A prognosis prediction system for neoadjuvant chemotherapy of gastric cancer based on habitat and elastography imaging

Abstract

The application discloses a stomach cancer neoadjuvant chemotherapy prognosis prediction system based on habitat and elastic imaging, relates to the medical image prognosis field, and comprises the following modules: an image labeling module, which is used for acquiring a pre-treatment CT image of a locally advanced gastric cancer patient and labeling a tumor region of interest; a habitat imaging module, which is used for extracting local imageomics features of each voxel in the tumor region of interest, dividing the voxels into multiple habitat sub-regions, and extracting imageomics features; an elastic imaging module, which is used for constructing elastic invariants that are invariant to non-rigid deformation, performing signal suppression and dynamic range compression on the elastic invariants, and then extracting elastic texture features; and a prognosis prediction module, which is used for constructing a prognosis prediction model for outputting postoperative survival prediction results. The application can realize accurate prediction of the overall survival period of a locally advanced gastric cancer patient after neoadjuvant chemotherapy, and improve the individualized risk stratification capability.

Description

Technical Field

[0001] This invention relates to the field of medical imaging prognosis, and in particular to a prognostic prediction system for neoadjuvant chemotherapy in gastric cancer based on habitat and elastography. Background Technology

[0002] Gastric cancer is a highly prevalent malignant tumor of the digestive tract worldwide, with most patients already progressing to locally advanced gastric cancer at the time of clinical diagnosis. Neoadjuvant chemotherapy combined with radical surgery is currently the standard treatment strategy for locally advanced gastric cancer. Its core objective is to reduce tumor burden and eliminate micrometastases through preoperative chemotherapy, thereby improving the radical resection rate and prolonging overall survival. However, due to the high heterogeneity of tumors at the molecular and cellular levels, even patients at the same pathological stage can show significant differences in treatment response and long-term prognosis after receiving the same neoadjuvant chemotherapy regimen, with some patients still facing a high risk of postoperative recurrence and metastasis. Therefore, there is an urgent clinical need for a method to accurately stratify patients' risks early after neoadjuvant chemotherapy. In recent years, with the development of radiomics, quantitative features extracted from high-throughput CT images have been attempted for tumor prognostic assessment. However, most of these studies rely on global imaging features reflecting the static anatomical structure of the tumor, lacking in-depth characterization of the spatial heterogeneity distribution and tissue elasticity properties within the tumor. This makes it difficult to comprehensively capture biological information closely related to the tumor's invasive and metastatic potential, limiting further improvement in prognostic predictive efficacy.

[0003] Existing prognostic assessment methods for neoadjuvant chemotherapy in locally advanced gastric cancer have the following shortcomings: First, most studies use the overall imaging features of a single tumor region as model input, failing to analyze the spatial heterogeneity within the tumor and ignoring the differential impact of different biological behaviors of tumor subregions on prognosis, resulting in insufficient ability of predictive models to characterize the complex biological features of tumors. Second, existing prognostic studies based on CT images mainly rely on static grayscale features reflecting tissue density and morphology, failing to integrate the tissue elasticity information contained in CT elastography. However, the elastic properties of tumor tissue, such as stiffness and heterogeneity, are closely related to its invasiveness and metastatic ability; the lack of such features weakens the accuracy of identifying malignant biological behaviors of tumors. Third, current prognostic models mostly focus on predicting short-term treatment response and do not systematically incorporate key clinicopathological factors that determine long-term survival outcomes, such as postoperative pathological staging and vascular invasion, making it difficult to establish a direct correlation between imaging features and long-term survival outcomes. The clinical relevance and prognostic robustness of the models need to be improved. Therefore, based on the above-mentioned challenges, this invention proposes a prognostic prediction system for neoadjuvant chemotherapy in gastric cancer based on habitat and elastography. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a prognostic prediction system for neoadjuvant chemotherapy in gastric cancer based on habitat and elastography. This system addresses the limitations in prediction accuracy of existing prognostic assessment methods after neoadjuvant chemotherapy for gastric cancer, which neglect internal spatial heterogeneity of the tumor, lack information on tissue elasticity properties, and have insufficient correlation with long-term survival outcomes. The system aims to achieve accurate prediction of overall survival after neoadjuvant chemotherapy in patients with locally advanced gastric cancer and improve individualized risk stratification capabilities.

[0005] To achieve the above objectives, this invention provides a prognostic prediction system for neoadjuvant chemotherapy in gastric cancer based on habitat and elastography. The system first acquires pre-treatment CT images of patients with locally advanced gastric cancer, labels the tumor region of interest, and extracts local radiomics features of each voxel within the region. Gaussian mixture model clustering is used to divide the tumor region into multiple habitat subregions with different biological characteristics to quantify intratumoral heterogeneity. Simultaneously, based on the gradient field and Hessian matrix of the CT images, a differential affine elastic invariant invariant to non-rigid deformation is constructed. After signal suppression and dynamic range compression through nonlinear monotonic mapping, an elastic feature map is generated, from which multi-directional elastic texture features are extracted. Subsequently, principal component analysis is used to reduce the dimensionality of the radiomics features of each habitat subregion, and LASSO-Cox regression is used to generate habitat risk scores and elasticity risk scores, respectively. Finally, a multivariate Cox proportional hazards regression model is used to integrate the above risk scores with the patient's postoperative pathological stage, vascular invasion, and other clinicopathological information to construct a prognostic prediction model and output individualized survival prediction results.

[0006] In a first aspect, the present invention provides a prognostic prediction system for neoadjuvant chemotherapy in gastric cancer based on habitat and elastography, comprising: The image annotation module is used to acquire pre-treatment CT images of patients with locally advanced gastric cancer and annotate the tumor region of interest. The habitat imaging module is used to extract local radiomics features of each voxel within the region of interest of the tumor, and to divide the voxels into multiple habitat subregions with different biological characteristics through Gaussian mixture model clustering, and to extract radiomics features of each habitat subregion. The elastic imaging module constructs elastic invariants that are invariant to non-rigid deformation based on the gradient field and Hessian matrix of the CT image, and uses nonlinear monotonic mapping to perform signal suppression and dynamic range compression on the elastic invariants to generate an elastic feature map, and then extracts elastic texture features from the elastic feature map. The prognostic prediction module generates habitat risk scores by aggregating radiomics features of each habitat subregion and generates elasticity risk scores by utilizing the elastic texture features. It then integrates the habitat risk scores, elasticity risk scores, and patient clinicopathological information using a multivariate survival regression model to construct a prognostic prediction model for outputting postoperative survival prediction results.

[0007] Furthermore, the local image omics features extracted by the habitat imaging module include at least first-order entropy features and multiple texture features, and the optimal number of clusters in the Gaussian mixture model clustering is automatically determined based on the Bayesian information criterion, so as to avoid oversegmentation or undersegmentation while ensuring the model fit, and to ensure the objectivity of habitat subregion division.

