Cervical cancer assisted chemoradiotherapy reaction prediction method based on multi-mode MRI deep radiomics
By combining imaging omics and deep learning technology, the features of multimodal MRI images are extracted and multiple feature screening and classification models are adopted, the limitations of feature extraction and screening methods in the existing technology are solved, and the prediction accuracy of adjuvant chemoradiation and treatment response in cervical cancer is improved.
Patent Information
- Application Number
- CN202411686327.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-22
- Publication Date
- 2025-03-14
AI Technical Summary
The prior art has limitations in the prediction of adjuvant chemoradiotherapy response in cervical cancer, the singularity of feature screening methods and the limitations of prediction models, resulting in low prediction accuracy.
By combining imaging omics and deep learning technology, artificial imaging features and deep learning features are extracted separately, and a variety of feature screening methods and classification models are used to predict the response to adjuvant chemoradiation and treatment of cervical cancer.
The prediction accuracy of neoadjuvant chemoradiation and chemotherapy response in cervical cancer has been improved, and the accuracy and stability of the model have been improved through multi-level feature extraction, diversity feature screening and integrated learning.
Smart Images
Figure CN119626541A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer artificial intelligence, and specifically relates to a method for predicting the response of cervical cancer to adjuvant chemoradiotherapy based on multimodal MRI deep radiomics. Background Art
[0002] Magnetic resonance imaging (MRI) is a diagnostic technology that uses the magnetic resonance phenomenon of certain atomic nuclei in human tissue to reconstruct images of a certain layer of the human body by electronic computer processing of radio frequency signals. It provides high-resolution soft tissue contrast images under non-invasive conditions and is an important tool for diagnosing cervical cancer. Radiomics extracts quantitative features from medical images, including extracting information from complex patterns that are difficult for the human eye to recognize or quantify. This information can be used for diagnosis, grading, and prediction of treatment response. As a branch of artificial intelligence, deep learning automatically learns complex features through multi-layer neural networks and has been widely used in image processing and analysis.
[0003] The existing technical solutions only use radiomics to extract artificial image features from MRI, and use a single feature screening method to screen image features, and finally use machine learning classification methods to predict the results. This application proposal uses radiomics and deep learning models to extract artificial image features and deep learning features respectively, and combines multiple feature screening methods, and uses multiple machine learning classification methods to predict, so as to achieve the best prediction effect.
[0004] The existing technology has defects:
[0005] 1. Limitations of artificial feature extraction: Existing technologies only use radiomics methods to extract artificially defined features, while the deep abstract features of the image are not extracted.
[0006] 2. Single feature screening method: Using only one feature screening method may not be able to effectively distinguish important features from noise features, affecting the accuracy of the model.
[0007] 3. Limitations of prediction models: Most existing technologies use a single machine learning classification method and cannot fully combine the advantages of different models to improve prediction results. Summary of the invention
[0008] The purpose of the present invention is to provide a method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics. By combining radiomics with deep learning technology, artificial image features and deep learning features are extracted respectively, and a variety of feature screening methods and classification models are applied to improve the prediction accuracy of the response to neoadjuvant chemoradiotherapy for cervical cancer.
[0009] To achieve the above objectives, the technical solution of the present invention is: a method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics, which combines radiomics with deep learning technology to extract artificial image features and deep learning features respectively, and applies a variety of feature screening methods and classification models to achieve the prediction of the response to adjuvant chemoradiotherapy for cervical cancer.
[0010] In one embodiment of the present invention, the method includes:
[0011] (1) Multimodal MRI data acquisition: Image data of cervical cancer patients are acquired through multimodal MRI, including: T1-weighted contrast-enhanced imaging and T2-weighted imaging image data;
[0012] (2) Data preprocessing: preprocessing the image data of T1-weighted contrast-enhanced imaging and T2-weighted imaging;
[0013] (3) Feature extraction, including:
[0014] Radiomics feature extraction: extract a series of artificial image features from preprocessed images;
[0015] Deep learning feature extraction: Use pre-trained deep neural networks to extract deep learning features from pre-processed images;
[0016] (4) Feature screening: Combine multiple feature screening methods to screen the extracted features;
[0017] (5) Multimodal feature combination: combining the selected multimodal features;
[0018] (6) Classification model training: Use a variety of machine learning classification algorithms to train the screened features, and select the best classifier combination through cross-validation.
