Non-small cell lung cancer EGFR mutation state prediction method

Abstract

The invention belongs to the field of intelligent medical treatment, and provides a non-small cell lung cancer EGFR mutation state prediction method which comprises the steps of patient information collection, 3D deep learning model prediction, multi-instance learning model prediction and joint model prediction. According to the invention, through a 3D deep learning model, radiology features are captured, and the spatial position of a lung focus and the integrity of extracted information are reserved; through a multi-instance learning model, the process of manually sketching a region of interest is omitted by processing a tissue slice WSI image, and a large amount of time and manpower consumption are saved; through the joint model, the prediction efficiency of the model is improved, the macroscopic and microscopic characteristics of the tumor are comprehensively captured, the heterogeneity of the tumor tissue is supplemented, the limitations of invasive biopsy and biopsy tissue taking are avoided, and the expensive gene detection cost is saved.

Description

Technical Field

[0001] This invention relates to the field of intelligent healthcare, and in particular to a method for predicting EGFR mutation status in non-small cell lung cancer. Background Technology

[0002] Non-small cell lung cancer is abbreviated as NSCLC, epidermal growth factor receptor is abbreviated as EGFR, computed tomography (CT) is abbreviated as CT, convolutional neural network is abbreviated as CNN, and full-view digital imaging (WSI) is abbreviated as WSI.

[0003] Lung cancer is a leading cause of cancer-related deaths worldwide, and the incidence of advanced lung cancer is declining, closely related to early detection and more targeted, individualized treatment. Non-small cell lung cancer (NSCLC) accounts for approximately 80% of all lung cancers. In recent years, with the rapid development of molecular medicine and personalized precision medicine, targeted therapy has provided a viable treatment option for NSCLC patients. For example, EGFR tyrosine kinase inhibitors have significantly improved the survival and prognosis of patients with EGFR mutations. Identification of EGFR mutation status mainly relies on gene testing of tissue biopsies or surgical specimens. However, tumor tissue exhibits heterogeneity, so biopsy tissue cannot fully represent the entire tumor. Biopsies have inherent limitations, including insufficient quantity and quality of tissue samples, improper technique, and increased potential risk of tumor metastasis. Furthermore, EGFR status may change throughout treatment, making repeated biopsies impossible, especially for advanced-stage patients. In addition, the substantial costs associated with gene testing increase the financial burden on patients. Therefore, there is an urgent need for an accurate, effective, inexpensive, and reusable predictive tool to predict the EGFR mutation status of NSCLC patients, guiding personalized clinical treatment decisions. Computed tomography (CT) images contain complete information about the tumor region and its surrounding microenvironment, making them an essential examination in routine lung diagnosis and treatment. Radiomics, by extracting and screening medical image features at high throughput, uses machine learning algorithms to build models to predict a range of clinical events, including disease classification and staging, recurrence and metastasis, progression, and survival, and is increasingly widely used in clinical practice. Since the inception of radiomics research, scientists have achieved some success in predicting EGFR mutation status using different algorithmic models. With the rapid development of artificial intelligence, deep learning has demonstrated more effective predictive capabilities. Many researchers focus on 2D deep learning methods based on the maximum cross-sectional level of the lesion, while rarely using 3D convolutional neural networks, which may affect the capture of comprehensive and three-dimensional spatial information of the lesion. In recent years, the rapid development of the integration of artificial intelligence and digital pathology has yielded encouraging results and shows great potential for future development. The diagnosis of almost all cancer patients requires histopathological slides, which can be digitized by scanners into high-resolution, full-view digital images for permanent storage. Pathomics can extract tumor region features from pathological tissue slide images at the microscopic level, capturing information on various tissue phenotypes invisible to the naked eye. These features can be used to predict a range of clinical outcomes. Deep learning methods for identifying complex visual features in high-resolution digitized histopathological images have shown great promise in predicting the staging, pathological subtypes, lymph node metastasis, and treatment response of various malignant lesions. However, previous studies have largely relied on supervised classification learning methods that require manual annotation, resulting in significant time and labor costs.In recent years, studies have confirmed that multi-omics prediction models have achieved satisfactory results in risk stratification and treatment response in ovarian cancer, breast cancer, and rectal cancer.