[0008] Furthermore, the elastic invariant is constructed by forming a quadratic form from the gradient vector of the CT image and the inverse of the Hessian matrix to obtain a differential affine elastic invariant that reflects the local elastic properties of the tissue.

[0009] The nonlinear monotonic mapping includes a combination of square root operation and arctangent transformation, which is used to compress the numerical range of the elastic invariant to a preset finite interval, and suppress noise and enhance weak elastic signals during the compression process.

[0010] Furthermore, the process of generating the elastic feature map also includes: The elastic invariants processed by nonlinear monotonic mapping are scaled to grayscale and combined with the mask of the region of interest of the tumor to generate an elastic feature map that retains only the elastic information inside the tumor. This eliminates the interference of non-lesion tissue around the tumor on the extraction of elastic features and improves the correlation between elastic features and tumor biological behavior.

[0011] Furthermore, the extraction process of the elastic texture features includes: A gray-level co-occurrence matrix is ​​constructed on the elastic feature map along multiple preset three-dimensional spatial directions, and extended Haralick texture feature parameters are extracted from the gray-level co-occurrence matrix in each direction.

[0012] Furthermore, the specific methods by which the prognostic prediction module generates habitat risk scores include: For each patient, the radiomics feature vectors belonging to the same habitat subregion were reduced to a single principal component through principal component analysis, and the corresponding feature values ​​for missing habitat subregions were set to zero. Subsequently, the LASSO-Cox regression model was used to screen features and the habitat risk score was obtained by weighted summation of coefficients.

[0013] Furthermore, the prognostic prediction module receives post-treatment CT images of the same patient, uses the habitat imaging module to divide the tumor region of the post-treatment images into habitat subregions and extract post-treatment habitat features, calculates the relative change of habitat features before and after treatment, and integrates the relative change into the multivariate survival regression model to capture the dynamic information of tumor imaging feature state space reconstruction during neoadjuvant chemotherapy, providing incremental prognostic value for long-term survival prediction.

[0014] Furthermore, when generating prognostic prediction results, the prognostic prediction module employs an inter-feature attention mechanism to deeply fuse the radiomics features and elasticity features of the habitat subregions. This mechanism uses the biological phenotypic features of each habitat subregion as query conditions and the elasticity features of each local location in the tumor region as matching objects. It adaptively calculates attention weights to filter elasticity-habitat coupling patterns that are significantly related to prognosis, and generates elasticity context-enhanced fusion features based on these coupling patterns for prognostic risk score calculation.

[0015] Furthermore, the prognostic prediction module is also configured to construct a tumor imaging feature state evolution model based on CT images of the same patient before and after treatment. This model defines the tumor imaging feature state by combining radiomics features and elasticity features of each habitat subregion, describes the dynamic evolution trajectory of each habitat subregion under elastic driving and habitat interaction during neoadjuvant chemotherapy using continuous temporal differential equations, and extracts evolutionary feature parameters that characterize the evolution rate and direction from the evolutionary trajectory. The evolutionary feature parameters are then incorporated into the prognostic prediction as temporal evolution features.

[0016] Furthermore, the prognostic prediction module is also configured to assess the individualized treatment benefits of patients. It constructs counterfactual survival prediction models under different treatment pathways to infer the conditional survival probability of the same patient under the assumption of receiving neoadjuvant chemotherapy and not receiving neoadjuvant chemotherapy, and quantifies the degree of individualized treatment benefit of the patient based on the difference between the two counterfactual survival curves, which is used to guide the individualized decision-making of postoperative adjuvant therapy plans.

[0017] Furthermore, the prognostic prediction module also constructs a visual nomogram for individualized survival probability assessment, the nomogram being composed of independent prognostic factors in the multivariate survival regression model.

[0018] Secondly, a method for predicting the prognosis of neoadjuvant chemotherapy for gastric cancer based on habitat and elastography is also provided. This method is based on the system described in the first aspect above, and includes: Acquire pre-treatment CT images of patients with locally advanced gastric cancer and annotate the tumor region of interest; Local radiomics features of each voxel within the region of interest of the tumor are extracted, and Gaussian mixture model clustering is used to divide the voxels into multiple habitat subregions, and radiomics features of each habitat subregion are extracted. Based on the gradient field and Hessian matrix of the CT image, an elastic invariant that is invariant to non-rigid deformation is constructed. The elastic invariant is then subjected to signal suppression and dynamic range compression using a nonlinear monotonic mapping to generate an elastic feature map. Elastic texture features are then extracted from the elastic feature map. The image omics features of each habitat subregion are aggregated to generate a habitat risk score, and the elastic texture features are used to generate an elastic risk score. A multivariate survival regression model was used to integrate the habitat risk score, elasticity risk score and patient clinicopathological information to construct a prognostic prediction model and output postoperative survival prediction results.

[0019] This invention proposes a prognostic prediction system for neoadjuvant chemotherapy in gastric cancer based on habitat and elastography. By organically combining tumor spatial heterogeneity analysis, tissue elasticity property characterization and clinicopathological factors, it achieves a more accurate prognostic assessment than existing systems based on a single imaging modality or traditional pathological indicators. Habitat imaging technology divides tumors into multiple biological subregions and extracts features from each, breaking the limitations of relying solely on global image features and significantly improving the ability to characterize the heterogeneity of tumors. Elastography technology extracts elastic features that are invariant to non-rigid deformation from conventional CT images, adding tissue elasticity information missing from traditional CT grayscale features and enhancing the discriminative basis related to tumor invasion and metastasis potential. The composite risk score constructed through principal component analysis and regularized regression strategies can retain the most valuable discriminative information while reducing feature redundancy and obtain prognostic prediction capabilities independent of traditional clinicopathological factors. The synergistic modeling of multidimensional image features and postoperative key pathological factors establishes a direct correlation from image phenotype to long-term survival outcomes. Validated by multi-center data, it demonstrates good predictive accuracy and generalization performance, and can effectively assist in individualized risk stratification and clinical decision-making for patients with locally advanced gastric cancer after neoadjuvant chemotherapy.

[0020] Beneficial effects By implementing the above-mentioned prognostic prediction system for neoadjuvant chemotherapy of gastric cancer based on habitat and elastography provided by the present invention, the following technical effects are achieved: (1) By extracting first-order entropy features and multi-class texture features at the voxel level, and using a Gaussian mixture model based on the Bayesian information criterion to automatically determine the optimal number of clusters for unsupervised classification of voxels, the macroscopic tumor region can be divided into multiple habitat subregions with significant differences in gray-level distribution uniformity, texture complexity, and structural connectivity. This division objectively reflects the spatial distribution pattern of biological heterogeneity within the tumor caused by factors such as differences in cell density, unequal degrees of necrosis, and uneven stromal response. Subsequently, radiomics features are extracted for each subregion, transforming the fuzzy global assessment into a refined representation at the subregion level. This makes the mapping relationship between tumor biological behavior and radiographic features more accurate, providing a more discriminative basic feature space for the construction of subsequent prognostic risk scores.

