Clinical parotid gland classification method based on CT image omics
By combining CT image-omics, deep learning features and clinical data in the diagnosis of parotid tumors, and using a three-phase fusion model, the problem of low diagnostic accuracy in parotid tumors in the prior art is solved, achieving higher diagnostic accuracy and timeliness of clinical applications.
Patent Information
- Application Number
- CN202510063982.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-15
- Publication Date
- 2025-05-13
AI Technical Summary
The prior art has low accuracy in the diagnosis of benign and malignant parotid tumors, is susceptible to subjective experience, and imagingomics methods have problems such as difficulty in extracting features and insufficient classification accuracy.
The parotid gland classification method based on CT image omics and clinical practice is adopted. By combining imaging omics features, deep learning features and clinical data, a three-phase fusion model is used to improve the accuracy of benign and malignant classification of parotid tumors.
It realizes accurate preoperative classification of benign and malignant diseases of parotid tumors, improves the accuracy and reliability of diagnosis, and provides a more reliable basis for clinical decision-making.
Smart Images
Figure CN119992185A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of CT image processing, and in particular to a parotid gland classification method based on CT image omics and clinical practice. Background Art
[0002] Parotid gland tumors are the most common type of salivary gland tumors, and the diagnosis of benign and malignant tumors is of great significance for clinical treatment. At present, preoperative diagnostic methods mainly rely on fine needle aspiration biopsy (FNAB) and traditional imaging examinations (such as CT, MRI, ultrasound, etc.), but these methods have the disadvantages of low accuracy and susceptibility to subjective experience. Therefore, how to accurately judge the benign and malignant nature of parotid gland tumors before surgery has become a difficult problem in clinical medicine.
[0003] As an emerging technology, radiomics provides a new direction for tumor classification and prognosis assessment by quantitatively analyzing microscopic features in medical images and establishing diagnostic models with deep learning methods. However, existing technologies still have problems such as single methods, difficulty in feature extraction, and insufficient classification accuracy.
[0004] In order to solve the above technical problems, the present invention provides a method for classifying benign and malignant parotid tumors based on arterial phase CT images. By integrating radiomics and CT image deep learning technology and combining clinical characteristics, accurate benign and malignant classification can be achieved before surgery. Summary of the invention
[0005] The purpose of the present invention is to provide a parotid gland classification method based on CT image omics and clinical data. The method combines imaging omics features, deep learning features and clinical data, and improves the accuracy of benign and malignant classification of parotid gland tumors through a three-phase fusion model, providing a more reliable basis for clinical preoperative decision-making. The specific steps are:
[0006] Preferably, in step 1, the method for labeling one of the parotid gland tumor CT images is specifically as follows:
[0007] 1.1. Use ITK-SNAP software to mark the tumor area and obtain the marked area within the tumor;
[0008] 1.2. Use Python software to expand the marked area within the tumor by 2 mm to obtain the marked area around the tumor.
[0009] Preferably, in step 2, the omics feature extraction method is specifically:
[0010] 2.1. The PyRadiomics library was used to extract radiomic features from the intratumoral area, including first-order statistical features, texture features (such as GLCM, GLRLM, etc.) and other high-order texture features. A total of 1,874 intratumoral radiomic features were extracted.
[0011] 2.2. The PyRadiomics library was used to extract radiomic features from the peritumoral area, including first-order statistical features, texture features (such as GLCM, GLRLM, etc.) and other high-order texture features. A total of 1,874 peritumoral radiomic features were extracted.
[0012] 2.3. The PyRadiomics library was used to extract radiomic features from the intratumoral and peritumoral areas, including first-order statistical features, texture features (such as GLCM, GLRLM, etc.) and other high-order texture features. A total of 1,874 intratumoral and peritumoral radiomic features were extracted.
[0013] Preferably, in step 3, the method of feature screening is:
[0014] 3.1. Use SPSS software to perform univariate and multivariate logistic regression on clinical characteristics to screen features with p < 0.05, and screen the remaining 3 features;
[0015] 3.2. Pearson correlation analysis was used to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features, respectively, to eliminate redundant features and retain features without high correlation;
[0016] 3.3. Use T test to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.2, and screen out the features with significant differences;
[0017] 3.4. Use LASSO regression to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.3, and screen out the features that are closely related to benign and malignant classification;
[0018] 3.5. Use RFE to calculate the feature importance of the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.4, and screen out the top 20 features with the highest importance.
