Nasopharynx cancer local recurrence risk prediction method, model construction method, device, equipment and medium
Through the methods of unsupervised clustering and multimodal feature fusion, the problems of heterogeneity and lack of multimodal information in the prediction of local recurrence risk of NPC were solved, and more accurate prediction was achieved, especially the local recurrence risk assessment of NPC through subregional segmentation and dual-gating model.
Patent Information
- Application Number
- CN202510808485.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-16
- Publication Date
- 2025-10-03
AI Technical Summary
Existing methods for predicting the risk of local recurrence of nasopharyngeal carcinoma ignore the influence of internal heterogeneity and peritumoral microenvironment of nasopharyngeal carcinoma and lack the integration of multimodal information, resulting in insufficient prediction accuracy.
An unsupervised linear iterative clustering algorithm was used for subregional segmentation. Deep learning and omics feature extraction were combined to extract global and subregional features from CT images and radiotherapy dose distribution maps. A double-gated hybrid expert model was used for prediction, integrating imaging and dose features and combining clinical information to assess the risk of local recurrence of nasopharyngeal carcinoma.
It improves the accuracy and comprehensiveness of local recurrence risk prediction for nasopharyngeal carcinoma, can quantify the internal heterogeneity of the tumor, reveal complex structural and functional differences, and identify high-risk areas that may recur due to insufficient dose.
Smart Images
Figure BDA0005454204920000031 
Figure BDA0005454204920000092 
Figure BDA0005454204920000093
Abstract
Description
Technical Field
[0001] The present application relates to the field of medical image processing technology, and in particular to a method for predicting the risk of local recurrence of nasopharyngeal carcinoma, a model construction method, an apparatus, equipment, and a medium. Background Art
[0002] As a malignant tumor of the head and neck with insidious onset and poor specificity of early symptoms, nasopharyngeal carcinoma is often diagnosed in the middle or late stages of the disease, which brings many difficulties and challenges to subsequent treatment. Since nasopharyngeal carcinoma is sensitive to radiotherapy and chemotherapy, and the anatomical structure of the nasopharynx is complex and adjacent to many important tissue structures, it is difficult to completely remove it surgically. Radiotherapy has become the primary radical cure for nasopharyngeal carcinoma. Early-stage diseases can usually achieve good results with radiotherapy alone, while for advanced diseases, a comprehensive treatment model of radiotherapy combined with chemotherapy, targeted therapy or immunotherapy is required. Even so, some patients will still experience local recurrence and / or distant metastasis after receiving radical treatment, and the overall prognosis is still not ideal, which seriously threatens the physical and mental health of patients.
[0003] Accurate assessment of nasopharyngeal carcinoma (NPC) recurrence risk is crucial for personalized treatment decisions and improved prognosis in high-risk individuals. Computed tomography (CT), as an important imaging modality, plays a multifaceted role in NPC radiotherapy, encompassing diagnosis, treatment planning, treatment implementation, and efficacy evaluation. In recent years, the rapid development of machine learning technologies has provided new insights into NPC recurrence risk prediction. Researchers rely on computers to automatically or semi-automatically optimize tumor CT image features and input them into classifier training, thereby constructing reasonable prediction models. Radiomics and deep learning, two mainstream algorithms, have become effective tools for NPC recurrence risk prediction. However, existing methods still have limitations. First, previous studies have focused on global lesion characteristics, ignoring the impact of internal NPC heterogeneity and the peritumoral microenvironment on its biological behavior and treatment response. Second, existing prediction models are mostly based on single-modality imaging data and lack the integration of multimodal information. Summary of the Invention
[0004] The purpose of this application is to provide a method for predicting the risk of local recurrence of nasopharyngeal carcinoma, a model construction method, an apparatus, equipment and a medium, which can improve the accuracy of predicting local recurrence of nasopharyngeal carcinoma.
