Glioma molecular subtype prediction method based on magnetic resonance image and machine learning

Abstract

The invention discloses a glioma molecular subtype prediction method based on a magnetic resonance image and machine learning, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining magnetic resonance image data and clinical data of a to-be-predicted object, and constructing domain attribute information; according to the magnetic resonance image data and the domain attribute information, outputting biological phenotypic features and domain related features; mapping the biological phenotypic characteristics into image area characteristics according to the distribution of the biological phenotypic characteristics in the image space, constructing a response relationship between each image area and molecular subtype prediction, and dynamically adjusting the image area characteristics; coding the clinical data, and carrying out conditional modulation on the adjusted image region features to form multi-modal joint features; predicting the molecular subtype of the glioma according to the multi-modal joint features, and outputting a molecular subtype prediction result; according to the invention, through domain attribute modeling and multi-modal feature fusion, the problems of multi-source heterogeneous data and low prediction precision are solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for predicting molecular subtypes of gliomas based on magnetic resonance imaging and machine learning. Background Technology

[0002] Gliomas are a type of highly aggressive central nervous system tumor, with significant individual variations in clinical presentation and prognosis. In recent years, molecular subtyping has become a crucial basis for glioma diagnosis, treatment, and prognostic assessment. Different molecular subtypes exhibit significant differences in gene mutations, epigenetic modifications, and signaling pathway activity, directly impacting patient treatment response and survival. However, in clinical practice, relying solely on traditional imaging or pathological methods is insufficient for accurately and non-invasively determining glioma molecular subtypes. Therefore, how to utilize magnetic resonance imaging and clinical data for molecular subtype prediction has become a cutting-edge issue in precision oncology research.

[0003] With the rapid development of computer vision and deep learning technologies, machine learning-based radiomics methods have gradually become a research hotspot. Magnetic resonance imaging, especially MRI and multi-sequence, multimodal imaging, provides rich spatial information on tumor tissue structure, angiogenesis, and microenvironment. By extracting features and recognizing patterns from image data, morphological features, texture features, and dynamic changes of tumors can be obtained. There is a potential correlation between these imaging phenotypes and molecular features, providing possibilities for non-invasive prediction. At the same time, clinical data such as patient age, gender, disease course, and laboratory indicators also contain prior information related to tumor molecular characteristics. Integrating imaging information with clinical information is expected to improve the reliability and biological interpretability of predictions.

[0004] For example, the invention patent with publication number CN117766124A discloses a method for predicting the genotype of glioma IDH based on multimodal deep learning. This method preprocesses multiple MRI sequences, extracts features of the glioma region in each MRI sequence image, and performs feature fusion. Based on the fused features, it predicts the genotype of glioma IDH. Finally, it trains the model and optimizes the parameters. This invention designs a feature extractor to obtain local-global features of each MRI sequence; it uses a glioma segmentation task to guide the feature extractor to automatically capture features of the glioma region in each MRI sequence; it calculates the feature similarity between multiple MRI sequences to perform cross-modal feature enhancement and uses an attention mechanism to fuse the features. This not only removes redundancy but also fully utilizes the complementary information between multimodal MRI sequences, achieving more accurate IDH genotype prediction.

[0005] For example, the invention patent with publication number CN112215826A discloses a method for predicting molecular subtypes and prognosis of gliomas based on deep image features, including the following steps: Step 1, constructing a dataset including magnetic resonance imaging (MRI) scans of multiple patients with primary gliomas; Step 2, constructing a feature extractor based on the RA-UNet neural network; Step 3, constructing a feature optimizer; Step 4, constructing a DeepRA model based on deep image features, including a primary glioma molecular subtype predictor and a patient survival time predictor; Step 5, evaluating the performance of the primary glioma molecular subtype predictor and the patient survival time predictor; Step 6, using the molecular subtype predictor and the survival time predictor to predict the MRI scans of patients with primary gliomas. Using deep image features to predict the molecular subtypes of primary gliomas and patient survival time improves the accuracy of these predictions compared to previous methods.

[0006] The above-disclosed technical solutions have at least the following technical problems: Traditional techniques often ignore domain-related biases in images during feature extraction, mixing all features before inputting them into the prediction model. This makes the prediction results susceptible to domain noise and lack interpretability. Furthermore, traditional methods typically input overall image features into the model, neglecting the different contributions of local regions to molecular subtypes, resulting in insufficient sensitivity of the model to local lesions. To address these issues, this invention proposes a solution. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for predicting molecular subtypes of glioma based on magnetic resonance imaging and machine learning. By using domain attribute modeling and multimodal feature fusion, the method addresses the problems of multi-source heterogeneous data and low prediction accuracy.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning includes: acquiring magnetic resonance imaging data and clinical data of the object to be predicted, and constructing domain attribute information of the magnetic resonance imaging data source; performing separate extraction processing of image features based on the magnetic resonance imaging data and domain attribute information, and outputting biological phenotypic features and domain-related features; mapping biological phenotypic features to image region features based on the distribution of biological phenotypic features in the image space, and constructing the response relationship between each image region and molecular subtype prediction, and dynamically adjusting the image region features according to the response relationship; encoding clinical data, and conditionally modulating the adjusted image region features to form multimodal joint features; predicting the molecular subtype of glioma based on the multimodal joint features, and outputting the molecular subtype prediction result.