[0021] (2) By constructing a quadratic elastic invariant composed of gradient vectors and the inverse of the Hessian matrix, the extracted elastic features are theoretically invariant to complex non-rigid deformations, ensuring feature stability and repeatability across individuals and scanning conditions. Simultaneously, a nonlinear monotonic mapping mechanism based on square root operation and arctangent transformation is introduced. This effectively attenuates background noise and enhances the extraction of weak elastic signals while compressing the elastic invariant to a suitable bounded interval for feature analysis, significantly improving the tissue elasticity difference resolution of the elastic feature map. Subsequently, a gray-level co-occurrence matrix is ​​constructed along multiple directions in three-dimensional space, and extended Haralick texture features are extracted. This comprehensively quantifies the distribution patterns and heterogeneity of tumor tissue elastic properties in different directions, providing a set of complementary discriminative information independent of traditional gray-level texture features for the prognostic model.

[0022] (3) By using the biological phenotype of habitat subregions as query conditions and the elasticity attribute of corresponding spatial locations as matching objects, this mechanism can adaptively assign higher attention weights to elasticity-habitat coupling patterns that are highly correlated with prognosis, while effectively suppressing the interference of background elasticity signals unrelated to survival outcomes on the feature fusion process. The resulting elasticity context-enhanced fusion features are significantly better than simple feature vector concatenation in prognostic discrimination, enabling the prognostic model to more accurately distinguish patient subgroups with different risks of recurrence and metastasis. In particular, for difficult-to-discriminate cases with complex habitat subregion spatial composition and high heterogeneity of elasticity attributes, the discriminative power of the fusion features is more significantly improved.

[0023] (4) By describing the continuous evolution trajectory of each habitat subregion under elastic driving and habitat interaction using ordinary differential equations, and extracting evolutionary characteristic parameters such as evolution rate constant and steady-state asymptotic value from the evolutionary trajectory, this method can effectively capture the ebb and flow of tumor regions with different biological characteristics under chemotherapy stress. The prognostic model after incorporating evolutionary characteristic parameters significantly improved the accuracy of long-term survival prediction. Among them, the steady-state asymptotic value showed independent prognostic significance in multivariate survival analysis, and the survival difference between the high evolution rate group and the low evolution rate group was statistically significant.

[0024] (5) By constructing counterfactual survival prediction models under different treatment pathways, a quantitative assessment of the individualized survival benefit of each patient receiving neoadjuvant chemotherapy was achieved, filling the gap in existing prognostic models that can only output the survival probability under a given treatment plan but cannot answer the question of treatment value. This method uses the patient's pre-treatment imaging features as a bridge, and uses a shared representation space to correct the feature distribution offset between different treatment pathways, ensuring that the two counterfactual survival curves are comparable in the same scale space, and quantifying the individual benefit magnitude by the difference in restricted mean survival time. The benefit stratification decision rule constructed based on this can effectively identify patient subgroups that are judged as low-risk by traditional pathological stratification but can actually benefit significantly from postoperative adjuvant chemotherapy, avoiding undertreatment; at the same time, it screens out low-benefit patients with limited benefit, avoiding unnecessary treatment toxicity, and providing a quantitative basis for individualized upgrading and downgrading decisions of postoperative adjuvant therapy plans. Attached Figure Description

[0025] To make the above-described prognostic prediction system for neoadjuvant chemotherapy of gastric cancer based on habitat and elastography of the present invention more apparent and understandable, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the method described in this application; Figure 2 This is a flowchart illustrating the generation of an elastic feature map and the extraction of elastic features. Figure 3 A nomogram visualizing the prognostic prediction model; Figure 4 This plot represents the ROC curve representing the time dependence of the prognostic model on the external test set. Figure 5 This represents the Kaplan-Meier survival curves on the external test set of the prognostic model. Figure 6 This represents the calibration curve of the external test set for the prognostic model. Figure 7 This diagram illustrates the principle of the elastic-habitat mutual attention mechanism. Detailed Implementation

[0027] Example 1: This embodiment provides a system and method for predicting the prognosis after neoadjuvant chemotherapy for gastric cancer based on CT habitat imaging and elastography. The procedure is as follows: Figure 1 As shown.

[0028] This study included 1021 patients with locally advanced gastric cancer who underwent neoadjuvant chemotherapy and radical gastrectomy at three hospitals. Hospital 1 had 572 patients, Hospital 2 had 195 patients (as supplementary test set 1), and Hospital 3 had 254 patients (as supplementary test set 2). Inclusion criteria were age over 18 years, clinical diagnosis of locally advanced gastric cancer, completion of complete preoperative neoadjuvant chemotherapy and radical gastrectomy, complete baseline clinical data prior to neoadjuvant chemotherapy, and a chest CT scan performed within one month prior to neoadjuvant therapy. Exclusion criteria included a history of one or more primary malignancies, preoperative radiotherapy or chemotherapy, inability to assess treatment response, lack of postoperative histopathological examination results, and poor CT image quality leading to unclear lesions. Preoperative portal venous phase CT images, preoperative clinicopathological information, and postoperative pathological features were collected from the above patients. Postoperative portal venous phase CT images of patients from Hospital 1 were also collected.

[0029] During the data acquisition and annotation phase, experienced imaging experts annotated the largest and superior / inferior planes of the tumor region in pre-treatment CT images, identifying the annotated areas as regions of interest for subsequent analysis. Using pre-treatment portal venous phase CT images and annotation masks as input data, the CT images were first processed... Spline interpolation is resampled at equal intervals to And set the window width and window level to unify the consistency of image information and amplify the contrast of a specific grayscale range.

[0030] The next stage involves habitat subregion segmentation, employing voxel-based habitat imaging analysis to cluster the labeled tumor regions. Specifically, the CT images are represented as a three-dimensional matrix. ,in , , These are width, height, and depth, respectively. In real number space. A segmentation mask for the tumor region. , voxels The mark, This voxel belongs to the region of interest of the tumor. This is the background. For each voxel within the mask area... Define the neighborhood around it. We used PyRadiomics to extract local image omics features.

[0031] In the voxel feature extraction process, entropy, a first-order statistical feature, is an indicator of the uniformity of image gray-level distribution, reflecting the non-uniformity and complexity of the internal structure of a tumor, and is defined as:

[0032] In the formula, , , These are the coordinates along the x, y, and z axes of the image, respectively. voxels The entropy value of the neighborhood; The gray value within the neighborhood The probability of voxels appearing. Higher entropy indicates more heterogeneous and complex tumor cells or tumor structures, potentially suggesting high heterogeneity, treatment resistance, or poor prognosis. For example, CT images are denoted as... A three-dimensional matrix, taking tumor endogenous voxels , build Large and small neighborhoods are used to statistically analyze the voxel CT values ​​within the neighborhood. If a CT value of 20 HU appears... The second occurrence was 30HU. The second occurrence was 40HU. Then the probabilities of each gray level are respectively , , Substituting into the formula, the entropy value is approximately .

