Device and method for predicting EGFR gene mutation of lung cancer CT thin and thick layer image
By using Copy-Paste data augmentation and DenseNet model feature extraction, combined with orthogonal decomposition and dual loss optimization, the problem of adapting thin-slice and thick-slice CT images was solved, enabling accurate prediction of EGFR gene mutations and improving the model's accuracy and robustness.
Patent Information
- Application Number
- CN202511470975.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-15
- Publication Date
- 2026-01-20
AI Technical Summary
Existing technologies struggle to adapt to both thin-slice and thick-slice CT images simultaneously, resulting in insufficient accuracy and robustness in EGFR gene mutation prediction. Furthermore, there is a lack of effective data augmentation methods to alleviate model overfitting issues.
The data sample was expanded using Copy-Paste data augmentation technology, and features were extracted using the DenseNet model. An EGFR gene mutation prediction model was constructed using orthogonal decomposition and dual loss optimization strategies.
It significantly improves the prediction accuracy and generalization ability in CT images with different slice thicknesses, ensuring accurate prediction of EGFR gene mutations.
Smart Images
Figure CN121366136A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of medical image processing, and particularly relates to a lung cancer CT thin-thick layer image EGFR gene mutation prediction device and method. BACKGROUND
[0002] Lung cancer is one of the malignant tumors with the highest incidence and mortality worldwide, and its accurate diagnosis and treatment are of great significance to the treatment effect and prognosis of patients. EGFR gene mutation is an important marker of lung cancer molecular typing and is closely related to the efficacy of targeted drug treatment. Therefore, how to efficiently predict the EGFR gene mutation state of patients through non-invasive means has become a key problem in clinical attention.
[0003] CT images are an important tool for lung cancer diagnosis and have a wide range of applications. Thin layer CT images can show more detailed information due to their high resolution, while thick layer CT images are suitable for screening due to their lower radiation dose. However, there are significant differences in image resolution, noise level, and texture features between thin layer and thick layer CT data, making it difficult to model and analyze uniformly based on traditional methods. In addition, due to the uneven quality of lung cancer CT image acquisition and inconsistent sample data distribution, the generalization ability of the model is often limited.
[0004] The existing technology has the following difficulties when analyzing lung cancer CT images: 1) the heterogeneity of thin layer and thick layer CT image data distribution, making it difficult for traditional methods to adapt to different layer thickness data simultaneously; 2) existing image analysis techniques cannot fully exploit the potential gene mutations in CT images, resulting in insufficient accuracy and robustness of EGFR gene mutation prediction; 3) in the case of limited data volume, there is a lack of effective data augmentation methods to alleviate the model overfitting problem and improve prediction performance. SUMMARY
[0005] The technical problem to be solved by the present application is to provide a lung cancer CT thin-thick layer image EGFR gene mutation prediction device and method, which accurately predicts the EGFR gene mutation state and provides an important decision basis for targeted treatment and personalized diagnosis and treatment of lung cancer.
[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows: A lung cancer CT thin-thick layer image EGFR gene mutation prediction device, characterized in that it comprises: A first processing module for acquiring lung cancer CT thin-thick layer images of a patient; A second processing module for performing data enhancement processing on the lung cancer CT thin-thick layer images; A third processing module for performing feature extraction on the lung cancer CT thin-thick layer images after data enhancement processing; The fourth processing module is configured to perform orthogonal decomposition feature refinement and double loss optimization on the extracted features to obtain an EGFR gene mutation prediction model. The fifth processing module is configured to input the tested lung cancer CT thin-thick layer image into the EGFR gene mutation prediction model to perform EGFR gene mutation detection.
[0007] Preferably, the lung cancer CT thin-thick layer image comprises EGFR gene mutation information. Preferably, the second processing module performs Copy-Paste data enhancement processing on the lung cancer CT thin-thick layer image.
[0008] Preferably, the third processing module is configured to perform feature extraction on the lung cancer CT thin-thick layer image after data enhancement processing by using a DenseNet model.
[0009] Preferably, the loss function of the EGFR gene mutation prediction model comprises a classification loss and a feature consistency loss.