[0009] In a preferred technical solution, the acquisition of magnetic resonance imaging (MRI) data and clinical data of the object to be predicted, and the construction of domain attribute information of the MRI data source, are specifically as follows: MRI data of the object to be predicted are acquired from different data acquisition sources; for each MRI data, corresponding image acquisition condition information is extracted, and the image acquisition condition information is digitized into an image acquisition condition data set; based on the image acquisition condition data set, the MRI data is classified into several data source domains; based on each data source domain, corresponding domain attribute information is constructed, and the domain attribute information is bound to the corresponding MRI data; based on the domain attribute information, corresponding preprocessing strategies are selected for the MRI data of different data source domains to reduce the differences in image distribution between different data source domains; clinical data corresponding to the MRI data is acquired, and the clinical data is associated and stored with its corresponding MRI data and domain attribute information.

[0010] In a preferred technical solution, the step of constructing corresponding domain attribute information based on each data source domain is as follows: within each data source domain, output the statistical features of the image data; based on the statistical features, perform cluster analysis on the images within the domain using the K-means clustering algorithm to obtain the cluster category and corresponding cluster center for each image; generate a domain-internal representation distribution vector based on the clustering results; input the domain-internal representation distribution vector of each data source domain into a lightweight neural network for mapping to generate a corresponding domain attribute vector; perform enhancement processing on the domain attribute vector of each domain by fusing the statistical features of different sequence images within the domain to generate an enhanced domain attribute vector.

[0011] In a preferred embodiment, the step of performing a separate extraction process of image features based on magnetic resonance image data and domain attribute information to output biological phenotypic features and domain-related features is as follows: Based on the magnetic resonance image data, basic feature values ​​of pixels are extracted in the image space, and the basic feature values ​​are combined to form an image feature representation; the domain attribute vectors in the domain attribute information are aligned with the image feature representation; based on the aligned domain attribute vectors, the feature components in the image feature representation are weighted and differentiated to form candidate domain-related feature components and candidate biological phenotypic feature components; based on the labeling results of the candidate biological phenotypic feature components and candidate domain-related feature components, the image feature representation is grouped to form a first feature set and a second feature set that are mutually distinguishable; a cross-data source domain consistency constraint is applied to the first feature set to form biological phenotypic features; a domain differentiation constraint is applied to the second feature set to obtain domain-related features; a feature independence constraint is applied between the biological phenotypic features and the domain-related features to output the final biological phenotypic features and domain-related features.

[0012] In a preferred embodiment, the step of weighting and adjusting the feature components in the image feature representation based on the aligned domain attribute vector, and distinguishing them based on the correlation between each feature component and the domain attribute vector, is as follows: After aligning the domain attribute vector with the features of the image feature representation, a correspondence is established between the aligned domain attribute vector and each feature component in the image feature representation; based on the correspondence between the feature components and the domain attribute vector, the correlation value between the feature components and the domain attribute vector is output using the cosine similarity method; the feature components in the image feature representation are weighted and adjusted based on the correlation value; based on the magnitude of the correlation value, the feature components in the image feature representation are distinguished; feature components with a correlation value higher than a preset threshold or at the top of the correlation distribution are marked as candidate domain-related feature components; feature components with a correlation value lower than a preset threshold or at the bottom of the correlation distribution are marked as candidate biological phenotypic feature components.

[0013] In a preferred embodiment, the application of domain-discriminating constraints to the second feature set to obtain domain-related features specifically involves: extracting the second feature set from different data source domains into a matrix representation; establishing a domain association mapping relationship for each feature component in the second feature set based on the corresponding domain attribute vector; grouping the feature components in the second feature set according to their corresponding data source domains based on the domain association mapping relationship; analyzing the distribution information of each feature component in the feature group within the data source domain to form the intra-domain statistical features of the corresponding data source domain; comparing the intra-domain statistical features of each data source domain with the feature groupings between different data source domains pairwise to output the inter-domain statistical features; using the intra-domain statistical features of each data source domain as the distribution reference center; outputting the distance between each feature component in the second feature set and the intra-domain statistical reference center of its corresponding data source domain based on the feature space distance measurement method to obtain the same-domain deviation, and outputting the distance difference between the intra-domain statistical reference centers of other data source domains to obtain the inter-domain relative difference; weighting the feature components in the second feature set according to the same-domain deviation and inter-domain relative difference; and outputting the second feature set adjusted by the domain-discriminating constraints as the domain-related features.

[0014] In a preferred embodiment, the step of mapping biological phenotypic features to image region features based on their distribution in image space, constructing a response relationship between each image region and molecular subtype prediction, and dynamically adjusting the image region features according to the response relationship is as follows: Based on the spatial coordinate information of biological phenotypic features in image space, a one-to-one correspondence is established between biological phenotypic features and pixel positions in magnetic resonance imaging data, forming a positional mapping relationship between biological phenotypic features in image space; according to the positional mapping relationship, the magnetic resonance imaging data is divided into regions, forming several mutually distinguishable image regions; biological phenotypic features located within the same image region are aggregated, and image region features corresponding to the image region are generated by statistically analyzing the biological phenotypic features within the image region; based on the image region features, a response relationship is established between each image region and the molecular subtype prediction result; and the image region features are dynamically adjusted according to the response relationship.

