Prognosis assessment method and system for GBM patients based on magnetic resonance imaging and pathological image fusion

Abstract

The present application relates to the field of image analysis technology, and in particular to a method and system for evaluating the prognosis of GBM patients based on magnetic resonance-pathology image fusion, the method comprising: obtaining a data set of GBM patients, and obtaining MRI images and pathology WSI, and obtaining transcriptome sequencing data; obtaining image-pathology fusion features based on MRI images and pathology WSI, and constructing an image-pathology omics model; obtaining image-pathology omics analysis results based on the image-pathology omics model and transcriptome sequencing data; using weighted gene co-expression network analysis and gene set enrichment analysis methods to analyze transcriptome sequencing data, obtain cross-biological pathways, and determine the key biological pathways behind the image-pathology omics features. The present application uses magnetic resonance and pathology images to obtain image-pathology fusion features and construct a fusion model to evaluate the incremental prognostic predictive value of image-pathology fusion omics labels for IDH wild-type GBM patients relative to traditional clinical pathological factors.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of image analysis technology, and in particular to a method and system for evaluating the prognosis of GBM patients based on magnetic resonance imaging and pathological image fusion. Background Art

[0002] Glioblastoma (GBM) is the most malignant, fastest-progressing, and least effective malignant tumor among central nervous system gliomas. GBM originates from glial cells or their precursor cells, and its incidence accounts for more than half of primary malignant brain tumors. Rapid recurrence and short survival are its most distinctive characteristics. The current consensus treatment for GBM is the largest surgical resection of the tumor, followed by standardized radiotherapy and temozolomide (TMZ) chemotherapy. Although the above-mentioned active treatments have effectively prolonged survival, the median survival time of GBM patients is still less than 15 months. Studies have shown that the shortened life expectancy of GBM patients is related to multiple factors, the most important of which is the complexity between different cells within the tumor and the immune microenvironment in which they are located. This high degree of heterogeneity on multiple levels jointly contributes to the malignant progression of the tumor and resistance to treatment.

[0003] Magnetic resonance imaging (MRI) is one of the most widely used medical imaging tools in clinical practice. MRI can provide detailed information about the location, size, range and adjacent brain tissue structures of tumors, and is also used to monitor treatment response and detect tumor recurrence. As a structural imaging sequence in MRI, Diffusion Tensor Imaging (DTI) is currently the only advanced sequence in MRI that can show the structure of brain fiber bundles in vivo. It uses the anisotropy of the diffusion movement of water molecules in brain tissue to show the microstructural characteristics of brain white matter fiber bundles and reflect the degree of destruction and infiltration of brain white matter fiber bundles by tumors.

[0004] Existing machine learning studies on medical images of isocitrate dehydrogenase (IDH) wild-type GBM usually focus on the limitations of a single magnetic resonance image or pathological image. However, it is unclear whether a comprehensive study of the two images is more accurate in predicting patient prognosis. In addition, the comprehensive relationship between magnetic resonance images or pathological images and potential biological mechanisms and pathways needs further study. The imaging genomics analysis of IDH wild-type GBM requires a large data set to train and validate the imaging-pathology fusion omics model. Due to the rarity of IDH wild-type GBM, it is difficult to obtain a large data set. Summary of the invention

[0005] The embodiments of the present application provide a method and system for assessing the prognosis of GBM patients based on magnetic resonance and pathological image fusion, which utilizes magnetic resonance and pathological images to extract imaging genomics features and pathological genomics features respectively and construct a fusion model to evaluate the incremental prognostic predictive value of imaging-pathological fusion genomics labels of IDH wild-type GBM patients relative to traditional clinical pathological factors.

[0006] To solve the above technical problems, in the first aspect, an embodiment of the present application provides a prognosis assessment method for GBM patients based on magnetic resonance imaging and pathological image fusion, comprising the following steps: first, obtaining a data set of GBM patients; GBM patients include IDH wild-type glioblastoma patients; then, performing image acquisition on the data set to obtain MRI images and pathological WSIs; and based on the data set, obtaining transcriptome sequencing data; the pathological WSI is a full-field digital pathological slice image; next, based on the MRI image and the pathological WSI, obtaining image-pathological fusion features; then, based on the image-pathological fusion features, constructing an image-pathological omics model; and based on the image-pathological omics model and the transcriptome sequencing data, obtaining image-pathological omics analysis results; next, using weighted gene co-expression network analysis and gene set enrichment analysis methods, respectively, to analyze the transcriptome sequencing data to obtain cross-biological pathways for revealing the biological meaning of image-pathological omics features; finally, the image-pathological omics model and the cross-biological pathways are used to determine the key biological pathways behind the image-pathological omics features.

[0007] In some exemplary embodiments, image acquisition is performed on a data set to obtain an MRI image and a pathological WSI, including: performing magnetic resonance imaging acquisition and pathological image acquisition on the data set respectively to obtain an MRI image and a pathological WSI respectively; obtaining an image-pathological fusion feature based on the MRI image and the pathological WSI, including: first, standardizing the MRI image and the pathological WSI respectively, and extracting imaging genomics features and pathological genomics features from the standardized MRI image and pathological WSI respectively to obtain an imaging genomics feature data set and a pathological genomics feature data set; then, fusing the imaging genomics feature data set with the pathological genomics feature data set to obtain an image-pathological fusion feature data set; finally, performing feature screening on the features in the image-pathological fusion feature data set to obtain an image-pathological fusion feature.

[0008] In some exemplary embodiments, the MRI image is normalized, including: first, the MRI image is corrected using an N4 bias field to remove the influence of the skull; then, all voxels are isotropically resampled to 1×1×1 mm using a trilinear interpolation method. 3voxels; next, taking fluid attenuated inversion recovery imaging as a template, and adopting a mutual information-based similarity measurement method to rigidly register the MRI images to generate a registered image; specifically, extracting radiomics features from the standardized MRI images, including: first, extracting quantitative radiomics features from the standardized images, the quantitative radiomics features including shape, intensity and texture features; then, using a filter to transform or derive the original MRI images to obtain wavelet features and Gaussian-Laplace features.

[0009] In some exemplary embodiments, the pathological WSI is standardized, including: first, accurately separating the tissue area and non-tissue area in the pathological WSI, and eliminating the blank non-tissue area; then, dividing the tissue area into tissue patches of 1024×1024 size; next, clustering the tissue patches according to the main features and patterns in the image using the K-means clustering algorithm to obtain several plaque clusters, and selecting the optimal plaque from each plaque cluster; specifically, extracting pathological genomic features from the standardized pathological WSI, including: performing comprehensive feature extraction on the optimal plaque selected from each plaque cluster to obtain pathological genomic features.

[0010] In some exemplary embodiments, comprehensive feature extraction is performed on the optimal plaque selected from each plaque cluster to obtain pathological omics features, including: first, using an unmixing color module to unmix representative images; then, three modules with adaptive thresholds are used to automatically identify cell nuclei, cell bodies, and cytoplasm; the three modules are primary identification module, secondary identification module, and tertiary identification module; next, a series of functional modules are used to perform in-depth analysis of the image to measure the features of objects in the image; the object features include the size, shape, texture, density, distribution, and intensity features of the object; finally, the measured object features are statistically analyzed to obtain different statistical indicators to comprehensively and objectively describe the features and properties of the objects in the image; the statistical indicators include mean, median, and standard deviation.

[0011] In some exemplary embodiments, an imaging genomics feature data set and a pathology genomics feature data set are fused to obtain an imaging-pathology fusion feature data set; the features in the imaging-pathology fusion feature data set are screened to obtain imaging-pathology fusion features, including: first, the imaging genomics features obtained in the imaging genomics feature data set are screened to obtain an imaging genomics feature data set with high consistency; the judgment criterion for high consistency is: the ICC value is greater than or equal to 0.8; then, the imaging genomics feature data set with high consistency is originally fused with the pathology feature data set to form a fusion feature set covering imaging features and pathology features; next, a univariate Cox regression model is used to analyze the features of the fusion feature set to obtain the analysis results, and the consistency index is calculated to measure the accuracy of the analysis results; finally, the LASSO-Cox algorithm is used to perform secondary screening on the features of the fusion feature set, and finally the best features are screened out as imaging-pathology fusion features to form the final imaging-pathology fusion feature data set.