[0019] In one embodiment of the present invention, after multimodal MRI data acquisition, ITK-Snap is used to manually outline the tumor area.
[0020] In one embodiment of the present invention, the data preprocessing process includes image standardization, registration, and noise reduction processing.
[0021] In one embodiment of the present invention, in the radiomics feature extraction process, Pyradiomics is used to extract artificial image features from the pre-processed images.
[0022] In one embodiment of the present invention, the deep learning feature extraction process uses the pre-trained ResNet50 to extract deep learning features from the pre-processed image, that is, the deep learning features are extracted through the convolution operation, which is specifically implemented as follows:
[0023] The deep learning features are extracted by convolving the convolution kernel with the input preprocessed image. The convolution formula is as follows:
[0024]
[0025] Among them, f(x,y) is the value of the output feature map at position (x,y); I(x+i,y+j) is the pixel value of the input image at position (x+i,y+j); K(i,j) is the value of the convolution kernel at position (i,j); m and n are the half-width and half-height of the convolution kernel, rounded down.
[0026] In one embodiment of the present invention, the feature screening process adopts ElasticNet and SVM-RFE feature screening methods, wherein ElasticNet combines the advantages of L1 and L2 regularization for feature selection; the ElasticNet regression formula is as follows:
[0027]
[0028] Where N is the number of samples, p is the number of features, and X ij is the jth eigenvalue of sample i, y i is the actual output value of sample i, β j is the regression coefficient of the jth feature, λ1 and λ2 are the weight parameters of L1 and L2 regularization respectively;
[0029] SVM-RFE stands for Support Vector Machine-Recursive Feature Elimination: It uses support vector machine to perform recursive feature elimination and gradually remove the least important features. The SVM-RFE selection criterion is to select the feature with the smallest absolute value of weight for elimination based on the absolute value of the weight in the support vector machine.
[0030] After being filtered by ElasticNet and SVM-RFE, the features are then subjected to a circular binary search to test the optimal number of features.
[0031] In one embodiment of the present invention, the multimodal feature combination process is specifically implemented as follows:
[0032] The features of the image data of T1-weighted contrast-enhanced imaging and T2-weighted imaging after preprocessing and subjected to radiomics feature extraction and deep learning feature extraction are expressed as: T1 radiomics feature, T2 radiomics feature, T1 deep learning feature, T2 deep learning feature, respectively;
[0033] Weighted fusion: The weighted fusion formula (1) is used to fuse T1 radiomics features and T2 radiomics features, and the T1 deep learning features and T2 deep learning features are fused. The weighted fusion formula (2) is then used to fuse the weighted fusion radiomics features and deep learning features.
[0034] F combinedT =αF T1 +βF T2 (1)
[0035] F combinedF =αF RF +βF HF (2)
[0036] Among them, F T1 and F T2 are the eigenvectors of T1 and T2 modes, respectively, and F RF and F HF are the feature vectors of radiomics features and deep learning features, respectively, and α and β are weighting coefficients;
[0037] Through the weighted fusion process, the fused radiomics feature F after the fusion of T1 radiomics feature and T2 radiomics feature is obtained. combinedT1 , the fused deep learning feature F after the fusion of T1 deep learning feature and T2 deep learning feature combinedT2 , the image after fusion of radiomics features, deep learning features and deep fusion features F combinedF ;
[0038] The T1 radiomics features, T2 radiomics features, T1 deep learning features, T2 deep learning features, and fusion radiomics features F combinedT1 , fusion deep learning features F combinedT2 , image and depth fusion features F combinedF Seven feature sets are formed for training classification models.
[0039] In one embodiment of the present invention, the classification model training process adopts machine learning classification algorithms such as support vector machine, random forest, and GradientBoosting.
[0040] In one embodiment of the present invention, the implementation formulas of the machine learning classification algorithms of support vector machine, random forest, and GradientBoosting are respectively as follows:
[0041] Support Vector Machines:
[0042]
[0043] Among them, K(x i ,x) is the kernel function, α i and b are model parameters, sign is the sign function, y i is the label of the i-th sample in the training set, x i is the feature vector of the i-th sample in the training set, and x is the feature vector of the newly input sample;
[0044] Random Forest:
[0045]
[0046] Where T is the number of trees, h t (x) is the prediction result of the tth tree, x is the input parameter;
[0047] GradientBoosting:
[0048]
[0049] in, is the loss function, Ω(f k ) is the model complexity penalty term, y i is the true value, is the model prediction value, f k is the kth tree, N is the total number of samples in the training data, and K is the total number of trees in the model.