[0033] In addition to entropy features, multiple texture features were extracted. Gray-level co-occurrence matrix feature set. Features including contrast, dissimilarity, homogeneity, energy, and correlation are used to describe the relationships between pixels at different gray levels in an image. Contrast measures the difference in gray levels between pixels; high contrast indicates large differences in gray levels within a tumor. Dissimilarity measures the degree of change in gray values; it is positively correlated with contrast but emphasizes local changes. Homogeneity reflects the smoothness of texture; high homogeneity indicates a relatively uniform internal structure of the tumor and low resistance to drug penetration. Energy represents the uniformity of gray level distribution. Correlation reflects the degree of correlation between the gray values ​​of adjacent pixels; low correlation often suggests disordered cell arrangement and severe stromal damage.

[0034] Gray-scale run-length matrix feature set This term describes the continuous length of pixels with the same grayscale value. Short run length emphasizes fine-grained texture, and high values ​​are often associated with higher recurrence rates and poorer survival prognosis; long run length emphasizes large, continuous areas of the same grayscale value; grayscale nonuniformity reflects the degree of unevenness in the distribution of grayscale values; and run length nonuniformity reflects the degree of unevenness in the distribution of run lengths.

[0035] Gray-scale region size matrix feature set This describes the size distribution of connected regions with the same grayscale value. The normalized region size nonuniformity measures the degree of unevenness in the region size distribution; a higher value indicates extremely irregular tumor growth patterns. Small areas with high grayscale emphasize capturing fine and high-density texture structures.

[0036] Neighborhood gray-level difference matrix feature set Texture characteristics are described by quantifying the grayscale differences between pixels and their neighboring pixels. Roughness reflects the coarseness of texture particles; busyness describes the frequency of local texture changes; complexity reflects the complexity of texture distribution; and intensity measures texture saliency and recognizability. These multi-dimensional texture features synergistically characterize the microscopic and macroscopic heterogeneity of tumors from different dimensions, providing important quantitative evidence for non-invasive prediction of treatment sensitivity and prognosis assessment.

[0037] Each voxel The entropy value and various texture features are combined to form a feature vector. The feature map is generated by mapping the feature value of each voxel to the image pixel value. Then, the feature vectors of all voxels within the region of interest are combined into a feature matrix. ,in Let voxels be the number of voxels within the region of interest. The total number of features. After standardization by the mean and standard deviation of each feature, voxels are assigned to different categories using a Gaussian mixture model clustering method. The probability density function of the Gaussian mixture model is... P(x) Defined as Weighted sum of Gaussian distributions:

[0038] In the formula, x The eigenvectors of a single voxel, i.e., the eigenmatrix. A certain row (dimension is) ); For feature vectors x Generation probability under Gaussian mixture model; The Gaussian component number is the preset number of cluster categories; For the first The mixing coefficient of the Gaussian component indicates that any voxel belongs to the . The prior probability of the first subregion simultaneously reflects the first... The proportion of each subregion within the overall region of interest in the tumor; For the first The mean vector of Gaussian components (dimension) ); For the first The probability density function of Gaussian components; The covariance matrix (dimensions) of the k-th Gaussian component ).

[0039] To determine the optimal number of clusters in a Gaussian mixture model The Bayesian Information Criterion (BIC) is used to balance model complexity and goodness of fit. A smaller BIC value indicates a better model. The formula for BIC is:

[0040] In the formula, Let be the likelihood probability of the Gaussian mixture model for all voxel feature data. The larger the value, the better the model fits the data; for The total number of all freely adjustable parameters for each Gaussian component; The number of samples, i.e., the total number of voxels in the tumor region of interest, is also the feature matrix. The number of rows. After clustering, each voxel obtains a category label. The clustering results are mapped back to the original 3D space to obtain habitat segmentation images. , defined as: when hour ,otherwise The output format is NIfTI, and the trained Gaussian mixture model is saved for habitat subregion segmentation on additional test sets and post-treatment CT images.

[0041] Building upon this, radiomics features were extracted from each habitat subregion in the CT images using the PyRadiomics tool. The extraction process was configured using a YAML file, and the image grayscale values ​​were normalized to [value missing]. A fixed range, and with The bin width is discretized. It should be noted that the image grayscale values ​​are normalized to... The fixed range does not refer to a direct mapping of the Henle unit (HU) values ​​from the original CT acquisition, but is based on the following two steps: First, a clinical standard soft tissue window (window level) is used. Window width The original CT image is truncated to grayscale, and the HU values ​​within the window width are linearly mapped to... Extremely high HU values ​​(such as bone tissue and calcifications) and extremely low HU values ​​(such as air and gas) outside the window width range are reasonably pruned because they do not contain effective information about tumor tissue. Secondly, using 500 as the upper limit of the mapping ensures that all voxels within the tumor region are effectively resolved, while also leaving some redundancy to avoid information loss due to minor fluctuations in the window level. This parameter range has been validated with multi-center data to demonstrate stable predictive performance. Wavelet transform and Laplace-Gaussian filtering algorithms (LoG) are applied. By combining various nonlinear transformations such as square, square root, and exponentiation, a comprehensive feature set is extracted from the original and multi-filtered spaces, covering first-order statistics, texture features (including the selected GLCM and the complete GLRLM, GLSZM, GLDM, NGTDM features), and three-dimensional shape features, in order to quantify the spatial heterogeneity and microstructural complexity of each habitat subregion.

[0042] In the elasticity image generation and feature extraction stage, local deformation invariants are first extracted from a single CT scan image to generate an elastic feature map that reflects the elastic properties of soft tissue. The process flow is as follows: Figure 2 As shown. It should be noted that the elastography described here does not refer to traditional imaging modalities such as ultrasound elastography and magnetic resonance elastography, which apply mechanical excitation to tissues and measure deformation responses. Instead, it is a calculation method based on the grayscale field of conventional CT images, which derives the local elastic properties of tissues from the perspective of non-rigid deformation invariance by calculating the differential geometric invariants formed by the gradient vector and the Hessian matrix. This method does not rely on any external excitation and generates feature maps reflecting the elastic properties of tissues solely from a single CT scan. The CT image data... pass spline interpolation resampling to The window width and level are set, and Gaussian smoothing filtering is used to suppress noise in the image. The smoothed image is denoted as... ,in A three-dimensional Gaussian kernel; is the Gaussian smoothing coefficient, representing the convolution operation.

[0043] Subsequently, based on the principles of differential geometry and local affine transformation, the first and second partial derivatives of the image are calculated respectively. The first partial derivative is the gradient vector. The Sobel operator is used to characterize the local variation trend of image grayscale in three-dimensional space.

[0044] In the formula, , , .

[0045] The second-order partial derivatives are calculated using a Deriche filter (second-order derivative) to construct the Hessian matrix. :

[0046] In the formula, , and It is a pure second-order partial derivative; , and These are the mixed partial derivatives. The Deriche filter is a one-dimensional recursive filter, and its coefficients are determined by the smoothing parameter. Decide, Preferably, it is 1, so as to accurately capture elastic fluctuation information and suppress noise.