[0010] The application also provides a lung cancer CT thin-thick layer image EGFR gene mutation prediction method, comprising the following steps: Step S1, obtaining a lung cancer CT thin-thick layer image of a patient; Step S2, performing data enhancement processing on the lung cancer CT thin-thick layer image; Step S3, performing feature extraction on the lung cancer CT thin-thick layer image after data enhancement processing; Step S4, performing orthogonal decomposition feature refinement and double loss optimization on the extracted features to obtain an EGFR gene mutation prediction model. Step S5, inputting the tested lung cancer CT thin-thick layer image into the EGFR gene mutation prediction model to perform EGFR gene mutation detection.
[0011] Preferably, the lung cancer CT thin-thick layer image comprises EGFR gene mutation information.
[0012] Preferably, in step S2, Copy-Paste data enhancement processing is performed on the lung cancer CT thin-thick layer image.
[0013] Preferably, in step S3, feature extraction is performed on the lung cancer CT thin-thick layer image after data enhancement processing by using a DenseNet model.
[0014] Preferably, the loss function of the EGFR gene mutation prediction model comprises a classification loss and a feature consistency loss.
[0015] The application is based on deep learning for lung cancer CT thin and thick layer image EGFR gene mutation prediction, aims to solve the model adaptation problem caused by the heterogeneity of thin and thick layer data in the existing lung cancer CT image analysis, and realizes accurate prediction of EGFR gene mutation through efficient feature extraction and optimization method. Through data enhancement, feature refinement and double loss optimization strategy, the prediction accuracy and generalization ability in different layer thickness CT images are significantly improved. BRIEF DESCRIPTION OF DRAWINGS
[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.
[0017] Figure 1 The principle diagram of the lung cancer CT thin and thick layer image EGFR gene mutation prediction device of the embodiment of the present application is shown. Figure 2 The data processing diagram of the lung cancer CT thin and thick layer image EGFR gene mutation prediction of the present application is shown. Figure 3 The feature diagram of the DensNet feature extraction used in the present application is shown. DETAILED DESCRIPTION
[0018] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0019] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail in combination with the drawings and specific embodiments.
[0020] Embodiment 1: As shown in Figure 1 , 2 , 3, the present application provides a lung cancer CT thin and thick layer image EGFR gene mutation prediction device, comprising: A first processing module is used for acquiring lung cancer CT thin and thick layer images of a patient; wherein the lung cancer CT thin and thick layer images contain EGFR gene mutation information. A second processing module is used for Copy-Paste data enhancement processing of the lung cancer CT thin and thick layer images. The third processing module is configured to perform feature extraction on the lung cancer CT thin-thick layer image processed by the data enhancement, by using a DenseNet model. The fourth processing module is configured to perform orthogonal decomposition feature refinement and double loss optimization on the extracted features, to obtain an EGFR gene mutation prediction model. The fifth processing module is configured to input the tested lung cancer CT thin-thick layer image into the EGFR gene mutation prediction model, to perform EGFR gene mutation detection.
[0021] As an embodiment of the present application, the second processing module expands the data samples of the lung cancer CT thin-thick layer image by using a Copy-Paste data enhancement method, increases the diversity of the data, and reduces the heterogeneity between the thin layer and thick layer CT images. Assuming that the data samples are wherein n represents the thin layer image, k represents the thick layer image, N represents the total amount of data, there are N data in total, i represents the current data, and {i=1, N} represents the total amount of data, which can be i, and i does not exceed N; is the CT thin layer scanning image, is the CT thick layer scanning image, is the EGFR mutation label corresponding to the scanning image of the patient. If the Copy-Paste operation of is applied to the corresponding , by randomly cutting an image block and pasting it to the same position of , an enhanced image is obtained. Similarly, if the Copy-Paste operation of is applied to the corresponding , an enhanced image is obtained. The above can be represented as: wherein and are the original images, and the corresponding enhanced images obtained by the copy-paste operation are respectively; is the image block exchanged by the copy-paste operation; is the mask image of the image block, which represents the specific position and size information of the image block.
[0022] As an embodiment of the present application, the third processing module takes the original thin-thick layer image pair and the data enhanced thin-thick layer image pair as inputs, and extracts the features thereof by using a DenseNet model. Given the input thin-thick layer image pair, the feature expression output by the DenseNet model is: wherein, is a DenseNet feature extraction model, is a model parameter.