[0050] Compared with the prior art, the present invention has the following beneficial effects:
[0051] 1. Multi-level feature extraction: This model not only uses traditional imaging genomics to extract artificial features, but also extracts complex deep-level features through deep learning, so that the prediction model can capture more comprehensive image information.
[0052] 2. Diversity of feature screening: Introducing a variety of feature screening methods to effectively remove redundant features, retain the most valuable features for prediction, and improve the accuracy of the model.
[0053] 3. Integration advantages of classification models: By combining the advantages of multiple classification models through integrated learning, the stability and robustness of the prediction results are improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0055] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.
[0056] like Figure 1 As shown, the present invention provides a method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics. By combining radiomics with deep learning technology, artificial image features and deep learning features are extracted respectively, and a variety of feature screening methods and classification models are applied to achieve the prediction of the response to adjuvant chemoradiotherapy for cervical cancer.
[0057] In one embodiment of the present invention, the method includes:
[0058] (1) Multimodal MRI data acquisition: Image data of cervical cancer patients were collected through multimodal MRI (T1-weighted contrast-enhanced imaging, T2-weighted imaging), and the tumor area was manually delineated using ITK-Snap;
[0059] (2) Data preprocessing: Preprocess the image data of T1-weighted contrast-enhanced imaging and T2-weighted imaging; preprocessing includes image standardization, registration, and noise reduction to eliminate variability in the imaging process and ensure the accuracy of subsequent analysis;
[0060] (3) Feature extraction, including:
[0061] Radiomics feature extraction: extract a series of artificial image features from preprocessed images; such as texture, shape, intensity, etc.
[0062] Pyradiomics is an open source python package that can extract radiomics features from medical images. The image type and feature type used for feature extraction are specified by custom extraction features, as shown in Table 1 for image types and Table 2 for feature types:
[0063] Table 1
[0064]
[0065] Table 2
[0066] Feature Type explain FirstOrderFeatures First-order features ShapeFeatures(2D) 2D shape features ShapeFeatures(3D) 3D shape features GLCM GrayLevelCooccurrenceMatrixFeatures Gray Level Cooccurrence Matrix Features GLS GrayLevelSizeZoneMatrixFeatures Gray level size zone matrix features GLRLM GrayLevelRunLengthMatrixFeatures Gray Level Run Length Matrix Features NGTDM NeighbouringGrayToneDifferenceMatrixFeatures adjacent grayscale difference matrix GLDM GrayLevelDependenceMatrixFeatures Gray Level Dependence Matrix Features
[0067] Deep learning feature extraction: Use a pre-trained deep neural network (pre-trained ResNet50) to automatically extract deep features from the image, that is, extract deep learning features through convolution operations. The specific implementation is as follows:
[0068] The deep learning features are extracted by convolving the convolution kernel with the input preprocessed image. The convolution formula is as follows:
[0069]
[0070] Among them, f(x,y) is the value of the output feature map at position (x,y); I(x+i,y+j) is the pixel value of the input image at position (x+i,y+j); K(i,j) is the value of the convolution kernel at position (i,j); m and n are the half-width and half-height of the convolution kernel, rounded down.
[0071] (4) Feature screening: Combine multiple feature screening methods to screen the extracted features; Combine multiple feature screening methods (ElasticNet, SVM-RFE) to screen the extracted features and retain the features with the most predictive value.
[0072] ElasticNet combines the advantages of L1 and L2 regularization for feature selection; the ElasticNet regression formula is as follows:
[0073]
[0074] Where N is the number of samples, p is the number of features, and X ij is the jth eigenvalue of sample i, y i is the actual output value of sample i, β j is the regression coefficient of the jth feature, λ1 and λ2 are the weight parameters of L1 and L2 regularization respectively;
[0075] SVM-RFE stands for Support Vector Machine-Recursive Feature Elimination: It uses support vector machine to perform recursive feature elimination and gradually remove the least important features. The SVM-RFE selection criterion is to select the feature with the smallest absolute value of weight for elimination based on the absolute value of the weight in the support vector machine.