[0005] To achieve the above objectives, this application provides the following solutions:
[0006] In a first aspect, the present application provides a method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma, comprising:
[0007] Collect sample data from multiple nasopharyngeal carcinoma patients, each including clinical information, CT images, and radiotherapy dose distribution maps;
[0008] For each CT image and radiotherapy dose distribution map, an unsupervised linear iterative clustering algorithm was used to perform subregion segmentation to obtain the subregion subgraphs of each CT image and radiotherapy dose distribution map;
[0009] Each sample data, as well as the sub-regional subgraph of the CT image and the sub-regional subgraph of the radiotherapy dose distribution map corresponding to each sample data, is used as input data, and the local recurrence results of nasopharyngeal carcinoma are used as label data to train a target model to obtain a local recurrence risk prediction model for nasopharyngeal carcinoma; the target model is used to extract global features from the CT image and the radiotherapy dose distribution map, obtain sub-regional features through the sub-regional subgraph of the CT image and the sub-regional subgraph of the radiotherapy dose distribution map, and predict local recurrence of nasopharyngeal carcinoma based on the global features and sub-regional features.
[0010] Optionally, the target model includes a deep learning feature extraction module, an omics feature extraction module and a multi-region analysis module, and the outputs of the deep learning feature extraction module and the omics feature extraction module are both connected to the input of the multi-region analysis module;
[0011] The deep learning feature extraction module is used to extract global depth features from CT images and radiotherapy dose distribution maps, and map the global depth features to sub-regional sub-maps of the CT images and sub-regional sub-maps of the radiotherapy dose distribution maps to obtain sub-regional depth features;
[0012] The omics feature extraction module is used to extract global omics features from CT images and radiotherapy dose distribution maps, wherein the global omics features include CT imaging features and dose characteristics, and both the CT imaging features and the dose characteristics include histogram features, texture features, shape features, and transformed image features; and the global omics features are mapped to subregional subgraphs of the CT images and subregional subgraphs of the radiotherapy dose distribution maps to obtain subregional omics features;
[0013] The multi-region analysis module is used to predict local recurrence of nasopharyngeal carcinoma based on the global features and sub-regional features.
[0014] Optionally, the multi-region analysis module includes a dual-gated hybrid expert model, a global branch and a sub-regional branch, the dual-gated hybrid expert model is used to input global features and sub-regional features, the global features include global depth features and global omics features, and the sub-regional features include sub-regional depth features and sub-regional omics features; the first output end of the dual-gated hybrid expert model is connected to the input of the global branch, and the second output end of the dual-gated hybrid expert model is connected to the input of the sub-regional branch; the global branch and the sub-regional branch also input the clinical information respectively, and the predicted probability output by the global branch and the predicted probability output by the sub-regional branch are weighted averaged, and the output of the target model is determined based on the weighted average result.
[0015] Optionally, the deep learning feature extraction module uses Swin-Transformer as the backbone network.
[0016] Optionally, the loss function used when training the target model is expressed as:
[0017]
[0018] in, is the total loss value, is the global branch loss, is the sub-region branch loss, w1 and w2 are loss weight factors; both the global branch loss and the sub-region branch loss are cross entropy losses, 2logσ global is the uncertainty of the global branch, σ global is the characteristic variance of the global branch output, 2logσ subregion is the uncertainty of the sub-regional branch, σ subregion is the feature variance of the sub-region branch output.
[0019] Optionally, the clinical information includes age, TNM stage, chemotherapy, targeted or immunotherapy, epidermal growth factor receptor expression, vascular endothelial growth factor expression and peripheral blood deoxyribonucleic acid expression.
[0020] In a second aspect, the present application provides a method for predicting the risk of local recurrence of nasopharyngeal carcinoma, the method comprising:
[0021] Acquiring data to be predicted, wherein the data to be predicted includes clinical information, CT images, and radiotherapy dose distribution maps;
[0022] An unsupervised linear iterative clustering algorithm is used to perform sub-region segmentation on the CT images and radiotherapy dose distribution maps in the target prediction data;
[0023] The data to be predicted and the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution maps corresponding to the data to be predicted are input into the nasopharyngeal carcinoma local recurrence risk prediction model described in any one of claims 1-6 to obtain a nasopharyngeal carcinoma local recurrence prediction result.
