Endometrial cancer lymph node metastasis prediction method and system based on whole uterus MRI image driving
Patent Information
- Application Number
- CN202610867182.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-16
- Publication Date
- 2026-10-20
AI Technical Summary
[0004]针对现有模型泛化不理想的问题，本发明提供一种基于全子宫MRI图像驱动的子宫内膜癌淋巴结转移预测方法及系统，基于子宫内膜癌患者的术前盆腔MRI图像中的全子宫组织，采用深度学习自动分割技术获取全子宫感兴趣体积，结合放射组学和机器学习技术构建淋巴结转移预测模型，通过模型输出风险评分，可以快速、准确地进行术前患者风险分层，从而为临床医生制定个性化手术方案提供客观依据
[0037]本发明采用深度学习自动分割技术获取全子宫感兴趣体积，结合放射组学和机器学习技术构建淋巴结转移预测模型，通过模型输出风险评分，并基于临床影像学评估与模型预测的共识投票策略实现患者风险分层，同时依据SHAP可解释性分析探索预测结果的影像学与生物学可解释性。本发明能够解决现有子宫内膜癌淋巴结转移评估方法主观性强、常规MRI敏感性低（仅）、以及基于肿瘤病灶的放射组学模型分割不稳定（DSC仅约0.70）且泛化能力差等问题，可以快速、准确地进行术前患者风险分层，从而为临床医生制定个性化手术方案提供客观依据，能够通过共识投票策略和SHAP可解释性分析相结合的方式，促进人工智能模型的临床转化；通过创新性地采用全子宫作为感兴趣区域而非不稳定的肿瘤病灶，本发明提升了分割的鲁棒性和特征提取的稳定性，为构建更加精准的预测模型提供了可靠的基础，并采用XGBoost机器学习技术，自动挖掘海量放射组学特征中的关键预后信息，充分捕捉全子宫内的肿瘤异质性及肿瘤-间质微环境特征，可以为子宫内膜癌患者的术前淋巴结转移风险预测提供高效、无创的辅助工具，可对预测结果进行影像学和生物学的双重可解释性分析，有望推广至多中心临床实践，能够促进子宫内膜癌的个性化治疗和精准医疗，在影像医学及妇科肿瘤学领域具有较好的应用前景。
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer-aided prognostic technology for endometrial cancer, and in particular to a method and system for predicting lymph node metastasis in endometrial cancer based on whole-uterine MRI images. Background Technology
[0002] Magnetic resonance imaging (MRI), a non-invasive imaging technique with excellent soft tissue contrast, has become an indispensable tool in the preoperative evaluation of endometrial cancer. However, the sensitivity of conventional MRI imaging evaluation for detecting lymph node metastasis (LNM) is only 59%-65%, which is insufficient to meet the needs of precise clinical decision-making.
[0003] In recent years, with the development of artificial intelligence and medical image processing technology, it has become possible to establish prognostic prediction models for endometrial cancer by combining MRI images and radiomics technology. Radiomics can extract quantitative features imperceptible to the human eye from medical images in high throughput, showing great potential in tumor grading, staging, and metastasis risk assessment. The entire uterine volume contains tumor tissue and surrounding tissue, with the surrounding tissue consisting of a complex mixture of non-tumor cells that can influence disease progression and tumor metastasis through interaction with cancer cells. Existing studies have shown that integrating information from intratumoral and peritumoral regions can improve the prediction accuracy of lymph node metastasis. Therefore, systematically quantifying the entire uterine tissue in MRI images of endometrial cancer will help to develop more reasonable management plans and treatment decisions for patients. Preliminary studies have predicted lymph node metastasis in endometrial cancer based on MRI radiomics, with the area under the receiver operating characteristic (AUC) reaching 0.80-0.90. However, these studies generally suffer from small sample sizes, are mostly single-center designs, and lack independent external validation, leading to doubts about the generalization ability of the models and making it difficult to translate them into clinical practice. Therefore, further exploring the characteristics of whole uterine imaging, establishing a precise radiomics prediction model based on whole uterine MRI, and conducting multi-center external validation of the model will help to make personalized postoperative treatment decisions for endometrial cancer patients, reduce unnecessary lymph node dissections, and promote the clinical translation of deep learning models and artificial intelligence-assisted diagnostic systems. Summary of the Invention
[0004] To address the issue of poor generalization in existing models, this invention provides a method and system for predicting lymph node metastasis in endometrial cancer based on whole-uterine MRI images. Using preoperative pelvic MRI images of endometrial cancer patients, the system employs deep learning-based automatic segmentation technology to obtain the volume of interest (LOI) of the entire uterus. A lymph node metastasis prediction model is constructed by combining radiomics and machine learning techniques. The model outputs a risk score, enabling rapid and accurate preoperative patient risk stratification, thus providing objective evidence for clinicians to develop personalized surgical plans.
