Gastric cancer targeted therapy efficacy prediction method, device, equipment and medium

Abstract

The present application provides a gastric cancer targeted therapy efficacy prediction method, device, equipment and medium, the method comprises: obtaining the CT image data to be identified before and after gastric cancer targeted therapy; the CT image data to be identified is input into the gastric cancer targeted therapy efficacy prediction model, and the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model is obtained; wherein, the CT image data to be identified includes the image corresponding to the baseline time and the image corresponding to multiple follow-up time points, and the CT image data to be identified corresponding to each time point includes multiple lesions. The present application is used to solve the defects of the prior art, such as low accuracy of gastric cancer targeted therapy efficacy prediction, poor effect of gastric cancer targeted therapy efficacy prediction, improve the accuracy of gastric cancer targeted therapy efficacy prediction, and improve the effect of gastric cancer targeted therapy efficacy prediction.

Description

Technical Field

[0001] This invention relates to the field of efficacy prediction technology, and in particular to a method, device, equipment and medium for predicting the efficacy of targeted therapy for gastric cancer. Background Technology

[0002] Gastric cancer ranks among the top five cancers in terms of incidence and mortality worldwide, and is the second most common cancer in my country. It is highly malignant and has a poor prognosis, with patients often diagnosed at an advanced stage. Previous studies have found that the status of human epidermal growth factor receptor 2 (HER2) is associated with the occurrence and prognosis of gastric cancer, with HER2 positivity found in approximately 6%–34% of gastric cancer patients. A large-scale randomized, controlled, international, multicenter phase III clinical trial (ToGA trial) demonstrated that, compared with systemic chemotherapy alone, the anti-HER2 antibody trastuzumab combined with standard chemotherapy provides a survival benefit for HER2-positive cases, extending overall survival to 16 months. This confirms the first-line treatment status of trastuzumab in HER2-positive advanced gastric cancer.

[0003] Despite numerous clinical trials supporting its efficacy, the response rate in HER2-positive gastric cancer patients treated with trastuzumab remains limited. In the ToGA study, the objective response rate for trastuzumab was 47.3%. Furthermore, some patients who initially respond to trastuzumab eventually develop resistance within one year. Therefore, early identification of patients who benefit from trastuzumab treatment and the development of personalized treatment plans for those with poor responses remain crucial steps in improving gastric cancer survival rates.

[0004] Studies have reported that trastuzumab appears to offer limited benefit in patients with heterogeneous HER2-expressing metastatic gastric cancer, with patients having more homogeneous HER2-positive gastric cancer exhibiting longer survival. Endoscopic biopsy with immunohistochemical staining is the primary method for determining HER2 expression status in advanced gastric cancer; however, routine biopsy specimens are merely small and static snapshots of the entire heterogeneous tumor. Compared to invasive biopsies, CT images obtained through regular clinical follow-up can provide objective, longitudinal information from a comprehensive, macroscopic perspective.

[0005] In recent years, with the continuous development of deep learning, neural network-based models have made significant progress in clinical medical applications. Deep learning technology can extract a large amount of information from CT images that is invisible to the naked eye, thus complementing the cognition and diagnosis of clinicians, which provides great potential for the quantitative assessment of tumors. Existing research has shown that selecting appropriate treatment plans can effectively improve the prognosis of gastric cancer patients. A 2019 study demonstrated that using a combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) can effectively uncover the changing trends in CT images of lung nodules, thereby making relatively accurate predictions of early treatment efficacy for lung cancer patients; another 2021 study proved that using convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) can predict the efficacy of treatment for metastatic rectal cancer patients based on sequence image data from early treatment.

[0006] However, there are currently no studies on time series and multifocal subjects analyzing biomarkers for predicting survival in advanced HER2-positive gastric cancer. Existing studies are limited to CT images at a single time point or a single lesion, and cannot effectively integrate other information from the images. As a result, the accuracy of current gastric cancer targeted therapy efficacy prediction is low, and the predictive effect of gastric cancer targeted therapy is poor. Summary of the Invention

[0007] This invention provides a method, device, equipment, and medium for predicting the efficacy of targeted therapy for gastric cancer, which addresses the shortcomings of existing technologies in predicting the efficacy of targeted therapy for gastric cancer, namely, low accuracy and poor predictive effect. This invention aims to improve the accuracy and effectiveness of predicting the efficacy of targeted therapy for gastric cancer.