[0023] Further, the DenseNet model is composed of a 3D convolution layer, a plurality of Dense blocks, a maximum pooling layer, a global average pooling layer, and a full connection layer. Each Dense block includes a 3D batch normalization, a ReLU activation function, and a 3D convolution operation connected in sequence, and realizes cross-layer feature reuse and efficient gradient transmission through dense connection. In the process of extracting features, the DenseNet model realizes excellent feature expression ability through multi-level 3D convolution operation and dense connection mechanism. In each Dense block (DenseBlock), the input feature map is processed through the 3D batch normalization (BN) layer, the ReLU activation function, and the 3D convolution layer connected layer by layer, thereby constructing a nonlinear feature representation space. Each layer of convolution operation not only extracts the spatial features of the local region, but also inputs the feature maps of all previous layers as input through dense connection. This inter-layer feature reuse significantly improves the efficiency of information flow, reduces the gradient vanishing problem, and ensures the trainability of the network depth.
[0024] As an embodiment of the embodiment of the present application, the orthogonal decomposition feature refinement process of the fourth processing module is: In order to effectively avoid the information attenuation caused by deepening the network and the original information loss caused by the copy-paste operation, in the fourth processing module, the extracted original image features are finely integrated through feature fusion to generate a template feature. The template feature not only retains the key information in the network, but also serves as a representative of the original feature, ensuring the integrity and efficient transmission of information.
[0025] wherein, represents the finally obtained template feature, and respectively represent 3D convolution and image channel dimension splicing operation; the template feature and other features are flattened into one-dimensional vectors in the network, and this process is realized by "flatten" operation on the features: wherein, indicates that the feature is stretched into a one-dimensional vector, and in the process of orthogonal decomposition, the orthogonal projection and fine operation of the feature are combined to ensure accurate transmission of feature information, and and the feature projection of , that is, The projection feature Then the view With In the EGFR gene mutation prediction model, in order to generate more stable and refined features, other features will be reduced by this part of the redundant features to obtain With To achieve feature refinement: (7) Finally, the fusion of its , The feature is the output feature of the EGFR gene mutation prediction model, and the feature is used to predict the EGFR gene mutation.
[0026] Further, the double loss optimization process of the fourth processing module is: When constructing the loss function of the EGFR gene mutation prediction model, the classification loss and the feature consistency loss are introduced to constrain the consistency between the prediction accuracy of the EGFR gene mutation prediction model and the feature. The classification loss uses cross-entropy loss, that is: (8) Wherein, is the true label of the data sample EGFR gene mutation, is the probability value of the EGFR gene mutation prediction model predicted as a positive sample.
[0027] The feature extracted from the enhanced image through Copy-Paste is again subjected to the enhanced feature extracted from the original image through Copy-Paste operation in the same position, so the feature consistency loss is introduced to control the feature extracted by the DenseNet model always corresponds to the original image. The mutual information constraint is used in the loss, which can be expressed as: Wherein, is the feature map extracted from the original image, is the feature of the CP enhanced image upsampled to Size CP operation on the feature map in the same position, is the joint probability distribution between the feature maps, is the edge probability distribution of the feature map itself, if q=n then m is k, if q=k then m is n, that is, q and m represent a pair of thin and thick layer images.
[0028] In summary, the double loss optimization of the EGFR gene mutation prediction model is: Wherein, A hyperparameter that weighs the contribution of each loss.
[0029] The application adopts the Copy-Paste data enhancement technology to expand the lung cancer CT thin and thick layer images, increases the diversity of samples, and reduces the heterogeneity influence between different layer thickness data. The DenseNet model is used to extract multi-level features from multiple lung cancer CT thin and thick layer images at the same time, so as to ensure the comprehensiveness, accuracy and efficiency of the feature extraction process. The features extracted by the DenseNet are subjected to orthogonal decomposition processing, and the feature representation is refined, so that the features can more accurately reflect the potential information of the image. Then, a dual loss optimization process based on image and feature information is constructed, the prediction accuracy and robustness are considered in the optimization process, and the performance of the model on different data sets is improved.