[0005] According to the design scheme provided by the present invention, on one hand, a method for predicting lymph node metastasis of endometrial cancer based on whole uterine MRI images is provided, comprising:
[0006] We acquired and preprocessed pelvic MRI clinical image data from multiple medical centers that were publicly authorized. We then randomly selected N images from the clinical image data of each medical center for whole-uterine target annotation. The annotated clinical image data was used as the first sample, and the remaining unannotated clinical image data was used as the second sample.
[0007] A deep learning segmentation network is trained using the first sample to obtain a whole uterus target segmentation model; the second sample is then input into the whole uterus target segmentation model, and the whole uterus volume of interest is extracted from the second sample using the whole uterus target segmentation model. The extracted whole uterus volume of interest is then used as the third sample.
[0008] Radiomics features were extracted from the third sample, and the parameters of the radiomics prediction model were optimized based on the radiomics features to obtain the radiomics target prediction model. The radiomics prediction model was constructed using the XGBoost algorithm to assess the risk of lymph node metastasis.
[0009] The whole uterus target segmentation model was used to extract the whole uterus volume of interest from the pelvic MRI of the target patient, and the radiomics features of the whole uterus volume of interest of the target patient were extracted. Based on the extracted radiomics features, the radiomics target prediction model was used to assess the risk of lymph node metastasis of the whole uterus volume of interest of the target patient.
[0010] As part of the present invention's method for predicting lymph node metastasis in endometrial cancer based on whole uterine MRI images, the method further includes preprocessing the pelvic MRI clinical image data, comprising:
[0011] Clinical image data undergoes data cleaning and normalization.
[0012] The normalized image data is cropped into three-dimensional voxel blocks of a specified pixel size, so as to construct sample data using the three-dimensional voxel blocks.
[0013] As part of the present invention's method for predicting lymph node metastasis in endometrial cancer based on whole uterine MRI images, a deep learning segmentation network is further trained using the first sample to obtain a whole uterine target segmentation model, comprising:
[0014] The first sample is divided into a training set and a test set according to a preset ratio;
[0015] A deep learning segmentation network is constructed based on the SegFormer3D network. The deep learning segmentation network includes: a hierarchical encoder for extracting multi-scale features from input image data, and a decoder for aggregating multi-scale features and outputting segmentation results.
[0016] The deep learning segmentation network is trained using the training set, and the network parameters are updated using the backpropagation algorithm.
[0017] The trained deep learning segmentation network was evaluated using a test set to obtain a whole uterus target segmentation model based on the evaluation results.
[0018] As part of the present invention's method for predicting lymph node metastasis in endometrial cancer based on whole uterine MRI images, further, radiomics features are extracted from the third sample, including:
[0019] Radiomics features were extracted from the third sample using an open-source medical imaging omics feature extraction tool. The radiomics features included at least: first-order statistics, shape features, first texture features described by a gray-level co-occurrence matrix, and second texture features described by a gray-level run-length matrix.