[0004] Previous studies on predicting EGFR mutations in NSCLC patients have mainly focused on extracting features from single types of medical images, which is not comprehensive enough in capturing tumor biological information. In addition, there are currently no studies on multimodal combined models that integrate radiomics, pathomics and clinical features to predict the EGFR mutation status of NSCLC. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method for predicting EGFR mutation status in non-small cell lung cancer, which solves the problems of insufficient consideration of radioactive characteristics, the need for extensive manual annotation, and insufficient consideration of macroscopic and microscopic characteristics in existing methods.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for predicting EGFR mutation status in non-small cell lung cancer includes:

[0008] Acquire CT images, WSI images, and clinical characteristics of the target patient;

[0009] The CT images are input into a pre-trained 3D deep learning model for prediction to obtain radiomics features;

[0010] The WSI image is input into a pre-trained multi-instance learning model for prediction to obtain pathological features;

[0011] The clinical features, radiomics features, and pathological features were predicted using a joint model based on nomograms to obtain the EGFR mutation status prediction results.

[0012] Preferably, the training process of the 3D deep learning model includes:

[0013] The pre-collected CT images were imported into ITK-SNAP software in DICOM format for manual annotation to obtain raw CT data;

[0014] The raw CT data is resampled and spatially normalized using a fixed resolution to obtain CT preprocessed data.

[0015] The CT preprocessed data is processed using data augmentation technology to obtain CT enhanced data;

[0016] Based on the cosine annealing learning rate scheduling mechanism and stochastic gradient descent, the ResNet50-3D model, ShuffleNet-3D model, and Densenet121-3D model are trained using CT-enhanced data with softmax cross-entropy as the loss function.

[0017] The 3D deep learning model is obtained by fusing the trained ResNet50-3D model, ShuffleNet-3D model, and Densenet121-3D model.

[0018] Preferably, the training process of the multi-instance learning model includes:

[0019] The pre-collected WSI images are divided into 512×512 pixel tiles. The same EGFR tag is added to the tiles belonging to the same WSI image. The white background of the tiles is removed using the OnekeyAI OKT-patch2predict tool to obtain the original tile set.

[0020] The ResNet50 model is trained using the original tile set to obtain the multi-instance learning model.

[0021] Preferably, the construction process of the joint model includes:

[0022] Univariate and multivariate regression analyses were performed on the pre-collected clinical characteristics;

[0023] The clinical features, radiomics features, and pathological features with p-values ​​less than 0.05 in the regression analysis results are integrated into a nomogram to obtain the combined model.

[0024] Preferably, the expression for the cosine annealing learning rate scheduling mechanism is:

[0025]

[0026] Where, η t Let be the learning rate at step t; Minimum learning rate; T represents the maximum learning rate; i T is the number of iterations. cur This represents the current iteration number.

[0027] Preferably, the voxel spacing of the fixed resolution is standardized to 1mm×1mm×1mm.

[0028] Preferably, the fusion methods of the ResNet50-3D model, the ShuffleNet-3D model, and the Densenet121-3D model include: maximum value fusion, minimum value fusion, and mean value fusion.

[0029] Preferably, the ResNet50 model is trained using the original tile set to obtain the multi-instance learning model, including:

[0030] The ResNet50 model is used to predict the small magnetic piece, and the predicted probability and predicted label are obtained.

[0031] The predicted probability and the frequency with which the predicted label falls into each preset container are statistically analyzed to obtain multi-instance learning features; the multi-instance learning features include: multi-instance learning probability and multi-instance learning label;

[0032] A dictionary is obtained by identifying the unique element in the predicted probability and the predicted label. Each small magnetic piece is represented as a vector recording the frequency of each element in the small magnetic piece that belongs to the dictionary. The vector is then subjected to word frequency inverse document frequency transformation to obtain bag-of-words model features. The bag-of-words model features include: bag-of-words probability and bag-of-words label.

[0033] The multi-instance learned features and the bag-of-words model features are concatenated to obtain a comprehensive feature vector; the expression of the comprehensive feature vector is:

[0034]

[0035] Among them, feature fusion The comprehensive feature vector; Histo prob Histo pred They are respectively; Bow prob Bow pred They are respectively; This indicates element-wise addition.

[0036] The present invention discloses the following technical effects:

[0037] This invention provides a method for predicting EGFR mutation status in non-small cell lung cancer. By using a 3D deep learning model, it solves the problem of insufficient consideration of radioactive features in existing methods and realizes the extraction of spatial features. By using a multi-instance learning model, it solves the problem of requiring a lot of manual annotation in existing methods and realizes weakly supervised learning of the model. By using a joint model, it solves the problem of insufficient consideration of macroscopic and microscopic features in existing methods and realizes the sum analysis of clinical features, radiomics features, and pathological features. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the EGFR mutation status prediction process for non-small cell lung cancer provided in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of a multi-mode joint model provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] The purpose of this invention is to provide a method for predicting EGFR mutation status in non-small cell lung cancer, which solves the problems of insufficient consideration of radioactive characteristics, the need for a large amount of manual annotation, and insufficient consideration of macroscopic and microscopic characteristics in existing methods.