[0024] In a third aspect, the present application provides a device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma, wherein the device applies any of the above-described methods for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma, and the device comprises:
[0025] The data acquisition module is used to collect sample data from multiple nasopharyngeal carcinoma patients, each of which includes clinical information, CT images, and radiotherapy dose distribution maps;
[0026] The subregion segmentation module is used to perform subregion segmentation on each CT image and radiotherapy dose distribution map using an unsupervised linear iterative clustering algorithm to obtain subregion subgraphs of each CT image and radiotherapy dose distribution map;
[0027] The model training module is used to use each sample data, as well as the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution map corresponding to each sample data as input data, and the local recurrence results of nasopharyngeal carcinoma as label data to train the target model to obtain a local recurrence risk prediction model for nasopharyngeal carcinoma; the target model is used to extract global features from the CT images and the radiotherapy dose distribution map, obtain sub-regional features through the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution map, and predict local recurrence of nasopharyngeal carcinoma based on the global features and sub-regional features.
[0028] In a fourth aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma as described in any one of the above-mentioned methods.
[0029] In a fifth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma as described above.
[0030] According to the specific embodiments provided in this application, this application discloses the following technical effects:
[0031] The present application provides a method, model construction method, apparatus, device, and medium for predicting the risk of local recurrence of nasopharyngeal carcinoma. For each CT image and radiotherapy dose distribution map, an unsupervised simple linear iterative clustering (SLIC) algorithm is used to perform sub-region segmentation to obtain sub-regional subgraphs of each CT image and radiotherapy dose distribution map, thereby quantifying the internal heterogeneity of the tumor and revealing the complex structural and functional differences within the tumor, thereby providing richer information for predicting the risk of local recurrence of nasopharyngeal carcinoma and improving the accuracy of recurrence risk prediction. In addition, global features are extracted from the CT images and radiotherapy dose distribution maps, and sub-regional features are obtained through the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution maps. Local recurrence of nasopharyngeal carcinoma is predicted based on the global features and sub-regional features. The imaging features and dosimetric features are integrated with clinical information to more comprehensively assess the recurrence risk of nasopharyngeal carcinoma tumors and identify high-risk areas that may relapse due to insufficient dose, thereby further improving the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0033] Figure 1 A flowchart of a method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma provided in one embodiment of the present application.
[0034] Figure 2 This is a flowchart of the principle of constructing a local recurrence risk prediction model for nasopharyngeal carcinoma provided in one embodiment of the present application.
[0035] Figure 3 This is a schematic diagram of the structure of a local recurrence risk prediction model for nasopharyngeal carcinoma provided in one embodiment of the present application.
[0036] Figure 4 A schematic diagram of the functional modules of a device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma provided in one embodiment of the present application.
[0037] Figure 5 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.
[0039] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0040] In an exemplary embodiment, the present application provides a method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma, such as Figure 1 and Figure 2 As shown, the method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma includes steps 101 to 103.
[0041] Step 101: Collect sample data of multiple nasopharyngeal carcinoma patients, each sample data including clinical information, CT images and radiotherapy dose distribution maps.
[0042] Among them, the CT image is the CT image of nasopharyngeal carcinoma.
[0043] Step 102: For each CT image and radiotherapy dose distribution map, an unsupervised linear iterative clustering algorithm is used to perform sub-region segmentation to obtain a sub-region sub-graph of each CT image and radiotherapy dose distribution map.
[0044] Step 103: Using each sample data, as well as the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution maps corresponding to each sample data as input data, and the local recurrence results of nasopharyngeal carcinoma as label data to train a target model, thereby obtaining a local recurrence risk prediction model for nasopharyngeal carcinoma; the target model is used to extract global features from the CT images and the radiotherapy dose distribution maps, obtain sub-regional features through the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution maps, and predict local recurrence of nasopharyngeal carcinoma based on the global features and the sub-regional features.
[0045] The results of local recurrence of nasopharyngeal carcinoma were recurrence or no recurrence.
[0046] In an exemplary embodiment, step 101 specifically includes acquiring nasopharyngeal carcinoma CT images and radiation therapy plan dose distribution maps (radiotherapy dose distribution maps), and establishing a multimodal dataset after preprocessing. The sample data in the multimodal dataset includes clinical information, CT images, and radiation dose distribution maps. The clinical information includes collecting and screening clinical features that are significantly associated with local recurrence of nasopharyngeal carcinoma.
[0047] In an exemplary embodiment, the construction of the multimodal dataset specifically includes:
[0048] ① Determine the location and area of the primary nasopharyngeal carcinoma lesion for each enrolled patient, outline the three-dimensional contour of the lesion, and expand it by 3 mm.