[0020] The extracted radiomics features were screened to determine the optimal radiomics features to be used as input to the prediction model.
[0021] As part of the present invention's method for predicting lymph node metastasis in endometrial cancer based on whole uterine MRI images, the extracted radiomics features are further screened, including:
[0022] The Mann-Whitney U test was used to assess the univariate association between each radiomics feature and lymph node metastasis status, and radiomics features that met the expected association were retained.
[0023] Calculate the pairwise Pearson correlation coefficients between the retained radiomics features, and retain the radiomics features with strong correlations in the feature pairs based on the Pearson correlation coefficients;
[0024] LASSO regression was used to optimize the retained radiomics features, and cross-validation was used to determine the optimal regularization parameter to determine the final radiomics features obtained from the screening.
[0025] As part of the present invention's method for predicting lymph node metastasis in endometrial cancer based on whole uterine MRI images, a lymph node metastasis risk assessment and prediction model is further constructed using the XGBoost algorithm, comprising:
[0026] Decision trees are used to model the input radiomics features. In each iteration, the best features and split points are selected by splitting nodes, and the trees are built step by step. Each tree is modeled on the residual of the previous tree. In the modeling process, the features for building the prediction model are selected by evaluating the importance of the features for predicting the risk of lymph node metastasis.
[0027] As a method for predicting lymph node metastasis in endometrial cancer based on whole uterine MRI images, this invention further includes assessing the risk of lymph node metastasis in the whole uterine volume of interest of the target patient using a radiomics target prediction model, and also includes:
[0028] Consensus voting was conducted on the outputs of the clinical impact medical assessment and radiomics target prediction model to classify target patients into high-risk, intermediate-risk, and low-risk groups.
[0029] The radiomics features were analyzed using the SHAP interpretable machine learning framework to obtain the shapely value corresponding to each feature and to build an interpretive model.
[0030] The contribution of each feature to the prediction of lymph node metastasis risk is explained by interpreting the model and the shapely value corresponding to each feature, so as to generate a feature importance ranking chart.
[0031] Furthermore, this invention also provides a system for predicting lymph node metastasis in endometrial cancer based on whole-uterine MRI images, comprising: a medical sample acquisition module, a segmentation model training module, a prediction model training module, and a risk prediction output module, wherein...
[0032] The medical sample acquisition module is used to acquire and preprocess pelvic MRI clinical image data publicly authorized by multiple medical centers, and randomly select N images from the clinical image data of each medical center for whole uterine target annotation. The annotated clinical image data is used as the first sample, and the remaining unannotated clinical image data is used as the second sample.
[0033] The segmentation model training module is used to train a deep learning segmentation network using the first sample to obtain a whole uterus target segmentation model; and inputs the second sample into the whole uterus target segmentation model, uses the whole uterus target segmentation model to extract the whole uterus volume of interest in the second sample, and uses the extracted whole uterus volume of interest as the third sample;
[0034] The prediction model training module is used to extract radiomics features from the third sample, optimize the parameters of the radiomics prediction model based on the radiomics features, and obtain the radiomics target prediction model. The radiomics prediction model is a lymph node metastasis risk assessment prediction model constructed using the XGBoost algorithm.
[0035] The risk prediction output module is used to extract the volume of interest (LOI) of the entire uterus in the pelvic MRI of the target patient using a whole uterine target segmentation model, extract the radiomics features of the whole uterine LIO of the target patient, and perform lymph node metastasis risk assessment on the whole uterine LIO of the target patient based on the extracted radiomics features and using a radiomics target prediction model.