[0015] In a preferred embodiment, the process of encoding clinical data and conditionally modulating the adjusted image region features to form multimodal joint features is as follows: The clinical data of the object to be predicted is preprocessed by encoding discrete features, normalizing continuous features, and assigning different weights based on the relative importance of features in the prediction task, thus forming a clinical condition vector; the adjusted image region features are used as basic image information, and a one-to-one correspondence is established between them and the corresponding clinical condition vector for each image region; for each image region feature, multi-level modulation is performed based on the corresponding clinical condition vector; during the conditional modulation process, features of adjacent image regions are smoothly fused according to the spatial neighborhood relationship between image regions; the conditionally modulated image region features and the corresponding clinical condition vectors are statistically converged to generate multimodal joint features containing image phenotypic information and clinical prior information.

[0016] In a preferred technical solution, the multi-level modulation of each image region feature based on the corresponding clinical condition vector is performed as follows: Each feature value in the clinical condition vector is multiplied by a preset weight using a linear weighting method, and applied to the corresponding dimension of the image region feature to form a first-level modulation feature; for the feature dimensions in the clinical conditions that contribute significantly to image region prediction, the image region feature is non-linearly scaled to form a second-level modulation feature; during the modulation process, based on the spatial neighborhood relationship between image regions, the modulation features of adjacent image regions are weighted and averaged to obtain a spatial fusion feature; the first-level modulation feature, the second-level modulation feature, and the spatial fusion feature are fused using a feature stitching method to form the conditionally modulated image region feature.

[0017] In a preferred embodiment, the step of predicting the molecular subtype of glioma based on multimodal joint features and outputting the molecular subtype prediction result is as follows: A molecular subtype prediction model is established based on the multimodal joint features, and the output is the prediction probability of each candidate molecular subtype; during the model training phase, supervised training is performed using labeled glioma case data, enabling the model to learn the mapping relationship between the multimodal joint features and known molecular subtypes, thereby predicting new samples; based on the prediction probability, the final molecular subtype prediction result is determined using the maximum probability selection method.

[0018] The technical effects and advantages of the glioma molecular subtype prediction method based on magnetic resonance imaging and machine learning in this invention are as follows: 1. This invention achieves high-precision and interpretable prediction of glioma molecular subtypes through magnetic resonance imaging data fusion and feature separation processing, demonstrating significant technical effects and advantages. Firstly, this invention acquires magnetic resonance imaging data and clinical data of the target object and constructs domain attribute information of the magnetic resonance imaging data sources, effectively solving the heterogeneity problem of multi-source image data in terms of acquisition conditions, imaging parameters, and distribution. By modeling domain attributes of different data source domains and combining them with preprocessing strategies to unify image distribution, this invention can significantly reduce distribution differences between multi-source data, ensuring the reliability and consistency of subsequent feature extraction, thereby improving the model's generalization ability.

[0019] 2. This invention employs a segregated extraction strategy for image features, dividing them into biological phenotypic features and domain-related features. Cross-domain consistency constraints and domain-discrimination constraints are applied to each category, while feature independence constraints ensure that the two types of features do not interfere with each other. This method not only preserves information highly correlated with the biological characteristics of the disease in the image but also eliminates domain bias introduced by different data sources, achieving efficient feature expression and reliable separation, thus providing a solid foundation for accurate prediction. Furthermore, this invention utilizes image spatial information to map biological phenotypic features to image region features and establishes a response relationship between image regions and molecular subtype prediction. By dynamically adjusting region features, the model can capture the sensitivity of local features in the image to molecular subtypes, improving the spatial resolution and accuracy of prediction. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of clinical data for the present invention; Figure 3 This is a schematic diagram of MRI image feature extraction according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1, Figure 1 This invention presents a method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning, including: S1, acquire the magnetic resonance imaging data and clinical data of the object to be predicted, and construct the domain attribute information of the magnetic resonance imaging data source; In this embodiment, magnetic resonance imaging data and clinical data of the object to be predicted are obtained, and domain attribute information of the source of the magnetic resonance imaging data is constructed, as follows: Magnetic resonance imaging data of the object to be predicted are obtained from different data acquisition sources. The magnetic resonance imaging data includes image data formed under different medical institutions, different imaging equipment or different scanning protocols, including but not limited to multimodal magnetic resonance imaging data such as MRI, DWI and DSC-PWI, and a unique data identifier is assigned to each magnetic resonance imaging data. For each magnetic resonance imaging data, the corresponding image acquisition condition information is extracted, and the image acquisition condition information is digitized into an image acquisition condition data set, which includes the imaging device type and scanning protocol; Based on the data set of image acquisition conditions, the magnetic resonance image data are classified into domain categories. Magnetic resonance image data with similar image acquisition conditions are classified into the same data source domain, forming several data source domains. For each data source domain, corresponding domain attribute information is constructed. The domain attribute information is used to describe the overall imaging characteristics of the data source domain and is represented in vector form. The domain attribute information is bound to the corresponding magnetic resonance imaging data, so that each magnetic resonance imaging data carries its corresponding data source domain attribute when it is input. Based on the domain attribute information, select the corresponding preprocessing strategy for magnetic resonance image data from different data source domains to reduce the differences in image distribution between different data source domains; Acquire clinical data corresponding to magnetic resonance imaging data, and associate and store the clinical data with its corresponding magnetic resonance imaging data and domain attribute information, such as... Figure 2 As shown, the clinical data includes patient age, gender, and tumor clinical manifestation information, wherein the tumor clinical manifestation information includes tumor grade information and tumor location information.

