value and medium ADC value; The second type of physiological habitat area: tumor area with low blood supply and high cell density, low e-THRIVE value and low ADC value; The third type of physiological habitat area: tumor area with low blood supply and low cell density, low e-THRIVE value and high ADC value.

9. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 1, characterized in that: Feature extraction and selection adopts a multi-stage feature dimensionality reduction method, which includes: The radiomics features extracted from different MRI images based on the habitat area are normalized to obtain the normalized radiomics features of each MRI image. Balance the number of positive and negative samples in the training set, including the number of relapsed patients and non-relapsed patients; Univariate analysis, based on positive and negative samples, performed separate statistical tests on each radiomics feature to screen out features related to efficacy prediction; The multimodal MRI imaging features selected by univariate analysis were further screened to optimize the input feature set, including iteratively training the classification model to gradually eliminate the features with the smallest contribution from the selected features to obtain the optimal features; The optimal features after screening are tested for correlation, and the Pearson correlation coefficients between the two features are calculated. Feature pairs with correlation coefficients less than 0.9 are retained to obtain the optimal feature subset.

10. The method for predicting the efficacy of glioma radiotherapy based on habitat analysis imaging omics according to claim 9, characterized in that: The Pearson correlation coefficient between two features measures the degree of linear correlation, and the formula is: ； in, is the covariance, , is the standard deviation.

Images