[0049] ② A region of interest (VOI-CT) was captured from the lesion center and mapped onto the radiotherapy dose distribution map to form a region of interest (VOI-dose). The VOI was scaled to 128 × 128 × 32 using bilinear interpolation.
[0050] ③ The VOI-CT grayscale was intercepted in the range of (-1024HU, 2048HU) and normalized to (0, 1), and the VOI-dose dose distribution was adjusted to the interval of (0, 1).
[0051] Each sample in the multimodal dataset contains CT images of NPC patients, radiation therapy dose distribution maps, local recurrence labels of the primary lesion after radiotherapy, and clinical information. The diagnostic criteria for local recurrence of NPC is defined as the recurrence of the nasopharyngeal lesion confirmed by pathology six months after the end of radiotherapy. Patients are considered to be in the recurrence group if the lesion reappears in the nasopharynx; otherwise, they are considered to be in the non-recurrence group.
[0052] Normally distributed continuous variables were expressed as mean ± standard deviation (x ± s), and independent sample t-tests were used for intergroup comparisons. Skewed continuous variables were expressed as median (interquartile range), and Mann-Whitney U tests were performed. Categorical variables were presented as case numbers (n) and percentages (%), and intergroup comparisons were performed using chi-square tests or Fisher's exact test. A P < 0.05 in univariate analysis was considered a significant association between the clinical indicator and local recurrence of NPC. Continuous variables were standardized, and discrete variables were one-hot encoded.
[0053] The clinical information includes age, tumor node metastasis (TNM) stage, chemotherapy, targeted or immunotherapy, epidermal growth factor receptor expression (Ati-EGFR), vascular endothelial growth factor expression (VGEF) and peripheral blood deoxyribonucleic acid expression (EBV-DNA) and other clinical characteristics.
[0054] The clinical information is a univariate analysis of clinical information related to each nasopharyngeal carcinoma to screen out features that are significantly correlated with local recurrence of nasopharyngeal carcinoma.
[0055] The clinical information is input into the target model specifically including: performing continuous variable standardization and discrete variable one-hot encoding processing on each clinical feature in the clinical information through a clinical parameter encoder to obtain each encoded clinical feature, and inputting each encoded clinical feature into the target model.
[0056] In an exemplary embodiment, step 102 specifically includes: performing multi-scale supervoxel sub-region segmentation on the CT image-dose map using an unsupervised linear iterative clustering algorithm. The dose map is a radiation therapy plan dose distribution map.
[0057] An unsupervised linear iterative clustering algorithm is used to perform multi-scale supervoxel sub-region segmentation on CT images and radiotherapy dose distribution maps. The process is as follows:
[0058] ① Initialize the cluster centers. Evenly distribute k initial seed points in the CT image. In this example, k = 32.
[0059] ② Define the search area. Let the square area centered on each cluster center be the search area, with a side length of 2S. Where N represents the number of pixels.
[0060] ③Supervoxel generation. Calculate the comprehensive distance between each pixel in the search area and the cluster center to achieve pixel clustering. The formula is as follows:
[0061] D=d c +d s / S;
[0062]
[0063] Among them, d c is the grayscale distance, d s is the spatial distance; I represents the pixel gray value, I c Represents the grayscale value of the cluster center; (x, y) is the pixel coordinate, (x c ,y c ) are the cluster center coordinates.
[0064] ④ Update the cluster center. Update the grayscale value and spatial position of the new seed point. The formula is as follows:
[0065]
[0066] Among them, I i is the gray value of the i-th pixel, C represents the search area, N c is the number of pixels in the search area, (x i ,y i ) is the pixel coordinate of the i-th pixel, x i is the X-axis coordinate, y i is the Y-axis coordinate.
[0067] ⑤Iterative optimization to ensure the stability and accuracy of clustering results.
[0068] ⑥For each cluster, calculate the minimum and maximum coordinate values of all cluster points in the area and generate a candidate box.
[0069] In an exemplary embodiment, step 103 specifically includes: extracting CT imaging omics features and dose omics features based on the entire region and the sub-regions, respectively.
[0070] The preprocessed CT image-dose map is input into the target model, and the obtained global depth feature is mapped to the supervoxel segmentation area to obtain the sub-region depth feature.