[0012] In some exemplary embodiments, an image-pathology omics model is constructed based on image-pathology fusion features, including: using a machine learning method combining a lasso algorithm and a Cox risk regression analysis, constructing an image-pathology omics model based on image-pathology fusion features, and verifying the image-pathology omics model; wherein the verification process includes: calculating a multi-omics risk score for each patient, determining an optimal cutoff value, and applying the image-pathology omics model on an internal validation set; the internal validation set is obtained by dividing the data set; applying the image-pathology omics model on an external validation set to verify the validity and stability of the image-pathology omics model; the external validation set is a data set of GBM patients used to evaluate the quality of the model; the external validation set and the internal validation set are different data sets; and the predictive effect of the image-pathology omics model is evaluated by survival curves, consistency indexes, and decision curves, respectively.

[0013] In some exemplary embodiments, weighted gene co-expression network analysis and gene set enrichment analysis methods are used to analyze transcriptome sequencing data, including: using cluster analysis to identify and eliminate possible outlier samples; constructing gene modules and calculating the GSVA value of each gene module, and then performing Pearson correlation analysis with the multi-omics risk score to screen gene modules with a false discovery rate of less than 0.05 to ensure that the retained gene modules have significant biological significance and reliability; for the retained gene modules, performing WGCNA analysis of the gene modules to analyze the relationships between genes; and performing different gene expression analyses to ensure that the enrichment analysis of the gene set obtains accurate and reliable results.

[0014] In some exemplary embodiments, the imaging-pathology omics model and the cross-biological pathways are used to determine the key biological pathways behind the imaging-pathology omics features, including: first, using two analysis methods, weighted gene co-expression network analysis and gene set enrichment analysis, to obtain the respective enrichment pathway results; then, using the cross-biological pathways of the two analysis methods, weighted gene co-expression network analysis and gene set enrichment analysis, to calculate the GSVA score corresponding to each pathway, and perform Pearson correlation analysis with the imaging-pathology fusion features to study the relationship between each feature value in the multi-omics risk score and the cross-biological pathway, so as to compare the expression differences between high-risk and low-risk patient groups in key biological pathways, so as to gain a deeper understanding of the differences between different risk groups in imaging-pathology genomics.

[0015] In a second aspect, the embodiment of the present application further provides a GBM patient prognosis assessment system based on magnetic resonance-pathological image fusion, comprising: a data set module, an image and transcriptome sequencing data acquisition module, a fusion feature extraction module, a fusion model construction module and a data analysis module connected in sequence; the data set module is used to obtain a data set of GBM patients; GBM patients include IDH wild-type glioblastoma patients; the image and transcriptome sequencing data acquisition module is used to collect images of the data set to obtain MRI images and pathological WSI; and based on the data set, transcriptome sequencing data is obtained; the pathological WSI is a full-field digital pathological section image; the fusion feature extraction module is used to obtain image-pathology fusion features based on MRI images and pathological WSI; the fusion model construction module is used to construct an image-pathology omics model based on the image-pathology fusion features; and based on the image-pathology omics model and transcriptome sequencing data, the image-pathology omics analysis results are obtained; the data analysis module is used to analyze the transcriptome sequencing data using weighted gene co-expression network analysis and gene set enrichment analysis methods, respectively, to obtain cross-biological pathways used to reveal the biological meaning of image-pathology omics features; and the image-pathology omics model and cross-biological pathways are used to determine the key biological pathways behind the image-pathology omics features.

[0016] The technical solution provided by the embodiments of the present application has at least the following advantages:

[0017] The embodiment of the present application provides a method and system for assessing the prognosis of GBM patients based on magnetic resonance imaging and pathological images, the method comprising the following steps: first, obtaining a data set of GBM patients; GBM patients include IDH wild-type glioblastoma patients; then, performing image acquisition on the data set to obtain MRI images and pathological WSIs; and based on the data set, obtaining transcriptome sequencing data; the pathological WSIs are full-field digital pathological slice images; next, based on the MRI images and pathological WSIs, obtaining image-pathological fusion features; then, based on the image-pathological fusion features, constructing an image-pathological omics model; and based on the image-pathological omics model and the transcriptome sequencing data, obtaining image-pathological omics analysis results; next, using weighted gene co-expression network analysis and gene set enrichment analysis methods, respectively, to analyze the transcriptome sequencing data to obtain cross-biological pathways for revealing the biological meaning of image-pathological omics features; finally, the image-pathological omics model and the cross-biological pathways are used to determine the key biological pathways behind the image-pathological omics features.

[0018] The prognosis assessment method for GBM patients based on magnetic resonance-pathology image fusion provided in this application collects MRI images and pathological WSI data of GBM patients, aligns them, and extracts image-pathology fusion features, and selects the best prognostic image-pathology fusion omics features to construct an image-pathology omics model. Secondly, the transcriptome sequencing data were analyzed using two methods, weighted gene co-expression network analysis and gene set enrichment analysis, to identify the cross-biological pathways between the two. Finally, the image-pathology fusion features were associated with the biological pathways, and the key biological pathways behind the fusion omics features were determined. By integrating imaging, pathology omics and gene expression data, this application can provide a deep understanding of biological mechanisms and pathways, and provide more specific guidance and predictions for patient treatment and prognosis assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] One or more embodiments are exemplarily described by pictures in the corresponding drawings, and these exemplifications do not constitute limitations on the embodiments. Unless otherwise stated, the pictures in the drawings do not constitute proportional limitations.

[0020] Figure 1 A flowchart of a method for prognosis assessment of GBM patients based on magnetic resonance imaging and pathological image fusion is provided in accordance with an embodiment of the present application.

[0021] Figure 2A A schematic diagram of obtaining image-pathology fusion features based on an IDH wild-type glioblastoma patient dataset provided in one embodiment of the present application.

[0022] Figure 2BThis is a curve showing the change of partial likelihood deviation with Log(λ) in Lasso regression provided in one embodiment of the present application.

[0023] Figure 2C This is a decision curve analysis curve of the internal validation set model provided in one embodiment of the present application.

[0024] Figure 2D Survival curves obtained by internal validation and external validation respectively provided in an embodiment of the present application.

[0025] Figure 2E A nomogram is provided for one embodiment of the present application to show the probability of OS at 12 months, 24 months and 36 months predicted by Multiscore and clinical variables in an internal validation set.

[0026] Figure 3A A schematic diagram of gene enrichment results in GSEA analysis provided in one embodiment of the present application.

[0027] Figure 3B A heat map of pathway GSVA scores obtained by GSEA analysis provided in one embodiment of the present application.

[0028] Figure 3C A ridge plot of a gene set enrichment pathway provided in one embodiment of the present application.

[0029] Figure 3D A bar chart of a gene set enrichment pathway provided in one embodiment of the present application.

[0030] Figure 4A A WGCNA analysis clustering dendrogram provided in an embodiment of the present application.

[0031] Figure 4B A heat map of the correlation between gene modules and Multiscore provided in one embodiment of the present application.

[0032] Figure 4C A heat map of pathway GSVA scores obtained through WGCNA analysis provided in an embodiment of the present application.

[0033] Figure 4D A bar chart of the relevant gene module enrichment pathway provided in one embodiment of the present application.

[0034] Figure 4E A bubble diagram of the enriched pathway of related gene modules provided in one embodiment of the present application.

[0035] Figure 5A A schematic diagram of cross-pathways and biological classification of GSEA results and WGCNA results provided in one embodiment of the present application.

[0036] Figure 5BA GSVA score heat map of the cross-path of the GSEA result and the WGCNA result provided in one embodiment of the present application.

[0037] Figure 5C A schematic diagram of the distribution of biological pathways related to prognostic features provided in one embodiment of the present application.

[0038] Figure 5D A schematic diagram of the distribution of biological pathways related to prognostic features provided in another embodiment of the present application.

[0039] Figure 5E A schematic diagram of the relationship between a classical biological pathway and a single feature provided in one embodiment of the present application.

