Multi-modal feature-based lung cancer targeted drug treatment result prediction method

Abstract

The invention discloses a lung cancer targeted drug treatment result prediction method based on multi-modal characteristics, and belongs to the technical field of medical image analysis, deep learning and precision medical treatment. According to the method, effective feature extraction is performed on pathological images, CT images and clinical text data by using different encoders, and deep fusion and information complementation of different modal data are realized by designing a multi-modal collaborative modeling module and a cross-modal feature sensing module. Meanwhile, a multi-instance learning method is introduced to better process heterogeneity among different modal data, contribution weights of different instances to an overall prediction result are automatically learned through an attention mechanism, and the capture and expression ability of the model to important features is enhanced. According to the method disclosed by the invention, the accuracy and the stability of prediction of the targeted drug treatment effect of the lung cancer patient are remarkably improved, meanwhile, the calculation complexity of multi-modal data processing is effectively reduced, and the method has important clinical application value and technical innovation.

Description

Technical Field

[0001] The present invention relates to a method for predicting the results of lung cancer targeted drug treatment based on multimodal features, and belongs to the technical fields of medical image analysis, deep learning and precision medicine. Background Art

[0002] Lung cancer is one of the most lethal malignancies in the world, especially non-small cell lung cancer (NSCLC). Due to the lack of obvious symptoms in the early stages, most patients are already in the advanced stage when diagnosed, and the prognosis is poor. In recent years, targeted therapy with epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs) has made significant progress, significantly improving the survival rate of patients with EGFR mutations. However, EGFR-TKIs are not suitable for all patients, and some patients, even those with EGFR mutations, may not respond well to treatment due to drug resistance or other factors. In addition, EGFR-TKIs may cause adverse reactions such as rash and diarrhea. Therefore, how to screen patients who can truly benefit from EGFR-TKIs has become a major challenge in current precision medicine.

[0003] Currently, the markers used clinically to predict the efficacy of EGFR-TKI treatment mainly include EGFR mutation status, programmed death ligand-1 (PD-L1) expression level, and tumor mutation burden (TMB). However, the predictive ability of these markers is still limited. For example, although EGFR mutation status is a core biomarker for EGFR-TKI treatment, there are still large differences in the response to treatment among patients with the same mutation. In addition, radiological imaging assessments (such as CT) can provide information on tumor size and morphological changes, but their ability to predict early efficacy is limited, and it is difficult to accurately distinguish between resistant and sensitive patients.

[0004] In recent years, deep learning technology has demonstrated tremendous potential in medical imaging analysis. By extracting high-level features from imaging data, deep learning models can capture biological characteristics of tumors that are difficult to identify with traditional imaging, providing new insights into the efficacy prediction of EGFR-TKIs. However, most current deep learning models still rely on single-modality data (such as imaging or genetic data), making it difficult to fully integrate clinical information, pathological features, and imaging characteristics of lung cancer patients, resulting in limited predictive capabilities.

[0005] The integration of multimodal data offers a new approach to addressing this problem. Combining clinical data, pathological features, and imaging information can provide a more comprehensive picture of a patient's condition and improve the ability to predict the response to EGFR-TKI therapy. However, the heterogeneity between multimodal data and the fusion method place higher demands on the accuracy and generalization capabilities of the model. Therefore, developing a deep learning model based on multimodal features to fully exploit EGFR-TKI-related biological information and improve the accuracy and stability of efficacy prediction has become an important research direction in the current fields of medical imaging analysis and precision medicine. Summary of the Invention

[0006] In order to improve the accuracy and robustness of lung cancer targeted drug treatment outcome prediction, the present invention provides a lung cancer targeted drug treatment outcome prediction method based on multimodal features. The technical solution is as follows:

[0007] Step 1: Obtain the patient's clinical data, CT images, and pathological images as multimodal input data;

[0008] Step 2: Use different feature encoders to extract clinical data features, CT image features, and pathological image features respectively to obtain different modality features;

[0009] Step 3: Use the multimodal collaborative modeling module to align different modal features into a unified feature space;

[0010] Step 4: Use the cross-modal feature perception module to model the interaction between different modalities and highlight key features through the attention mechanism;

[0011] Step 5: Generate the final treatment outcome prediction using a multi-instance learning-based prediction strategy.