[0047] Based on gradient vectors and Hessian matrices, an elastic invariant that integrates first-order and second-order derivative information of the image is constructed. It simultaneously encodes shape changes and fluctuation information, enabling it to qualitatively and quantitatively reflect the local elastic properties of soft tissues, and possesses theoretical invariance to complex non-rigid deformations.

[0048] In the formula, It is the inverse of the Hessian matrix; This is the transpose of the gradient vector. Because... The numerical range is extremely wide ( Direct quantization leads to information loss; therefore, a signal suppression method based on the arctangent function is used to reduce information loss. Mapping to a finite interval Simultaneously achieving noise suppression and weak signal amplification, the processed parameters are denoted as follows. The signal suppression formula is:

[0049] In the formula, The root-order parameter is used to adjust the degree of signal compression. It is selected based on the difference in CT values ​​between normal gastric tissue and cancerous tissue, with the parameter being the closest in magnitude. ; This is a numerical stability constant to prevent division-by-zero errors. The signal is then suppressed... The value is mapped to a specified grayscale range using linear grayscale scaling. Generate elastic feature map:

[0050] In the formula, This is the rounding function; This refers to the grayscale level, with values ​​ranging from 104 to 128. The parameter with the closest difference in CT values ​​between normal and cancerous gastric tissue is selected. Preferred Finally, the pre-drawn and pre-processed region of interest mask for the lesion is read. Only the elastic feature information of the region of interest of the lesion is retained to obtain the final CT elastic feature map. :when hour ,otherwise Save the final elastic feature map in NIfTI format.

[0051] When extracting elastic features from an elasticity feature map, the elasticity feature map is first loaded, normalized, and the voxel values ​​are quantized. Each gray level. In Calculate the gray-level co-occurrence matrix in different three-dimensional spatial directions, including Each axis direction , , , Each face is slanted , , , , , ,as well as Individual oblique , , , The neighborhood distance is set to For each direction's gray-level co-occurrence matrix, calculate a set of... An extended Haralick texture feature, in In addition to the basic Haralick features (contrast, dissimilarity, homogeneity, second moment, and correlation), new features have been added, including mean, standard deviation, joint entropy, marginal entropy, difference entropy, sum entropy, cluster shading, cluster salience, inverse difference moment, information measure coefficient, and autocorrelation. Features, total One characteristic. (To be continued) Each direction A combination of features is used to ultimately extract one. A 3D feature vector comprehensively quantifies the spatial distribution pattern of elastic properties within a tumor.

[0052] In the prognostic prediction model construction and validation phase, habitat characteristics and resilience features of each habitat subregion before treatment were used to generate prognostic-related habitat risk scores and resilience risk scores, respectively. The screening of radiomic features for each habitat subregion employed an independent prognostic factor screening strategy. First, univariate Cox regression analysis was used to screen for radiomic features significantly associated with prognosis, retaining only those... Value less than To avoid overfitting, features are selected, and then LASSO regression analysis is applied for feature selection. LASSO automatically selects features by introducing an L1 penalty term to compress the coefficients of some features to zero. Its objective function is to minimize the following expression:

[0053] In the formula, This is a real label; The characteristic matrix; The coefficients are characteristic coefficients; To adjust the parameter, i.e., the regularization strength. With... The larger the value, the stronger the regularization, the smaller the feature coefficients will become, ultimately retaining only features with coefficients greater than zero. A 10-fold cross-validation strategy is used to determine the optimal approach. The value that minimizes the average cross-validation error. The value is used as the optimal parameter. Subsequently, features belonging to the same subregion from each patient are aggregated to form the feature vector of that subregion. Principal component analysis is then applied to the feature vector of each subregion to reduce it to a single principal component. For missing subregions, the corresponding feature value is set to zero, resulting in... Each principal component feature is used. Finally, LASSO-Cox regression is used to... The habitat risk score is obtained by weighted summation of the principal component features.

[0054] The same strategy for screening independent prognostic factors was used to process elasticity features. First, univariate Cox regression analysis was used to screen for independent risk factors. Then, LASSO-Cox regression was used to screen for features most relevant to prognosis, and the elasticity risk score was obtained by weighted summation of coefficients. For the collected key clinicopathological features that determine long-term prognosis, including serum tumor markers, preoperative clinical characteristics, postoperative pathological stage, vascular invasion, and nerve invasion, univariate Cox regression analysis was also used to screen for independent risk factors, retaining only those relevant to long-term prognosis. Value less than The features were then analyzed, and LASSO regression was used to screen for the features most relevant to prognosis. The cross-validation strategy retains features with coefficients greater than zero.

[0055] Hospital data was divided into training and validation sets in a 7:3 ratio. Habitat risk scores, elasticity risk scores, and screened clinicopathological features from the training set were incorporated into a multivariate Cox proportional hazards regression model to construct a prognostic prediction model. This model predicted overall survival after neoadjuvant chemotherapy for gastric cancer, yielding a survival score. The survival score output by the model was used, and the `surv_cutpoint()` function of the `survminer` package was employed to maximize the difference to determine the optimal threshold, thus dividing patients into high-risk and low-risk groups. Survival outcomes for both groups were presented using Kaplan-Meier survival curves and evaluated using the log-rank test. Subsequently, independent prognostic factors ( ) were analyzed based on the multivariate Cox regression. Constructing a nomogram of overall survival transforms the prognostic prediction model into a quantifiable scoring system. This involves building a nomogram based on habitat risk score, resilience risk score, and independent clinical predictors, such as... Figure 3 As shown, clinical nomograms were constructed based solely on independent clinical predictors. The predictive accuracy, robustness, and net clinical benefit of the multidimensional prognostic model were comprehensively evaluated using time-dependent ROC curves, C-index, survival curves, calibration curves, and decision curves in the validation set, Hospital 2, and Hospital 3 datasets. Furthermore, stratified analysis was performed on the entire dataset, and the results are presented in forest plot form. The time-dependent ROC curves of the external test set for the prognostic model are shown below. Figure 4 As shown, the Kaplan-Meier survival curve is as follows: Figure 5 As shown in Figure 6, the calibration curve is also shown.

[0056] In the prognostic analysis phase combining pre- and post-treatment CT images, based on the habitat subregion segmentation model of pre-treatment CT images, habitat subregions were segmented in the post-treatment pre-operative images, and habitat and elasticity features were extracted. In addition to pre- and post-treatment features, the same habitat subregion and the same elasticity feature map were also calculated. Quantitative imaging features were used to capture the relative changes in tumor features during neoadjuvant chemotherapy. The feature is calculated as the percentage change between the feature values ​​before and after treatment, using the following formula:

[0057] In the formula, and These are characteristic values ​​before and after neoadjuvant chemotherapy; A tiny positive real number is added to ensure numerical stability; positive This indicates a decrease in eigenvalues ​​after treatment; a negative value indicates a reduction in eigenvalues. This indicates an increase. Subsequently, the characteristics of the habitat, resilience, and habitat after treatment were analyzed. Features and resilience The features undergo the same filtering and dimensionality reduction process, resulting in elasticity after feature filtering. Features not included, resulting in Key components of habitat ( Principal component characteristics of the pre-treatment habitat; Post-treatment habitat principal component characteristics and individual habitats Principal component features were used to obtain habitat risk scores by weighted summation of these principal component features using LASSO-Cox regression, and then incorporated into multivariate Cox proportional hazards regression to construct a prognostic prediction model.