[0040] Fig. 5F A schematic diagram of the relationship between a classical biological pathway and a single feature provided in another embodiment of the present application.

[0041] Figure 5G A schematic diagram of the relationship between a classical biological pathway and a single feature provided in yet another embodiment of the present application.

[0042] Figure 6 A schematic diagram of the structure of a GBM patient prognosis assessment system based on magnetic resonance imaging and pathological image fusion provided in one embodiment of the present application. DETAILED DESCRIPTION

[0043] As can be seen from the background technology, in the existing machine learning research on IDH wild-type GBM medical images, it is not clear whether the comprehensive study of the two images is more accurate in predicting the prognosis of patients. Moreover, the comprehensive relationship between magnetic resonance images or pathological images and potential biological mechanisms and pathways needs further study.

[0044] In recent years, emerging therapies for GBM, such as targeted therapy, immunotherapy, gene editing, and small molecule inhibitors, are in the ascendant, but these treatment improvements and explorations have not contributed to a significant increase in patient survival. In addition, according to the 2021 CNS5 GBM diagnostic criteria, the IDH-mutant GBM previously classified in the 2016 version of CNS4 was excluded from the GBM category, and the new criteria included some diffuse astrocytomas that were previously identified as IDH wild-type. This latest classification emphasizes the key role of biomarkers in the diagnosis and treatment of GBM, and the invasiveness of these markers further suggests the importance of developing a non-invasive and efficient biomarker to predict the prognosis of CNS5-based IDH wild-type GBM and risk stratify it.

[0045] Radiomics extracts a large number of quantitative image features from MRI, revealing hidden information that cannot be detected by naked eye observation alone, and provides tumor heterogeneity information in a non-invasive manner, thus providing new help for clinical diagnosis and treatment. Due to the unique biological behavior of gliomas and the challenges associated with diagnosis and treatment, people are increasingly interested in applying radiomics to gliomas. After years of development, it has been proven that non-invasive biomarkers for GBM constructed using radiomics can effectively guide risk stratification and prognosis.

[0046] Pathogenomics, which uses artificial intelligence methods to extract morphological structures in whole slide imaging (WSI), generates multi-scale microscopic data. The structure it displays can be represented by features such as size, shape, texture, etc. AI-based pathogenomics can extract tissue pathological features independently associated with risk through cutting-edge technology to identify complex patterns and biological features, thereby predicting prognosis, etc., especially in the field of tumors. Pathogenomics has been applied to multiple tumor disciplines, including breast cancer, lung adenocarcinoma, central nervous system lesions, and hematological malignancies.

[0047] At present, the research on machine learning of glioma mainly focuses on its classification prediction and prognosis assessment, but most of them are limited to limited sample applications for a single magnetic resonance image or a single pathological image. In addition, these studies also lack biological explanations for the models. Therefore, many scholars at home and abroad have conducted in-depth research on a series of advanced and effective multi-omics data integration methods, aiming to achieve integrated analysis of multi-omics data. The methods for integrated analysis of multi-omics data in recent years can be classified according to the timing of data integration, including three significantly independent categories: early integration, mid-term integration, and late integration.

[0048] Radiomics extracts quantitative features from radiomics (such as MRI, CT, and a variety of new imaging methods), mainly conducting high-throughput studies of diseases at a macroscopic scale, while pathomic omics focuses on the lesion microenvironment and uses WSI to provide in-depth analysis of tissue structural changes at the microscopic level. The two are similar, but their differences still indicate that joint learning can be used to create better non-invasive clinical prediction models. Combined with imaging-pathological genomics analysis, that is, through cross-analysis of pathways closely related to multi-omics risk scores (Multiscore) generated by Gene Set Enrichment Analysis (GSEA) and Weighted Gene Co-expression Network Analysis (WGCNA), reliable pathways that reveal the biological significance of prognostic phenotypes were obtained. A similar study selected 14 of the 1,834 radiomic features extracted from CT images to build a radiomics model. Among the three algorithms of support vector machine (SVM), logistic regression (LR) and multi-layer perceptron (MLP), the MLP model training and test groups were 0.843 and 0.797, respectively, which showed the most robust prediction performance, expressed by the model evaluation criterion AUC (Area Under Curve). Similarly, 10 of the 512 pathology omics features extracted from the entire slide image were selected to establish the pathology omics model, among which the SVM model training and test groups were 0.937 and 0.792, respectively, with the highest prediction efficiency. Finally, the radio-pathology joint model showed better discrimination ability (AUC, training cohort: 0.951; test cohort: 0.837), which shows that pathology omics and imaging omics can indeed complement each other.

[0049] In summary, the existing technologies have the following defects: On the one hand, the existing technologies focus on the limitations of a single magnetic resonance image or pathological image, and it is unclear whether the comprehensive study of the two images is more accurate in predicting the prognosis of patients. In addition, the comprehensive relationship between magnetic resonance images or pathological images and potential biological mechanisms and pathways needs further study. On the other hand, the imaging genomics analysis of IDH wild-type GBM requires a large data set to train and verify the imaging-pathology fusion omics model. Due to the rarity of IDH wild-type GBM, it is difficult to obtain a large data set.

[0050] In order to solve the above technical problems, an embodiment of the present application provides a method and system for prognosis assessment of GBM patients based on magnetic resonance imaging and pathological image fusion, the method comprising the following steps: first, obtaining a data set of GBM patients; GBM patients include IDH wild-type glioblastoma patients; then, performing image acquisition on the data set to obtain MRI images and pathological WSI; and based on the data set, obtaining transcriptome sequencing data; the pathological WSI is a full-field digital pathological slice image; next, based on the MRI image and the pathological WSI, obtaining image-pathological fusion features; then, based on the image-pathological fusion features, constructing an image-pathological omics model; and based on the image-pathological omics model and the transcriptome sequencing data, obtaining image-pathological omics analysis results; next, using weighted gene co-expression network analysis and gene set enrichment analysis methods, respectively, to analyze the transcriptome sequencing data to obtain cross-biological pathways for revealing the biological meaning of image-pathological omics features; finally, the image-pathological omics model and the cross-biological pathways are used to determine the key biological pathways behind the image-pathological omics features.

[0051] This application provides a prognostic evaluation method and system for GBM patients based on magnetic resonance-pathology image fusion, adopts a strategy of combining imaging-pathology omics, and successfully constructs a prognostic risk scoring model based on the patient's multi-omics risk score (Multiscore), namely, the imaging-pathology omics model. This application makes full use of the combined advantages of medical imaging and pathological information, which helps to more comprehensively and accurately evaluate the prognosis of IDH wild-type GBM patients. In addition, paired IDH wild-type glioblastoma MRI and RNA sequencing data obtained by high-throughput sequencing technology are used to explore the biological mechanisms behind the prognostic imaging-pathology omics features. By integrating imaging, pathology, and gene expression data, we can gain a deep understanding of biological mechanisms and pathways, and provide more specific guidance and predictions for patient treatment and prognostic evaluation.

[0052] The following will describe the various embodiments of the present application in detail with reference to the accompanying drawings. However, it will be appreciated by those skilled in the art that in the various embodiments of the present application, many technical details are provided in order to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solution claimed in the present application can be implemented.

[0053] See also Figure 1 The present application embodiment provides a method for evaluating the prognosis of GBM patients based on magnetic resonance imaging and pathological image fusion, comprising the following steps:

[0054] Step S1, obtaining a data set of GBM patients; GBM patients include IDH wild-type glioblastoma patients.

[0055] Step S2: collect images of the data set to obtain MRI images and pathological WSI; and based on the data set, obtain transcriptome sequencing data; the pathological WSI is a full-field digital pathological slice image.

[0056] Step S3: Obtain image-pathology fusion features based on the MRI image and pathology WSI.

[0057] Step S4: construct an image-pathology omics model based on the image-pathology fusion features; and obtain the image-pathology omics analysis results based on the image-pathology omics model and transcriptome sequencing data.

[0058] Step S5: using weighted gene co-expression network analysis and gene set enrichment analysis methods to analyze the transcriptome sequencing data, and obtain cross-biological pathways for revealing the biological meaning of imaging-pathology features.