[0012] Optionally, in step 2, the CT image features are extracted using a MobileNet encoder, the pathological image features are extracted using a ResNet-18 encoder, and the clinical data features are extracted using a BERT encoder.

[0013] Optionally, the multimodal collaborative modeling module includes a feature mapping module and a feature aggregation module;

[0014] The feature mapping module includes a normalization layer and a global average pooling layer for nonlinear mapping and abstraction of different modal features; the feature aggregation module includes a feature splicing operation and a fully connected layer for generating a unified multimodal joint feature representation.

[0015] Optionally, the cross-modal feature perception module includes a cross-modal correlation calculation unit, an attention mapping module, a residual connection and normalization unit;

[0016] The cross-modal correlation calculation unit is used to calculate the correlation matrix between different modal data; the attention mapping unit is used to highlight key features with diagnostic value based on the correlation matrix; the residual connection and normalization unit is used to enhance the stability of feature transfer and prevent network degradation.

[0017] Optionally, the multi-instance learning strategy includes a transformer module and a nonlinear change function sigmoid;

[0018] The transformer module is used to capture the long-range dependencies between different instances and automatically learn the importance weight of each instance based on the self-attention mechanism; the nonlinear change function sigmoid is used to convert the final features into predicted probabilities, outputting the predicted score for each instance, which explicitly reflects the probability or risk level of response to targeted drug treatment.

[0019] A second object of the present invention is to provide a system for predicting the results of lung cancer targeted drug treatment, the system comprising:

[0020] A data acquisition module is used to obtain the patient's clinical data, CT images, and pathological images as multimodal input data;

[0021] The multimodal information encoding module uses different feature encoders to extract clinical data features, CT image features, and pathological image features to obtain different modality features;

[0022] Multimodal collaborative modeling module, used to align different modal features into a unified feature space;

[0023] Cross-modal feature perception module, used to model the interaction between different modalities and highlight key features through the attention mechanism;

[0024] The prediction module uses a prediction strategy based on multi-instance learning to generate the final treatment outcome prediction.

[0025] Optionally, the multimodal collaborative modeling module includes a feature mapping module and a feature aggregation module;

[0026] The feature mapping module includes a normalization layer and a global average pooling layer for nonlinear mapping and abstraction of different modal features; the feature aggregation module includes a feature splicing operation and a fully connected layer for generating a unified multimodal joint feature representation.

[0027] Optionally, the cross-modal feature perception module includes a cross-modal correlation calculation unit, an attention mapping module, a residual connection and normalization unit;

[0028] The cross-modal correlation calculation unit is used to calculate the correlation matrix between different modal data; the attention mapping unit is used to highlight key features with diagnostic value based on the correlation matrix; the residual connection and normalization unit is used to enhance the stability of feature transfer and prevent network degradation.

[0029] Optionally, the prediction module includes a transformer module and a nonlinear change function sigmoid;

[0030] The transformer module is used to capture the long-range dependencies between different instances and automatically learn the importance weight of each instance based on the self-attention mechanism; the nonlinear change function sigmoid is used to convert the final features into predicted probabilities, outputting the predicted score for each instance, which explicitly reflects the probability or risk level of response to targeted drug treatment.

[0031] The third object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting the results of lung cancer targeted drug treatment as described in any one of the above items is implemented.

[0032] The beneficial effects of the present invention are:

[0033] The present invention has achieved a significant technological breakthrough in the field of targeted lung cancer treatment prediction by integrating multimodal medical data such as CT images, pathological images and clinical data. By deeply integrating the morphological features of CT images, the microscopic features of pathological images and the comprehensive information of clinical data, a more comprehensive patient characterization model is constructed, which significantly improves the prediction accuracy compared with the traditional single-modality method. The multi-instance learning architecture effectively solves the inherent heterogeneity problem of medical data. Through the dynamic weighted fusion of different modal features, the model can still maintain stable prediction performance when facing missing data or uneven quality, and the model robustness is significantly improved.