[0058] Meanwhile, using principal component analysis and t-SNE dimensionality reduction, the characteristic distributions of the three habitats (Habitat1, 2, and 3) before and after treatment were compared. The unique characteristic spatial distributions of different habitats and the distribution patterns of samples with different treatment responses in the characteristic space were visualized. The impact of treatment on sample separation and clustering in habitat subregions was assessed by the changes in the characteristic space.

[0059] Example 2: Existing radiomics methods, when fusing habitat and elasticity features, typically employ simple vector concatenation or linear weighting, neglecting the spatial correspondence between the elasticity attributes of different habitat subregions and their prognostic significance. Different habitat subregions within a tumor exhibit drastically different tissue elasticity response characteristics due to variations in cell density, stromal composition, and vascularization. Furthermore, this elasticity-habitat coupling relationship exhibits significant heterogeneity among different patients, directly determining the tumor's invasive potential and treatment resistance. This embodiment proposes an elasticity-habitat mutual attention mechanism. Using the spatial distribution of habitat subregions as query conditions and the elasticity attributes of corresponding regions in the elasticity feature map as key-value pairs, it adaptively enhances the expression of prognostic-sensitive elasticity-habitat coupling patterns by calculating cross-modal attention weights between habitat subregions and elastic attributes, while suppressing background elasticity signals irrelevant to prognosis. This achieves deep fusion of habitat and elasticity features in the feature space. The principle of the elasticity-habitat mutual attention mechanism is as follows: Figure 7 As shown.

[0060] For the K habitat subregions obtained by clustering through Gaussian mixture model, the radiomics features of each subregion are extracted. The feature vector of each subregion is mapped to a query vector of fixed dimension through a fully connected layer. The query vector encodes the biological phenotypic information of the habitat subregion, including its grayscale texture characteristics and spatial morphological features.

[0061] The elasticity feature map is divided into several local regions corresponding to the spatial location of habitat subregions. The distribution of elasticity invariants in each local region is statistically modeled, and its first-order statistics and Haralick texture features are extracted. These are then mapped to key vectors and value vectors through another fully connected layer, which respectively represent the elasticity attribute identifiers of the region and the prognostic information they contain.

[0062] For each habitat subregion's query vector, its similarity to the key vectors of all local elastic regions is calculated and normalized to obtain the attention weight matrix. The calculation of this attention weight follows the formula below:

[0063] In the formula, For the first Each habitat subregion is related to the first The normalized attention weights for each local elastic region represent the relative importance of the elastic-habitat coupling relationship between the habitat subregion and the elastic region. Their values ​​range from (0, 1) and satisfy the condition that for all... Normalization constraints that sum to 1; For the natural constant An exponential function with base 1 maps the scaled similarity score to a positive real number, ensuring the non-negativity of the attention weights; For the first The query vector for each habitat subregion is obtained by mapping the radiomics features of that subregion through a fully connected layer. It encodes the biological phenotypic information of that habitat subregion, such as grayscale texture characteristics and spatial morphological features. Its dimension is determined by the hidden layer dimension preset by the model. It is a learnable linear transformation matrix that acts on the key vector to map the query vector and the key vector to the same embedding subspace to calculate their inner product similarity. Its dimension is determined by the query vector dimension and the embedding space dimension. For the first The key vector of each local elastic region is obtained by mapping the statistical features of the elastic invariants within that region through a fully connected layer. It represents the elastic attribute identifier of the region and is used for similarity matching with the query vector. To determine the dimensions of the query vector and the key vector; The gated activation function is applied to the output of the gated channel. A nonlinear transformation is introduced to adaptively filter the elastic habitat coupling mode that is significantly related to the prognosis and suppress the interference of irrelevant modes. It is a learnable gated transformation matrix; This represents the total number of local elastic regions, i.e., the number of regions divided by the elastic feature map corresponding to the spatial location of the habitat subregion.

[0064] Using the obtained attention weight matrix, the value vectors of each local elastic region are weighted and aggregated to obtain the elastic context enhancement vector of each habitat subregion. The enhancement vector is then residually connected with the original radiomics features of the habitat subregion to form an elastic-habitat fusion feature, which is used for subsequent prognostic risk score generation.

[0065] During model training, in addition to the survival analysis loss, a supervised contrast loss (temperature coefficient of 0.1 and loss weight of 0.1) is introduced to constrain patients with the same prognostic outcome to be close to each other in the fusion feature space and patients with different prognostic outcomes to be far apart, so as to further enhance the ability of fusion features to distinguish survival outcomes.

[0066] To verify the effectiveness of the feature fusion method based on the elasticity-habitat mutual attention mechanism, data from 572 patients with locally advanced gastric cancer were used as experimental subjects. These patients were randomly divided into a training set of 400 cases and a validation set of 172 cases at a 7:3 ratio. A multi-center dataset of 254 cases was used as an independent external test set. The comparative method employed a conventional feature vector direct concatenation fusion strategy, where radiomics features of each habitat subregion were concatenated with global elasticity features and directly input into the LASSO-Cox regression model. Both methods were trained on the same training set, and the same 10-fold cross-validation strategy was used to determine the optimal hyperparameters. Performance was evaluated on the same validation and external test sets. The concordance index and the area under the receiver operating characteristic (ROC) curve for three-year overall survival prediction were used as evaluation metrics. Experimental results show that the consistency index on the validation set using the conventional splicing fusion method is 0.714, and the area under the curve (AUC) for three-year survival prediction is 0.80. However, after adopting the elastic-habitat mutual attention mechanism fusion method, the consistency index on the validation set increases to 0.763, the AUC for three-year survival prediction increases to 0.84, and the consistency index on the external test set increases from 0.727 to 0.775. Further ablation experiments show that removing the gated activation function results in the most significant performance degradation, with the validation set consistency index falling back to 0.731. This confirms that the gating mechanism plays a crucial role in adaptively selecting prognostic-related elastic-habitat coupling patterns, effectively suppressing the interference of background elastic signals unrelated to prognosis on the feature fusion process, thereby significantly enhancing the prognostic discriminative ability of the fused features.