[0019] Preferably, in step 4, the specific steps of performing image processing and data augmentation on the CT image are:
[0020] 4.1. Use Python software to read CT image files and perform 40-400 window pixel capture, as well as read annotation files, and extract intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices in depth direction;
[0021] 4.2. Use Python software to normalize the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices, specifically by subtracting the mean and then dividing by the difference between the maximum value and the minimum value, and normalizing to the range of -1 to 1;
[0022] 4.3. Use Python software to rotate the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices of the training set within the range of -10° to 10°, flip them horizontally and vertically, and add Gaussian noise with a mean of 0 and a maximum of 0.01;
[0023] 4.4. Use Python software to regularize the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices to a size of 256*256, and fill the insufficient parts with 0.
[0024] Preferably, in step 5, the method for constructing a benign and malignant classification model for parotid tumors is specifically as follows:
[0025] 5.1. Use the torchvision.models.mobilenet_v3_small model and import the torchvision.models.MobileNet_V3_Small_Weights.DEFAULT default weights to build a basic model.
[0026] 5.2. Modify the basic model features[0][0] = nn.Conv2d(1,16,kernel_size = (3,3),stride = (2,2),padding = (1,1),bias = False). The modification point is to change the default input channel 3 to input channel 1.
[0027] 5.3. Modify the base model’s classifier[3]=nn.Linear(self.base_model.classifier[3].in_features,100). The modification point is to change the number of output features to 100.
[0028] 5.4. Use 3 fully connected layers to build a perceptron classifier; the first layer inputs the features output by the basic model + omics features, and the output is 1000; the second layer inputs 1000 and the output is 100; the third layer inputs 100 and the output is 1.
[0029] 5.5. Use the weighted Binary Cross-Entropy with Logits Loss loss function to calculate the loss and set the pos_weight value to 3.
[0030] 5.6. Use intratumoral sections + intratumoral omics features, peritumoral sections + peritumoral omics features, and intratumoral + peritumoral sections + intratumoral + peritumoral omics features to train three fusion models, and use AUC as the evaluation index to select the best model;
[0031] 5.7. Use 3 fully connected layers to build a perceptron classifier; the first layer inputs the features output by the basic model + omics features, and the output is 1000; the second layer inputs 1000 and the output is 100; the third layer inputs 100 and the output is 1.
[0032] 5.8. Based on the best model in 5.6, the final model is trained using the corresponding CT image slices + omics features + clinical features.
[0033] Preferably, in step six, the performance evaluation parameters for obtaining the benign or malignant prediction classification results include: accuracy, recall and precision.
[0034] Compared with related technologies, the parotid gland classification method based on CT image omics and clinical practice provided by the present invention has the following beneficial effects:
[0035] 1. The method of the present invention expands the peritumoral annotated area based on the intratumoral annotated area, and is not limited to the intratumoral area. More information is extracted from the intratumoral and peritumoral areas;
[0036] 2. A fusion model was established based on CT images, CT radiomics features and clinical characteristics, combining multimodal information to improve the prediction accuracy of the model and ensure the timeliness and practicality of the parotid tumor classification model in clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 This is a graph of the omics feature screening results in the present invention.
[0038] Figure 2 This is a model structure diagram in the present invention.
[0039] Figure 3 It is the ROC curve diagram of the prediction results in the present invention. DETAILED DESCRIPTION
[0040] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.
[0041] The present invention provides a parotid gland classification method based on CT image omics and clinical practice, comprising the following steps:
[0042] Step 1: Based on the parotid gland tumor CT image, mark the tumor area and expand the surrounding area to mark the tumor area; the specific steps of the marking processing method for one of the parotid gland tumor CT images are as follows:
[0043] 1.1. Use ITK-SNAP software to mark the tumor area and obtain the marked area within the tumor;
[0044] 1.2. Use Python software to expand the marked area within the tumor by 2 mm to obtain the marked area around the tumor.
[0045] Step 2: Extract omics features from the tumor area and peritumoral area. The specific steps are as follows:
[0046] 2.1. The PyRadiomics library was used to extract radiomic features from the intratumoral area, including first-order statistical features, texture features (such as GLCM, GLRLM, etc.) and other high-order texture features. A total of 1,874 intratumoral radiomic features were extracted.
[0047] 2.2. The PyRadiomics library was used to extract radiomic features from the peritumoral area, including first-order statistical features, texture features (such as GLCM, GLRLM, etc.) and other high-order texture features. A total of 1,874 peritumoral radiomic features were extracted.