[0034] The multimodal feature alignment technology and cross-modal attention mechanism adopted in this invention can automatically identify key diagnostic features in different data sources. The generated prediction results have better interpretability and can provide clinicians with multi-dimensional decision-making basis including imaging feature weights, key pathological indicators, etc. It can not only predict the possibility of treatment response, but also indicate potential influencing factors, providing a quantitative reference basis for the formulation of personalized treatment plans.

[0035] In addition, in terms of computational efficiency, the present invention designs different feature extraction networks for different data types, achieving a reasonable allocation of computing resources while ensuring the quality of feature extraction, so that the system can run efficiently on conventional medical servers. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 Schematic diagram of the basic process of the method for predicting the treatment results of lung cancer targeted drugs of the present invention.

[0038] Figure 2 This is a structural diagram of a lung cancer targeted drug treatment outcome prediction model based on multimodal features in Example 2 of the present invention.

[0039] Figure 3 This is a schematic diagram of the structure of the multimodal collaborative modeling module in the second embodiment of the present invention.

[0040] Figure 4 Schematic diagram of the cross-modal feature perception module structure in the second embodiment of the present invention.

[0041] Figure 5 Schematic diagram of the multi-instance learning strategy in the second embodiment of the present invention. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0043] Example 1:

[0044] This embodiment provides a method for predicting the treatment results of lung cancer targeted drugs based on multimodal features, including:

[0045] Step 1: Collect the patient's clinical data, including but not limited to age, gender, smoking history, TNM stage, tumor location, and cell type; obtain the patient's imaging data, including CT images and pathology images; CT images can provide the morphological characteristics of the tumor, and pathology images can provide the microscopic characteristics of tumor cells.

[0046] Step 2: Use different feature encoders to extract clinical data features, CT image features, and pathological image features respectively to obtain different modality features.

[0047] Specifically, in this embodiment, the MobileNet encoder is used to extract CT image features, the ResNet-18 encoder is used to extract pathological image features, and the BERT encoder is used to extract clinical data features.

[0048] Step 3: Use the multimodal collaborative modeling module to align different modal features into a unified feature space.

[0049] The multimodal collaborative modeling module includes a feature mapping block and a feature aggregation block; the feature mapping block includes a normalization layer and a global average pooling layer, which are used to perform nonlinear mapping and abstraction of different modal features; the feature aggregation block includes a feature splicing operation and a fully connected layer to generate a unified multimodal joint feature representation.

[0050] Step 4: Use the cross-modal feature perception module to model the interaction relationship between different modalities and highlight key features through the attention mechanism.

[0051] The cross-modal feature perception module includes a cross-modal correlation calculation unit, an attention mapping module, a residual connection and normalization unit; the cross-modal correlation calculation unit is used to calculate the correlation matrix between different modal data; the attention mapping unit highlights key features with diagnostic value based on the above correlation matrix; the residual connection and normalization unit is used to enhance the stability of feature transfer and prevent network degradation.

[0052] Step 5: Generate the final treatment outcome prediction using a multi-instance learning-based prediction strategy.

[0053] The multi-instance learning strategy is suitable for processing heterogeneous data and can optimize the contribution ratio of different modal features in the prediction process.

[0054] The multi-instance learning strategy includes a transformer module and a nonlinear change function sigmoid; the transformer module is an attention mechanism used to capture long-range dependencies between different instances and automatically learn the importance weights of each instance based on the self-attention mechanism; the nonlinear change function sigmoid is used to convert the final features into predicted probabilities, outputting the predicted score for each instance, explicitly reflecting the probability or risk level of response to targeted drug treatment.

[0055] Example 2:

[0056] This embodiment provides a method for predicting the treatment results of lung cancer targeted drugs based on multimodal features. Figures 1-4 , the method comprising:

[0057] Step 1: Obtain a dataset of CT, pathology, and clinical text for lung cancer targeted drug treatment. In this dataset, 500 cases with good responses to lung cancer targeted drug treatment and 500 cases with no obvious response were used for training; 80 cases with good responses and 80 cases with no obvious response were used for testing. The images were all resized to 256×256 pixels. Use Python OpenCV to read the original images and convert them into a three-dimensional pixel matrix.