[0067] Example 3: Neoadjuvant chemotherapy is not only a treatment to reduce tumor volume, but also a dynamic perturbation process that exerts selective pressure on the tumor's imaging features. During this process, chemotherapy-sensitive tumor cells are eliminated, while drug-resistant clonal subpopulations may survive and undergo spatial remodeling. The relative volume, texture features, and elastic properties of different habitat subregions exhibit regular evolutionary trajectories. These trajectories contain key prognostic information such as the clonal composition, invasive potential, and recurrence risk of residual lesions after chemotherapy. However, existing methods only use static imaging features before or after treatment for prediction in isolation, failing to capture the continuous evolution of tumor imaging features during neoadjuvant chemotherapy and losing dynamic prognostic signals during the treatment response process. This embodiment proposes a tumor imaging feature evolution modeling method based on ordinary differential equations. Using habitat features before and after treatment as observation nodes, it fits the continuous evolutionary trajectory of each habitat subregion's features in phase space and extracts evolutionary feature parameters characterizing the evolution speed and direction, thereby extending static imaging features into dynamic evolutionary features.

[0068] The habitat subregion feature vectors of corresponding spatial locations in the CT images of the same patient before and after treatment are arranged in chronological order to form a multidimensional time series. The time stamp before treatment is defined as time zero, and the time stamp after treatment is defined as the normalized treatment cycle length. At the same time, the actual number of neoadjuvant chemotherapy cycles for each patient is recorded.

[0069] The first-order statistical features and texture features of each habitat subregion are used as state variables, and the normalized average elasticity coefficient derived from the elasticity feature map is defined as the tissue sclerosis index. Together, they constitute a phase space coordinate system describing the characteristic state of tumor imaging, where the horizontal axis represents the biological heterogeneity dimension and the vertical axis represents the elastic sclerosis dimension.

[0070] Assuming the evolution trajectory of tumor imaging features under chemotherapy perturbation can be described by a system of first-order ordinary differential equations, and defining the time derivative of the state vector as the sum of the elastic driving term and the habitat interaction term, its evolution equation satisfies the following formula:

[0071] In the formula, for The tumor imaging feature state vector at time t, whose components are the tumor image features of each habitat subregion at time t. The feature set, including first-order statistical features and texture feature parameters, is used to characterize the real-time biological state of each subregion within the tumor. The elastic-biological coupling coefficient is used to regulate the relative contribution of elastic driving terms in the overall evolution process, reflecting the extent to which tissue elastic properties regulate the remodeling of tumor imaging features. The gradient of the elastic potential energy function with respect to the state vector characterizes the sensitivity and direction of elastic properties to changes in habitat state, indicating the evolutionary trend driven by elastic potential difference between different habitat subregions. This is an interaction matrix between habitat subregions, where the diagonal elements represent the self-limiting growth coefficient of each habitat subregion, and the off-diagonal elements represent the promoting or competitive inhibitory effects between different habitat subregions. It is used to model the mutual influence between tumor cell populations with different biological characteristics under chemotherapy stress. It is a pure vector with all elements being 1, and its dimension is the same as that of the state vector. It is used to construct the relative growth rate factor. The ratio of the state vector to the carrying capacity vector is the element-wise ratio, representing the current biological load of each habitat subregion.

[0072] Using observational data from two time points before and after treatment as boundary conditions, the adjoint sensitivity analysis method was employed to optimize the solution of evolutionary characteristic parameters such as the elastic-biological coupling coefficient, the interaction matrix, and the environmental carrying capacity vector, so that the evolutionary trajectory best fits the observational data in the least squares sense.

[0073] Based on the evolutionary characteristic parameters obtained from the solution, the evolutionary rate constant and steady-state asymptotic value of each habitat subregion are calculated and used as time-series evolutionary characteristics, which are then incorporated into the prognostic prediction model together with static habitat characteristics and elasticity characteristics.

[0074] To validate the incremental prognostic value of a temporal modeling method based on the evolution of tumor imaging features, 572 patients with complete pre- and post-treatment portal venous phase CT images were selected as experimental subjects and randomly divided into a training set of 457 patients and a test set of 115 patients at an 8:2 ratio. The comparative method involved directly inputting pre- and post-treatment static habitat and elasticity features into the prognostic prediction model without introducing evolutionary equations for continuous trajectory modeling. Both methods employed identical image preprocessing procedures, habitat subregion division models, and elasticity feature extraction parameters. Features were selected using LASSO-Cox regression and then incorporated into a multivariate Cox proportional hazards regression model. The independent prognostic significance of temporal features in the multivariate Cox regression, the increment of the likelihood ratio test statistic, and the consistency index of the test set were used as evaluation indicators. Experimental results show that the method of directly splicing static features has a consistency index of 0.721 on the test set, and the hazard ratio of time-series features in multivariate Cox regression is 1.46, but it does not reach the level of independent prognostic significance. However, after introducing evolutionary modeling, the steady-state asymptotic value feature extracted from the evolutionary trajectory has a hazard ratio of 2.13 in multivariate Cox regression and reaches the level of independent prognostic significance. The likelihood ratio test statistic increases significantly, and the test set consistency index improves to 0.768. Patients were divided into high-evolution and low-evolution groups using the evolution rate constant and the median of the steady-state asymptotic value as thresholds. Kaplan-Meier survival curve analysis showed that the survival difference between the two groups was statistically significant, indicating that evolutionary feature parameters can effectively distinguish different remodeling patterns of tumor imaging features after chemotherapy, providing dynamic prognostic information that static features cannot capture for long-term survival prediction.

[0075] Example 4: Existing prognostic prediction models can only output the survival probability of patients under a given treatment regimen, but cannot answer the more clinically valuable counterfactual question: how would the survival outcome change if the same patient underwent surgery directly without neoadjuvant chemotherapy? Therefore, they cannot quantify the individualized survival benefit of neoadjuvant chemotherapy for each patient. Since it is impossible to observe the outcomes of both treatment pathways simultaneously for the same patient in reality, this embodiment proposes an individualized treatment benefit assessment method based on a counterfactual reasoning framework. Utilizing the distribution of imaging and pathological characteristics of patients with different treatment pathways in observational data, and through propensity score matching and a counterfactual outcome prediction model, it infers the conditional survival probability of each patient under the two conditions of receiving and not receiving neoadjuvant chemotherapy. The difference between the two is the individualized treatment benefit, providing a direct basis for decisions on upgrading or downgrading postoperative adjuvant therapy.

[0076] Using pre-treatment CT image habitat and elasticity characteristics, pre-treatment clinical stage, and serum tumor markers as covariates, a binary propensity score model was constructed to estimate the conditional probability of each patient being assigned to the neoadjuvant chemotherapy pathway. Since all patients in this dataset received neoadjuvant chemotherapy, the untreated cohort was constructed as follows: a cohort of patients who met the same inclusion / exclusion criteria but did not receive neoadjuvant chemotherapy and underwent radical gastrectomy was selected from publicly available gastric cancer databases and matched with this dataset using propensity score matching to construct a comparable pseudo-control cohort of treatment and untreated groups. For cases where external matching samples were unavailable, a synthetic control method was used: based on the pre-treatment clinicopathological and imaging characteristics of patients in this dataset, a virtual untreated control population was synthesized using inverse probability weighting or covariate-balanced propensity score weighting. The survival outcome distribution of this virtual untreated control population was determined by referencing survival data of untreated patients in similar real-world studies.