[0036] The beneficial effects of this invention are:
[0037] This invention employs deep learning-based automatic segmentation technology to obtain the volume of interest (LOI) of the entire uterus. It combines radiomics and machine learning techniques to construct a lymph node metastasis prediction model. The model outputs a risk score, and patient risk stratification is achieved based on a consensus voting strategy combining clinical imaging assessment and model prediction. Furthermore, the imaging and biological interpretability of the prediction results is explored using SHAP interpretability analysis. This invention addresses the shortcomings of existing methods for assessing lymph node metastasis in endometrial cancer, which suffer from high subjectivity and low sensitivity of conventional MRI (only...). This invention addresses the problems of unstable segmentation (DSC of only about 0.70) and poor generalization ability of radiomics models based on tumor lesions. It can quickly and accurately perform preoperative patient risk stratification, thus providing objective evidence for clinicians to formulate personalized surgical plans. It can promote the clinical translation of artificial intelligence models through a combination of consensus voting strategy and SHAP interpretability analysis. By innovatively using the entire uterus as the region of interest rather than unstable tumor lesions, this invention improves the robustness of segmentation and the stability of feature extraction, providing a reliable foundation for building more accurate prediction models. It also uses XGBoost machine learning technology to automatically mine key prognostic information from massive radiomics features, fully capturing the tumor heterogeneity and tumor-stromal microenvironment characteristics within the entire uterus. It can provide an efficient and non-invasive auxiliary tool for predicting the preoperative lymph node metastasis risk of endometrial cancer patients. It can perform dual interpretability analysis of the prediction results from both imaging and biological perspectives, and is expected to be promoted to multi-center clinical practice, which can promote personalized treatment and precision medicine for endometrial cancer. It has good application prospects in the fields of imaging medicine and gynecologic oncology. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of the endometrial cancer lymph node metastasis prediction process driven by whole uterine MRI images in the embodiment.
[0039] Figure 2 This is a schematic diagram illustrating the construction process of the whole uterus segmentation model in the embodiment;
[0040] Figure 3 This is a schematic diagram of the radiomics feature extraction and screening process in the example.
[0041] Figure 4This is a schematic diagram illustrating the effect of automatic uterine segmentation in the embodiment;
[0042] Figure 5 This is a schematic diagram of the receiver operating characteristic (ROC) curves of the XGBoost prediction model in the training set, internal validation set, and external validation set in the embodiment.
[0043] Figure 6 This is a schematic diagram of the patient risk stratification results based on the consensus voting strategy in the example.
[0044] Figure 7 This is a schematic diagram illustrating the importance analysis of SHAP features in the example. Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.
[0046] For the quantification of the entire uterine tissue in preoperative MRI images of endometrial cancer, see the embodiments of the present invention. Figure 1 As shown, a method for predicting lymph node metastasis in endometrial cancer based on whole-uterine MRI images is provided, comprising:
[0047] S101. Obtain publicly authorized pelvic MRI clinical image data from multiple medical centers and preprocess them. Randomly select N images from the clinical image data of each medical center for whole uterine target annotation. Use the annotated clinical image data as the first sample and the remaining unannotated clinical image data as the second sample.
[0048] First, the clinical image data can be cleaned and normalized. Then, the normalized image data can be cropped into three-dimensional voxel blocks of a specified pixel size to construct sample data.
[0049] For example, Sun Yat-sen University Cancer Center (Center A), Sun Yat-sen Memorial Hospital (Center B), and Shantou Central Hospital (Center C) can all be considered medical centers participating in the lymph node metastasis risk assessment study. Sample data was constructed based on the three datasets from Centers A, B, and C, with MRI images including at least T1-weighted enhanced sequences. 300 pathological images were randomly selected from these three centers, and the entire uterine lining was manually annotated by radiologists and other experts to obtain the first sample data; the remaining images were used as the second sample data. All MRI images were cropped after preprocessing. A three-dimensional voxel block of pixels.
[0050] Because different medical centers use different MRI scanner models, field strengths (1.5T or 3.0T), and acquisition parameters, the grayscale distribution in the original MRI images may have slight differences, leading to biases in feature extraction. In this embodiment, intensity normalization methods can be used for image normalization and resampling preprocessing, thereby reducing errors and improving the model's generalization ability.