[0008] This invention provides a method for predicting the efficacy of targeted therapy for gastric cancer, comprising:

[0009] Acquire CT image data of the target cancer before and after targeted therapy;

[0010] The CT image data to be identified is input into the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model.

[0011] The CT image data to be identified includes images corresponding to the baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified at each time point includes multiple lesions.

[0012] According to the present invention, a method for predicting the efficacy of targeted therapy for gastric cancer is provided, wherein the prediction model for targeted therapy for gastric cancer is trained based on time-series imaging data of anti-HER2 targeted therapy for metastatic gastric cancer;

[0013] The gastric cancer targeted therapy efficacy prediction model includes a temporal heterogeneity module and a spatial heterogeneity module;

[0014] The temporal heterogeneity module is used to model the longitudinal changes of the same lesion in temporal heterogeneity, and the spatial heterogeneity module is used to model the interaction relationships between multiple lesions in spatial heterogeneity.

[0015] According to the present invention, a method for predicting the efficacy of targeted therapy for gastric cancer includes inputting the CT image data to be identified into a gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model, including:

[0016] The CT image data to be identified is input into the convolutional feature extractor in the gastric cancer targeted therapy efficacy prediction model to obtain the image features output by the convolutional feature extractor;

[0017] The image features are input into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module.

[0018] The temporal heterogeneity features are input into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the spatial heterogeneity features.

[0019] The temporal heterogeneity features and the spatial heterogeneity features are input into the classifier in the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the classifier.

[0020] According to a method for predicting the efficacy of targeted therapy for gastric cancer provided by the present invention, the image features are input into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module, including:

[0021] The temporal heterogeneity module includes two MHA modules, which are used to integrate information from the same lesion at different time points in the CT image data to be identified based on the image features.

[0022] The offset in days of each time point of the CT image data to be identified relative to the baseline time is mapped onto a multidimensional feature space to obtain multidimensional mapping features;

[0023] The multidimensional mapping features and the image features are input into the first MHA module to obtain the output of the first MHA module;

[0024] The output of the first MHA module is input to the second MHA module to obtain the output of the second MHA module;

[0025] Based on the output of the first MHA module and the output of the second MHA module, the temporal heterogeneity characteristics of the output of the temporal heterogeneity module are determined.

[0026] According to a method for predicting the efficacy of targeted therapy for gastric cancer provided by the present invention, the temporal heterogeneity features are input into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the spatial heterogeneity features, including:

[0027] The temporal heterogeneity features are added to the integrated token to obtain the integrated features;

[0028] The integrated features are input into the third MHA module to obtain the output of the third MHA module;

[0029] The output of the third MHA module is input to the first feature normalization layer to obtain the output of the first feature normalization layer.

[0030] The integrated feature and the output of the feature normalization layer are input into the fully connected layer to obtain the output of the fully connected layer;

[0031] The output of the fully connected layer is input to the second feature normalization layer to obtain the output of the second feature normalization layer.

[0032] The spatial heterogeneity features are determined based on the output of the first feature normalization layer and the output of the second feature normalization layer.

[0033] According to the present invention, a method for predicting the efficacy of targeted therapy for gastric cancer is provided, wherein the training data of the gastric cancer targeted therapy efficacy prediction model is divided into low-risk group data and high-risk data;

[0034] The low-risk group data and the high-risk data were obtained by dividing the patients' total survival time into one-year survival periods.

[0035] The gastric cancer targeted therapy efficacy prediction model is trained based on a survival loss function, which includes a cross-entropy loss function and a risk probability ratio loss function.

[0036] The present invention also provides a device for predicting the efficacy of targeted therapy for gastric cancer, comprising:

[0037] The data acquisition module is used to acquire CT image data to be identified before and after targeted therapy for gastric cancer.

[0038] The risk probability prediction module is used to input the CT image data to be identified into the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model.