[0076] After being filtered by ElasticNet and SVM-RFE, the features are then subjected to a circular binary search to test the optimal number of features.
[0077] (5) Multimodal feature combination: combining the selected multimodal features;
[0078] The multimodal feature combination process is specifically implemented as follows:
[0079] The features of the image data of T1-weighted contrast-enhanced imaging and T2-weighted imaging after preprocessing and subjected to radiomics feature extraction and deep learning feature extraction are expressed as: T1 radiomics feature, T2 radiomics feature, T1 deep learning feature, T2 deep learning feature, respectively;
[0080] Weighted fusion: The weighted fusion formula (1) is used to fuse T1 radiomics features and T2 radiomics features, and the T1 deep learning features and T2 deep learning features are fused. The weighted fusion formula (2) is then used to fuse the weighted fusion radiomics features and deep learning features.
[0081] F combinedT =αF T1 +βF T2 (1)
[0082] F combinedF =αF RF +βF HF (2)
[0083] Among them, F T1 and F T2 are the eigenvectors of T1 and T2 modes, respectively, and FRF and F HF are the feature vectors of radiomics features and deep learning features, respectively, and α and β are weighting coefficients;
[0084] Through the weighted fusion process, the fused radiomics feature F after the fusion of T1 radiomics feature and T2 radiomics feature is obtained. combinedT1 , the fused deep learning feature F after the fusion of T1 deep learning feature and T2 deep learning feature combinedT2 , the image after fusion of radiomics features, deep learning features and deep fusion features F combinedF ;
[0085] The T1 radiomics features, T2 radiomics features, T1 deep learning features, T2 deep learning features, and fusion radiomics features F combinedT1 , fusion deep learning features F combinedT2 , image and depth fusion features F combinedF Seven feature sets are formed for training classification models.
[0086] (6) Classification model training: Use a variety of machine learning classification algorithms (support vector machine, random forest, extreme tree, GradientBoosting) to train the screened features, and select the best classifier combination through cross-validation. The implementation formulas of the machine learning classification algorithms of support vector machine, random forest, and GradientBoosting are as follows:
[0087] Support Vector Machines:
[0088]
[0089] Among them, K(x i ,x) is the kernel function, α i and b are model parameters, sign is the sign function, y i is the label of the i-th sample in the training set, x i is the feature vector of the i-th sample in the training set, and x is the feature vector of the newly input sample;
[0090] Random Forest:
[0091]
[0092] Where T is the number of trees, h t (x) is the prediction result of the tth tree, x is the input parameter;
[0093] Gradient Boosting:
[0094]
[0095] in, is the loss function, Ω(f k ) is the model complexity penalty term, y i is the true value, is the model prediction value, f k is the kth tree, N is the total number of samples in the training data, and K is the total number of trees in the model.
[0096] (7) Model validation and evaluation: Use an independent test data set to verify the model performance, and evaluate its prediction effect through indicators such as ROC curve and AUC value.
[0097] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.
Claims
1. A method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics, characterized in that: By combining radiomics and deep learning technology, artificial image features and deep learning features are extracted respectively, and a variety of feature screening methods and classification models are applied to achieve the prediction of adjuvant chemoradiotherapy response for cervical cancer.
2. The method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics according to claim 1, characterized in that: include: (1) Multimodal MRI data acquisition: Image data of cervical cancer patients are acquired through multimodal MRI, including: T1-weighted contrast-enhanced imaging and T2-weighted imaging image data; (2) Data preprocessing: preprocessing the image data of T1-weighted contrast-enhanced imaging and T2-weighted imaging; (3) Feature extraction, including: Radiomics feature extraction: extract a series of artificial image features from preprocessed images; Deep learning feature extraction: Use pre-trained deep neural networks to extract deep learning features from pre-processed images; (4) Feature screening: Combine multiple feature screening methods to screen the extracted features; (5) Multimodal feature combination: combining the selected multimodal features; (6) Classification model training: Use a variety of machine learning classification algorithms to train the screened features, and select the best classifier combination through cross-validation.
3. The method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics according to claim 2, characterized in that: After multimodal MRI data acquisition, the tumor region was manually delineated using ITK-Snap.
4. The method for predicting the response to adjuvant chemoradiotherapy of cervical cancer based on multimodal MRI deep radiomics according to claim 2, characterized in that: The data preprocessing process includes image standardization, registration, and noise reduction.