[0071] The target model was trained jointly by CT image-dose map and supervoxel subregion submap, and combined with clinical characteristics to obtain a local recurrence risk prediction model for nasopharyngeal carcinoma.
[0072] Among them, the nasopharyngeal carcinoma CT images and radiotherapy dose distribution maps in the sample data are obtained by preprocessing the collected data. The preprocessing specifically includes target area delineation, target area segmentation, intensity normalization and dose normalization operations.
[0073] The target model includes a deep learning feature extraction module, an omics feature extraction module and a multi-region analysis module, and the outputs of the deep learning feature extraction module and the omics feature extraction module are both connected to the input of the multi-region analysis module.
[0074] The deep learning feature extraction module is used to extract global depth features from CT images and radiotherapy dose distribution maps, and map the global depth features to sub-regional sub-maps of the CT images and sub-regional sub-maps of the radiotherapy dose distribution maps to obtain sub-regional depth features.
[0075] The omics feature extraction module is used to extract global omics features from CT images and radiotherapy dose distribution maps. The global omics features include CT imaging features and dose features. Both CT imaging features and dose features include histogram features, texture features, shape features and transformation image features. The global omics features are mapped to sub-regional subgraphs of CT images and sub-regional subgraphs of radiotherapy dose distribution maps to obtain sub-regional omics features.
[0076] The multi-region analysis module is used to predict local recurrence of nasopharyngeal carcinoma based on the global features and sub-regional features.
[0077] The multi-region analysis module includes a dual-gate mixture of experts model (Multi-gate Mixture of Experts, MMoE), a global branch and a sub-regional branch. The dual-gate mixture of experts model is used to input global features and sub-regional features. The global features include global depth features and global omics features, and the sub-regional features include sub-regional depth features and sub-regional omics features. The first output end of the dual-gate mixture of experts model is connected to the input of the global branch, and the second output end of the dual-gate mixture of experts model is connected to the input of the sub-regional branch. The global branch and the sub-regional branch also input the clinical information respectively, and the predicted probability output by the global branch and the predicted probability output by the sub-regional branch are weighted averaged. The output of the target model is determined based on the weighted average result to obtain a local recurrence prediction result of nasopharyngeal carcinoma, and the local recurrence prediction result of nasopharyngeal carcinoma is recurrence or no recurrence.
[0078] More specifically, if the weighted average result is greater than the set probability value, the prediction result of local recurrence of nasopharyngeal carcinoma is recurrence, otherwise it is no recurrence.
[0079] The global branch and subregional branch perform the local recurrence prediction task of the entire NPC tumor and the local recurrence prediction task of the subregion in parallel.
[0080] The sub-regional deep features were resized to a uniform size using the region of interest alignment (ROI-Align) method. CT image-dose deep features and omics features were fused based on the tumor global and sub-regions, respectively, and used as inputs for the global and sub-regional branches of the target model.
[0081] The encoded clinical features are input into the target model. Specifically, the encoded clinical features are fused with the features output by the double-gated hybrid expert model before the fully connected layer (FC) of the global branch and sub-regional branch of the target model, and the predicted probability is output through the decision layer of each branch (global branch and sub-regional branch).
[0082] Extracting CT imaging features and dosimetric features from CT images and radiotherapy dose distribution maps, specifically including: extracting CT imaging features and dosimetric features from CT images and radiotherapy dose distribution maps using a quantitative radiomics analysis method.
[0083] More specifically, the omics feature extraction module uses the Python third-party library pyRadiomics to implement quantitative radiomics analysis, extracting CT imaging features and dose features from VOI-CT and VOI-dose based on the entire domain and sub-region, respectively, including histogram features, texture features, shape features, and transformation image features.
[0084] like Figure 3As shown in the figure, the deep learning feature extraction module uses the Swin-Transformer as the backbone network to extract global deep features. The backbone network consists of three layers: Stage I, Stage II, and Stage III. The three layers are composed of two, two, and six Swin-Transformer units connected in series, respectively. The global deep features are mapped to the supervoxel segmented regions, and the feature size is adjusted using ROI-Align to obtain sub-regional deep features.