[0059] Step S6: Combine the imaging-pathology omics model and the cross-biological pathways to determine the key biological pathways behind the imaging-pathology omics features.

[0060] The prognosis assessment method for GBM patients provided in this application adopts an early integration approach to fuse imaging and pathological features, combines high-throughput data in pathological images and magnetic resonance images with disease-related gene sequencing data, and conducts genomic research on magnetic resonance-pathological image fusion phenotypes based on machine learning, aiming to explore the relationship between medical image data and genomic data, and provide new insights into the pathogenesis of GBM.

[0061] Step S1 is mainly the process of obtaining a data set of GBM patients. This application obtains a data set of GBM patients by setting "patient cohort selection criteria". The main cohort studied in this application is cases that underwent brain tumor surgical resection in the neurosurgery department of the hospital and were diagnosed with IDH wild-type GBM. Among them, the patient cohort selection criteria include case inclusion criteria and case exclusion criteria, which are shown below respectively.

[0062] The inclusion criteria for the cases were as follows: (1) patients aged 18 years and above; (2) recent preoperative MRI imaging data, including T1-weighted imaging (T1), T2-weighted imaging (T2), fluid-attenuated inversion recovery imaging (FLAIR), T1-weighted contrast-enhanced imaging (T1c) imaging examinations and corresponding derived ADC images (an imaging form in MRI, which stands for apparent diffusion coefficient image); (3) patients had not received tumor-related treatment before magnetic resonance imaging (MRI); (4) reliable tumor tissue pathology images (Whole Slide Images, WSI) were available after surgery; and (5) patients had a clear pathological diagnosis.

[0063] The exclusion criteria for cases were as follows: (1) the MRI images collected through non-high-field magnetic resonance imaging did not meet the standards; (2) there were other neurological diseases (such as cerebral hemorrhage, meningioma, etc.); (3) there were problems with the MRI images, such as obvious abnormal artifacts. The quality of the postoperative pathological slice images was poor and the features could not be accurately extracted; (5) the basic clinical information of the patient could not be fully obtained, such as loss of follow-up; (6) the patient's death was not caused by GBM itself, but by other unexpected related factors.

[0064] After the above screening process, patients who met the criteria were included in the study cohort. Among them, some patients in the study cohort had paired RNA sequencing data. In addition, the clinical pathological data of each patient were obtained through the clinical medical record system, including: gender, age, radiotherapy, chemotherapy, tumor grade, preoperative Karnofsky Performance Status (KPS), and tumor resection degree (complete or incomplete). Follow-up data came from telephone follow-up and patient review.

[0065] In some embodiments, step S2 performs image acquisition on the data set to obtain an MRI image and a pathological WSI, including: performing magnetic resonance imaging acquisition and pathological image acquisition on the data set to obtain an MRI image and a pathological WSI, respectively. Figure 2A shown.

[0066] In some embodiments, after obtaining the data set of GBM patients, molecular information and transcriptome RNA sequencing operations are performed. Among them, there are two sources of molecular information. One is that some patients checked the IDH gene mutation status in postoperative pathological testing, and the molecular information of these patients came from the hospital electronic archive system. Another part of the patients did not undergo this examination. The team members borrowed the embedded tissue wax blocks of the corresponding patients from the pathology department, and used Sanger sequencing and fluorescence in situ hybridization (FISH) technology to detect the corresponding molecular information. In addition, this application also randomly selected some fresh frozen tumor specimens of the enrolled patients for transcriptome sequencing (RNA Sequencing, RNA-Seq) to obtain transcriptome sequencing data of tumor specimens.

[0067] The image acquisition of the data set in step S2 mainly includes MRI image acquisition and pathological WSI image acquisition. During the MRI image acquisition process, all patients in the study cohort need to undergo a 3.0T MR scanner for examination before surgery and collect relevant images. Before entering the examination room, the metal objects carried by the patient are removed to avoid metal artifacts and major safety accidents. Before image acquisition, the patient's head is placed in the center of the coil and fixed, and the patient is advised to avoid movement to prevent motion artifacts. T1-weighted contrast-enhanced imaging (T1c) sequences are obtained immediately after intravenous injection of gadolinium-based contrast agents at a dose of 0.1mmol / kg. Magnetic resonance imaging schemes include T1WI, T2WI, T1c, FLAIR, and ADC. In the process of pathological image acquisition, this application adopts strict experimental operations, from cutting to staining, dehydration and sealing, etc., and then uses a high-resolution digital pathology slice scanner (KF-PRO-120-HI) to scan all WSIs at a magnification of 20 times. All WSIs are reviewed and screened by two pathologists in the field of brain tumor pathology diagnosis.

[0068] In step S3, based on the MRI image and the pathological WSI, image-pathology fusion features are obtained, including the following steps: step S301, the MRI image and the pathological WSI are standardized, and the imaging genomics features and pathological genomics features are extracted from the standardized MRI image and the pathological WSI, respectively, to obtain an imaging genomics feature data set and a pathological genomics feature data set. Step S302, the imaging genomics feature data set and the pathological genomics feature data set are fused to obtain an image-pathology fusion feature data set. Step S303, the features in the image-pathology fusion feature data set are feature screened to obtain image-pathology fusion features.

[0069] Step S3 mainly involves image preprocessing and feature extraction. Specifically, the process includes MRI image preprocessing and feature extraction, pathological WSI preprocessing and feature extraction.

[0070] In some embodiments, the MRI image is normalized in step S301, including: first, the MRI image is corrected using the N4 bias field to remove the influence of the skull; then, all voxels are isotropically resampled to 1×1×1 mm using the trilinear interpolation method. 3voxels; next, the fluid attenuated inversion recovery imaging (FLAIR) was used as a template, and the MRI images were rigidly registered using a mutual information-based similarity measurement method to generate a registered image. In this application, five registered images were obtained after standardizing the MRI images. The three-dimensional volume of interest (VOI) was defined as the tumor core and the edema area around the tumor. First, a radiologist manually drew the axial FLAIR image using ITK-SNAP software (version: 3.4.0). In order to select more robust features, another neurosurgeon with 12 years of clinical experience re-drew some patients randomly selected from the included cohort to generate an inter-rater dataset. Both doctors were blinded to the clinical pathological data and survival status of the patient cohort.

[0071] In some embodiments, step S301 extracts radiomics features from the standardized MRI image, including: first, extracting quantitative radiomics features from the standardized image, the quantitative radiomics features including shape, intensity and texture features; then, using a filter to transform or derive the original MRI image to obtain wavelet features and Gaussian-Laplace features.

[0072] Specifically, in the process of extracting radiomic features from the standardized MRI image in step S301, a large number of quantitative radiomic features, including shape, intensity and texture features, are extracted from the standardized image using the open source Python software package PyRadiomics (version: 2.2.0). In order to make the image features more comprehensive and accurate and improve the effect and accuracy of image analysis, a filter is used to transform or derive the original image to obtain wavelet features and Laplacian of Gaussian (LoG) features. Wavelet features decompose the image by wavelet transform, and then extract wavelet coefficients of different frequencies as features, usually including low-frequency coefficients and high-frequency coefficients. LoG features are feature extraction methods based on Gaussian filtering and Laplacian operators. Noise and detail information can be suppressed by Gaussian filtering the image, and then the Laplacian operator is used to extract the overall structure and morphological information of the image. The Laplacian operator includes four levels: 2.0, 3.0, 4.0 and 5.0. This application performs feature extraction according to the imaging biomarker standardization initiative.

[0073] In some embodiments, the pathological WSI is standardized in step S301, including: first, accurately separating the tissue area and non-tissue area in the pathological WSI, and eliminating the blank non-tissue area; then, dividing the tissue area into tissue patches of 1024×1024 size; next, using the K-means clustering algorithm to cluster the tissue patches according to the main features and patterns in the image to obtain several plaque clusters, and selecting the optimal plaque from each plaque cluster.