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Figure 1 This is a schematic diagram of the EGFR mutation status prediction process for non-small cell lung cancer provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a method for predicting EGFR mutation status in non-small cell lung cancer, comprising:

[0045] Step 100: Acquire CT images, WSI images, and clinical features of the target patient;

[0046] Step 200: Input the CT image into a pre-trained 3D deep learning model for prediction to obtain radiomics features;

[0047] Step 300: Input the WSI image into a pre-trained multi-instance learning model for prediction to obtain pathological features;

[0048] Step 400: Use a joint model based on nomograms to predict the clinical features, radiomics features, and pathological features to obtain EGFR mutation status prediction results.

[0049] Furthermore, the training process of the 3D deep learning model includes:

[0050] The pre-collected CT images were imported into ITK-SNAP software in DICOM format for manual annotation to obtain raw CT data;

[0051] The raw CT data is resampled and spatially normalized using a fixed resolution to obtain CT preprocessed data.

[0052] The CT preprocessed data is processed using data augmentation technology to obtain CT enhanced data;

[0053] Based on the cosine annealing learning rate scheduling mechanism and stochastic gradient descent, the ResNet50-3D model, ShuffleNet-3D model, and Densenet121-3D model are trained using CT-enhanced data with softmax cross-entropy as the loss function.

[0054] The 3D deep learning model is obtained by fusing the trained ResNet50-3D model, ShuffleNet-3D model, and Densenet121-3D model.

[0055] Specifically, the training process of the multi-instance learning model includes:

[0056] The pre-collected WSI images are divided into 512×512 pixel tiles. The same EGFR tag is added to the tiles belonging to the same WSI image. The white background of the tiles is removed using the OnekeyAI OKT-patch2predict tool to obtain the original tile set.

[0057] The ResNet50 model is trained using the original tile set to obtain the multi-instance learning model.

[0058] Furthermore, the construction process of the joint model includes:

[0059] Univariate and multivariate regression analyses were performed on the pre-collected clinical characteristics;

[0060] The clinical features, radiomics features, and pathological features with p-values ​​less than 0.05 in the regression analysis results are integrated into a nomogram to obtain the combined model.

[0061] Specifically, the expression for the cosine annealing learning rate scheduling mechanism is:

[0062]

[0063] Where, η t Let be the learning rate at step t; Minimum learning rate; T represents the maximum learning rate; i T is the number of iterations. cur This represents the current iteration number.

[0064] Optionally, the voxel spacing of the fixed resolution is standardized to 1mm×1mm×1mm.

[0065] Preferably, the fusion methods of the ResNet50-3D model, the ShuffleNet-3D model, and the Densenet121-3D model include: maximum value fusion, minimum value fusion, and mean value fusion.

[0066] Furthermore, the ResNet50 model is trained using the original tile set to obtain the multi-instance learning model, including:

[0067] The ResNet50 model is used to predict the small magnetic piece, and the predicted probability and predicted label are obtained.

[0068] The predicted probability and the frequency with which the predicted label falls into each preset container are statistically analyzed to obtain multi-instance learning features; the multi-instance learning features include: multi-instance learning probability and multi-instance learning label;

[0069] A dictionary is obtained by identifying the unique element in the predicted probability and the predicted label. Each small magnetic piece is represented as a vector recording the frequency of each element in the small magnetic piece that belongs to the dictionary. The vector is then subjected to word frequency inverse document frequency transformation to obtain bag-of-words model features. The bag-of-words model features include: bag-of-words probability and bag-of-words label.

[0070] The multi-instance learned features and the bag-of-words model features are concatenated to obtain a comprehensive feature vector; the expression of the comprehensive feature vector is:

[0071]

[0072] Among them, feature fusion The comprehensive feature vector; Histo prob Histo pred They are respectively; Bow prob Bow pred They are respectively; This indicates element-wise addition.