[0085] The ROI-Align algorithm process is as follows:
[0086] ① Sub-region mapping: Based on the size relationship between the original image and the global feature, all sub-region coordinates are mapped to the global feature map coordinate system.
[0087] ② Sampling point positioning. This example unifies the sub-region depth features to 7×7, thus subdividing each sub-region into 49 sub-regions. Four evenly distributed sampling points are positioned in each sub-region, and the coordinate values are not rounded up.
[0088] ③ Eigenvalue calculation: Use bilinear interpolation to obtain the eigenvalues of the sampling points in each sub-region, and perform a maximum pooling operation to obtain the final eigenvalue of the sub-region.
[0089] In the multi-region analysis module, MMoE consists of two specific expert subnets (E1, E2) and a general expert subnet E c Composition, and design of dual gate unit (G A , G B ) flexibly adjusts the contribution of each expert subnetwork, then connects the global branch and the subregional branch, and performs the global local recurrence prediction task and the subregional local recurrence prediction task in parallel. Specifically, the specific expert subnetwork E1 is used to learn global features, the specific expert subnetwork E2 is used to learn subregional features, and the general expert subnetwork E c Used to learn global features and sub-region features simultaneously; the output of the specific expert subnet E1 is combined with the general expert subnet E c The output of the gate control unit G B The first addition result is obtained by weighted average, and the first addition result is integrated with clinical information and input into the global branch; the output of the specific expert subnet E2 is combined with the general expert subnet E c The output of the gate control unit G A The second summation result is obtained by weighted averaging, and the second summation result is integrated with clinical information and input into the sub-regional branch. Both the global branch and the sub-regional branch are composed of fully connected layers.
[0090] After the clinical parameter encoder generates clinical features, they are concatenated with the output features of each branch before the fully connected layers of the global and subregional branches. The predicted probabilities are then output through the decision layers of each branch. The weighted average of the predicted probabilities is then taken to determine the local recurrence of NPC.
[0091] When training the target model, the classification errors of the global branch and the sub-regional branch are calculated based on the cross entropy loss function. The loss function is expressed as:
[0092]
[0093] in, is the total loss value, is the global branch loss, is the sub-region branch loss, w1 and w2 are loss weight factors; both the global branch loss and the sub-region branch loss are cross entropy losses, 2logσ global is the uncertainty of the global branch (logarithmic variance), σ global is the characteristic variance of the global branch output, 2logσ subregion is the uncertainty of the sub-regional branch (logarithmic variance), σ subregion is the feature variance of the sub-region branch output.
[0094] To improve the convergence rate of the joint loss function, this example uses the Adam optimizer. During training, the number of training epochs (epochs) is set to 150, the batch size (batch_size) is set to 16, and the initial learning rate (lr) is set to 2e-3. Furthermore, to further optimize the training results, the learning rate is automatically reduced by 10% after every 10 training epochs. The dataset is partitioned using a ten-fold cross-validation method, with 10% of the patient data randomly selected as the validation set and the remaining data used for training. During training, if the model's loss value does not show a downward trend after 150 epochs or 40 fine-tuning cycles, training is terminated immediately, and the weight parameters corresponding to the lowest loss value are saved as the optimal parameters for the nasopharyngeal carcinoma local recurrence risk prediction model.
[0095] This application fully utilizes the correlation between internal tumor heterogeneity and global characteristics to improve the accuracy and efficiency of local recurrence risk prediction of nasopharyngeal carcinoma.
[0096] In an exemplary embodiment, after the construction and training of the nasopharyngeal carcinoma local recurrence risk prediction model is completed, the collected nasopharyngeal carcinoma CT images and dose distribution data are used to accurately predict the risk of local recurrence of the primary lesion, and analyze the correlation between the local recurrence area and the radiotherapy dose distribution.
[0097] A method for predicting the risk of local recurrence of nasopharyngeal carcinoma includes steps 201 to 203.
[0098] Step 201: Acquire data to be predicted, wherein the data to be predicted includes clinical information, CT images, and radiotherapy dose distribution maps.
[0099] Step 202: Use an unsupervised linear iterative clustering algorithm to perform sub-region segmentation on the CT image and radiotherapy dose distribution map in the data to be predicted.