[0077] In the treatment cohort, the aforementioned multidimensional prognostic prediction model was used to estimate the conditional survival function after neoadjuvant chemotherapy, while in the untreated cohort, a simplified prognostic model containing only preoperative imaging and clinical features was used to estimate the conditional survival function after direct surgery. Both models adopted a time-series prediction architecture with equidistant nodes to ensure consistency in output scale.

[0078] By using a multi-task learning framework, the feature extraction parts of the two treatment pathway-specific prognostic prediction models are shared. At the same time, a maximum mean difference constraint term is introduced into the loss function to correct the feature distribution shift between different treatment groups, ensuring that the image features extracted by the two models are comparable at the same implicit spatial scale.

[0079] For each patient receiving neoadjuvant chemotherapy, their pre-treatment imaging features were input into the prognostic models of the treatment group and the untreated group, respectively, to obtain two counterfactual survival curves. The difference in the limiting mean survival time within a specified time range was calculated, which is the individualized treatment benefit for that patient.

[0080] The calculated individualized treatment benefits are jointly modeled with the patient's clinicopathological characteristics and habitat resilience risk score. Subgroup characteristic combinations of the benefit population are determined through recursive partitioning analysis. Clinically operable stratification rules are established to guide postoperative adjuvant chemotherapy decisions, so that high-benefit patients continue to receive adjuvant therapy while low-benefit patients reduce the intensity of treatment or are only followed up.

Claims

1. A prognostic prediction system for neoadjuvant chemotherapy in gastric cancer based on habitat and elastography, characterized in that, include: The image annotation module is used to acquire pre-treatment CT images of patients with locally advanced gastric cancer and annotate the tumor region of interest. The habitat imaging module is used to extract local radiomics features of each voxel within the region of interest of the tumor, and to divide the voxels into multiple habitat subregions with different biological characteristics through Gaussian mixture model clustering, and to extract radiomics features of each habitat subregion. The elastic imaging module constructs elastic invariants that are invariant to non-rigid deformation based on the gradient field and Hessian matrix of the CT image, and uses nonlinear monotonic mapping to perform signal suppression and dynamic range compression on the elastic invariants to generate an elastic feature map, and then extracts elastic texture features from the elastic feature map. The prognostic prediction module generates habitat risk scores by aggregating radiomics features of each habitat subregion and generates elasticity risk scores by utilizing the elastic texture features. It then integrates the habitat risk scores, elasticity risk scores, and patient clinicopathological information using a multivariate survival regression model to construct a prognostic prediction model for outputting postoperative survival prediction results.

2. The system according to claim 1, characterized in that: The local image omics features extracted by the habitat imaging module include at least first-order entropy features and multiple texture features, and the optimal number of clusters in the Gaussian mixture model clustering is automatically determined based on the Bayesian information criterion.

3. The system according to claim 1, characterized in that: The elastic invariant is constructed by forming a quadratic form from the gradient vector of the CT image and the inverse of the Hessian matrix to obtain a differential affine elastic invariant that reflects the local elastic properties of the tissue. The nonlinear monotonic mapping includes a combination of square root operation and arctangent transformation, which is used to compress the numerical range of the elastic invariant to a preset finite interval, and suppress noise and enhance weak elastic signals during the compression process.

4. The system according to claim 3, characterized in that, The process of generating the elastic feature map also includes: The elastic invariants processed by nonlinear monotonic mapping are grayscaled and combined with a mask of the region of interest of the tumor to generate an elastic feature map that retains only the elastic information inside the tumor.

5. The system according to claim 1, characterized in that, The extraction process of the elastic texture features includes: A gray-level co-occurrence matrix is ​​constructed on the elastic feature map along multiple preset three-dimensional spatial directions, and extended Haralick texture feature parameters are extracted from the gray-level co-occurrence matrix in each direction.

6. The system according to claim 1, characterized in that, The specific methods by which the prognostic prediction module generates habitat risk scores include: For each patient, the radiomics feature vectors belonging to the same habitat subregion were reduced to a single principal component through principal component analysis, and the corresponding feature values ​​for missing habitat subregions were set to zero. Subsequently, the LASSO-Cox regression model was used to screen features and the habitat risk score was obtained by weighted summation of coefficients.

7. The system according to claim 1, characterized in that: The prognosis prediction module receives post-treatment CT images of the same patient, uses the habitat imaging module to divide the tumor region of the post-treatment images into habitat subregions and extract post-treatment habitat features, calculates the relative change of habitat features before and after treatment, and integrates the relative change into the multivariate survival regression model.

8. The system according to claim 1, characterized in that: When generating prognostic prediction results, the prognostic prediction module employs an inter-feature attention mechanism to deeply fuse the radiomics features and elasticity features of the habitat subregions. This mechanism uses the biological phenotypic features of each habitat subregion as query conditions and the elasticity features of each local location in the tumor region as matching objects. It adaptively calculates attention weights to filter elasticity-habitat coupling patterns that are significantly related to prognosis, and generates elasticity context-enhanced fusion features based on these coupling patterns for prognostic risk score calculation.

9. The system according to claim 1, characterized in that: The prognostic prediction module is also configured to construct a tumor imaging feature state evolution model based on CT images of the same patient before and after treatment. This model defines the tumor imaging feature state by combining radiomics features and elasticity features of each habitat subregion. It describes the dynamic evolution trajectory of each habitat subregion under elastic driving and habitat interaction during neoadjuvant chemotherapy using continuous temporal differential equations. It extracts evolutionary feature parameters that characterize the evolution rate and direction from the evolutionary trajectory and incorporates the evolutionary feature parameters as temporal evolution features into the prognostic prediction.

10. A method for predicting the prognosis of neoadjuvant chemotherapy for gastric cancer based on habitat and elastography, characterized in that: The method is implemented based on the system described in any one of claims 1-9: The method includes: Acquire pre-treatment CT images of patients with locally advanced gastric cancer and annotate the tumor region of interest; Local radiomics features of each voxel within the region of interest of the tumor are extracted, and Gaussian mixture model clustering is used to divide the voxels into multiple habitat subregions, and radiomics features of each habitat subregion are extracted. Based on the gradient field and Hessian matrix of the CT image, an elastic invariant that is invariant to non-rigid deformation is constructed. The elastic invariant is then subjected to signal suppression and dynamic range compression using a nonlinear monotonic mapping to generate an elastic feature map. Elastic texture features are then extracted from the elastic feature map. The image omics features of each habitat subregion are aggregated to generate a habitat risk score, and the elastic texture features are used to generate an elastic risk score. A multivariate survival regression model was used to integrate the habitat risk score, elasticity risk score and patient clinicopathological information to construct a prognostic prediction model and output postoperative survival prediction results.

Images