[0039] The CT image data to be identified includes images corresponding to the baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified at each time point includes multiple lesions.

[0040] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the gastric cancer targeted therapy efficacy prediction method as described above.

[0041] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the gastric cancer targeted therapy efficacy prediction method as described above.

[0042] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the gastric cancer targeted therapy efficacy prediction method as described above.

[0043] The present invention provides a method, device, equipment, and medium for predicting the efficacy of targeted therapy for gastric cancer. Through a gastric cancer targeted therapy efficacy prediction model, this model predicts the probability of gastric cancer risk from CT image data of any number of follow-ups and any number of lesions. It effectively integrates information from different follow-up data and different lesions, and can handle image data from any number of follow-ups and any number of lesions. Therefore, it can be effectively applied in actual clinical scenarios. This method of predicting patient efficacy and prognosis by integrating information from different follow-up data and different lesions improves the accuracy of gastric cancer targeted therapy efficacy prediction, thereby enhancing the predictive effect of gastric cancer targeted therapy. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0045] Figure 1 This is one of the flowcharts of the gastric cancer targeted therapy efficacy prediction method provided by the present invention;

[0046] Figure 2 This is the second flowchart of the gastric cancer targeted therapy efficacy prediction method provided by the present invention;

[0047] Figure 3 This is an overall framework diagram of the gastric cancer targeted therapy efficacy prediction method provided by the present invention;

[0048] Figure 4 This is the third flowchart of the gastric cancer targeted therapy efficacy prediction method provided by the present invention;

[0049] Figure 5 This is a schematic diagram of the structure of the MHA module, temporal heterogeneity module, and spatial heterogeneity module in the gastric cancer targeted therapy efficacy prediction method provided by the present invention.

[0050] Figure 6 This is the fourth flowchart of the gastric cancer targeted therapy efficacy prediction method provided by the present invention;

[0051] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0053] The following is combined Figures 1-6 This invention describes a method for predicting the efficacy of targeted therapy for gastric cancer.

[0054] Please refer to Figure 1 The method for predicting the efficacy of targeted therapy for gastric cancer proposed in this invention includes:

[0055] Step 10: Obtain the CT image data to be identified before and after targeted therapy for gastric cancer;

[0056] Step 20: Input the CT image data to be identified into the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model;

[0057] The CT image data to be identified includes images corresponding to the baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified at each time point includes multiple lesions.

[0058] The primary focus is on the model's ability to predict treatment efficacy based on CT image data from early targeted therapy. Before implementation, baseline CT images of the patient and several (or zero) follow-up CT images are required. For each CT image, a radiologist needs to delineate the lesion region (at least one) (or ROI region). For each lesion at each time point, its corresponding ROI region is then delineated.

[0059] Suppose the image of the nth lesion at time point t is represented as follows: Then it can be adopted This represents the image sequence of the i-th lesion at time point T. (Using...) This represents the number of days the T time points are offset from the baseline time. Then, the image is cropped, scaled, and filled according to the delineated ROI region, and standardized using statistical information of different lesions to obtain the CT image data to be identified.

[0060] Then, the CT image data to be identified is input into the gastric cancer targeted therapy efficacy prediction model for prediction, and the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model is obtained.

[0061] The gastric cancer targeted therapy efficacy prediction method provided by this invention uses a gastric cancer targeted therapy efficacy prediction model. This model predicts the risk probability of gastric cancer from CT image data of any number of follow-ups and any number of lesions. It can effectively integrate information from different follow-up data and different lesion information, and can handle image data of any number of follow-ups and any number of lesions. Therefore, it can be effectively applied in actual clinical scenarios. This method of predicting the patient's efficacy and prognosis by integrating information from different follow-up data and different lesion information improves the accuracy of gastric cancer targeted therapy efficacy prediction, thereby improving the efficacy prediction effect of gastric cancer targeted therapy.