[0023] In this embodiment, corresponding domain attribute information is constructed based on each data source domain, as detailed below: Based on the image acquisition conditions, multi-source images are divided into data source domains, and images with similar acquisition conditions are assigned to the same data source domain. Within each data source domain, the statistical characteristics of the image data are output, including the mean, variance, and histogram distribution. Based on statistical characteristics, the K-means clustering algorithm is used to perform cluster analysis on the images within the domain, so that each image is assigned to the nearest cluster center, and the cluster category and corresponding cluster center of each image are obtained. Based on the clustering results, an intra-domain representation distribution vector is generated. The intra-domain representation distribution vector includes cluster centers, sample weights within clusters, and statistical features such as mean and variance. Specifically, based on the clustering results, the cluster center vectors and category labels of all images within the domain are statistically summarized. By calculating the sample proportion, mean vector, and variance vector of each cluster, an intra-domain representation distribution vector is generated. The internal representation distribution vector of each data source domain is input into a lightweight neural network for mapping to generate a corresponding domain attribute vector. Specifically, for each data source domain, its internal representation distribution vector is input into the lightweight neural network. This neural network can consist of at least one fully connected layer, each layer containing several neurons, and can use non-linear activation functions (such as ReLU, Sigmoid). By mapping the input internal representation distribution of the domain to a fixed-dimensional domain attribute vector, the lightweight neural network compresses and encodes the overall image characteristics of each domain into a vector representation that can be used by subsequent models. The domain attribute vector of each domain is enhanced by fusing the statistical features of different image sequences within the domain to generate the enhanced domain attribute vector, which is the final domain attribute information.

[0024] In this embodiment, based on the domain attribute information, corresponding preprocessing strategies are selected for magnetic resonance image data from different data source domains, as follows: Based on the domain attribute vector in the domain attribute information, the data source domain to which each magnetic resonance image data belongs is identified, and the data source domain to which each magnetic resonance image data belongs is determined. For different data source domains, the preprocessing strategy mapping rule corresponding to the domain attribute information is invoked to determine the image preprocessing strategy that matches the data source domain. The image preprocessing strategy consists of a preprocessing data set associated with the domain attribute vector. The preprocessing data set includes image intensity normalization method, grayscale distribution matching degree, spatial resampling scale, and denoising. For magnetic resonance image data belonging to the same data source domain, image preprocessing is performed using a consistent preprocessing data set according to the determined image preprocessing strategy to ensure the consistency of the statistical distribution of image data within the data source domain. For magnetic resonance image data belonging to different data source domains, image preprocessing is performed according to their corresponding image preprocessing strategies to make magnetic resonance images from different data source domains tend to be consistent in terms of grayscale distribution, spatial resolution or statistical characteristics. Based on a defined preprocessing strategy, magnetic resonance image data within the corresponding data source domain are preprocessed to reduce the differences in grayscale distribution, spatial resolution, or statistical characteristics of image data from different data source domains, thereby reducing the differences in image distribution between different data source domains.

[0025] S2, based on magnetic resonance imaging data and domain attribute information, performs separate extraction processing of image features, and outputs biological phenotypic features and domain-related features; In this embodiment, based on magnetic resonance imaging data and domain attribute information, image feature extraction is performed separately to output biological phenotypic features and domain-related features, as detailed below: like Figure 3 As shown, based on magnetic resonance imaging data, specifically MRI image data, the basic feature values ​​of pixels are extracted in the image space, and the basic feature values ​​are combined to form an image feature representation. The basic feature values ​​include the image's grayscale information, spatial position relationship, and local structure. The domain attribute vectors in the domain attribute information are aligned with the image feature representation to make the domain attribute vectors and the image feature representation have a dimensional correspondence, so that the image feature representation can be modulated subsequently. Based on the aligned domain attribute vector, the feature components in the image feature representation are weighted and adjusted, and distinguished according to the correlation between each feature component and the domain attribute vector. Among them, the feature components with high correlation with the domain attribute vector are marked as candidate domain-related feature components, and the feature components with low correlation with the domain attribute vector are marked as candidate biological phenotypic feature components. Based on the labeling results of candidate biological phenotypic feature components and candidate domain-related feature components, the image feature representation is subjected to feature grouping processing to form a first feature set and a second feature set that are mutually distinguishable in the same image feature representation. If a feature component is labeled as a candidate biological phenotypic feature component, it is added to the first feature set; if a feature component is labeled as a candidate domain-related feature component, it is added to the second feature set. A cross-data source domain consistency constraint is applied to the first feature set to ensure that the distribution of the first feature set remains consistent in the corresponding images of different data source domains, thereby suppressing the differential information introduced by imaging equipment or scanning conditions and forming biological phenotypic features. A domain-discriminative constraint is applied to the second feature set so that the second feature set can maintain the feature differences between different data source domains, thereby extracting domain-related features that reflect differences in imaging conditions or image distribution; By imposing feature independence constraints between biological phenotypic features and domain-related features, the information overlap between the two types of features is reduced, thereby enhancing the stability and reproducibility of feature separation results and outputting the final biological phenotypic features and domain-related features.

