System and method for predicting endometrial cancer molecular subtypes based on magnetic resonance deep learning radiomics
By using the Swin UNETR deep learning segmentation model and multimodal MRI data fusion technology, fully automated prediction of molecular subtypes of endometrial cancer has been achieved, solving the problems of invasiveness and time consumption in existing technologies, improving diagnostic efficiency and accuracy, and supporting personalized treatment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGDA HOSPITAL SOUTHEAST UNIV
- Filing Date
- 2025-12-25
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies for identifying molecular subtypes of endometrial cancer rely on invasive tissue specimen analysis, which is costly and time-consuming, making it difficult to popularize in resource-limited environments. Traditional radiomics and deep learning methods are time-consuming, labor-intensive, and have poor reproducibility, making large-scale automation impossible.
The Swin UNETR deep learning segmentation model was used for fully automated tumor segmentation. Combined with multimodal MRI scan data, end-to-end molecular subtype prediction of endometrial cancer was achieved through the fusion of radiomics, deep learning and clinical features.
It enables non-invasive, fully automated prediction of molecular subtypes of endometrial cancer, improving diagnostic efficiency and accuracy, reducing the burden on doctors, and providing support for personalized treatment decisions.
Smart Images
Figure CN121962025A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical device technology, and particularly relates to the application of medical image processing, artificial intelligence, deep learning and radiomics technology in cancer diagnosis, especially the molecular subtyping of endometrial cancer, specifically a system and method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics. Background Technology
[0002] Endometrial cancer (EC) is one of the most common gynecological malignancies in women. According to the Cancer Genome Atlas (TCGA), EC can be classified into four subtypes with different prognostic significance based on its molecular characteristics: POLE hypermutation (POLEmut), mismatch repair deficiency (MMRd), p53 abnormality (p53abn), and no specific molecular signature (NSMP). Accurate molecular subtype identification is crucial for guiding diagnosis, assessing prognosis, and developing personalized treatment plans (such as whether aggressive adjuvant therapy is needed). Therefore, several authoritative guidelines, including those from the World Health Organization (WHO), the European Society for Gynecologic Oncology (ESGO), and the International Federation of Gynecologic and Obstetric Cancer (FIGO), have incorporated molecular classification into the management and staging criteria for EC.
[0003] Currently, the mainstream technology for identifying EC molecular subtypes in the industry mainly relies on the analysis of tumor tissue samples from patients. Specifically, this involves using the gold standard method of complex gene sequencing (such as the method used by TCGA), or using more clinically practical alternatives, such as the ProMisE classifier. ProMisE combines targeted sequencing (for POLE mutations) with immunohistochemical (IHC) detection (for MMR and p53 protein status).
[0004] However, existing technologies have the following shortcomings:
[0005] 1. Invasiveness and limitations: These methods all rely on invasive tissue specimens obtained through biopsy or surgery.
[0006] 2. High cost and time-consuming: Gene sequencing and multi-platform testing (such as ProMisE) are complex, costly and time-consuming processes.
[0007] 3. Limited application: Due to the above-mentioned shortcomings, these methods (especially in resource-limited environments) are difficult to widely implement before surgery, which prevents doctors from obtaining key molecular subtyping information before making surgical and treatment decisions, thus limiting personalized EC management.
[0008] Limitations of radiological methods: As alternatives, some studies have attempted to use radiomics or deep learning (DL) methods, but existing methods also have limitations:
[0009] 1. Traditional radiomics relies on predefined "handcrafted" features, making it difficult to capture complex tumor heterogeneity;
[0010] 2. Many existing deep learning (DL) methods still require experienced radiologists to manually segment tumors, a process that is extremely time-consuming, labor-intensive, and has poor reproducibility, making it unsuitable for large-scale automated clinical workflows. Summary of the Invention
[0011] To address the aforementioned technical challenges, this invention proposes a system and method for predicting molecular subtypes of endometrial cancer based on magnetic resonance imaging deep learning radiomics. By using the Swin UNETR deep learning segmentation model, not only is fully automated tumor segmentation achieved, but end-to-end automation from MRI scanning to subtype prediction is also realized, greatly reducing the workload of clinicians and improving efficiency and repeatability.