[0062] In one possible embodiment, the gastric cancer targeted therapy efficacy prediction model is trained based on time-series imaging data of anti-HER2 targeted therapy for metastatic gastric cancer;

[0063] The gastric cancer targeted therapy efficacy prediction model includes a temporal heterogeneity module and a spatial heterogeneity module;

[0064] The temporal heterogeneity module is used to model the longitudinal changes of the same lesion in temporal heterogeneity, and the spatial heterogeneity module is used to model the interaction relationships between multiple lesions in spatial heterogeneity.

[0065] In this embodiment, for each lesion in the training data, assuming there are T time points, the data is first standardized using pre-calculated mean and standard deviation. Then, a shared convolutional feature extractor is used to obtain T image features. Next, a temporal heterogeneity module is used to model the changes in the current lesion over time, thus obtaining the representative features of that lesion. Then, a spatial heterogeneity module is used to mine the interactions between N lesions, thereby obtaining a global representative feature. Finally, the global representative feature is used by a classifier to generate a risk coefficient. A higher risk coefficient indicates a worse response to targeted therapy for the patient, while a lower risk coefficient indicates a better response.

[0066] This embodiment proposes a predictive model for the efficacy of targeted therapy for gastric cancer based on a spatiotemporal heterogeneity module. The temporal heterogeneity module is used to mine the changing trends of the same lesion at different time points; the spatial heterogeneity module is used to model the interaction relationships between different lesions in the same patient.

[0067] In one possible embodiment, please refer to Figure 2 Step 20: Input the CT image data to be identified into the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model, including:

[0068] Step 21: Input the CT image data to be identified into the convolutional feature extractor in the gastric cancer targeted therapy efficacy prediction model to obtain the image features output by the convolutional feature extractor;

[0069] Step 22: Input the image features into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module;

[0070] Step 23: Input the temporal heterogeneity features into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the spatial heterogeneity features;

[0071] Step 24: Input the temporal heterogeneity features and the spatial heterogeneity features into the classifier in the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the classifier.

[0072] The overall architecture diagram of the proposed method is as follows: Figure 3 As shown, for a patient, assuming that their baseline (BL) and first follow-up (1) data can be collected... st Follow-up, 1F), second follow-up (2) nd Follow-up (2F) and other follow-up studies at a total of T time points, with CT image data for N lesions at each time point. For each lesion at each time point, the region containing it (called the Region of Interest, ROI) is delineated. Assume the image of the nth lesion at time point t is represented as... Then it can be adopted This represents the image sequence of the i-th lesion at time point T. (Using...) This represents the number of days the T time points are offset from the baseline time. Then, the image is cropped, scaled, and filled according to the delineated ROI region, and standardized using statistical information of different lesions to obtain the CT image data to be identified.

[0073] Image features are extracted using the shared convolutional feature extractor ε (ResNet-18), and are represented as follows: The characteristic sequence F of the nth lesion n The temporal heterogeneity module will be used to obtain the temporal heterogeneity features of the lesion image at different time points. For N lesions, N temporal heterogeneity features can be obtained, represented as follows:

[0074] The temporal heterogeneity characteristics of different lesions will be obtained through the spatial heterogeneity module to obtain the interaction relationships between multiple lesions, thereby obtaining the spatial heterogeneity characteristics. Then, a classifier consisting of two layers of sensing mechanisms is used to predict the probability of gastric cancer risk. The formula for calculating the probability of gastric cancer risk includes:

[0075]

[0076] in

[0077] By leveraging the characteristics of Transformer, the number of follow-up times and the number of lesions are decoupled from the model structure, enabling the model to adapt to any number of follow-up data and any number of lesion images, thus allowing the model to be applied to complex scenarios in actual clinical practice.

[0078] We propose a predictive model for the efficacy of targeted therapy in gastric cancer based on the spatiotemporal heterogeneity Transformer. This model simultaneously mines the interaction information between different lesions (spatial heterogeneity) and effectively integrates information from different follow-up data (temporal heterogeneity). This model can handle data from any number of follow-ups and any number of lesions, thus making it effectively applicable in real-world clinical scenarios. We will also validate the effectiveness of the proposed model using retrospective and prospective cohort studies.