[0051] S102. Use the first sample to train a deep learning segmentation network to obtain a whole uterus target segmentation model; and input the second sample into the whole uterus target segmentation model, use the whole uterus target segmentation model to extract the whole uterus volume of interest in the second sample, and use the extracted whole uterus volume of interest as the third sample.
[0052] Specifically, a deep learning segmentation network is trained using the first sample to obtain a whole uterus target segmentation model, which can be designed to include:
[0053] The first sample is divided into a training set and a test set according to a preset ratio;
[0054] A deep learning segmentation network is constructed based on the SegFormer3D network. The deep learning segmentation network includes: a hierarchical encoder for extracting multi-scale features from input image data, and a decoder for aggregating multi-scale features and outputting segmentation results.
[0055] The deep learning segmentation network is trained using the training set, and the network parameters are updated using the backpropagation algorithm.
[0056] The trained deep learning segmentation network was evaluated using a test set to obtain a whole uterus target segmentation model based on the evaluation results.
[0057] like Figure 2 As shown, an automatic whole uterus segmentation model was constructed using the SegFormer3D network. The final segmentation model achieved an average Dice similarity coefficient of 0.83 on the independent test set. This segmentation model was then used to extract the whole uterus volume of interest from each central dataset.
[0058] S103. Extract radiomics features from the third sample, optimize the parameters of the radiomics prediction model based on the radiomics features, and obtain the radiomics target prediction model. The radiomics prediction model is a lymph node metastasis risk assessment prediction model constructed using the XGBoost algorithm.
[0059] Specifically, radiomics features were extracted from the third sample, including:
[0060] Radiomics features were extracted from the third sample using an open-source medical imaging omics feature extraction tool. The radiomics features included at least: first-order statistics, shape features, first texture features described by a gray-level co-occurrence matrix, and second texture features described by a gray-level run-length matrix.
[0061] The extracted radiomics features were screened to determine the optimal radiomics features to be used as input to the prediction model.
[0062] The extracted radiomics features were screened, including:
[0063] use The univariate association between each radiomics feature and lymph node metastasis status was examined and evaluated, and radiomics features that met the expected association were retained.
[0064] Calculate the pairwise Pearson correlation coefficients between the retained radiomics features, and retain the radiomics features with strong correlations in the feature pairs based on the Pearson correlation coefficients;
[0065] LASSO regression was used to optimize the retained radiomics features, and cross-validation was used to determine the optimal regularization parameter to determine the final radiomics features obtained from the screening.
[0066] like Figure 3 As shown, a radiomics feature extraction workflow was constructed using the PyRadiomics tool. The datasets from each center were divided into training, internal, and external validation sets. A prediction model was then built using the training set. First, radiomics features were extracted from the whole uterine volume of interest for each patient in the third sample dataset. The extracted features included first-order statistics, shape features, texture features based on the gray-level co-occurrence matrix, and texture features based on the gray-level run-length matrix.
[0067] When determining the optimal feature subset, one can first use Test the univariate association between each feature and lymph node metastasis status, retaining... The features were then analyzed; subsequently, the pairwise Pearson correlation coefficients between the retained features were calculated, for highly correlated features (…). For each feature pair, the one with the stronger association to the outcome is retained; finally, LASSO regression is used to further optimize the feature set, and the optimal regularization parameter is determined through 10-fold cross-validation. Only features with non-zero coefficients are retained.
[0068] After the above multi-step screening, 40 compact and information-rich radiomics feature subsets were selected from the initially extracted 1130 features for model construction.
[0069] Based on 40 selected radiomics features, hyperparameters were optimized using five-fold cross-validation, ultimately outputting the patient's lymph node metastasis risk score and classification results. The XGBoost model's loss function can employ binary logistic loss.