[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0013] A system for predicting molecular subtypes of endometrial cancer based on magnetic resonance imaging deep learning radiomics includes a data acquisition module, a segmentation module, a feature extraction module, a fusion module, and a prediction module, which are sequentially connected.
[0014] The data acquisition module is used to acquire the patient's preoperative pelvic multimodal MRI scan data;
[0015] The segmentation module includes the Swin UNETR deep learning segmentation model, which consists of an encoder and a decoder, used to segment the input MRI scan data.
[0016] The feature extraction module is used to extract features from different dimensions of MRI scan data;
[0017] The fusion module is used to concatenate and fuse features extracted from different dimensions by the feature extraction module to form a feature set;
[0018] The prediction module includes a machine learning model, which is trained and used to make predictions using a feature set formed by fusion.
[0019] A method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics includes the following steps:
[0020] S1. Data Acquisition and Input;
[0021] S2, Fully Automated Tumor Segmentation;
[0022] S3, Multimodal Feature Extraction;
[0023] S4, Feature Engineering and Integration;
[0024] S5, Machine Learning Classification and Prediction.
[0025] As a preferred technical solution of the present invention, step S1 is specifically as follows:
[0026] S11. Data Reception: The system receives the patient's preoperative pelvic multimodal MRI scan data;
[0027] S12. Source data extraction: Extract 3D volume data, including at least T2-weighted imaging sequences and diffusion-weighted imaging sequences;
[0028] S13. Data resampling: 3D volumetric data is resampled to standardized voxel spacing and uniform input dimensions.
[0029] As a preferred technical solution of the present invention, step S2 is specifically as follows:
[0030] S21. Data transmission: The volume data of the T2 weighted imaging sequence and the diffusion-weighted imaging sequence from step S1 are transmitted to a Swin UNETR deep learning segmentation model respectively.
[0031] S22. Data Processing: The input T2-weighted imaging sequence and diffusion-weighted imaging sequence volume data are processed in the Swin UNETR deep learning segmentation model to automatically identify and depict the three-dimensional region of the tumor.
[0032] S23. Results Data: Automatically generate tumor segmentation masks for each patient on T2-weighted imaging sequences and diffusion-weighted imaging sequences.
[0033] As a preferred technical solution of the present invention, step S3 is as follows:
[0034] The feature extraction module consists of three parallel sub-modules for extracting features from different dimensions, as follows:
[0035] S31, Radiomics Feature Extraction Module:
[0036] Source data: the raw MRI scan data from step S1 and the tumor mask generated in step S2;
[0037] Data processing: Using the PyRadiomics library, high-dimensional “handmade” radiomics features were extracted from the raw MRI scan data based on the three-dimensional region defined by the tumor mask;
[0038] Results data: A set of image omics feature vectors was generated, which includes first-order statistical features, shape features and second-order texture features;
[0039] S32, Deep Learning Feature Extraction Module:
[0040] Source data: the raw MRI scan data in step S1 and the encoder of the pre-trained Swin UNETR deep learning segmentation model in step S2;
[0041] Data processing: The encoder of the Swin UNETR deep learning segmentation model is used as a feature extractor. Image data is captured by the encoder through complex deep semantics and hierarchical patterns that cannot be obtained by hand-designing.
[0042] Result data: A set of deep learning feature vectors is generated;
[0043] S33, Clinical Feature Collection Module:
[0044] Source data: The patient's electronic medical record;
[0045] Data processing: Retrospective collection of clinicopathological data related to EC subtypes;
[0046] Results data: A set of clinical feature vectors was generated, including age, BMI, menopausal status, serum tumor markers, and FIGO stage.