[0100] Step 203: Input the data to be predicted and the sub-regional sub-graphs of the CT image and the sub-regional sub-graphs of the radiotherapy dose distribution map corresponding to the data to be predicted into a nasopharyngeal carcinoma local recurrence risk prediction model to obtain a nasopharyngeal carcinoma local recurrence prediction result.
[0101] Compared with the prior art, this application has the following beneficial effects:
[0102] 1. This application uses an unsupervised clustering method to divide nasopharyngeal carcinoma tumors into multiple subregions, thereby quantifying the internal heterogeneity of the tumor. These subregions not only show differences in imaging but may also be associated with different biological behaviors, such as cell proliferation, invasion, and sensitivity to treatment. By deeply analyzing the characteristics of these subregions, the model can reveal the complex structural and functional differences within the tumor, thereby providing richer information for predicting the risk of local recurrence of nasopharyngeal carcinoma.
[0103] 2. Considering that imaging features can indicate the biological properties of tumors, while dosimetric features accurately reflect the physical properties of radiotherapy doses, this application fully integrates the two. The proposed multimodal omics prediction method can more comprehensively assess the recurrence risk of nasopharyngeal carcinoma tumors and identify high-risk areas that may relapse due to insufficient dose.
[0104] 3. This application uses a dual-gated hybrid expert model with global and subregional branches to implement multi-task learning for local recurrence risk prediction in nasopharyngeal carcinoma. It utilizes a variance uncertainty optimization model to dynamically adjust the weights of global and subregional tasks. This results in higher prediction accuracy and robustness compared to separately trained models.
[0105] Based on the same inventive concept, the embodiments of the present application also provide a device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma, which is used to implement the aforementioned method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma provided below can be found in the above-mentioned limitations of the method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma, and will not be repeated here.
[0106] In an exemplary embodiment, Figure 4As shown, a device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma is provided. The device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma applies the above-mentioned method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma. The device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma comprises:
[0107] The data acquisition module is used to collect sample data of multiple nasopharyngeal carcinoma patients, each sample data including clinical information, CT images and radiotherapy dose distribution map.
[0108] The subregion segmentation module is used to perform subregion segmentation on each CT image and radiotherapy dose distribution map using an unsupervised linear iterative clustering algorithm to obtain a subregion subgraph of each CT image and radiotherapy dose distribution map.
[0109] The model training module is used to use each sample data, as well as the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution map corresponding to each sample data as input data, and the local recurrence results of nasopharyngeal carcinoma as label data to train the target model to obtain a local recurrence risk prediction model for nasopharyngeal carcinoma; the target model is used to extract global features from the CT images and the radiotherapy dose distribution map, obtain sub-regional features through the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution map, and predict local recurrence of nasopharyngeal carcinoma based on the global features and sub-regional features.
[0110] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma is implemented.
[0111] Those skilled in the art will understand that Figure 5The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application and does not constitute a limitation on the computer device to which the solution of the present application is applied. A specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the above-mentioned method embodiments when executing the computer program.
[0112] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.
[0113] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0114] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.
[0115] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0116] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, data processing logic of programmable logic devices, and the like.
[0117] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0118] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma, characterized in that: The method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma comprises: Collect sample data from multiple nasopharyngeal carcinoma patients, each including clinical information, CT images, and radiotherapy dose distribution maps; For each CT image and radiotherapy dose distribution map, an unsupervised linear iterative clustering algorithm was used to perform subregion segmentation to obtain the subregion subgraphs of each CT image and radiotherapy dose distribution map; Each sample data, as well as the sub-regional subgraph of the CT image and the sub-regional subgraph of the radiotherapy dose distribution map corresponding to each sample data, is used as input data, and the local recurrence results of nasopharyngeal carcinoma are used as label data to train a target model to obtain a local recurrence risk prediction model for nasopharyngeal carcinoma; the target model is used to extract global features from the CT image and the radiotherapy dose distribution map, obtain sub-regional features through the sub-regional subgraph of the CT image and the sub-regional subgraph of the radiotherapy dose distribution map, and predict local recurrence of nasopharyngeal carcinoma based on the global features and sub-regional features.