[0058] Step 2: Use the lung cancer targeted drug treatment outcome prediction model based on multimodal features to train the labeled image data to obtain a trained lung cancer targeted drug treatment prediction model.

[0059] Specifically, refer to Figure 2 ,The prediction model for lung cancer targeted drug treatment includes : a multimodal information encoding module, a multimodal feature ,joint module, a cross-modal feature perception module and a multi-instance learning ,strategy.

[0060] The feature encoders include: ResNet-18 encoder, MobileNet encoder and Bert encoder, which are used to extract different multimodal features and obtain the extracted CT features M respectively. CT Pathological characteristics Pth and clinical text features M text .

[0061] Further, refer to Figure 3 ,The multimodal collaborative modeling module includes a feature mapping block and a ,feature aggregation block.

[0062] The feature mapping block includes a normalization layer and a global average pooling layer, which are used to perform nonlinear mapping and abstraction of different modal features. Taking the CT feature as an example, the calculation formula is expressed as:

[0063] F CT =GAP(LN(M CT ))

[0064] Where LN(·) represents the normalization layer, GAP(·) is the global average pooling layer, and M CT is the CT feature extracted by the CT feature encoder. Similarly, the pathology and text modalities obtain the corresponding enhanced features F Pth , F text .

[0065] This feature mapping block effectively reduces the dimension and computational complexity of multimodal features and enhances the multimodal feature fusion effect.

[0066] The feature aggregation block includes feature concatenation operations and fully connected layers.

[0067] The feature concatenation operation is expressed as:

[0068] J=Concat(F CT ,F Pth ,F text )

[0069] Concat is a vector concatenation operation to obtain the combined multimodal feature J.

[0070] The fully connected layer operation is:

[0071] A=FC(J)

[0072] Where FC(·) represents a fully connected layer, which is used to map the concatenated features of different dimensions to a consistent feature dimension. The obtained result A represents the final aggregated feature representation, which will be used as one of the inputs of the cross-modal attention mechanism.

[0073] For further reference, Figure 4 The cross-modal feature perception module includes a cross-modal correlation calculation unit, an attention mapping module, a residual connection and normalization unit; the cross-modal correlation calculation unit is used to calculate the correlation matrix between different modal data; the attention mapping unit highlights the key features with diagnostic value based on the above correlation matrix; the residual connection and normalization unit is used to enhance the stability of feature transfer and prevent network degradation.

[0074] Specifically, after completing the aforementioned aggregated feature representation, further modeling of the interactions and correlations between different modalities is required. By calculating the dependencies between each modality's features and the joint features within a shared semantic space, the team aims to highlight key features that are helpful in predicting the outcomes of targeted lung cancer drug treatment, thereby achieving higher discrimination accuracy.

[0075] Taking the CT modality as an example, the calculation formula of the cross-modality correlation calculation unit is expressed as:

[0076]

[0077] in W is the transposed expression of the enhanced features obtained after the CT modality is processed by global average pooling, linear mapping, etc.; A,CT is a learnable projection matrix used to match the feature dimension with A; is the feature dimension size or other scaling factors to avoid gradient problems caused by excessive values; tanh(·) maps the values to [-1,1], so that the model has a suppressive effect on outliers; C CT is the cross-modal correlation matrix between CT and multimodal joint features, which can measure the degree of matching between the two in the semantic space.

[0078] The same processing can be done for the pathological mode and the text mode, and we can get:

[0079]

[0080]

[0081] The attention mapping unit is used to highlight important features between modalities. Taking the CT modality as an example, the calculation formula is expressed as:

[0082] H CT =ReLU(FCT W C,CT C CT )

[0083] Where W C,CT For the learnable mapping matrix, the correlation matrix C CT Introducing CT features; ReLU (·) activation function truncates negative values to 0, ensuring the output is non-negative and increasing the nonlinear expression ability of the network; the obtained H CT Represents the intermediate attention map after fusing cross-modal information.

[0084] The same processing can be done for the pathological mode and the text mode, and we can get:

[0085] H Pth =ReLU(F Pth W C,Pth C Pth )

[0086] H text =ReLU(F text W C,text C text )

[0087] Subsequently, the cross-modal enhanced features of the CT modality are updated by combining them with the original features through residual or additive operations:

[0088]

[0089] Among them, Norm(·) can use LayerNorm for numerical normalization; is a learnable mapping matrix that controls the influence of attention mapping on the original features.