[0070] This study limits the maximum tree depth to 2 to 6, the learning rate to 0.01 to 0.2, and the maximum number of trees to 500. The optimal hyperparameter combination is determined through a combination of five-fold cross-validation and grid search. Training uses binary logistic loss as the objective function, generating decision trees one by one. Each new tree continues to learn from the fitting residuals of the previous tree, ultimately outputting a lymph node metastasis risk probability mapped to 0 to 1 via a logistic function. During modeling, the importance of features in predicting lymph node metastasis risk is evaluated based on the information gain of features at split nodes.
[0071] like Figure 4 As shown, the effect of automatic uterine segmentation was verified by qualitative visual examination. The automatic segmentation contour was highly consistent with the manual drawing by experts, especially at the outer edge of the uterine myometrium and the junction with adjacent pelvic organs.
[0072] like Figure 5 As shown, the XGBoost model exhibits excellent discriminative ability on both the internal and external validation sets. The model achieves an AUC of 0.852 on the external validation set 1 (center B). On the external validation set 2 (center C), an AUC of 0.848 was achieved. ).
[0073] The trained XGBoost model is used to output a lymph node metastasis risk score for each patient, and patients are divided into high-risk and low-risk categories according to a preset threshold.
[0074] S104. Using a whole uterine target segmentation model, extract the volume of interest (ROI) of the whole uterus in the pelvic MRI of the target patient, extract the radiomics features of the whole uterine ROI of the target patient, and use the extracted radiomics features and radiomics target prediction model to assess the risk of lymph node metastasis of the whole uterine ROI of the target patient.
[0075] Specifically, consensus voting can be conducted on the output of the clinical impact medical assessment and radiomics target prediction model to classify target patients into high-risk, intermediate-risk, and low-risk groups.
[0076] Generally, routine clinical imaging assessment (based on the short diameter of lymph nodes) Standard methods have low sensitivity to lymph node metastasis, while relying solely on model predictions lacks clinical intuitiveness. Therefore, in this embodiment, a consensus voting strategy is adopted to combine the two.
[0077] like Figure 6 As shown, a consensus vote was held based on clinical imaging assessments and radiomics model predictions, dividing patients into high-risk, intermediate-risk, and low-risk groups. The low-risk group comprised 64.4% of the total patients, with a pathologically confirmed negative lymph node metastasis rate of 97.3%, thus recommending exemption from lymph node dissection. The high-risk group, although only 4.1%, had a high positive lymph node metastasis rate of 75.0%, recommending thorough lymph node dissection. Overall, based on this consensus strategy, it is estimated that approximately 62.7% of the total patient population can safely avoid unnecessary lymph node dissection.
[0078] To enhance the interpretability of the model's prediction results, this embodiment also employs the SHAP method to interpret the prediction results. The SHAP interpretability machine learning framework is used to analyze radiomics features, obtaining the shapely value corresponding to each feature, and establishing an interpretive model. The interpretive model and the shapely value corresponding to each feature are used to explain the contribution of each feature to the prediction of lymph node metastasis risk, thereby generating a feature importance ranking chart.
[0079] like Figure 7 As shown, the SHAP analysis results indicate that the texture features after wavelet transform were consistently identified as the main contributing factors to the model predictions. These features reflect the heterogeneity within the tumor and the tumor-stromal microenvironment, and are related to biological processes such as tumor dedifferentiation, proliferation, and metastatic potential.
[0080] Furthermore, based on the above method, this embodiment of the invention also provides a system for predicting lymph node metastasis of endometrial cancer based on whole uterine MRI images, comprising: a medical sample acquisition module, a segmentation model training module, a prediction model training module, and a risk prediction output module, wherein,
[0081] The medical sample acquisition module is used to acquire and preprocess pelvic MRI clinical image data publicly authorized by multiple medical centers, and randomly select N images from the clinical image data of each medical center for whole uterine target annotation. The annotated clinical image data is used as the first sample, and the remaining unannotated clinical image data is used as the second sample.