[0074] Specifically, in the process of standardizing the pathological WSI in step S301, the present application uses an image processing engine called "Yottixel" to process the tumor WSI. This processing process includes three key steps. The first is to accurately separate the tissue and non-tissue areas in the WSI and eliminate the blank non-tissue areas. Secondly, automatic segmentation is performed to divide the tissue area into 1024×1024 tissue patches (plaque images). Finally, the present application uses the K-means clustering algorithm to cluster tissue patches according to the main features and patterns in the image. In this process, the present application successfully obtained three plaque clusters (K0, K1, K2). Subsequently, the present application invited pathology experts to participate and carefully selected the five plaques with the most satisfactory quality from each cluster.

[0075] Specifically, in step S301 , pathological omics features are extracted from the standardized pathological WSI, including: performing comprehensive feature extraction on the optimal plaque selected from each plaque cluster to obtain pathological omics features.

[0076] In some embodiments, comprehensive feature extraction is performed on the optimal plaque selected from each plaque cluster to obtain pathological omics features, including the following steps: first, a representative image is unmixed using an unmixing color module; then, the cell nucleus, cell body and cytoplasm are automatically identified through three modules with adaptive thresholds; the three modules are respectively an identification primary module, an identification secondary module and an identification tertiary module; next, a series of functional modules are used to perform in-depth analysis on the image to measure the features of the objects in the image; the object features include the size, shape, texture, density, distribution and intensity features of the object; finally, the measured object features are statistically analyzed to obtain different statistical indicators to comprehensively and objectively describe the features and properties of the objects in the image; the statistical indicators include mean, median and standard deviation.

[0077] Specifically, the present application uses CellProfiler software to perform comprehensive feature extraction on all selected plaques in each plaque cluster. The process first uses the Unmix Colors module to unmix representative images. Then, through three modules with adaptive thresholds, the three modules are Identify Primary Objects, Identify Secondary Objects, and Identify Tertiary Objects, the nucleus, cell body, and cytoplasm are automatically identified. Subsequently, the present application uses a series of functional modules to conduct in-depth analysis of the image, including "Measure Size and Shape", "Measure Texture", "Measure Granularity", "Measure Neighbors", "Measure Occupancy Area", "Measure Intensity", and "Measure Intensity Distribution". Through these modules, the present application obtains a rich and diverse set of features. In this process, the present application can not only capture information such as the size, shape, and texture of objects in the image, but also accurately measure features such as density, distribution, and intensity of objects. Finally, this application statistically analyzes the features measured by these modules and obtains statistical indicators such as mean, median and standard deviation to comprehensively and objectively describe the characteristics and properties of objects in the image.

[0078] In some embodiments, in step S302, the imaging genomics feature data set and the pathological genomics feature data set are fused to obtain an imaging-pathology fusion feature data set; and in step S303, the features in the imaging-pathology fusion feature data set are screened to obtain imaging-pathology fusion features, including the following steps: first, the imaging genomics features obtained in the imaging genomics feature data set are screened to obtain an imaging genomics feature data set with higher consistency; the judgment criterion for higher consistency is: the ICC value is greater than or equal to 0.8; then, the imaging genomics feature data set with higher consistency is originally fused with the pathological feature data set to form a fusion feature set covering imaging features and pathological features; next, a univariate Cox regression model is used to analyze the features of the fusion feature set to obtain analysis results, and a consistency index is calculated to measure the accuracy of the analysis results; finally, the LASSO-Cox algorithm is used to perform secondary screening on the features of the fusion feature set, and finally the best features are screened out as imaging-pathology fusion features to form the final imaging-pathology fusion feature data set.

[0079] It should be noted that in the process of imaging-pathology feature screening, this application first screened imaging features with high consistency. By comparing the imaging omics data of 97 patients undergoing secondary tumor delineation, features with high consistency (ICC value ≥ 0.8) were finally left for the next step of analysis. Then, the imaging and pathological features were fused. By means of early fusion, the obtained imaging feature data set was originally fused with the pathological feature data set to form a fusion feature set covering imaging features and pathological features. Next, a univariate Cox regression model was used for analysis, and the consistency index (Concordance index, C-index) was calculated to measure its accuracy. A C-index value of more than 0.5 means that the feature is positively correlated with the patient's good prognosis, while a value below 0.5 implies a negative correlation. Based on statistical significance, this application sets screening criteria to retain only those features with a C-index above 0.55 or below 0.45 and a P value less than 0.05. Finally, the features of the preliminary screening were screened for a second time, and the LASSO-Cox algorithm was used for secondary feature screening and modeling, and finally satisfactory features were screened out to form the final feature set.

[0080] In some embodiments, step S4 constructs an image-pathology omics model based on image-pathology fusion features, including: using a machine learning method combining a lasso algorithm and a Cox risk regression analysis to construct an image-pathology omics model based on image-pathology fusion features, and verifying the image-pathology omics model; wherein the verification process includes: calculating the multi-omics risk score of each patient, determining the optimal cutoff value, and applying the image-pathology omics model on an internal validation set; the internal validation set is obtained from the data set; applying the image-pathology omics model on an external validation set to verify the effectiveness and stability of the image-pathology omics model; the external validation set is a data set of GBM patients used to evaluate the quality of the model; the external validation set and the internal validation set are different data sets; and the predictive effect of the image-pathology omics model is evaluated by survival curves, consistency indexes, and decision curves, respectively.

[0081] It should be noted that this application collects data from 899 IDH wild-type glioblastoma patients from the First Affiliated Hospital of Zhengzhou University as the data set for step S1 of this application, and divides the data set into a training set (N=631) and an internal validation set (N=268); during the model validation process, the multi-omics risk score of each patient is calculated, and the optimal cutoff value is determined, and then the imaging-pathology omics model is applied to the internal validation set to validate the model; in order to further verify the quality of the model, this application prepares an external validation set in advance, and applies the imaging-pathology omics model to the external validation set to verify the effectiveness and stability of the imaging-pathology omics model. The data of the external validation set comes from the data of 98 IDH wild-type glioblastoma patients in the hospital, and this part of the data is included in the external validation set. This application uses internal and external data sets to validate the model separately, so that multi-center validation can better evaluate the quality of the model.

[0082] Specifically, in the process of constructing the imaging-pathology omics model (prognostic model), the present application adopts a machine learning method combining the lasso algorithm and the Cox risk regression analysis, and successfully constructs a satisfactory imaging-pathology omics prognostic model based on the best features selected by the present application. Then, the present application uses a specific formula to calculate the multi-omics risk score of each patient, namely Multiscore, and uses R software (survminer toolkit (version: 0.4.9) in version: 4.3.0) to determine the optimal cutoff value and apply it to the internal validation set. Furthermore, the present application also applies the model to the external validation set to verify its effectiveness and stability in the new data set. Through the survival curve (KM) evaluation, the present application found that Multiscore has significant ability to distinguish high and low risk groups. The Log-rank test further verified the analysis results of the two groups of KM curves, and further confirmed the predictive effect of Multiscore. In addition, this application used C-index to evaluate the predictive performance of the model, and Akaike Information Criterion (AIC) was used to evaluate the overfitting risk and the classification improvement performance of the model. Finally, this application performed a decision curve analysis (DCA) to evaluate the practicality and value of the model in clinical practice.

[0083] FIG. 2B to FIG. 2E A schematic diagram of the construction process of the imaging-pathology omics model is provided, including the cross-validation calculation of the optimal LASSO parameter λ, the internal validation set DCA (decision curve analysis) curve, the internal and external validation set KM curve, and the nomogram. Figure 2B The curve of partial likelihood deviation in Lasso regression with Log(λ) is shown. Figure 2BThe curve shows that the vertical dashed line indicates that the optimal LASSO parameter λ screened is 0.188; Figure 2C The DCA (decision curve analysis) curve of the internal validation set model is shown, showing the curve of net benefit changing with risk threshold; Figure 2D The KM (survival) curves for internal validation and external validation are shown, showing the curve of the change of survival probability over time, and listing the number of low-risk and high-risk patients corresponding to the time point of the abscissa of the KM curve; Figure 2E The probability of OS at 12 months, 24 months and 36 months predicted by Multiscore and clinical variables in the internal validation set is shown, and the nomograms of different variables are shown, in which different variables correspond to scores on the same vertical line at different data points.