[0030] Embodiment 2: The application also provides a lung cancer CT thin and thick layer image EGFR gene mutation prediction method, comprising: Step S1, acquiring the lung cancer CT thin and thick layer image of the patient; Step S2, performing data enhancement processing on the lung cancer CT thin and thick layer image; Step S3, performing feature extraction on the lung cancer CT thin and thick layer image after data enhancement processing; Step S4, orthogonal decomposition feature refinement and dual loss optimization are performed on the extracted features to obtain an EGFR gene mutation prediction model: Step S5, inputting the tested lung cancer CT thin and thick layer image into the EGFR gene mutation prediction model for EGFR gene mutation detection.
[0031] As an embodiment of the application, the lung cancer CT thin and thick layer image comprises: EGFR gene mutation information.
[0032] As an embodiment of the application, in step S2, the lung cancer CT thin and thick layer image is subjected to Copy-Paste data enhancement processing.
[0033] As an embodiment of the application, in step S3, the lung cancer CT thin and thick layer image after data enhancement processing is subjected to feature extraction by the DenseNet model.
[0034] As an embodiment of the application, the loss function of the EGFR gene mutation prediction model comprises: a classification loss and a feature consistency loss.
[0035] The above described embodiments are only to illustrate the preferred modes of the present application, and are not intended to limit the scope of the present application. Any modification and improvement made by those skilled in the art to the technical solutions of the present application without departing from the design spirit of the present application shall fall within the protection scope of the present application as defined by the claims.
Claims
1. A device for predicting EGFR gene mutations in thin-slice and thick-slice images of lung cancer CT scans, characterized in that, include: The first processing module is used to acquire thin and thick slice images of the patient's lung cancer CT scan. The second processing module is used to perform data enhancement processing on thin and thick slice images of lung cancer CT. The third processing module is used to extract features from the thin and thick slice images of lung cancer CT after data augmentation. The fourth processing module is used to refine the extracted features through orthogonal decomposition and dual loss optimization to obtain the EGFR gene mutation prediction model: The fifth processing module is used to input the tested lung cancer CT thin and thick slice images into the EGFR gene mutation prediction model for EGFR gene mutation detection.
2. The lung cancer CT thin-slice and thick-slice image EGFR gene mutation prediction device as described in claim 1, characterized in that, The lung cancer CT slice images contain EGFR gene mutation information.
3. The lung cancer CT thin-slice and thick-slice image EGFR gene mutation prediction device as described in claim 2, characterized in that, The second processing module performs copy-paste data enhancement processing on thin and thick slice images of lung cancer CT.
4. The lung cancer CT thin-slice and thick-slice image EGFR gene mutation prediction device as described in claim 3, characterized in that, The third processing module is used to extract features from the lung cancer CT thin-slice images after data augmentation using the DenseNet model.
5. The lung cancer CT thin-slice and thick-slice image EGFR gene mutation prediction device as described in claim 4, characterized in that, The loss function of the EGFR gene mutation prediction model includes classification loss and feature consistency loss.
6. A method for predicting EGFR gene mutations in thin-slice and thick-slice CT images of lung cancer, characterized in that, include: Step S1: Obtain thin and thick slice images of the patient's lung cancer on CT. Step S2: Perform data enhancement processing on thin and thick slice images of lung cancer CT. Step S3: Extract features from the thin and thick slice images of lung cancer CT after data enhancement processing; Step S4: Perform orthogonal decomposition feature refinement and dual loss optimization on the extracted features to obtain the EGFR gene mutation prediction model: Step S5: Input the tested lung cancer CT thin and thick slice images into the EGFR gene mutation prediction model to detect EGFR gene mutations.
7. The method for predicting EGFR gene mutations in thin-slice and thick-slice images of lung cancer as described in claim 6, characterized in that, The lung cancer CT slice images contain EGFR gene mutation information.
8. The method for predicting EGFR gene mutations in thin-slice and thick-slice images of lung cancer as described in claim 7, characterized in that, In step S2, copy-paste data enhancement processing is performed on the thin and thick slice images of lung cancer CT.
9. The method for predicting EGFR gene mutations in thin-slice CT images of lung cancer as described in claim 8, characterized in that, In step S3, the DenseNet model is used to extract features from the lung cancer CT thin and thick slice images after data augmentation.
10. The method for predicting EGFR gene mutations in thin-slice and thick-slice images of lung cancer as described in claim 9, characterized in that, The loss function of the EGFR gene mutation prediction model includes classification loss and feature consistency loss.