[0079] In one possible embodiment, please refer to Figure 4 Step 22: Input the image features into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module, including:

[0080] The temporal heterogeneity module includes two MHA modules, which are used to integrate information from the same lesion at different time points in the CT image data to be identified based on the image features.

[0081] Step 221: Map the number of days of offset from the baseline time at each time point of the CT image data to be identified onto a multidimensional feature space to obtain multidimensional mapping features;

[0082] Step 222: Input the multidimensional mapping features and the image features into the first MHA module to obtain the output of the first MHA module;

[0083] Step 223: Input the output of the first MHA module to the second MHA module to obtain the output of the second MHA module;

[0084] Step 224: Based on the output of the first MHA module and the output of the second MHA module, determine the temporal heterogeneity characteristics of the output of the temporal heterogeneity module.

[0085] The core modules of both the temporal heterogeneity module TH-former and the temporal-spatial heterogeneity module SH-former are multi-head attention (MHA) modules. For example... Figure 5 As shown in (b), assuming its input is Where M represents the number of sequence objects, and c represents the feature dimension of each object. Assuming there are H heads, the h-th head is assigned C′ = C / H features by the network, denoted as . The calculation process of the MHA module can be described as follows:

[0086]

[0087]

[0088] Where α(·), β(·), and γ(·) are all fully connected layers. This can be denoted as attention map A. This represents a connection operation in dimensional space.

[0089] In one possible embodiment, please refer to Figure 6 Step 23: Input the temporal heterogeneity features into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the spatial heterogeneity features, including:

[0090] Step 231: Add the temporal heterogeneity feature to the integration token to obtain the integration feature;

[0091] Step 232: Input the integrated features into the third MHA module to obtain the output of the third MHA module;

[0092] Step 233: Input the output of the third MHA module into the first feature normalization layer to obtain the output of the first feature normalization layer;

[0093] Step 234: Input the integrated feature and the output of the feature normalization layer into the fully connected layer to obtain the output of the fully connected layer;

[0094] Step 235: Input the output of the fully connected layer into the second feature normalization layer to obtain the output of the second feature normalization layer;

[0095] Step 236: Determine the spatial heterogeneity feature based on the output of the first feature normalization layer and the output of the second feature normalization layer.

[0096] In this embodiment, the spatial heterogeneity module, namely the Spatial-heterogeneity Transformer (SH-former), is used to model the interaction relationships between multiple lesions. For example... Figure 5 As shown in (c), the temporal heterogeneity characteristics of N lesions can be expressed as follows: After adding the integration token, it can be represented as The computation process of the spatial heterogeneity module SH-former can be described as follows:

[0097]

[0098]

[0099]

[0100] in, Selected(·) means taking the vector containing the integrated token, i.e.

[0101] In one possible embodiment, the training data of the gastric cancer targeted therapy efficacy prediction model is divided into low-risk group data and high-risk data;

[0102] The low-risk group data and the high-risk data were obtained by dividing the patients' total survival time into one-year survival periods.

[0103] The gastric cancer targeted therapy efficacy prediction model is trained based on a survival loss function, which includes a cross-entropy loss function and a risk probability ratio loss function.

[0104] During the training process, patients' overall survival (OS) was divided into low-risk and high-risk groups based on a one-year survival time. In other words, the training data for the gastric cancer targeted therapy efficacy prediction model was obtained by dividing patients based on their overall survival (OS) with a one-year survival time as the dividing line.

[0105] During training, data from a maximum of five time points are included for each patient. To improve training efficiency and model generalization, four lesions are randomly sampled with replacement for each training iteration at each time point. Let x be the input data for the i-th patient. i The model proposed above is h θ Then the model predicts x i The output can be represented as h θ (x i The training of a gastric cancer targeted therapy efficacy prediction model is supervised by the cross-entropy loss function to learn the division between low-risk and high-risk groups, as follows:

[0106]

[0107] Furthermore, considering that even within the same patient group, overall survival (OS) can vary significantly. For example, two patients with an OS of 50 months and 13 months, though both belonging to the low-risk group, should have different risk probabilities. Therefore, inspired by the DeepSurv model, we calculate the loss function l representing the proportional relationship of risk probabilities among different patients in the same group. surv To supervise the model in generating different risk probabilities based on different survival levels, as follows:

[0108]

[0109] In summary, combining the cross-entropy loss function for binary classification and the survival loss function based on the risk probability ratio, the overall loss function can be described as follows:

[0110] l = l ce +l surv

[0111] The testing process simulated a clinical use scenario, considering only CT images from early targeted therapy (approximately within six months of treatment). Therefore, data from a maximum of three time points were included for each patient: baseline (BL), first follow-up (1... st Follow-up, 1F) and second follow-up (2) nd Follow-up (2F), where all lesions at each time point are considered. Assume the input data for the i-th patient is x. i The model trained as described above Its risk probability can be predicted The larger the value, the lower the one-year survival probability; the smaller the value, the higher the one-year survival probability.

[0112] The test results are shown in the table below (AUC performance comparison table (values ​​in parentheses represent 95% confidence intervals)):

[0113]

[0114] Wherein, TB: Tumor burden (total target lesion area) information using baseline data. DL-BS: The proposed model, using only baseline data. DL-1F: The proposed model, using baseline and first follow-up data. DL-2F: The proposed model, using baseline, first follow-up, and second follow-up data.

[0115] During model training, with the addition of follow-up data, the model was able to better distinguish between high-risk and low-risk groups. The low-risk group model focused more on the data from the first and second follow-up visits, while the high-risk group model focused more on the data from the first follow-up visit. The distribution pattern of treatment response suggests that high-risk and low-risk populations for anti-HER2 therapy can be distinguished early after treatment, especially for the high-risk population, where the images from the first follow-up visit are more revealing of treatment response information. This may be related to the presence of trastuzumab resistance patterns in the high-risk population. The low-risk group model focused more on the relationship between the primary tumor and lung nodules, while the high-risk group model focused more on the relationship between the primary tumor and liver metastases. Previous literature has reported that HER2-positive advanced gastric cancer is more likely to metastasize to the liver, and liver metastasis is significantly associated with poor progression-free survival (PFS). This may explain why the high-risk group model focuses more on the interaction between the primary tumor and liver metastases.

[0116] In this embodiment, by combining binary classification loss and survival loss, the model can not only distinguish between high and low risk groups, but also rank the risk probabilities among different patients, thereby providing more detailed guidance for clinical decision-making.

[0117] The following describes the gastric cancer targeted therapy efficacy prediction device provided by the present invention. The gastric cancer targeted therapy efficacy prediction device described below and the gastric cancer targeted therapy efficacy prediction method described above can be referred to in correspondence.

[0118] The gastric cancer targeted therapy efficacy prediction device proposed in this invention includes:

[0119] The data acquisition module is used to acquire CT image data to be identified before and after targeted therapy for gastric cancer.

[0120] The risk probability prediction module is used to input the CT image data to be identified into the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model.

[0121] The CT image data to be identified includes images corresponding to the baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified at each time point includes multiple lesions.

[0122] Furthermore, the gastric cancer targeted therapy efficacy prediction model is trained based on time-series imaging data of metastatic gastric cancer anti-HER2 targeted therapy;

[0123] The gastric cancer targeted therapy efficacy prediction model includes a temporal heterogeneity module and a spatial heterogeneity module;

[0124] The temporal heterogeneity module is used to model the longitudinal changes of the same lesion in temporal heterogeneity, and the spatial heterogeneity module is used to model the interaction relationships between multiple lesions in spatial heterogeneity.

[0125] Furthermore, the risk probability prediction module is also used for:

[0126] The CT image data to be identified is input into the convolutional feature extractor in the gastric cancer targeted therapy efficacy prediction model to obtain the image features output by the convolutional feature extractor;

[0127] The image features are input into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module.

[0128] The temporal heterogeneity features are input into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the spatial heterogeneity features.

[0129] The temporal heterogeneity features and the spatial heterogeneity features are input into the classifier in the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the classifier.

[0130] Furthermore, the risk probability prediction module is also used for:

[0131] The temporal heterogeneity module includes two MHA modules, which are used to integrate information from the same lesion at different time points in the CT image data to be identified based on the image features.