[0090] The same processing can be done for the pathological mode and the text mode, and we can get:

[0091]

[0092] Specifically, refer to Figure 5 , using a multi-instance learning strategy, the multimodal features obtained above are used as instance input, the most important instances are extracted through the attention mechanism, and multiple instances form a whole to predict the label Y∈{0,1}. The classification score of each instance is calculated as:

[0093]

[0094] Where G is the transformer module; S is a nonlinear transformation function sigmoid, which is used to output the predicted value; Z i Represents the fusion feature of the i-th instance, specifically Z i ∈{X att,CT ,Xatt,Pth ,X att,text}; is the prediction score of the i-th instance.

[0095] The final prediction result is obtained through weighted aggregation:

[0096]

[0097] Where n is the number of instances; w i The instance weights obtained during model training are used to highlight key instances; is the final prediction value obtained after integrating all instances.

[0098] Step 3: Use the trained lung cancer targeted drug treatment outcome prediction model based on multimodal feature fusion to predict the multimodal data (including CT images, pathological images and clinical text data) of patients in the test sample set to obtain the patient's targeted drug treatment response probability and compare it with the actual clinical efficacy to evaluate the accuracy and stability of the model prediction.

[0099] In summary, this embodiment provides a method for predicting the results of lung cancer targeted drug treatment based on multimodal feature fusion and multi-instance learning. Through the effective design of the multimodal collaborative modeling module and the cross-modal feature perception module, it can fully capture and fuse the global features of different modal data and the long-range correlation between modalities, solving the problems of limited feature expression and information independence between modalities in traditional single modality analysis. At the same time, by introducing the transformer module and the adaptive instance weight attention mechanism, the model can automatically capture important instances, thereby improving the accuracy and interpretability of the prediction results. While improving the accuracy and robustness of the prediction of targeted drug treatment responses in lung cancer patients, this embodiment also reduces the complexity and computational burden of multimodal data analysis. It has broad application prospects and important practical value in the fields of clinical lung cancer treatment decision support and precision medicine.

[0100] Example 3:

[0101] This embodiment provides a lung cancer targeted drug treatment outcome prediction system based on multimodal features, which is used to implement the lung cancer targeted drug treatment outcome prediction method described in Example 2, including:

[0102] The data acquisition module is used to obtain the patient's clinical data, CT images and pathological images as multimodal input data.

[0103] The multimodal information encoding module uses different feature encoders to extract clinical data features, CT image features and pathological image features respectively to obtain different modality features.

[0104] Multimodal collaborative modeling module is used to align different modal features into a unified feature space.

[0105] The cross-modal feature perception module is used to model the interaction relationship between different modalities and highlight key features through the attention mechanism.

[0106] The prediction module uses a prediction strategy based on multi-instance learning to generate the final treatment outcome prediction.

[0107] Among them, the multimodal collaborative modeling module includes a feature mapping module and a feature aggregation module; the feature mapping module includes a normalization layer and a global average pooling layer, which are used to perform nonlinear mapping and abstraction of different modal features; the feature aggregation module includes a feature splicing operation and a fully connected layer, which are used to generate a unified multimodal joint feature representation.

[0108] The cross-modal feature perception module includes a cross-modal correlation calculation unit, an attention mapping module, a residual connection and normalization unit; the cross-modal correlation calculation unit is used to calculate the correlation matrix between different modal data; the attention mapping unit highlights key features with diagnostic value based on the correlation matrix; the residual connection and normalization unit is used to enhance the stability of feature transfer and prevent network degradation.

[0109] The prediction module includes a transformer module and a nonlinear change function sigmoid; the transformer module is used to capture the long-range dependencies between different instances and automatically learn the importance weights of each instance based on the self-attention mechanism; the nonlinear change function sigmoid is used to convert the final features into predicted probabilities, outputting the predicted score for each instance, which explicitly reflects the probability or risk level of response to targeted drug treatment.