[0048] 2.3. The PyRadiomics library was used to extract radiomic features from the intratumoral and peritumoral areas, including first-order statistical features, texture features (such as GLCM, GLRLM, etc.) and other high-order texture features. A total of 1,874 intratumoral and peritumoral radiomic features were extracted.
[0049] Step 3: Screen the omics features and clinical features. The specific steps are as follows:
[0050] 3.1. Use SPSS software to perform univariate and multivariate logistic regression on clinical characteristics to screen features with p < 0.05, and screen the remaining 3 features;
[0051] 3.2. Pearson correlation analysis was used to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features, respectively, to eliminate redundant features and retain features without high correlation;
[0052] 3.3. Use T test to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.2, and screen out the features with significant differences;
[0053] 3.4. Use LASSO regression to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.3, and screen out the features that are closely related to benign and malignant classification;
[0054] 3.5. Use RFE to calculate the feature importance of the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.4, and screen out the top 20 features with the highest importance.
[0055] Step 4: Perform image processing and data augmentation on the CT image. The specific steps are as follows:
[0056] 4.1. Use Python software to read CT image files and perform 40-400 window pixel capture, as well as read annotation files, and extract intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices in depth direction;
[0057] 4.2. Use Python software to normalize the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices, specifically by subtracting the mean and then dividing by the difference between the maximum value and the minimum value, and normalizing to the range of -1 to 1;
[0058] 4.3. Use Python software to rotate the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices of the training set within the range of -10° to 10°, flip them horizontally and vertically, and add Gaussian noise with a mean of 0 and a maximum of 0.01;
[0059] 4.4. Use Python software to regularize the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices to a size of 256*256, and fill the insufficient parts with 0.
[0060] Step 5: Input the parotid CT images to be classified, omics features, and clinical features into the trained classification model to obtain the benign and malignant prediction classification results. The specific method for constructing the benign and malignant classification model for parotid tumors is as follows;
[0061] 5.1. Use the torchvision.models.mobilenet_v3_small model and import the torchvision.models.MobileNet_V3_Small_Weights.DEFAULT default weights to build a basic model.
[0062] 5.2. Modify the basic model features[0][0] = nn.Conv2d(1,16,kernel_size = (3,3),stride = (2,2),padding = (1,1),bias = False). The modification point is to change the default input channel 3 to input channel 1.
[0063] 5.3. Modify the base model’s classifier[3]=nn.Linear(self.base_model.classifier[3].in_features,100). The modification point is to change the number of output features to 100.
[0064] 5.4. Use 3 fully connected layers to build a perceptron classifier; the first layer inputs the features output by the basic model + omics features, and the output is 1000; the second layer inputs 1000 and the output is 100; the third layer inputs 100 and the output is 1.
[0065] 5.5. Use the weighted Binary Cross-Entropy with Logits Loss loss function to calculate the loss and set the pos_weight value to 3.
[0066] 5.6. Use intratumoral sections + intratumoral omics features, peritumoral sections + peritumoral omics features, and intratumoral + peritumoral sections + intratumoral + peritumoral omics features to train three fusion models, and use AUC as the evaluation index to select the best model;
[0067] 5.7. Use 3 fully connected layers to build a perceptron classifier; the first layer inputs the features output by the basic model + omics features, and the output is 1000; the second layer inputs 1000 and the output is 100; the third layer inputs 100 and the output is 1.
[0068] 5.8. Based on the best model in 5.6, use the corresponding CT image slices + omics features + clinical features to train the final model
[0069] Step 6: Input the parotid CT images to be classified, omics features, and clinical features into the trained classification model to obtain the benign and malignant prediction classification results. The performance evaluation parameters include accuracy, recall, and precision.
[0070] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A parotid gland classification method based on CT image omics and clinical features, characterized in that: The following steps are involved: Step 1: Based on the CT image of the parotid tumor, mark the tumor area and expand the surrounding area to mark the tumor area; Step 2: Extract omics features from the tumor area and peritumoral area; Step 3: Feature screening of omics features and clinical features; Step 4: Perform image processing and data augmentation on the CT image; Step 5: Construct a classification model for benign and malignant parotid tumors and train it using the dataset; Step 6: Input the parotid CT image to be classified, omics features, and clinical features into the trained classification model to obtain the benign and malignant prediction classification results.
2. The parotid gland classification method based on CT image omics and clinical practice as claimed in claim 1, characterized in that: In the step 1, the method for labeling one of the parotid gland tumor CT images is: 1.