[0132] The offset in days of each time point of the CT image data to be identified relative to the baseline time is mapped onto a multidimensional feature space to obtain multidimensional mapping features;

[0133] The multidimensional mapping features and the image features are input into the first MHA module to obtain the output of the first MHA module;

[0134] The output of the first MHA module is input to the second MHA module to obtain the output of the second MHA module;

[0135] Based on the output of the first MHA module and the output of the second MHA module, the temporal heterogeneity characteristics of the output of the temporal heterogeneity module are determined.

[0136] Furthermore, the risk probability prediction module is also used for:

[0137] The temporal heterogeneity features are added to the integrated token to obtain the integrated features;

[0138] The integrated features are input into the third MHA module to obtain the output of the third MHA module;

[0139] The output of the third MHA module is input to the first feature normalization layer to obtain the output of the first feature normalization layer.

[0140] The integrated feature and the output of the feature normalization layer are input into the fully connected layer to obtain the output of the fully connected layer;

[0141] The output of the fully connected layer is input to the second feature normalization layer to obtain the output of the second feature normalization layer.

[0142] The spatial heterogeneity features are determined based on the output of the first feature normalization layer and the output of the second feature normalization layer.

[0143] Furthermore, the training data for the gastric cancer targeted therapy efficacy prediction model is divided into low-risk group data and high-risk data;

[0144] The low-risk group data and the high-risk data were obtained by dividing the patients' total survival time into one-year survival periods.

[0145] The gastric cancer targeted therapy efficacy prediction model is trained based on a survival loss function, which includes a cross-entropy loss function and a risk probability ratio loss function.

[0146] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown, the electronic device may include a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a method for predicting the efficacy of targeted therapy for gastric cancer. This method includes: acquiring CT image data to be identified before and after targeted therapy for gastric cancer; inputting the CT image data to be identified into a gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model; wherein the CT image data to be identified includes images corresponding to a baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified at each time point includes multiple lesions.

[0147] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the gastric cancer targeted therapy efficacy prediction method provided by the above methods. The method includes: acquiring CT image data to be identified before and after gastric cancer targeted therapy; inputting the CT image data to be identified into a gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model; wherein the CT image data to be identified includes images corresponding to the baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified corresponding to each time point includes multiple lesions.

[0149] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the gastric cancer targeted therapy efficacy prediction method provided by the above methods. The method includes: acquiring CT image data to be identified before and after gastric cancer targeted therapy; inputting the CT image data to be identified into a gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model; wherein the CT image data to be identified includes images corresponding to the baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified corresponding to each time point includes multiple lesions.

[0150] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0151] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the efficacy of targeted therapy for gastric cancer, characterized in that, include: Acquire CT image data of the target cancer before and after targeted therapy; The CT image data to be identified is input into the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model. The CT image data to be identified includes images corresponding to the baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified at each time point includes multiple lesions. The CT image data to be identified is input into the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model, including: The CT image data to be identified is input into the convolutional feature extractor in the gastric cancer targeted therapy efficacy prediction model to obtain the image features output by the convolutional feature extractor; The image features are input into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module. The temporal heterogeneity features are input into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain spatial heterogeneity features; The temporal heterogeneity features and the spatial heterogeneity features are input into the classifier in the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the classifier. The image features are input into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module, including: The temporal heterogeneity module includes two MHA modules, which are used to integrate information from the same lesion at different time points in the CT image data to be identified based on the image features. The offset in days of each time point of the CT image data to be identified relative to the baseline time is mapped onto a multidimensional feature space to obtain multidimensional mapping features; The multidimensional mapping features and the image features are input into the first MHA module to obtain the output of the first MHA module; The output of the first MHA module is input to the second MHA module to obtain the output of the second MHA module; Based on the output of the first MHA module and the output of the second MHA module, the temporal heterogeneity characteristics of the output of the temporal heterogeneity module are determined. The temporal heterogeneity features are input into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain spatial heterogeneity features, including: The temporal heterogeneity features are added to the integrated token to obtain the integrated features; The integrated features are input into the third MHA module to obtain the output of the third MHA module; The output of the third MHA module is input to the first feature normalization layer to obtain the output of the first feature normalization layer. The integrated feature and the output of the feature normalization layer are input into the fully connected layer to obtain the output of the fully connected layer; The output of the fully connected layer is input to the second feature normalization layer to obtain the output of the second feature normalization layer. The spatial heterogeneity features are determined based on the output of the first feature normalization layer and the output of the second feature normalization layer.