[0026] In this embodiment, the feature components in the image feature representation are weighted and adjusted according to the aligned domain attribute vector, and distinguished based on the correlation between each feature component and the domain attribute vector, as follows: After aligning the domain attribute vector with the image feature representation, establish a correspondence between the aligned domain attribute vector and each feature component in the image feature representation; Based on the correspondence between feature components and domain attribute vectors, the correlation value between feature components and domain attribute vectors is output using the cosine similarity method; Based on the correlation value, the feature components in the image feature representation are weighted and adjusted, with the feature components with higher correlation values ​​being given larger adjustment weights and the feature components with lower correlation values ​​being given smaller adjustment weights, thereby changing the contribution ratio of different feature components in the image feature representation. Based on the magnitude of the correlation value, the feature components in the image feature representation are distinguished; Feature components with a relevance value higher than a preset threshold or at the top of the relevance distribution are marked as candidate domain-related feature components. Feature components at the top of the relevance distribution refer to feature components whose relevance values ​​are arranged in descending order and are within the preset proportion range. Feature components with correlation values ​​lower than a preset threshold or located at the end of the correlation distribution are marked as candidate biological phenotypic feature components. Feature components at the end of the correlation distribution refer to feature components whose correlation values ​​are arranged in descending order and fall within a preset range.

[0027] In this embodiment, a cross-data source domain consistency constraint is applied to the first feature set to ensure that the distribution of the first feature set remains consistent across images corresponding to different data source domains. This suppresses differential information introduced by imaging equipment or scanning conditions, thereby forming biological phenotypic features, as detailed below: The first feature set corresponding to different data source domains is extracted into a vector representation to describe the overall and local distribution of image features in each domain; Each feature component is associated with its corresponding domain attribute vector, so that each feature component can be driven by domain attribute information, forming a mapping relationship that can be used for weighted calculation. Based on the mapping relationship, weighted statistical features are calculated according to the domain attribute vector corresponding to each feature component to form dynamic distribution features. The weighted statistical features include weighted mean, weighted variance, and weighted covariance. The weighted statistical characteristics of different data source domains are compared, and the distribution differences between domains are output. Based on the differences in distribution between domains, local and global joint correction is performed on each image feature component of the first feature set. Specifically, global statistical features (mean vector, variance vector, or covariance matrix) are calculated for all images or feature components in the first feature set to form a global reference distribution. The image features are divided into local regions, and local statistical features (mean, variance, or covariance) are calculated for the feature components of each local region to form a local reference distribution. The differences between each image feature component and the global reference distribution and the reference distribution of its local region are calculated to obtain global correction coefficients and local correction coefficients. According to the preset weighting rules, each feature component is jointly corrected using a linear scaling method to make the feature distribution tend to be consistent globally while maintaining the structural characteristics of the local region. The first set of features, after weighted correction and local-global joint adjustment, is used as a consistent biological phenotypic feature across data source domains.

[0028] In this embodiment, a domain-discriminative constraint is applied to the second feature set to ensure that the second feature set maintains the feature differences between different data source domains, thereby extracting domain-related features that reflect differences in imaging conditions or image distribution, as detailed below: The second feature set from different data source domains is extracted into a matrix representation to characterize the feature distribution formed by each data source domain under imaging equipment, scanning parameters or acquisition conditions; Based on the corresponding domain attribute vector, a domain association mapping relationship is established for each feature component in the second feature set, so that the feature component and its data source domain maintain a clear correspondence. Based on the domain association mapping relationship, the feature components in the second feature set are grouped according to their corresponding data source domains, so that each group of feature components contains only image features from the same data source domain. The statistical data source domain contains the distribution information of each feature component of the feature group, including the mean, variance and covariance of the feature components, forming the domain statistical characteristics of the corresponding data source domain. The feature grouping is performed between different data source domains, and the intra-domain statistical features corresponding to each data source domain are compared pairwise to output the statistical differences between different data source domains, forming inter-domain statistical features. The statistical differences include mean differences, variance differences, and covariance differences. The statistical characteristics within each data source domain are used as the distribution reference center to characterize the concentrated distribution location of the second feature set within the same data source domain. Based on the feature space distance metric method, the distance between each feature component in the second feature set and the intra-domain statistical reference center corresponding to its data source domain is output to obtain the intra-domain bias, and the distance difference between the intra-domain statistical reference centers corresponding to other data source domains is output to obtain the inter-domain relative difference. Based on the same-domain bias and the relative difference between different domains, the feature components in the second feature set are weighted and adjusted so that the feature components from the same data source domain move closer to their corresponding domain statistical reference center. While performing weighted adjustments, feature components from different data source domains are kept spaced in the feature space to enhance feature differentiation between different data source domains. The second feature set, adjusted by the domain differentiation constraint, is output as a domain-related feature reflecting the characteristics of different data source domains.