5. The method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics according to claim 2, characterized in that: In the radiomics feature extraction process, Pyradiomics was used to extract artificial image features from the preprocessed images.
6. The method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics according to claim 2, characterized in that: The deep learning feature extraction process uses the pre-trained ResNet50 to extract deep learning features from the pre-processed image, that is, deep learning features are extracted through convolution operations. The specific implementation is as follows: The deep learning features are extracted by convolving the convolution kernel with the input preprocessed image. The convolution formula is as follows: Among them, f(x,y) is the value of the output feature map at position (x,y); I(x+i,y+j) is the pixel value of the input image at position (x+i,y+j); K(i,j) is the value of the convolution kernel at position (i,j); m and n are the half-width and half-height of the convolution kernel, rounded down.
7. The method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics according to claim 2, characterized in that: The feature screening process uses ElasticNet and SVM-RFE feature screening methods, where: ElasticNet combines the advantages of L1 and L2 regularization for feature selection; the ElasticNet regression formula is as follows: Where N is the number of samples, p is the number of features, and X ij is the jth eigenvalue of sample i, y i is the actual output value of sample i, β j is the regression coefficient of the jth feature, λ1 and λ2 are the weight parameters of L1 and L2 regularization respectively; SVM-RFE stands for Support Vector Machine-Recursive Feature Elimination: It uses support vector machine to perform recursive feature elimination and gradually remove the least important features. The SVM-RFE selection criterion is to select the feature with the smallest absolute value of weight for elimination based on the absolute value of the weight in the support vector machine. After being filtered by ElasticNet and SVM-RFE, the features are then subjected to a circular binary search to test the optimal number of features.
8. The method for predicting the response to adjuvant chemoradiotherapy of cervical cancer based on multimodal MRI deep radiomics according to claim 2, characterized in that: The multimodal feature combination process is specifically implemented as follows: The features of the image data of T1-weighted contrast-enhanced imaging and T2-weighted imaging after preprocessing and subjected to radiomics feature extraction and deep learning feature extraction are expressed as: T1 radiomics feature, T2 radiomics feature, T1 deep learning feature, T2 deep learning feature, respectively; Weighted fusion: The weighted fusion formula (1) is used to fuse T1 radiomics features and T2 radiomics features, and the T1 deep learning features and T2 deep learning features are fused. The weighted fusion formula (2) is then used to fuse the weighted fusion radiomics features and deep learning features. F combinedT =αF T1 +βF T2 (1) F combinedF =αF RF +βF HF (2) Among them, F T1 and F T2 are the eigenvectors of T1 and T2 modes, respectively, and F RF and F HF are the feature vectors of radiomics features and deep learning features, respectively, and α and β are weighting coefficients; Through the weighted fusion process, the fused radiomics feature F after the fusion of T1 radiomics feature and T2 radiomics feature is obtained. combinedT1 , the fused deep learning feature F after the fusion of T1 deep learning feature and T2 deep learning feature combinedT2 , the image after fusion of radiomics features, deep learning features and deep fusion features F combinedF ; The T1 radiomics features, T2 radiomics features, T1 deep learning features, T2 deep learning features, and fusion radiomics features F combinedT1 , fusion deep learning features F combinedT2 , image and depth fusion features F combinedF Seven feature sets are formed for training classification models.
9. The method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics according to claim 2, characterized in that: The classification model training process uses machine learning classification algorithms such as support vector machine, random forest, and GradientBoosting.
10. The method for predicting the response to adjuvant chemoradiotherapy for cervical cancer based on multimodal MRI deep radiomics according to claim 9, characterized in that: The implementation formulas of the machine learning classification algorithms of support vector machine, random forest, and GradientBoosting are as follows: Support Vector Machines: Among them, K(x i ,x) is the kernel function, α i and b are model parameters, sign is the sign function, y i is the label of the i-th sample in the training set, x i is the feature vector of the i-th sample in the training set, and x is the feature vector of the newly input sample; Random Forest: Where T is the number of trees, h t (x) is the prediction result of the tth tree, x is the input parameter; GradientBoosting: in, is the loss function, Ω(f k ) is the model complexity penalty term, y i is the true value, is the model prediction value, f k is the kth tree, N is the total number of samples in the training data, and K is the total number of trees in the model.