[0047] As a preferred technical solution of the present invention, step S4 is specifically as follows:
[0048] S41, Source Data: Radiomics feature vector, deep learning feature vector, and clinical feature vector output from step S3;
[0049] S42. Data Processing:
[0050] S421. Radiomics feature processing: Z-score standardization, low-variance filtering, and high-correlation redundancy removal are performed.
[0051] S422. Deep Learning Feature Processing: First, global average pooling is performed on the feature map output by the encoder. Then, principal component analysis is used for dimensionality reduction.
[0052] S423. Clinical feature processing: Fill missing values with the mean or mode within the subtype, and use one-hot encoding for categorical variables;
[0053] S424. Feature Fusion: The processed radiomics feature vectors, deep learning feature vectors, and clinical feature vectors are merged into a unified, multi-dimensional feature set through a splicing operation.
[0054] As a preferred technical solution of the present invention, step S5 is as follows:
[0055] S51, Source Data: The feature set generated in step S4;
[0056] S52, Data Processing:
[0057] S521. Handling Imbalanced Data: Before training the model, SMOTE is used to process the training data to generate synthetic minority class samples and balance the dataset.
[0058] S522, Model Training: Input the processed feature set into one or more machine learning classifiers;
[0059] S523, Model Inference: For new unknown patient data, the data is processed sequentially through steps S1 to S4, and the generated feature set is input into the trained machine learning model.
[0060] S53. Results Data: The machine learning model outputs the predicted probability that the patient belongs to each of the four molecular subtypes, achieving non-invasive preoperative molecular typing.
[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0062] 1. Non-invasive and economical: This invention is a non-invasive strategy that uses only conventional preoperative MRI scans, completely avoiding the high cost, invasiveness and time-consuming gene sequencing or IHC testing required by traditional methods.
[0063] 2. Fully automated and highly efficient: This invention realizes a fully automated process from tumor segmentation to subtype prediction, eliminating the need for time-consuming manual operations and greatly improving the efficiency and consistency of clinical diagnosis.
[0064] 3. High accuracy: By innovatively integrating clinical data, radiomics features and deep learning features, the comprehensive model (C+R+DL) of this invention significantly improves the accuracy of predicting EC molecular subtypes, providing a reliable clinical decision-making tool for precision oncology.
[0065] 4. High clinical value: This invention can provide accurate molecular subtyping prediction for EC patients before surgery, thereby assisting doctors in formulating personalized surgical plans (such as the extent of lymph node dissection), assessing whether to preserve fertility, and guiding subsequent adjuvant treatment decisions. Attached Figure Description
[0066] Figure 1 This is a schematic diagram of the patient enrollment process in an embodiment of the present invention.
[0067] Figure 2 This is a flowchart of a method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics in an embodiment of the present invention.
[0068] Figure 3This is a schematic diagram of the Swin UNETR deep learning segmentation model architecture in an embodiment of the present invention.
[0069] Figure 4 This is a schematic diagram of the automatic segmentation effect sample in an embodiment of the present invention. Detailed Implementation
[0070] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0071] The system for predicting molecular subtypes of endometrial cancer based on magnetic resonance imaging deep learning radiomics proposed in this invention includes a data acquisition module, a segmentation module, a feature extraction module, a fusion module, and a prediction module. These modules are sequentially connected, wherein:
[0072] The data acquisition module is used to acquire the patient's preoperative pelvic multimodal MRI scan data;
[0073] The segmentation module includes the Swin UNETR deep learning segmentation model, which consists of an encoder and a decoder, used to segment the input MRI scan data.
[0074] The feature extraction module is used to extract features from different dimensions of MRI scan data;
[0075] The fusion module is used to concatenate and fuse features extracted from different dimensions by the feature extraction module to form a feature set;
[0076] The prediction module includes a machine learning model, which is trained and used to make predictions using a feature set formed by fusion.