[0084] In some embodiments, step S5 uses weighted gene co-expression network analysis and gene set enrichment analysis methods to analyze transcriptome sequencing data, including the following steps: using cluster analysis to identify and eliminate possible outlier samples; constructing gene modules and calculating the GSVA value of each gene module, and then performing Pearson correlation analysis with the multi-omics risk score to screen gene modules with a false discovery rate of less than 0.05 to ensure that the retained gene modules have significant biological significance and reliability; for the retained gene modules, performing WGCNA analysis of the gene modules to analyze the relationships between genes; and performing different gene expression analyses to ensure that the enrichment analysis of the gene set obtains accurate and reliable results.

[0085] This application uses two methods, WGCNA and GSEA, in the analysis of image-pathology fusion omics at the gene level. The analysis results are as follows FIG. 3A to FIG. 4E As shown, the analysis results of GSEA include: Figure 3A The gene enrichment results in the GSEA analysis shown in the figure are as follows: Figure 3B The pathway GSVA score heat map obtained by GSEA analysis is shown in the figure. Figure 3C The ridge plots of gene set enriched pathways shown in Figure 3D Bar chart of gene set enriched pathways shown; WGCNA analysis results include: Figure 4A The WGCNA analysis clustering dendrogram shown in Figure 4B The heat map of the correlation between the gene module and the multi-omics risk score (Multiscore) shown in Figure 4C The pathway GSVA score heat map obtained by WGCNA analysis is shown in Figure 4D The bar chart of the relevant gene module enrichment pathways shown in Figure 4EThe bubble diagram of the enriched pathways of the related gene modules shown in the figure finally obtained the top 5 gene sets. This application analyzes disease-related biological pathways from different angles and cross-validates the analysis results. Then, the cross-biological pathways of the two were taken to reveal the biological basis behind the prognostic imaging-pathological features, including synapses, proliferation, DNA damage response, metabolism, transcription and translation, immunity and complex cell functions.

[0086] Specifically, in the WGCNA analysis process, cluster analysis is the first step to identify possible outlier samples and remove them to ensure the accuracy and reliability of subsequent analysis. Next, the present application constructs gene modules by calculation, calculates the GSVA value of each gene module, and then performs Pearson correlation analysis with the multi-omics risk score (Multiscore). The present application strictly screens gene modules with a false discovery rate (False discover, FDR) less than 0.05 to ensure that the retained modules have significant biological significance and reliability. For the several gene modules retained in this application, the present application analyzes the enrichment of the genes therein to identify biological pathways related to tumor properties. At this stage, the present application applies the WGCNA package (version 1.72-5) in the R language environment (version 4.3.0) to perform WGCNA analysis of gene modules to deeply explore and analyze the relationship between genes. At the same time, the present application also uses the clusterProfiler toolkit (version: 4.9.0) of the R software (version: 4.3.0) to perform enrichment analysis on the genes in the gene module. It is worth noting that the annotated gene set data used in this application are partially the same as those used in the previous GSEA analysis to ensure consistency and comparability of the results.

[0087] In the image-pathology genomics analysis of this application, samples of different risk groups were classified, and then their RNA-Seq data were analyzed for gene expression differences to obtain the corresponding Log2 difference fold values ​​(FoldChange, FC). Subsequently, this application ranked all genes according to these conversion values ​​and applied gene set-based enrichment analysis (GSEA) to discover biological pathways associated with disease risk. During the analysis, this application paid special attention to the false discovery rate (FDR) and ensured that the selected biological pathways had a significance level of FDR less than 0.05. Subsequently, Gene Set Variation Analysis (GSVA) ​​was used to analyze the activation level of these biological pathways, and Pearson correlation analysis was performed with the image-pathology fusion model Multiscore. In this step, this application paid special attention to pathways with a false discovery rate (FDR) less than 0.05 to ensure that biological pathways with significant differences were screened out. For different gene expression analyses, this application used the DESeq2 toolkit (version: 1.41.8) in R software (version: 4.3.0), and conducted a comprehensive analysis of different gene expression through this toolkit. For GSEA analysis, the Cluster Profiler toolkit (version: 4.9.0) in R software (version: 4.3.0) was used to ensure accurate and reliable results for gene set enrichment analysis. The annotated gene set databases are: KEEG, Pathway Interaction Database, Reactome, Hallmark, BioCarta, and WikiPathways.

[0088] In some embodiments, step S6 uses the imaging-pathology omics model and the cross-biological pathway to determine the key biological pathways behind the imaging-pathology omics features, including: first, using two analysis methods, weighted gene co-expression network analysis and gene set enrichment analysis, to obtain the respective enrichment pathway results; then, using the cross-biological pathways of the two analysis methods, weighted gene co-expression network analysis and gene set enrichment analysis, to calculate the GSVA score corresponding to each pathway, and perform Pearson correlation analysis with the imaging-pathology fusion features to study the relationship between each feature value in the multi-omics risk score and the cross-biological pathway, so as to compare the expression differences between high-risk and low-risk patient groups in key biological pathways, so as to gain a deeper understanding of the differences between different risk groups in imaging-pathology genomics.

[0089] It should be noted that in order to make the results more reliable, accurate and reproducible and reduce abnormal data and errors, this application uses a cross-biological pathway (also referred to as a cross-pathway) of two analytical methods. Figure 5AThe cross-pathway and biological classification diagram of GSEA results and WGCNA results are shown; Figure 5B The GSVA score heat map of the cross-pathway of the GSEA result and the WGCNA result provided in one embodiment of the present application. In the biological classification diagram of the cross-pathway diagram, different pathway categories have different proportions, such as Figure 5A As shown, proliferation accounts for 38.1%, synapses account for 16.3%, immunity accounts for 3.6%, DNA damage response accounts for 15.1%, complex cell function accounts for 16.4%, metabolism accounts for 5.7%, and transcription and translation account for 4.6%.

[0090] Figure 5C to Figure 5G A schematic diagram of the biological interpretation results of the imaging-genomics prognostic model is shown, in which: Figure 5C and Figure 5D The schematic diagram shows the distribution of biological pathways related to prognostic features. Figure 5E to Figure 5G The relationship between classic biological pathways and individual features is shown, where Complex_Cellular_Functions represents pathways in the complex cellular function category, DNA_damage_response represents pathways in the DNA damage response category, Immunity represents pathways in the immune category, Metabolism represents pathways in the metabolism category, Proliferation represents pathways in the proliferation category, Synapse represents pathways in the synapse category, and Transcription_and_Translation represents pathways in the transcription and translation category; R is the correlation coefficient, red represents positive correlation, and blue represents negative correlation. This application calculates the GSVA score corresponding to each pathway and performs correlation analysis with the prognostic imaging-pathology fusion omics label to explore the relationship between each feature value in the Multiscore score and the cross-biological pathways, which can provide a deep understanding of the biological mechanisms and pathways, and provide more specific guidance and predictions for patient treatment and prognosis evaluation.

[0091] The prognostic evaluation method for GBM patients based on magnetic resonance-pathological image fusion provided in this application adopts a strategy combining imaging-pathological omics and successfully constructs a Multiscore prognostic risk scoring model. This application makes full use of the comprehensive advantages of medical imaging and pathological information, which helps to more comprehensively and accurately evaluate the prognosis of IDH wild-type GBM patients. In addition, paired IDH wild-type glioblastoma MRI and RNA sequencing data obtained by high-throughput sequencing technology are used to explore the biological mechanisms behind the prognostic imaging-pathological omics features. By integrating imaging, pathology and gene expression data, we can gain a deep understanding of biological mechanisms and pathways, and provide more specific guidance and predictions for the treatment and prognosis evaluation of patients. Moreover, according to the comprehensive diagnostic reclassification proposed by CNS5, more attention is paid to the role of molecular markers in the classification of nervous system tumors. By combining the imaging-pathological omics features of preoperative traditional MRI routine sequence images and postoperative tumor pathology WSI, an imaging-pathological omics model is successfully constructed as a prognostic evaluation model, which has a high prognostic value for IDH wild-type GBM patients. Secondly, there may be at least two significantly different subtypes of IDH wild-type GBM. This application proposes a non-invasive and efficient biomarker, namely the multi-omics risk score (Multiscore), which provides a new way to predict the survival of IDH wild-type GBM patients. The biological manifestations behind it are also different. One subtype is mainly characterized by active tumor cell proliferation and DNA damage response, with a poor prognosis; while the other subtype is mainly characterized by active synaptic activity and a relatively good prognosis. Third, imaging-pathology fusion omics uses two methods, WGCNA and GSEA, at the gene level to analyze disease-related biological pathways from different angles and cross-validate the analysis results. Then, the cross-biological pathways of the two are taken to reveal the biological basis behind the prognostic imaging-pathology omics features, including synapses, proliferation, DNA damage response, metabolism, transcription and translation, immunity, and complex cell functions.