[0073] Specifically, radiomics based on 3D deep learning predicts EGFR mutations in non-small cell lung cancer:

[0074] Research cohorts and data collection;

[0075] Image acquisition and segmentation: All thin-section CT images were imported into ITK-SNAP software (version 3.8.0, www.itksnap.org) in DICOM format for manual annotation;

[0076] Data Processing: First, to ensure consistency in voxel spacing across different ROIs, a fixed-resolution resampling method was used for spatial normalization. This process normalized the voxel spacing to 1mm × 1mm × 1mm, allowing for precise comparison and evaluation by aligning spatial dimensions across images. Then, 3D convolutions were used to train the ROI regions. Unlike 2D CNNs, 3D convolutions cropped the entire ROI region to obtain the smallest bounding box, which was then used for training.

[0077] Data augmentation: To enhance the model's robustness, Z-score normalization was applied to the images before they were used as input. During training, real-time data augmentation techniques were employed, including random cropping, random 90° rotation, and random flipping along the x, y, and z axes, to increase the diversity of the training data. However, for test images, only normalization was performed to maintain consistency in data preprocessing.

[0078] Model Training: 3D deep learning was employed to identify EGFR mutation states in regions of interest (ROIs) of NSCLC. For the input data of these models, a min-max transformation was performed to normalize grayscale values, scaling them to a range between -1 and 1. Furthermore, each cropped sub-region image was resized to 96×96×96 using nearest-nearest interpolation and used as input to the 3DCNN model. Due to the lack of publicly available 3D model pre-training parameters, the models were randomly initialized. Considering the limited availability of image data, a cosine annealing learning rate schedule was employed, defined as:

[0079]

[0080] The parameter settings are as follows: T i =48, representing the minimum learning rate, maximum learning rate, and number of iteration cycles, respectively. Other key hyperparameters include using stochastic gradient descent as the optimizer and softmax cross-entropy as the loss function to optimize the training process and ensure effective learning;

[0081] Construction and Validation of Deep Learning Radiomics Models: For the cropped region, ResNet50-3D, ShuffleNet-3D, and DenseNet121-3D models were used to predict its EGFR status. The predicted probabilities of the CNN model were designated as deep learning radiomics models. This embodiment includes three different models, with a separate model trained for each. To understand the different characteristics of each modality, a detailed comparative analysis of the results for each modality was performed. Furthermore, to explore the integration of multimodal methods, three different methods were used to combine the model results: maximum, minimum, and mean fusion. The predictive performance of the model was evaluated using ROC and AUC, and the consistency between the predicted probabilities and actual results was assessed using the Hosmer-Lemeshow test (HL test) calibration curve. The Delongtest and DCA were used to evaluate the model's performance and clinical benefits.

[0082] Furthermore, pathomic prediction of EGFR mutations in non-small cell lung cancer based on weakly supervised deep learning and multi-instance learning:

[0083] Research cohorts and data collection;

[0084] WSI image acquisition;

[0085] Data processing: Considering the large size of WSI, typically around 100,000 × 50,000 pixels, with a pixel resolution of approximately 0.5 μm / pixel, WSI was divided into small 512 × 512 pixel patches. The OnekeyAIOKT-patch2predict tool was used to remove all white backgrounds, resulting in more than 40,000 different, non-overlapping patches.

[0086] Weakly supervised learning: All patches from a single sample share the same EGFR label. For patch prediction, three top-performing networks were evaluated: DenseNet121, ResNet50, and Inception_v3 (only one was retained after evaluation). The goal was to evaluate the probability that each patch would be classified into the category corresponding to its WSI. During model training, the data augmentation and normalization strategies described above were implemented. These strategies included random horizontal and vertical patch flipping to increase the diversity of the training data, centering the patch to 224×224 pixels (299×299 pixels for InceptionV3) to ensure consistent input size, and Z-score normalization across RGB channels to standardize the distribution of pixel values ​​and ensure consistent and robust model performance. To enhance the generalization ability of the pathology model, the learning rate was carefully tuned using a cosine decay algorithm. The characteristics of this approach are as follows:

[0087]

[0088] In this formula, the minimum learning rate Maximum learning rate The number of iterations T during iterative training i =50. This learning rate plan allows for a gradual decrease in the learning rate, thus facilitating fine-tuning of the model as training progresses. To optimize the training process and improve prediction accuracy, stochastic gradient descent is used as the optimizer. Furthermore, softmax cross-entropy is used as the loss function to calculate the probability distribution across the target class, enabling the model to make accurate predictions.