2. The method for predicting the efficacy of targeted therapy for gastric cancer according to claim 1, characterized in that, The gastric cancer targeted therapy efficacy prediction model was trained based on time-series imaging data from anti-HER2 targeted therapy for metastatic gastric cancer. The gastric cancer targeted therapy efficacy prediction model includes a temporal heterogeneity module and a spatial heterogeneity module; The temporal heterogeneity module is used to model the longitudinal changes of the same lesion in temporal heterogeneity, and the spatial heterogeneity module is used to model the interaction relationships between multiple lesions in spatial heterogeneity.

3. The method for predicting the efficacy of targeted therapy for gastric cancer according to claim 1, characterized in that, The training data for the gastric cancer targeted therapy efficacy prediction model is divided into low-risk group data and high-risk group data; The low-risk group data and the high-risk data were obtained by dividing the patients' total survival time into one-year survival periods. The gastric cancer targeted therapy efficacy prediction model is trained based on a survival loss function, which includes a cross-entropy loss function and a risk probability ratio loss function.

4. A device for predicting the efficacy of targeted therapy for gastric cancer, characterized in that, include: The data acquisition module is used to acquire CT image data to be identified before and after targeted therapy for gastric cancer. The risk probability prediction module is used to input the CT image data to be identified into the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model. The CT image data to be identified includes images corresponding to the baseline time and images corresponding to multiple follow-up time points, and the CT image data to be identified at each time point includes multiple lesions. Specifically, the CT image data to be identified is input into a gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the gastric cancer targeted therapy efficacy prediction model, including: The CT image data to be identified is input into the convolutional feature extractor in the gastric cancer targeted therapy efficacy prediction model to obtain the image features output by the convolutional feature extractor; The image features are input into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module. The temporal heterogeneity features are input into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain spatial heterogeneity features; The temporal heterogeneity features and the spatial heterogeneity features are input into the classifier in the gastric cancer targeted therapy efficacy prediction model to obtain the gastric cancer risk probability output by the classifier. The image features are input into the temporal heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain the temporal heterogeneity features output by the temporal heterogeneity module, including: The temporal heterogeneity module includes two MHA modules, which are used to integrate information from the same lesion at different time points in the CT image data to be identified based on the image features. The offset in days of each time point of the CT image data to be identified relative to the baseline time is mapped onto a multidimensional feature space to obtain multidimensional mapping features; The multidimensional mapping features and the image features are input into the first MHA module to obtain the output of the first MHA module; The output of the first MHA module is input to the second MHA module to obtain the output of the second MHA module; Based on the output of the first MHA module and the output of the second MHA module, the temporal heterogeneity characteristics of the output of the temporal heterogeneity module are determined. The temporal heterogeneity features are input into the spatial heterogeneity module of the gastric cancer targeted therapy efficacy prediction model to obtain spatial heterogeneity features, including: The temporal heterogeneity features are added to the integrated token to obtain the integrated features; The integrated features are input into the third MHA module to obtain the output of the third MHA module; The output of the third MHA module is input to the first feature normalization layer to obtain the output of the first feature normalization layer. The integrated feature and the output of the feature normalization layer are input into the fully connected layer to obtain the output of the fully connected layer; The output of the fully connected layer is input to the second feature normalization layer to obtain the output of the second feature normalization layer. The spatial heterogeneity features are determined based on the output of the first feature normalization layer and the output of the second feature normalization layer.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the gastric cancer targeted therapy efficacy prediction method as described in any one of claims 1 to 3.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the gastric cancer targeted therapy efficacy prediction method as described in any one of claims 1 to 3.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the gastric cancer targeted therapy efficacy prediction method as described in any one of claims 1 to 3.

Images