[0029] In this embodiment, based on the corresponding domain attribute vector, a domain association mapping relationship is established for each feature component in the second feature set, as follows: Each feature component in the second feature set is represented as a feature vector and mapped to the same feature space along with the domain attribute vectors corresponding to each data source domain. In the feature space, the correlation between each feature component and attribute vectors of different domains is calculated using the cosine similarity method; The feature components are assigned to the data source domain with the highest correlation, thereby establishing a one-to-one mapping relationship between the feature components and their respective data source domains.

[0030] S3, based on the distribution of biological phenotypic features in the image space, maps biological phenotypic features to image region features, constructs the response relationship between each image region and molecular subtype prediction, and dynamically adjusts the image region features according to the response relationship; In this embodiment, based on the distribution of biological phenotypic features in the image space, the biological phenotypic features are mapped to image region features, and a response relationship between each image region and molecular subtype prediction is constructed. The image region features are then dynamically adjusted based on the response relationship, as follows: Based on the spatial coordinate information of biological phenotypic features in the image space, a one-to-one correspondence is formed between biological phenotypic features and pixel positions in magnetic resonance imaging data, thus establishing a positional mapping relationship between biological phenotypic features and image space. Based on the location mapping relationship, the magnetic resonance image data is divided into regions. Pixels that are spatially adjacent and have similar distributions of corresponding biological phenotypic features are grouped into the same image region, thereby forming several mutually distinguishable image regions in the image space. The similarity of biological phenotypic feature distribution is obtained by the Euclidean distance algorithm. Biological phenotypic features located within the same image region are aggregated, and image region features corresponding to the image region are generated by statistically analyzing the biological phenotypic features within the image region. Specifically, the biological phenotypic feature vectors corresponding to all pixels within the region are collected to form a feature set within the region, and this set is statistically aggregated. The feature set within the region includes calculating the mean, variance, and distribution pattern information (such as skewness or kurtosis) of each feature dimension. At the same time, a feature importance weighting strategy is combined to weight the feature components according to the potential contribution of each pixel or voxel feature to the prediction of molecular subtype, so as to highlight key region features and suppress non-key region features. Finally, the multidimensional statistics and weighted features are combined to form the image region features. Based on image region features, a response relationship between each image region and the molecular subtype prediction results is established. The response relationship characterizes the impact of changes in different image region features on the molecular subtype prediction output. The response relationship is formed by changing the features of a single image region and observing the changes in the molecular subtype prediction output, thereby quantifying the contribution information of different image region features to molecular subtype prediction. Based on the response relationship, the image region features are dynamically adjusted so that the image region features that contribute more to the prediction of molecular subtypes are enhanced in the feature representation, while the image region features that contribute less to the prediction of molecular subtypes are suppressed in the feature representation.

[0031] In this embodiment, the image region features are dynamically adjusted according to the response relationship, as follows: Based on the response relationship between image regions and molecular subtype prediction results, the contribution value of image region features to the prediction output is extracted, and the adjustment weight is determined based on the contribution value. The feature vector of each image region is linearly scaled according to the adjusted weights, so that the features of image regions that contribute more to the prediction are enhanced in the feature representation, while the features of image regions that contribute less to the prediction are suppressed accordingly in the feature representation. During the adjustment, based on the spatial neighborhood relationship between image region features, a weighted fusion method is used to maintain the continuity and structural information of adjacent region features, thereby forming dynamically adjusted image region features.

[0032] S4 encodes clinical data and conditionally modulates the adjusted image region features to form multimodal joint features; In this embodiment, clinical data is encoded, and the adjusted image region features are conditionally modulated to form multimodal joint features, as detailed below: The clinical data of the target to be predicted are preprocessed by encoding discrete features and normalizing continuous features. Different weights are assigned to features according to their relative importance in the prediction task to form a clinical condition vector. Discrete features include variables that can take finite class values, such as gender, pathological grade, tumor location, and previous treatment plan. These are converted into numerical feature vectors through encoding methods (such as one-hot encoding or integer encoding). Continuous features include variables that can take continuous values, such as age, blood biochemical indicators, tumor volume, or imaging measurements. The adjusted image region features are used as the basic image information, and a one-to-one correspondence is established with the clinical condition vector corresponding to each image region, so as to realize the conditional modulation of image region features under the guidance of clinical information. For each image region feature, multi-level modulation is performed based on the corresponding clinical condition vector, including nonlinear scaling and offset adjustment, so that the image region features not only retain their own spatial structure information, but also reflect the potential impact of clinical conditions on molecular subtype prediction. During conditional modulation, features of adjacent image regions are smoothly fused based on the spatial neighborhood relationship between image regions, so that the multimodal joint features maintain structural consistency in space, while strengthening the guiding role of clinical information on key image regions. The conditionally modulated image region features are statistically aggregated with the corresponding clinical condition vectors to generate multimodal joint features that include image phenotypic information and clinical prior information.