2. The method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma according to claim 1, characterized in that: The target model includes a deep learning feature extraction module, an omics feature extraction module and a multi-region analysis module, and the outputs of the deep learning feature extraction module and the omics feature extraction module are both connected to the input of the multi-region analysis module; The deep learning feature extraction module is used to extract global depth features from CT images and radiotherapy dose distribution maps, and map the global depth features to sub-regional sub-maps of the CT images and sub-regional sub-maps of the radiotherapy dose distribution maps to obtain sub-regional depth features; The omics feature extraction module is used to extract global omics features from CT images and radiotherapy dose distribution maps, wherein the global omics features include CT imaging features and dose characteristics, and both the CT imaging features and the dose characteristics include histogram features, texture features, shape features, and transformed image features; and the global omics features are mapped to subregional subgraphs of the CT images and subregional subgraphs of the radiotherapy dose distribution maps to obtain subregional omics features; The multi-region analysis module is used to predict local recurrence of nasopharyngeal carcinoma based on the global features and sub-regional features.
3. The method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma according to claim 2, characterized in that: The multi-region analysis module includes a double-gated hybrid expert model, a global branch and a sub-region branch, the double-gated hybrid expert model is used to input global features and sub-region features, the global features include global depth features and global omics features, and the sub-region features include sub-region depth features and sub-region omics features; The first output end of the dual-gated hybrid expert model is connected to the input of the global branch, and the second output end of the dual-gated hybrid expert model is connected to the input of the sub-regional branch; the global branch and the sub-regional branch also input the clinical information respectively, and the predicted probability output by the global branch and the predicted probability output by the sub-regional branch are weighted averaged, and the output of the target model is determined based on the weighted average result.
4. The method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma according to claim 2, characterized in that: The deep learning feature extraction module uses Swin-Transformer as the backbone network.
5. The method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma according to claim 3, characterized in that: The loss function used when training the target model is expressed as: in, is the total loss value, is the global branch loss, is the sub-region branch loss, w1 and w2 are loss weight factors; both the global branch loss and the sub-region branch loss are cross entropy losses, 2logσ global is the uncertainty of the global branch, σ global is the characteristic variance of the global branch output, 2logσ subregion is the uncertainty of the sub-regional branch, σ subregion is the feature variance of the sub-region branch output.
6. The method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma according to claim 1, characterized in that: The clinical information includes age, TNM stage, chemotherapy, targeted or immunotherapy, epidermal growth factor receptor expression, vascular endothelial growth factor expression and peripheral blood deoxyribonucleic acid expression.
7. A method for predicting the risk of local recurrence of nasopharyngeal carcinoma, characterized in that: The method for predicting the risk of local recurrence of nasopharyngeal carcinoma comprises: Acquiring data to be predicted, wherein the data to be predicted includes clinical information, CT images, and radiotherapy dose distribution maps; An unsupervised linear iterative clustering algorithm is used to perform sub-region segmentation on the CT images and radiotherapy dose distribution maps in the target prediction data; The data to be predicted and the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution maps corresponding to the data to be predicted are input into the nasopharyngeal carcinoma local recurrence risk prediction model described in any one of claims 1-6 to obtain a nasopharyngeal carcinoma local recurrence prediction result.
8. A device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma, characterized in that: The device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma applies the method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma according to any one of claims 1 to 6, and the device for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma comprises: The data acquisition module is used to collect sample data from multiple nasopharyngeal carcinoma patients, each of which includes clinical information, CT images, and radiotherapy dose distribution maps; The subregion segmentation module is used to perform subregion segmentation on each CT image and radiotherapy dose distribution map using an unsupervised linear iterative clustering algorithm to obtain subregion subgraphs of each CT image and radiotherapy dose distribution map; The model training module is used to use each sample data, as well as the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution map corresponding to each sample data as input data, and the local recurrence results of nasopharyngeal carcinoma as label data to train the target model to obtain a local recurrence risk prediction model for nasopharyngeal carcinoma; the target model is used to extract global features from the CT images and the radiotherapy dose distribution map, obtain sub-regional features through the sub-regional subgraphs of the CT images and the sub-regional subgraphs of the radiotherapy dose distribution map, and predict local recurrence of nasopharyngeal carcinoma based on the global features and sub-regional features.
9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma according to any one of claims 1 to 6.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for constructing a local recurrence risk prediction model for nasopharyngeal carcinoma according to any one of claims 1 to 6 is implemented.