[0092] See also Figure 6, the embodiment of the present application also provides a GBM patient prognosis assessment system based on magnetic resonance-pathological image fusion, comprising: a data set module 101, an image and transcriptome sequencing data acquisition module 102, a fusion feature extraction module 103, a fusion model construction module 104 and a data analysis module 105 connected in sequence; the data set module 101 is used to obtain a data set of GBM patients; GBM patients include IDH wild-type glioblastoma patients; the image and transcriptome sequencing data acquisition module 102 is used to collect images of the data set to obtain MRI images and pathological WSI; and based on the data set, the transcriptome sequencing data is obtained; the pathological WSI is a full-field digital pathological section image; the fusion feature extraction module 103 is used to obtain image-pathology fusion features based on the MRI image and the pathological WSI; the fusion model construction module 104 is used to construct an image-pathology omics model based on the image-pathology fusion features; and based on the image-pathology omics model and the transcriptome sequencing data, the image-pathology omics analysis results are obtained; the data analysis module 105 is used to analyze the transcriptome sequencing data using weighted gene co-expression network analysis and gene set enrichment analysis methods, respectively, to obtain cross-biological pathways for revealing the biological meaning of image-pathology omics features; and the image-pathology omics model and the cross-biological pathways are used to determine the key biological pathways behind the image-pathology omics features.

[0093] Compared with the prior art, the prognosis assessment method and system for GBM patients based on magnetic resonance-pathology image fusion provided in this application have the following advantages: First, the currently published glioma classification criteria are often used as the basis for evaluation, but a certain proportion of IDH mutant brain GBM exists among the enrolled GBM patients. This mixed situation may affect the analysis results because IDH status plays an important role in the development and prognosis of GBM. This application uses the 2021 IDH wild-type GBM definition as the inclusion criteria and conducts a comprehensive investigation and research on this specific patient group. Secondly, unlike the machine learning research on medical images of IDH wild-type GBM, which focuses on the limitations of a single magnetic resonance image or pathological image, this application adopts a strategy combining imaging-pathology omics and successfully constructs the imaging-pathology omics model Multiscore as a prognostic risk scoring model. This innovative method makes full use of the combined advantages of medical imaging and pathological information, which helps to more comprehensively and accurately assess the prognosis of IDH wild-type GBM patients. The selection of these fusion features not only takes into account the diversity of imaging parameters, but also covers the rich information of pathological parameters, thus providing multi-faceted support and reliability for the construction of prognostic models. This result emphasizes the importance of comprehensive consideration of different parameter sources and provides important clues for a deeper understanding of patient prognosis.

[0094] WGCNA is a powerful method for analyzing radiogenomics data, identifying gene modules co-expressed with imaging genomics features, and providing a method for identifying the potential molecular mechanisms of imaging phenotypes. GSEA is a widely used bioinformatics tool for analyzing gene expression data in the context of predefined gene sets or pathways. Since the two methods have different focuses and analytical ideas on gene enrichment, this application combines WGCNA and GSEA to improve the repeatability and reliability of biological pathway identification in imaging genomics analysis. When studying the risk stratification of IDH wild-type GBM patients, this application noticed that patients with different risk levels had obvious differences in the activity of biological pathways. Specifically, patients in the high-risk group generally showed more active tumor cell proliferation and DNA damage response, while patients in the low-risk group showed more active synaptic function, which is generally associated with better prognosis.

[0095] Based on the above technical scheme, the embodiment of the present application provides a method and system for prognosis assessment of GBM patients based on magnetic resonance imaging and pathological image fusion, the method comprising the following steps: first, obtaining a data set of GBM patients; GBM patients include IDH wild-type glioblastoma patients; then, performing image acquisition on the data set to obtain MRI images and pathological WSI; and based on the data set, obtaining transcriptome sequencing data; the pathological WSI is a full-field digital pathological slice image; next, based on the MRI image and the pathological WSI, obtaining image-pathological fusion features; then, based on the image-pathological fusion features, constructing an image-pathological omics model; and based on the image-pathological omics model and the transcriptome sequencing data, obtaining image-pathological omics analysis results; next, using weighted gene co-expression network analysis and gene set enrichment analysis methods, respectively, to analyze the transcriptome sequencing data to obtain cross-biological pathways for revealing the biological meaning of image-pathological omics features; finally, the image-pathological omics model and the cross-biological pathways are used to determine the key biological pathways behind the image-pathological omics features. The prognosis assessment method for GBM patients based on magnetic resonance-pathology image fusion provided in this application collects MRI images and pathological WSI data of GBM patients, aligns them, and extracts image-pathology fusion features, and selects the best prognostic image-pathology fusion omics features to construct an image-pathology omics model. Secondly, the transcriptome sequencing data were analyzed using two methods, weighted gene co-expression network analysis and gene set enrichment analysis, to identify the cross-biological pathways between the two. Finally, the image-pathology fusion features were associated with the biological pathways, and the key biological pathways behind the fusion omics features were determined. By integrating imaging, pathology omics and gene expression data, this application can provide a deep understanding of biological mechanisms and pathways, and provide more specific guidance and predictions for patient treatment and prognosis assessment.

[0096] Those skilled in the art can understand that the above-mentioned embodiments are specific examples for implementing the present application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present application. Any person skilled in the art can make their own changes and modifications without departing from the spirit and scope of the present application, so the scope of protection of the present application shall be based on the scope defined in the claims.

Claims

1. A method for evaluating the prognosis of GBM patients based on magnetic resonance imaging and pathological image fusion, characterized in that: The following steps are involved: Obtaining a data set of GBM patients; the GBM patients include IDH wild-type glioblastoma patients; Performing image acquisition on the data set to obtain MRI images and pathological WSI; and obtaining transcriptome sequencing data based on the data set; The pathological WSI is a full-field digital pathological slice image; Based on the MRI image and the pathological WSI, obtaining an image-pathology fusion feature; Based on the image-pathology fusion features, an image-pathology omics model is constructed; and based on the image-pathology omics model and the transcriptome sequencing data, an image-pathology omics analysis result is obtained; The transcriptome sequencing data were analyzed using weighted gene co-expression network analysis and gene set enrichment analysis methods to obtain cross-biological pathways that reveal the biological meaning of imaging-pathology features; the cross-biological pathways include synapses, proliferation, DNA damage response, metabolism, transcription and translation, immunity, and complex cell functions; The imaging-pathology omics model and the cross-biological pathways are used to determine the key biological pathways behind the imaging-pathology omics features; The data set is imaged to obtain MRI images and pathological WSI, including: Performing magnetic resonance imaging acquisition and pathological image acquisition on the data set to obtain MRI images and pathological WSI respectively; Based on the MRI image and the pathological WSI, image-pathology fusion features are obtained, including: The MRI image and the pathological WSI are respectively standardized, and radiomics features and pathological omics features are respectively extracted from the standardized MRI image and pathological WSI to obtain a radiomics feature data set and a pathological omics feature data set; Fusion of the imaging genomics feature dataset and the pathology genomics feature dataset to obtain an imaging-pathology fusion feature dataset; Performing feature screening on the features in the image-pathology fusion feature data set to obtain image-pathology fusion features; The imaging genomics feature dataset and the pathology genomics feature dataset are fused to obtain an imaging-pathology fusion feature dataset; and the features in the imaging-pathology fusion feature dataset are screened to obtain imaging-pathology fusion features, including: The imaging features obtained in the imaging feature dataset are screened to obtain an imaging feature dataset with high consistency; the criterion for high consistency is: an ICC value greater than or equal to 0.8; The dataset of imaging features with high consistency is fused with the dataset of pathological features to form a fusion feature set covering both imaging features and pathological features; The univariate Cox regression model was used to analyze the features of the fusion feature set, obtain the analysis results, and calculate the consistency index to measure the accuracy of the analysis results; The LASSO-Cox algorithm was used to perform secondary screening on the features of the fusion feature set, and finally the best features were selected as the image-pathology fusion features to form the final image-pathology fusion feature dataset.