[0033] In this embodiment, for each image region feature, multi-level modulation is performed based on the corresponding clinical condition vector, as follows: By using a linear weighting method, each feature value in the clinical condition vector is multiplied by a preset weight and applied to the corresponding dimension of the image region feature to form the first-level modulation feature. The preset weight is assigned according to the relative importance of the feature in the prediction task. For feature dimensions that contribute significantly to image region prediction in clinical conditions, nonlinear scaling or offset adjustment (such as exponential scaling, logarithmic scaling, or translation offset) is applied to the image region features to form second-level modulation features, thereby enhancing key region features in multimodal joint features. During the modulation process, based on the spatial neighborhood relationship between image regions, the spatial fusion features are obtained by weighted averaging of the modulation features of adjacent image regions, which suppresses the influence of local outliers and maintains the spatial continuity and structural consistency of image region features. The adjacent image regions are determined based on the spatial location information of the image regions in the magnetic resonance image. Specifically, the spatial center coordinates of each image region are determined, and the spatial distance between each image region is output. When the spatial distance between two image regions is less than a preset distance threshold, the two are determined to be adjacent image regions. The first-level modulation features, the second-level modulation features, and the spatial fusion features are fused together by feature stitching to form conditionally modulated image region features.

[0034] S5, based on multimodal combined features, predicts the molecular subtype of glioma and outputs the molecular subtype prediction results; In this embodiment, the molecular subtype of glioma is predicted based on multimodal joint features, and the molecular subtype prediction results are output as follows: Based on multimodal joint features, a molecular subtype prediction model is established, and the output is the predicted probability of each candidate molecular subtype. The molecular subtype prediction model is trained using a deep learning method, with the multimodal joint features as input. The deep learning method is based on a deep neural network, which includes an input layer, several hidden layers, and an output layer. The input layer receives the multimodal joint features, the hidden layers perform feature mapping and fusion of the multimodal joint features through fully connected layers and nonlinear activation functions, and the output layer outputs the predicted probability of each candidate molecular subtype through a softmax function. During training, based on labeled glioma sample data, a supervised learning approach is adopted, and the network parameters are optimized using a cross-entropy loss function and a backpropagation algorithm. During the model training phase, supervised training is performed using labeled glioma case data, enabling the model to learn the mapping relationship between multimodal joint features and known molecular subtypes, thereby predicting new samples. Based on the predicted probabilities, the final molecular subtype prediction result is determined by selecting the one with the highest probability.

[0035] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0036] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0037] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0038] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0039] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning, characterized in that, include: Acquire magnetic resonance imaging data and clinical data of the object to be predicted, and construct domain attribute information of the source of magnetic resonance imaging data; Based on magnetic resonance imaging data and domain attribute information, image features are extracted separately to output biological phenotypic features and domain-related features. Based on the distribution of biological phenotypic features in the image space, biological phenotypic features are mapped to image region features, and the response relationship between each image region and molecular subtype prediction is constructed. The image region features are then dynamically adjusted based on the response relationship. Clinical data is encoded, and the adjusted image region features are conditionally modulated to form multimodal joint features; Based on multimodal combined features, molecular subtypes of gliomas are predicted, and the molecular subtype prediction results are output.

2. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 1, characterized in that, The process of acquiring magnetic resonance imaging (MRI) data and clinical data of the object to be predicted, and constructing domain attribute information of the MRI data source, is as follows: Obtain magnetic resonance imaging data of the object to be predicted from different data acquisition sources; For each magnetic resonance imaging data, the corresponding image acquisition condition information is extracted, and the image acquisition condition information is digitized into an image acquisition condition data set. Based on the data set of image acquisition conditions, magnetic resonance image data are classified into several data source domains. Based on each data source domain, construct corresponding domain attribute information and bind the domain attribute information to the corresponding magnetic resonance imaging data; Based on the domain attribute information, select the corresponding preprocessing strategy for magnetic resonance image data from different data source domains to reduce the differences in image distribution between different data source domains; Acquire clinical data corresponding to magnetic resonance imaging data, and associate and store the clinical data with its corresponding magnetic resonance imaging data and domain attribute information.

3. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 2, characterized in that, The process of constructing corresponding domain attribute information based on each data source domain is as follows: Within each data source domain, output the statistical characteristics of the image data; Based on statistical characteristics, the K-means clustering algorithm is used to perform cluster analysis on the images within the domain to obtain the cluster category and corresponding cluster center for each image; Generate domain-specific distribution vectors based on the clustering results; The domain-internal representation distribution vector of each data source domain is input into a lightweight neural network for mapping, generating the corresponding domain attribute vector; The domain attribute vectors of each domain are enhanced by fusing statistical features of different image sequences within the domain to generate enhanced domain attribute vectors.

4. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 1, characterized in that, The process involves separating and extracting image features based on magnetic resonance imaging data and domain attribute information, outputting biological phenotypic features and domain-related features, as detailed below: Based on magnetic resonance imaging data, the basic feature values ​​of pixels are extracted in the image space, and the basic feature values ​​are combined to form an image feature representation; The domain attribute vectors in the domain attribute information are aligned with the image feature representation; Based on the aligned domain attribute vector, the feature components in the image feature representation are weighted and adjusted, and then distinguished to form candidate domain-related feature components and candidate biological phenotypic feature components. Based on the labeling results of candidate biological phenotypic feature components and candidate domain-related feature components, the image feature representation is subjected to feature grouping processing to form a first feature set and a second feature set that are mutually distinguishable. A cross-data-source-domain consistency constraint is imposed on the first feature set to form biological phenotypic features; Applying domain-discriminating constraints to the second feature set yields domain-related features; Apply feature independence constraints between biological phenotypic features and domain-related features to output the final biological phenotypic features and domain-related features.

5. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 4, characterized in that, The process involves weighting and adjusting the feature components in the image feature representation based on the aligned domain attribute vector, and then distinguishing them according to the correlation between each feature component and the domain attribute vector, as detailed below: After aligning the domain attribute vector with the image feature representation, establish a correspondence between the aligned domain attribute vector and each feature component in the image feature representation; Based on the correspondence between feature components and domain attribute vectors, the correlation value between feature components and domain attribute vectors is output using the cosine similarity method; Based on the correlation value, the feature components in the image feature representation are weighted and adjusted; Based on the magnitude of the correlation value, the feature components in the image feature representation are distinguished; Feature components with a relevance value higher than a preset threshold or that are among the top features in the relevance distribution are marked as candidate domain-related feature components. Feature components with correlation values ​​below a preset threshold or located at the end of the correlation distribution are marked as candidate biological phenotypic feature components.

6. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 5, characterized in that, The application of domain-discriminating constraints to the second feature set yields domain-related features, as detailed below: The second feature sets from different data source domains are extracted into matrix representations; Based on the corresponding domain attribute vector, establish a domain association mapping relationship for each feature component in the second feature set; Based on the domain association mapping relationship, the feature components in the second feature set are grouped according to their corresponding data source domains; The distribution information of each feature component of the feature group within the data source domain forms the domain-specific statistical features of the corresponding data source domain; Group the features between different data source domains, compare the intra-domain statistical features of each data source domain pairwise, and output the inter-domain statistical features. The statistical characteristics within each data source domain are used as the distribution reference center. Based on the feature space distance metric method, the distance between each feature component in the second feature set and the intra-domain statistical reference center corresponding to its data source domain is output to obtain the intra-domain bias, and the distance difference between the intra-domain statistical reference centers corresponding to other data source domains is output to obtain the inter-domain relative difference. Based on the same-domain deviation and the relative difference between different domains, the feature components in the second feature set are weighted and adjusted. The second feature set, adjusted by the domain differentiation constraint, is output as the domain-related feature.

7. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 1, characterized in that, The process involves mapping biological phenotypic features to image region features based on their distribution in image space, constructing a response relationship between each image region and molecular subtype prediction, and dynamically adjusting the image region features based on this response relationship, as detailed below: Based on the spatial coordinate information of biological phenotypic features in the image space, a one-to-one correspondence is formed between biological phenotypic features and pixel positions in magnetic resonance imaging data, thus establishing a positional mapping relationship between biological phenotypic features and image space. Based on the location mapping relationship, the magnetic resonance imaging data is divided into regions to form several mutually distinguishable image regions; The biological phenotypic features located within the same image region are aggregated, and the corresponding image region features are generated by statistically analyzing the biological phenotypic features within the image region. Based on image region features, establish the response relationship between each image region and the molecular subtype prediction results; Based on the response relationship, the features of the image region are dynamically adjusted.

8. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 1, characterized in that, The process of encoding clinical data and conditionally modulating the adjusted image region features to form multimodal joint features is as follows: The clinical data of the target to be predicted are preprocessed by encoding discrete features, normalizing continuous features, and assigning different weights according to the relative importance of features in the prediction task to form a clinical condition vector. The adjusted image region features are used as the basic image information, and a one-to-one correspondence is established with the clinical condition vector corresponding to each image region. For each image region feature, multi-level modulation is performed based on the corresponding clinical condition vector; During conditional modulation, features of adjacent image regions are smoothly fused based on the spatial neighborhood relationship between image regions. The conditionally modulated image region features are statistically aggregated with the corresponding clinical condition vectors to generate multimodal joint features that include image phenotypic information and clinical prior information.

9. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 8, characterized in that, For each image region feature, multi-level modulation is performed based on the corresponding clinical condition vector, as follows: By using a linear weighting method, each feature value in the clinical condition vector is multiplied by a preset weight and applied to the corresponding dimension of the image region feature to form the first-level modulation feature; For the feature dimensions that contribute significantly to the prediction of image regions in clinical conditions, nonlinear scaling of image region features is applied to form second-level modulation features. During the modulation process, spatial fusion features are obtained by weighted averaging of the modulation features of adjacent image regions based on the spatial neighborhood relationship between image regions. The first-level modulation features, the second-level modulation features, and the spatial fusion features are fused together by feature stitching to form conditionally modulated image region features.

10. The method for predicting glioma molecular subtypes based on magnetic resonance imaging and machine learning according to claim 1, characterized in that, The molecular subtype of glioma is predicted based on multimodal joint features, and the molecular subtype prediction results are output as follows: Based on multimodal joint features, a molecular subtype prediction model is established, and the output is the predicted probability of each candidate molecular subtype; During the model training phase, supervised training is performed using labeled glioma case data, enabling the model to learn the mapping relationship between multimodal joint features and known molecular subtypes, thereby predicting new samples. Based on the predicted probabilities, the final molecular subtype prediction result is determined using the maximum probability selection method.

Images