[0077] like Figure 1-4 As shown, the method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics proposed in this invention includes the following steps:
[0078] S1. Data Acquisition and Input;
[0079] Specifically as follows:
[0080] S11. Data Reception: The system receives the patient's preoperative pelvic multimodal MRI scan data;
[0081] S12. Source data extraction: Extract 3D volumetric data that includes at least T2-weighted imaging (T2WI) sequences and diffusion-weighted imaging (DWI) sequences;
[0082] S13, Data Resampling: 3D volumetric data is resampled to a standardized voxel spacing (e.g., [1.5, 1.5, 6] mm) and a uniform input dimension (e.g., 256 × 256 × 32).
[0083] S2, Fully Automated Tumor Segmentation;
[0084] Specifically as follows:
[0085] S21. Data Transfer: Transfer the volumetric data of the T2-weighted imaging sequence and the diffusion-weighted imaging sequence from step S1 to a Swin UNETR deep learning segmentation model (e.g., ...). Figure 3 (as shown)
[0086] Swin UNETR is a hybrid architecture that combines the Swin Transformer and UNet encoder-decoder structures. Its encoder uses a hierarchical self-attention mechanism to capture long-range dependencies, and the decoder progressively upsamples through skip connections to refine spatial details. The model (trained on an NVIDIA GPU using the MONAI framework and the Dice Focal loss function) processes the input 3D image to automatically identify and delineate the three-dimensional region of interest (ROI) of the tumor.
[0087] S22. Data Processing: The input T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) sequence volume data are processed in the Swin UNETR deep learning segmentation model to automatically identify and depict the three-dimensional region of the tumor (ROI).
[0088] S23. Results Data: Automatically generate tumor segmentation masks for each patient based on T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) sequences, such as... Figure 4 As shown, this step is fully automated and requires no manual drawing.
[0089] S3, Multimodal Feature Extraction;
[0090] Specifically as follows:
[0091] The feature extraction module consists of three parallel sub-modules for extracting features from different dimensions, as follows:
[0092] S31, Radiomics Feature Extraction Module:
[0093] Source data: the raw MRI scan data (T2WI, DWI) from step S1 and the tumor mask generated in step S2;
[0094] Data processing: Using the PyRadiomics library, high-dimensional “handcrafted” radiomics features were extracted from raw MRI scan data based on three-dimensional regions of interest (ROIs) defined by tumor masks.
[0095] Results data: A set of radiomics (R) feature vectors were generated, which included first-order statistical features (such as average intensity and skewness), shape features (such as volume and sphericity), and second-order texture features (derived from GLCM, GLRLM, etc.).
[0096] S32, Deep Learning Feature Extraction Module:
[0097] Source data: the raw MRI scan data (T2WI, DWI) in step S1 and the encoder of the pre-trained SwinUNETR deep learning segmentation model in step S2;
[0098] Data processing: The encoder of the Swin UNETR deep learning segmentation model is used as a feature extractor. Image data is captured by the encoder (composed of multiple Swin Transformer blocks) to capture complex deep semantics and hierarchical patterns that cannot be achieved by hand-designing.
[0099] Result data: Generates a set of deep learning (DL) feature vectors;
[0100] S33, Clinical Feature Collection Module:
[0101] Source data: The patient's electronic medical record;
[0102] Data processing: Retrospective collection of clinicopathological data related to EC subtypes;
[0103] Results data: A set of clinical (C) feature vectors were generated, including age, BMI, menopausal status, serum tumor markers, and FIGO stage.
[0104] S4, Feature Engineering and Integration;
[0105] Specifically as follows:
[0106] S41, Source Data: Radiomics feature vectors, deep learning feature vectors, and clinical feature vectors output from steps S31, S32, and S33;
[0107] S42. Data Processing:
[0108] S421. Radiomics feature processing: Z-score standardization was performed to eliminate scale effects, low variance filtering was performed to remove invalid features, and high correlation redundancy was removed. Pearson correlation coefficient and ANOVA F test were used to remove high correlation redundancy.