2. The method for evaluating the prognosis of GBM patients based on magnetic resonance imaging and pathological image fusion according to claim 1, characterized in that: The MRI image is subjected to standardization processing, including: The MRI image is corrected using an N4 bias field to remove the influence of the skull; All voxels were isotropically resampled to 1×1×1 mm using trilinear interpolation. 3 of voxels; Using fluid attenuated inversion recovery imaging as a template and using a mutual information-based similarity measurement method to rigidly register the MRI image to generate a registered image; The step of extracting radiomics features from the standardized MRI images comprises: Extracting quantitative radiomics features from the standardized image, wherein the quantitative radiomics features include shape, intensity and texture features; Wavelet features and Gaussian-Laplace features are obtained by transforming or deriving the original MRI images using filters.

3. The method for evaluating the prognosis of GBM patients based on magnetic resonance imaging and pathological image fusion according to claim 1, characterized in that: The pathological WSI is standardized, including: Accurately separate the tissue area and the non-tissue area in the pathological WSI, and remove the blank non-tissue area; The tissue area was segmented into tissue patches of 1024 × 1024 size; K-means clustering algorithm is used to cluster tissue plaques according to the main features and patterns in the image to obtain several plaque clusters, and the best plaque is selected from each plaque cluster; The step of extracting pathological omics features from the standardized pathological WSI includes: Comprehensive feature extraction is performed on the optimal plaque selected from each plaque cluster to obtain pathological omics features.

4. The method for evaluating the prognosis of GBM patients based on magnetic resonance imaging and pathological image fusion according to claim 3, characterized in that: Comprehensive feature extraction is performed on the optimal plaque selected from each plaque cluster to obtain pathological omics features, including: Unmix the representative image using the Unmix Color module; The cell nucleus, cell body and cytoplasm are automatically identified through three modules with adaptive thresholds; the three modules are identification of primary module, identification of secondary module and identification of tertiary module; A series of functional modules are used to conduct in-depth analysis of the image and measure the characteristics of the objects in the image; the object characteristics include the size, shape, texture, density, distribution and intensity characteristics of the object; The measured object features are statistically analyzed to obtain different statistical indicators to comprehensively and objectively describe the features and properties of the objects in the image; the statistical indicators include mean, median and standard deviation.

5. The method for evaluating the prognosis of GBM patients based on magnetic resonance-pathological image fusion according to claim 1, characterized in that: Based on the image-pathology fusion features, an image-pathology omics model is constructed, including: Using a machine learning method combining the lasso algorithm and Cox risk regression analysis, an imaging-pathology omics model was constructed based on the imaging-pathology fusion features, and the imaging-pathology omics model was verified; The verification process includes: Calculating a multi-omics risk score for each patient, determining an optimal cutoff value, and applying the imaging-pathology omics model on an internal validation set; the internal validation set is obtained by dividing the data set; Applying the imaging-pathology omics model on an external validation set to verify the effectiveness and stability of the imaging-pathology omics model; the external validation set is a data set of GBM patients used to evaluate the quality of the model; the external validation set and the internal validation set are different data sets; The predictive effect of the imaging-pathology omics model was evaluated by survival curve, consistency index and decision curve.

6. The method for evaluating the prognosis of GBM patients based on magnetic resonance imaging and pathological image fusion according to claim 1, characterized in that: The weighted gene co-expression network analysis and gene set enrichment analysis methods were used to analyze the transcriptome sequencing data, including: Cluster analysis is used to identify and eliminate possible outlier samples; Gene modules were constructed and the GSVA value of each gene module was calculated. Then, Pearson correlation analysis was performed with the multi-omics risk score to screen gene modules with a false discovery rate of less than 0.05 to ensure that the retained gene modules had significant biological significance and reliability. For the retained gene modules, WGCNA analysis was performed to analyze the relationships between genes; and different gene expression analyses were performed to ensure that the enrichment analysis of the gene set obtained accurate and reliable results.

7. The method for evaluating the prognosis of GBM patients based on magnetic resonance imaging and pathological image fusion according to claim 1, characterized in that: The imaging-pathology omics model and the cross-biological pathways are used to determine the key biological pathways behind the imaging-pathology omics features, including: Two analysis methods, weighted gene co-expression network analysis and gene set enrichment analysis, were used to obtain the respective enrichment pathway results; The cross-biological pathways of weighted gene co-expression network analysis and gene set enrichment analysis were used to calculate the GSVA score corresponding to each pathway, and Pearson correlation analysis was performed with the imaging-pathology fusion features to study the relationship between each feature value in the multi-omics risk score and the cross-biological pathway, thereby comparing the expression differences in key biological pathways between high-risk and low-risk patient groups, in order to gain a deeper understanding of the differences in imaging-pathology genomics between different risk groups.

8. A GBM patient prognosis assessment system based on magnetic resonance imaging and pathological image fusion, characterized in that: include: A dataset module, an image and transcriptome sequencing data acquisition module, a fusion feature extraction module, a fusion model construction module, and a data analysis module connected in sequence; The data set module is used to obtain a data set of GBM patients; the GBM patients include IDH wild-type glioblastoma patients; The image and transcriptome sequencing data acquisition module is used to collect images of the data set to obtain MRI images and pathological WSI; and based on the data set, obtain transcriptome sequencing data; The pathological WSI is a full-field digital pathological slice image; The fusion feature extraction module is used to obtain image-pathology fusion features according to the MRI image and the pathological WSI; The fusion model building module is used to build an image-pathology omics model according to the image-pathology fusion features; and Based on the imaging-pathology omics model and the transcriptome sequencing data, obtaining imaging-pathology omics analysis results; The data analysis module is used to analyze the transcriptome sequencing data by using weighted gene co-expression network analysis and gene set enrichment analysis methods, respectively, to obtain cross-biological pathways for revealing the biological meaning of imaging-pathology features; and to determine the key biological pathways behind the imaging-pathology features by using the imaging-pathology model and the cross-biological pathways; the cross-biological pathways include synapses, proliferation, DNA damage response, metabolism, transcription and translation, immunity and complex cell functions; The data set is imaged to obtain MRI images and pathological WSI, including: Performing magnetic resonance imaging acquisition and pathological image acquisition on the data set to obtain MRI images and pathological WSI respectively; Based on the MRI image and the pathological WSI, image-pathology fusion features are obtained, including: The MRI image and the pathological WSI are respectively standardized, and radiomics features and pathological omics features are respectively extracted from the standardized MRI image and pathological WSI to obtain a radiomics feature data set and a pathological omics feature data set; Fusion of the imaging genomics feature dataset and the pathology genomics feature dataset to obtain an imaging-pathology fusion feature dataset; Performing feature screening on the features in the image-pathology fusion feature data set to obtain image-pathology fusion features; The imaging genomics feature dataset and the pathology genomics feature dataset are fused to obtain an imaging-pathology fusion feature dataset; and the features in the imaging-pathology fusion feature dataset are screened to obtain imaging-pathology fusion features, including: The imaging features obtained in the imaging feature dataset are screened to obtain an imaging feature dataset with high consistency; the criterion for high consistency is: an ICC value greater than or equal to 0.8; The dataset of imaging features with high consistency is fused with the dataset of pathological features to form a fusion feature set covering both imaging features and pathological features; The univariate Cox regression model was used to analyze the features of the fusion feature set, obtain the analysis results, and calculate the consistency index to measure the accuracy of the analysis results; The LASSO-Cox algorithm was used to perform secondary screening on the features of the fusion feature set, and finally the best features were selected as the image-pathology fusion features to form the final image-pathology fusion feature dataset.

Images