1. Use ITK-SNAP software to mark the tumor area and obtain the marked area within the tumor; 1.
2. Use Python software to expand the marked area within the tumor by 2 mm to obtain the marked area around the tumor.
3. The parotid gland classification method based on CT image omics and clinical practice as claimed in claim 1, characterized in that: In the step 2, the omics feature extraction method is: 2.
1. The PyRadiomics library was used to extract radiomic features from the intratumoral area, including first-order statistical features, texture features (such as GLCM, GLRLM, etc.) and other high-order texture features. A total of 1,874 intratumoral radiomic features were extracted. 2.
2. The PyRadiomics library was used to extract radiomic features from the peritumoral area, including first-order statistical features, texture features (such as GLCM, GLRLM, etc.) and other high-order texture features. A total of 1,874 peritumoral radiomic features were extracted.
4. The parotid gland classification method based on CT image omics and clinical practice as claimed in claim 1, characterized in that: In step 3, the method of feature screening is: 3.
1. Use SPSS software to perform univariate and multivariate logistic regression on clinical characteristics to screen features with p < 0.05, and screen the remaining 3 features; 3.
2. Pearson correlation analysis was used to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features, respectively, to eliminate redundant features and retain features without high correlation; 3.
3. Use T test to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.2, and screen out the features with significant differences; 3.
4. Use LASSO regression to screen the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.3, and screen out the features that are closely related to benign and malignant classification; 3.
5. Use RFE to calculate the feature importance of the intratumoral omics features, peritumoral omics features, and intratumoral + peritumoral omics features after screening in 3.4, and screen out the top 20 features with the highest importance.
5. The parotid gland classification method based on CT image omics and clinical practice as claimed in claim 1, characterized in that: In step 4, the specific steps of performing image processing and data augmentation on the CT image are as follows: 4.
1. Use Python software to read CT image files and perform 40-400 window pixel capture, as well as read annotation files, and extract intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices in depth direction; 4.
2. Use Python software to normalize the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices, specifically by subtracting the mean and then dividing by the difference between the maximum value and the minimum value, and normalizing to the range of -1 to 1; 4.
3. Use Python software to rotate the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices of the training set within the range of -10° to 10°, flip them horizontally and vertically, and add Gaussian noise with a mean of 0 and a maximum of 0.01; 4.
4. Use Python software to regularize the intratumoral slices, peritumoral slices, and intratumoral + peritumoral slices to a size of 256*256, and fill the insufficient parts with 0.
6. The parotid gland classification method based on CT image omics and clinical practice as claimed in claim 1, characterized in that: In the step 5, the method for constructing a benign and malignant classification model for parotid tumors is specifically as follows: 5.
1. Use the torchvision.models.mobilenet_v3_small model and import the torchvision.models.MobileNet_V3_Small_Weights.DEFAULT default weights to build a basic model. 5.
2. Modify the basic model features[0][0] = nn.Conv2d(1,16,kernel_size = (3,3),stride = (2,2),padding = (1,1),bias = False). The modification point is to change the default input channel 3 to input channel 1. 5.
3. Modify the base model’s classifier[3]=nn.Linear(self.base_model.classifier[3].in_features,100). The modification point is to change the number of output features to 100. 5.
4. Use 3 layers of fully connected layers to build a perceptron classifier; the first layer inputs the features output by the basic model + omics features + clinical features, and the output is 1000; the second layer inputs 1000 and the output is 100; the third layer inputs 100 and the output is 1. 5.
5. Use the weighted Binary Cross-Entropywith Logits Loss loss function to calculate the loss and set the pos_weight value to 3. 5.
6. Use intratumoral sections + intratumoral omics features, peritumoral sections + peritumoral omics features, and intratumoral + peritumoral sections + intratumoral + peritumoral omics features to train three fusion models, and use AUC as the evaluation index to select the best model; 5.
7. Use 3 fully connected layers to build a perceptron classifier; the first layer inputs the features output by the basic model + omics features, and the output is 1000; the second layer inputs 1000 and the output is 100; the third layer inputs 100 and the output is 1. 5.
8. Based on the best model in 5.6, the final model is trained using the corresponding CT image slices + omics features + clinical features.
7. The parotid gland classification method based on CT image omics and clinical practice as claimed in claim 1, characterized in that: In the step 6, the performance evaluation parameters for obtaining the benign or malignant prediction classification result include: accuracy, recall and precision.