[0110] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the treatment results of lung cancer targeted drugs, characterized in that: The method comprises: Step 1: Obtain the patient's clinical data, CT images, and pathological images as multimodal input data; Step 2: Use different feature encoders to extract clinical data features, CT image features, and pathological image features respectively to obtain different modality features; Step 3: Use the multimodal collaborative modeling module to align different modal features into a unified feature space; Step 4: Use the cross-modal feature perception module to model the interaction between different modalities and highlight key features through the attention mechanism; Step 5: Generate the final treatment outcome prediction using a multi-instance learning-based prediction strategy.

2. The method for predicting the results of lung cancer targeted drug treatment according to claim 1, wherein: In step 2, the CT image features are extracted using a MobileNet encoder, the pathological image features are extracted using a ResNet-18 encoder, and the clinical data features are extracted using a BERT encoder.

3. The method for predicting the results of lung cancer targeted drug treatment according to claim 1, wherein: The multimodal collaborative modeling module includes a feature mapping module and a feature aggregation module; The feature mapping module includes a normalization layer and a global average pooling layer, which is used to perform nonlinear mapping and abstraction of different modal features; The feature aggregation module includes a feature concatenation operation and a fully connected layer, which is used to generate a unified multimodal joint feature representation.

4. The method for predicting the treatment results of lung cancer targeted drugs according to claim 1, characterized in that: The cross-modal feature perception module includes a cross-modal correlation calculation unit, an attention mapping module, a residual connection and normalization unit; The cross-modal correlation calculation unit is used to calculate the correlation matrix between different modal data; The attention mapping unit highlights key features with diagnostic value based on the correlation matrix; The residual connection and normalization unit are used to enhance the stability of feature transfer and prevent network degradation.

5. The method for predicting the treatment results of lung cancer targeted drugs according to claim 1, characterized in that: The multi-instance learning strategy includes a transformer module and a nonlinear change function sigmoid; The transformer module is used to capture the long-range dependencies between different instances and automatically learn the importance weight of each instance based on the self-attention mechanism; The nonlinear change function sigmoid is used to transform the final features into predicted probabilities, outputting the predicted score for each instance, which explicitly reflects the probability or risk level of response to targeted drug treatment.

6. A lung cancer targeted drug treatment outcome prediction system, characterized in that: The system comprises: A data acquisition module is used to obtain the patient's clinical data, CT images, and pathological images as multimodal input data; The multimodal information encoding module uses different feature encoders to extract clinical data features, CT image features, and pathological image features to obtain different modality features; Multimodal collaborative modeling module, used to align different modal features into a unified feature space; Cross-modal feature perception module, used to model the interaction between different modalities and highlight key features through the attention mechanism; The prediction module uses a prediction strategy based on multi-instance learning to generate the final treatment outcome prediction.

7. The lung cancer targeted drug treatment result prediction system according to claim 6, characterized in that: The multimodal collaborative modeling module includes a feature mapping module and a feature aggregation module; The feature mapping module includes a normalization layer and a global average pooling layer, which is used to perform nonlinear mapping and abstraction of different modal features; The feature aggregation module includes a feature concatenation operation and a fully connected layer, which is used to generate a unified multimodal joint feature representation.

8. The lung cancer targeted drug treatment result prediction system according to claim 6, characterized in that: The cross-modal feature perception module includes a cross-modal correlation calculation unit, an attention mapping module, a residual connection and normalization unit; The cross-modal correlation calculation unit is used to calculate the correlation matrix between different modal data; The attention mapping unit highlights key features with diagnostic value based on the correlation matrix; The residual connection and normalization unit are used to enhance the stability of feature transfer and prevent network degradation.

9. The lung cancer targeted drug treatment result prediction system according to claim 6, characterized in that: The prediction module includes a transformer module and a nonlinear change function sigmoid; The transformer module is used to capture the long-range dependencies between different instances and automatically learn the importance weight of each instance based on the self-attention mechanism; The nonlinear change function sigmoid is used to transform the final features into predicted probabilities, outputting the predicted score for each instance, which explicitly reflects the probability or risk level of response to targeted drug treatment.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the method for predicting the results of lung cancer targeted drug treatment according to any one of claims 1 to 5 is implemented.

Images