[0109] S422. Deep Learning Feature Processing: First, global average pooling (GAP) is performed on the feature map output by the encoder. Then, principal component analysis (PCA) is used to reduce dimensionality (e.g., from 1536 dimensions to 40 dimensions) to solve the problems of high dimensionality and collinearity.
[0110] S423. Clinical feature processing: Missing values are filled with the mean or mode within the subtype, and one-hot encoding is used for categorical variables (such as FIGO stage).
[0111] S424. Feature Fusion: The processed radiomics (R) feature vectors, deep learning (DL) feature vectors, and clinical (C) feature vectors are fused into a unified, multi-dimensional feature set through a concatenation operation.
[0112] S5, Machine Learning Classification and Prediction;
[0113] Specifically as follows:
[0114] S51, Source Data: The feature set generated in step S4;
[0115] S52, Data Processing:
[0116] S521. Handling Imbalanced Data: Due to the uneven distribution of EC subtypes (especially POLEmut), SMOTE (Synthetic Minority Oversampling) is used to process the training data before training the model to generate synthetic minority class samples and balance the dataset.
[0117] S522, Model Training: The processed feature set is input into one or more machine learning classifiers. This invention can use various classifiers, such as logistic regression (LR), random forest (RF), gradient boosting decision tree (GBDT), support vector machine (SVM) or extreme gradient boosting (XGBoost). The training process uses 5-fold cross-validation to ensure the robustness of model performance.
[0118] S523, Model Inference: For new unknown patient data (inference stage), the data passes through steps S1 to S4 in sequence, and the generated feature set is input into the trained machine learning model;
[0119] S53. Results Data: The machine learning model outputs the predicted probability of the patient belonging to each of the four molecular subtypes (POLEmut, MMRd, p53abn, NSMP), achieving non-invasive preoperative molecular subtyping.
[0120] Based on the above technical solution, the present invention can achieve:
[0121] Fully automated workflow, eliminating manual dependence: The biggest improvement of this invention lies in the use of the Swin UNETR deep learning segmentation model to achieve fully automated tumor segmentation. Most existing technologies rely on time-consuming and labor-intensive manual delineation. This solution achieves end-to-end automation from MRI scanning to subtype prediction, greatly reducing the workload of clinicians and improving efficiency and repeatability.
[0122] Innovative Fusion of Multidimensional Features: This invention is not limited to single radiomics features or deep learning features, but innovatively designs a trimodal fusion framework. It combines clinical data (C), handcrafted radiomics features (R), and deep semantic features (DL) extracted by the encoder of the Swin UNETR deep learning segmentation model.
[0123] Significant improvement in predictive performance: The fusion of deep learning (DL) features significantly enhances the model's discriminative ability. Experimental results show that the comprehensive model incorporating all three features (C+R+DL) significantly outperforms models using only clinical features (AUC 0.622), models using only radiomics features (AUC 0.600), or models combining any two features (such as the C+R model, AUC 0.676). The comprehensive model of this invention achieved a macro-average AUC of 0.703 in the validation cohort.
[0124] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.
Claims
1. A system for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics, characterized in that, It includes a data acquisition module, a segmentation module, a feature extraction module, a fusion module, and a prediction module, which are connected sequentially. The data acquisition module is used to acquire the patient's preoperative pelvic multimodal MRI scan data; The segmentation module includes the Swin UNETR deep learning segmentation model, which consists of an encoder and a decoder, used to segment the input MRI scan data. The feature extraction module is used to extract features from different dimensions of MRI scan data; The fusion module is used to concatenate and fuse features extracted from different dimensions by the feature extraction module to form a feature set; The prediction module includes a machine learning model, which is trained and used to make predictions using a feature set formed by fusion.
2. The method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics as described in claim 1, characterized in that, Includes the following steps: S1. Data Acquisition and Input; S2, Fully Automated Tumor Segmentation; S3, Multimodal Feature Extraction; S4, Feature Engineering and Integration; S5, Machine Learning Classification and Prediction.