[0089] WSI-integrated multi-instance learning: After training the deep learning model, the focus is on predicting the label of a single patch and its associated probability. These probabilities are then aggregated using a classifier to derive WSI-level predictions. To effectively integrate patch probabilities, this embodiment employs a feature fusion method based on multi-instance learning, aiming to improve the model's prediction accuracy. This method involves integrating various data points or instances from a single sample to develop a comprehensive feature set, with the specific steps as follows:

[0090] Patch prediction: Each patch is predicted using the ResNet50 model to obtain its corresponding probability and label, denoted as Patch. prob and Patch pred The predicted probability is rounded to two decimal places.

[0091] Multi-instance learning feature aggregation (histogram features): Each distinct number is treated as a "box," and the frequency of each data type within these "boxes" is calculated. Patch statistics are also performed. prob and Patch pred The frequency of falling into each container. All features were subjected to min-max normalization. This process produced Histo. prob and Histo pred ;

[0092] Bag-of-words (Bow) feature aggregation: Initially, by identifying patches... prob and Patch pred A dictionary is created using unique elements from the dictionary. Each patch is then represented as a vector, recording the frequency of each dictionary element within the patch. The inverse document frequency transformation (IF-FFM) is applied to these vectors, emphasizing the importance of features with lower frequencies but higher information content. This results in a Bow feature representation for each patch, summarizing the presence and importance of features within the patch. The final Bow feature representation is... prob and Bow predIt provides a comprehensive weighted representation of the patch, which is applicable to subsequent analysis processes;

[0093] Early Feature Fusion: The final stage of feature fusion based on multi-instance learning involves the integration of previously derived features: Histo prob Histo pred Bow prob and Bow pred To achieve this, a feature-based connection method is employed, using... This method combines these individual feature sets into a single, comprehensive feature vector. The specific formula for this connection is as follows:

[0094]

[0095] Construction and validation of the pathomic model: A comprehensive pathomic model was developed by constructing personalized patient profiles through the merging of patch predictions, probability histograms, and Bow features. To optimize feature selection, Pearson correlation coefficients were used, retaining only one feature with a correlation greater than 0.9 from each pair; this step was crucial for determining the final feature set supporting the pathomic model. The pathomic model was constructed using various machine learning algorithms to predict mutation states, including LR, SVM, RF, and XGBoost. The predictive performance of the model was evaluated using ROC and AUC, and the consistency between predicted probabilities and actual results was assessed using calibration curves derived from the HL test. The Delong test and DCA were used to evaluate the model's performance and clinical benefits.

[0096] refer to Figure 2 A multimodal combined model integrating radiomics, pathomics, and clinical features predicts EGFR mutations in non-small cell lung cancer:

[0097] Research cohorts and data collection;

[0098] Clinical model construction: Univariate and multivariate regression analyses were used to examine clinical characteristics. Clinical characteristics with p-values ​​<0.05 were retained and included in the joint model. Similar to the construction of the pathomics model, LR, SVM, and RF methods were used to construct the clinical model.

[0099] Construction and Comparative Analysis of Multimodal Combined Models: In the final stage, clinical, radiomics, and pathological features are integrated into a comprehensive nomogram, termed the combined model. This model aims for broad interpretation and analysis. For practical clinical applications, clinical features with p-values ​​<0.05 in multivariate regression analysis are combined with radiomics and pathological features to form a combined model. This model is visualized using a nomogram for easy interpretation. The predictive performance of Nomogram, Clinical, DLRadiomics, and Pathomics models is compared. Calibration curves, Delong tests, and DCA are used to validate the clinical benefits of each model.

[0100] Predicting the distribution of EGFR mutation subtypes: Analysis of variance was used to examine the distribution of the prediction features of the joint model under different EGFR mutation subtypes (19Del, L858R and other mutation subtypes) to verify the predictive ability of the joint model for EGFR mutation subtypes.

[0101] Preferably, the above implementation method uses a series of software tools and custom code to ensure accuracy and efficiency for analysis; ITK-SNAPv3.8.0 is used for medical image segmentation and processing; the computational model and data analysis are mainly implemented in Pythonv3.7.12; the key Python libraries used include PyTorchv1.8.0 for deep learning algorithms, OnekeyAIv2.2.3 for patch extraction and processing, scikit-learnv1.0.2 for machine learning models, and PyRadiomicsv3.0 for extracting radioactive features.