[0082] The segmentation model training module is used to train a deep learning segmentation network using the first sample to obtain a whole uterus target segmentation model; and inputs the second sample into the whole uterus target segmentation model, uses the whole uterus target segmentation model to extract the whole uterus volume of interest in the second sample, and uses the extracted whole uterus volume of interest as the third sample;
[0083] The prediction model training module is used to extract radiomics features from the third sample, optimize the parameters of the radiomics prediction model based on the radiomics features, and obtain the radiomics target prediction model. The radiomics prediction model is a lymph node metastasis risk assessment prediction model constructed using the XGBoost algorithm.
[0084] The risk prediction output module is used to extract the volume of interest (LOI) of the entire uterus in the pelvic MRI of the target patient using a whole uterine target segmentation model, extract the radiomics features of the whole uterine LIO of the target patient, and perform lymph node metastasis risk assessment on the whole uterine LIO of the target patient based on the extracted radiomics features and using a radiomics target prediction model.
[0085] This approach utilizes preoperative pelvic MRI images of endometrial cancer patients to obtain the entire uterine tissue. It employs deep learning-based automatic segmentation technology to acquire the volume of interest in the entire uterus, and combines radiomics and machine learning techniques to construct a lymph node metastasis prediction model. The model outputs a risk score, and a consensus voting strategy based on clinical imaging assessment and model prediction is used to stratify patient risk. This approach can assist clinicians in developing more accurate preoperative surgical plans, reduce unnecessary lymph node dissections, and accelerate the clinical translation of artificial intelligence models.
[0086] Unless otherwise specifically stated, the relative steps, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of the invention.
[0087] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0088] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.
[0089] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.
[0090] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for predicting lymph node metastasis in endometrial cancer based on whole-uterine MRI images, characterized in that, Include: We acquired and preprocessed pelvic MRI clinical image data from multiple medical centers that were publicly authorized. We then randomly selected N images from the clinical image data of each medical center for whole-uterine target annotation. The annotated clinical image data was used as the first sample, and the remaining unannotated clinical image data was used as the second sample. A deep learning segmentation network is trained using the first sample to obtain a whole uterus target segmentation model; the second sample is then input into the whole uterus target segmentation model, and the whole uterus volume of interest is extracted from the second sample using the whole uterus target segmentation model. The extracted whole uterus volume of interest is then used as the third sample. Radiomics features were extracted from the third sample, and the parameters of the radiomics prediction model were optimized based on the radiomics features to obtain the radiomics target prediction model. The radiomics prediction model was constructed using the XGBoost algorithm to assess the risk of lymph node metastasis. The whole uterus target segmentation model was used to extract the whole uterus volume of interest from the pelvic MRI of the target patient, and the radiomics features of the whole uterus volume of interest of the target patient were extracted. Based on the extracted radiomics features, the radiomics target prediction model was used to assess the risk of lymph node metastasis of the whole uterus volume of interest of the target patient.
2. The method for predicting lymph node metastasis of endometrial cancer based on whole uterine MRI images according to claim 1, characterized in that, Preprocessing of pelvic MRI clinical image data includes: Clinical image data undergoes data cleaning and normalization. The normalized image data is cropped into three-dimensional voxel blocks of a specified pixel size, so as to construct sample data using the three-dimensional voxel blocks.
3. The method for predicting lymph node metastasis of endometrial cancer based on whole uterine MRI images according to claim 1, characterized in that, A deep learning segmentation network is trained using the first sample to obtain a whole uterus target segmentation model, which includes: The first sample is divided into a training set and a test set according to a preset ratio; A deep learning segmentation network is constructed based on the SegFormer3D network. The deep learning segmentation network includes: a hierarchical encoder for extracting multi-scale features from input image data, and a decoder for aggregating multi-scale features and outputting segmentation results. The deep learning segmentation network is trained using the training set, and the network parameters are updated using the backpropagation algorithm. The trained deep learning segmentation network was evaluated using a test set to obtain a whole uterus target segmentation model based on the evaluation results.