3. The method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics according to claim 2, characterized in that, Step S1 is as follows: S11. Data Reception: The system receives the patient's preoperative pelvic multimodal MRI scan data; S12. Source data extraction: Extract 3D volume data, including at least T2-weighted imaging sequences and diffusion-weighted imaging sequences; S13. Data resampling: 3D volumetric data is resampled to standardized voxel spacing and uniform input dimensions.
4. The method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics according to claim 2, characterized in that, Step S2 is as follows: S21. Data transmission: The volume data of the T2 weighted imaging sequence and the diffusion-weighted imaging sequence from step S1 are transmitted to a Swin UNETR deep learning segmentation model respectively. S22. Data Processing: The input T2-weighted imaging sequence and diffusion-weighted imaging sequence volume data are processed in the Swin UNETR deep learning segmentation model to automatically identify and depict the three-dimensional region of the tumor. S23. Results Data: Automatically generate tumor segmentation masks for each patient on T2-weighted imaging sequences and diffusion-weighted imaging sequences.
5. The method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics according to claim 2, characterized in that, Step S3 is as follows: The feature extraction module consists of three parallel sub-modules for extracting features from different dimensions, as follows: S31, Radiomics Feature Extraction Module: Source data: the raw MRI scan data from step S1 and the tumor mask generated in step S2; Data processing: Using the PyRadiomics library, high-dimensional "handmade" radiomics features were extracted from the raw MRI scan data based on the three-dimensional region defined by the tumor mask; Results data: A set of image omics feature vectors was generated, which includes first-order statistical features, shape features and second-order texture features; S32, Deep Learning Feature Extraction Module: Source data: the raw MRI scan data in step S1 and the encoder of the pre-trained Swin UNETR deep learning segmentation model in step S2; Data processing: The encoder of the Swin UNETR deep learning segmentation model is used as a feature extractor. Image data is captured by the encoder through complex deep semantics and hierarchical patterns that cannot be obtained by hand-designing. Result data: A set of deep learning feature vectors is generated; S33, Clinical Feature Collection Module: Source data: The patient's electronic medical record; Data processing: Retrospective collection of clinicopathological data related to EC subtypes; Results data: A set of clinical feature vectors was generated, including age, BMI, menopausal status, serum tumor markers, and FIGO stage.
6. The method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics according to claim 2, characterized in that, Step S4 is as follows: S41, Source Data: Radiomics feature vector, deep learning feature vector, and clinical feature vector output from step S3; S42. Data Processing: S421. Radiomics feature processing: Z-score standardization, low-variance filtering, and high-correlation redundancy removal are performed. S422. Deep Learning Feature Processing: First, global average pooling is performed on the feature map output by the encoder. Then, principal component analysis is used for dimensionality reduction. S423. Clinical feature processing: Fill missing values with the mean or mode within the subtype, and use one-hot encoding for categorical variables; S424. Feature Fusion: The processed radiomics feature vectors, deep learning feature vectors, and clinical feature vectors are merged into a unified, multi-dimensional feature set through a splicing operation.
7. The method for predicting molecular subtypes of endometrial cancer based on magnetic resonance deep learning radiomics according to claim 2, characterized in that, Step S5 is as follows: S51, Source Data: The feature set generated in step S4; S52, Data Processing: S521. Handling Imbalanced Data: Before training the model, SMOTE is used to process the training data to generate synthetic minority class samples and balance the dataset. S522, Model Training: Input the processed feature set into one or more machine learning classifiers; S523, Model Inference: For new unknown patient data, the data is processed sequentially through steps S1 to S4, and the generated feature set is input into the trained machine learning model. S53. Results Data: The machine learning model outputs the predicted probability that the patient belongs to each of the four molecular subtypes, achieving non-invasive preoperative molecular typing.