[0102] The beneficial effects of this invention are as follows:

[0103] This invention captures radiological features through a 3D deep learning model, preserving the spatial location of lung lesions and the integrity of extracted information. Through a multi-instance learning model, the processing of WSI images of tissue sections eliminates the need for manual delineation of regions of interest, saving significant time and manpower. By using a joint model, the predictive efficiency of the model is improved, comprehensively capturing both macroscopic and microscopic features of tumors, compensating for the heterogeneity of tumor tissue, avoiding the limitations of invasive biopsies and tissue collection, and saving on expensive gene testing costs.

[0104] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0105] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting EGFR mutation status in non-small cell lung cancer, characterized in that, include: Acquire CT images, WSI images, and clinical characteristics of the target patient; The CT images are input into a pre-trained 3D deep learning model for prediction to obtain radiomics features; The WSI image is input into a pre-trained multi-instance learning model for prediction to obtain pathological features; The clinical features, radiomics features, and pathological features were predicted using a joint model based on nomograms to obtain the EGFR mutation status prediction results.

2. The method for predicting EGFR mutation status in non-small cell lung cancer according to claim 1, characterized in that, The training process of the 3D deep learning model includes: The pre-collected CT images were imported into ITK-SNAP software in DICOM format for manual annotation to obtain raw CT data; The raw CT data is resampled and spatially normalized using a fixed resolution to obtain CT preprocessed data. The CT preprocessed data is processed using data augmentation technology to obtain CT enhanced data; Based on the cosine annealing learning rate scheduling mechanism and stochastic gradient descent, the ResNet50-3D model, ShuffleNet-3D model, and Densenet121-3D model are trained using CT-enhanced data with softmax cross-entropy as the loss function. The 3D deep learning model is obtained by fusing the trained ResNet50-3D model, ShuffleNet-3D model, and Densenet121-3D model.

3. The method for predicting EGFR mutation status in non-small cell lung cancer according to claim 1, characterized in that, The training process of the multi-instance learning model includes: The pre-collected WSI images are divided into 512×512 pixel tiles. The same EGFR tag is added to the tiles belonging to the same WSI image. The white background of the tiles is removed using the OnekeyAI OKT-patch2predict tool to obtain the original tile set. The ResNet50 model is trained using the original tile set to obtain the multi-instance learning model.

4. The method for predicting EGFR mutation status in non-small cell lung cancer according to claim 1, characterized in that, The construction process of the joint model includes: Univariate and multivariate regression analyses were performed on the pre-collected clinical characteristics; The clinical features, radiomics features, and pathological features with p-values ​​less than 0.05 in the regression analysis results are integrated into a nomogram to obtain the combined model.

5. The method for predicting EGFR mutation status in non-small cell lung cancer according to claim 2, characterized in that, The expression for the cosine annealing learning rate scheduling mechanism is: Where, η t Let be the learning rate at step t; Minimum learning rate; T represents the maximum learning rate; i T is the number of iterations. cur This represents the current iteration number.

6. The method for predicting EGFR mutation status in non-small cell lung cancer according to claim 2, characterized in that, The fixed resolution voxel spacing is standardized to 1mm×1mm×1mm.

7. The method for predicting EGFR mutation status in non-small cell lung cancer according to claim 2, characterized in that, The fusion methods of the ResNet50-3D model, the ShuffleNet-3D model, and the Densenet121-3D model include: maximum value fusion, minimum value fusion, and mean value fusion.

8. The method for predicting EGFR mutation status in non-small cell lung cancer according to claim 3, characterized in that, The ResNet50 model is trained using the original tile set to obtain the multi-instance learning model, which includes: The ResNet50 model is used to predict the small magnetic piece, and the predicted probability and predicted label are obtained. The predicted probability and the frequency with which the predicted label falls into each preset container are statistically analyzed to obtain multi-instance learning features; the multi-instance learning features include: multi-instance learning probability and multi-instance learning label; A dictionary is obtained by identifying the unique element in the predicted probability and the predicted label. Each small magnetic piece is represented as a vector recording the frequency of each element in the small magnetic piece that belongs to the dictionary. The vector is then subjected to word frequency inverse document frequency transformation to obtain bag-of-words model features. The bag-of-words model features include: bag-of-words probability and bag-of-words label. The multi-instance learned features and the bag-of-words model features are concatenated to obtain a comprehensive feature vector; the expression of the comprehensive feature vector is: feature fusion =Histo prob ⊕Histo pred ⊕Bow prob ⊕Bow pred ; Among them, feature fusion The comprehensive feature vector; Histo prob Histo pred They are respectively; Bow prob Bow pred They are respectively; ⊕ indicates element-wise addition.

Images