4. The method for predicting lymph node metastasis of endometrial cancer based on whole uterine MRI images according to claim 1, characterized in that, Extracting radiomics features from the third sample, including: Radiomics features were extracted from the third sample using an open-source medical imaging omics feature extraction tool. The radiomics features included at least: first-order statistics, shape features, first texture features described by a gray-level co-occurrence matrix, and second texture features described by a gray-level run-length matrix. The extracted radiomics features were screened to determine the optimal radiomics features to be used as input to the prediction model.
5. The method for predicting lymph node metastasis of endometrial cancer based on whole uterine MRI images according to claim 4, characterized in that, The extracted radiomics features were screened, including: The Mann-Whitney U test was used to assess the univariate association between each radiomics feature and lymph node metastasis status, and radiomics features that met the expected association were retained. Calculate the pairwise Pearson correlation coefficients between the retained radiomics features, and retain the radiomics features with strong correlations in the feature pairs based on the Pearson correlation coefficients; LASSO regression was used to optimize the retained radiomics features, and cross-validation was used to determine the optimal regularization parameter to determine the final radiomics features obtained from the screening.
6. The method for predicting lymph node metastasis of endometrial cancer based on whole uterine MRI images according to claim 1, characterized in that, A lymph node metastasis risk assessment and prediction model was constructed using the XGBoost algorithm, including: Decision trees are used to model the input radiomics features. In each iteration, the best features and split points are selected by splitting nodes, and the trees are built step by step. Each tree is modeled on the residual of the previous tree. In the modeling process, the features for building the prediction model are selected by evaluating the importance of the features for predicting the risk of lymph node metastasis.
7. The method for predicting lymph node metastasis of endometrial cancer based on whole uterine MRI images according to claim 1, characterized in that, The risk of lymph node metastasis in the entire uterine volume of interest of the target patient was assessed using a radiomics-based target prediction model, and also included: Consensus voting was conducted on the outputs of the clinical impact medical assessment and radiomics target prediction model to classify target patients into high-risk, intermediate-risk, and low-risk groups. The radiomics features were analyzed using the SHAP interpretable machine learning framework to obtain the shapely value corresponding to each feature and to build an interpretive model. The contribution of each feature to the prediction of lymph node metastasis risk is explained by interpreting the model and the shapely value corresponding to each feature, so as to generate a feature importance ranking chart.
8. A system for predicting lymph node metastasis in endometrial cancer based on whole-uterine MRI images, characterized in that, It includes: a medical sample acquisition module, a segmentation model training module, a prediction model training module, and a risk prediction output module. The medical sample acquisition module is used to acquire and preprocess pelvic MRI clinical image data publicly authorized by multiple medical centers, and randomly select N images from the clinical image data of each medical center for whole uterine target annotation. The annotated clinical image data is used as the first sample, and the remaining unannotated clinical image data is used as the second sample. The segmentation model training module is used to train a deep learning segmentation network using the first sample to obtain a whole uterus target segmentation model; and inputs the second sample into the whole uterus target segmentation model, uses the whole uterus target segmentation model to extract the whole uterus volume of interest in the second sample, and uses the extracted whole uterus volume of interest as the third sample; The prediction model training module is used to extract radiomics features from the third sample, optimize the parameters of the radiomics prediction model based on the radiomics features, and obtain the radiomics target prediction model. The radiomics prediction model is a lymph node metastasis risk assessment prediction model constructed using the XGBoost algorithm. The risk prediction output module is used to extract the volume of interest (LOI) of the entire uterus in the pelvic MRI of the target patient using a whole uterine target segmentation model, extract the radiomics features of the whole uterine LIO of the target patient, and perform lymph node metastasis risk assessment on the whole uterine LIO of the target patient based on the extracted radiomics features and using a radiomics target prediction model.
9. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 7.