Kd-net-based tumor radiotherapy response prediction method and model

Abstract

The embodiment of the application provides a kind of based on Kd-Net's tumor radiotherapy response prediction method and model, the method includes the following steps: extracting tumor feature fusion forms point cloud, according to SUV variation gradient, point cloud is classified and labeled, and is divided into training set, verification set and test set;Based on the network architecture of Kd-Net, construct deep learning model;Based on training set and verification set, the deep learning model is trained, and model parameter is adjusted;Based on the optimized model parameters and test set, the performance of the optimized deep learning model is evaluated, and the trained deep learning model is obtained;The tumor data set to be predicted is input into the trained deep learning model, and the tumor radiotherapy response is predicted.The prediction method and model provided in the application can predict whether the tumor responds after radiotherapy, thereby intelligently assisting the doctor to adjust the radiotherapy dose, implementing adaptive precision radiotherapy decision for the patient, reducing the harm of radiotherapy to the patient, and improving the work efficiency and diagnosis and treatment quality of the doctor.

Description

Technical Field

[0001] This application relates to the field of medical deep learning technology, and in particular to a method and model for predicting tumor radiotherapy response based on Kd-Net. Background Technology

[0002] Radiotherapy is one of the three major treatment methods for malignant tumors. It is a localized treatment that can be used to cure or alleviate the primary tumor or metastatic lesions in a given area. The side effects of radiotherapy vary from person to person and are related to the radiation dose, the treatment site, and the patient's individual health condition. The most common radiotherapy side effects include skin damage in the irradiated area, nausea and vomiting, fever, decreased peripheral blood cell count, and fatigue. In addition, depending on the irradiated site, it can also cause dry and painful mouth and throat, radiation pneumonitis, radiation esophagitis, and hair loss. Both high and low radiation doses can have a significant impact on patients. However, currently, the setting of radiation doses mainly relies on physician experience, and individual differences and accidental factors have a significant impact, making it impossible to guarantee scientific accuracy, rationality, and effectiveness. Therefore, the scientific setting of radiation doses for the tumor target area is of great significance for achieving precision radiotherapy. Summary of the Invention

[0003] This application provides a method and model for predicting tumor radiotherapy response based on Kd-Net. It can predict whether there will be a response after tumor radiotherapy, thereby intelligently assisting doctors in adjusting the radiotherapy dose, implementing adaptive and precise radiotherapy decisions for patients, reducing the harm of radiotherapy to patients, and improving doctors' work efficiency and treatment quality.

[0004] To address the aforementioned technical problems, in a first aspect, embodiments of this application provide a method for predicting tumor radiotherapy response based on Kd-Net. This method includes the following steps: First, tumor features are extracted and fused to form a point cloud. The point cloud is then classified and labeled according to the SUV variation gradient, and divided into training, validation, and test sets. Next, a deep learning model is constructed based on the Kd-Net network architecture. Then, the deep learning model is trained using the training and validation sets, and the model parameters are adjusted to obtain an optimized deep learning model and optimized model parameters. Finally, the optimized deep learning model is evaluated based on the optimized model parameters and the test set to obtain a trained deep learning model. Finally, the tumor dataset to be predicted is input into the trained deep learning model to predict the tumor radiotherapy response.

[0005] In some exemplary embodiments, extracting tumor features and fusing them to form a point cloud includes: extracting features related to the mid-radiotherapy voxel SUV value from the tumor features, fusing the features to form a three-dimensional point cloud; the features include voxel location, pre-radiotherapy voxel SUV value, and radiotherapy dose.

[0006] In some exemplary embodiments, classifying and labeling the point cloud according to the SUV change gradient includes: classifying each voxel according to the SUV change gradient to obtain two major categories of voxels; dividing the two major categories of voxels into 14 subcategories of voxels according to the RatioSUV gradient; segmenting the point cloud to complete the preparation of the dataset; wherein, classifying each voxel to obtain two major categories of voxels includes: determining whether the SUV value of each voxel before radiotherapy and the SUV value during radiotherapy satisfy a correspondence; if yes, the voxel is classified into the "responsive" category and labeled "1"; if no, the voxel is classified into the "non-responsive" category and labeled "0"; the correspondence is as follows:

[0007]

[0008] RatioSUV≤0.7

[0009] Among them, PreSUV represents the SUV value of a single voxel before radiotherapy, which reflects the degree of tumor lesion through the standard uptake value of voxels; MidSUV represents the SUV value during the mid-term of single voxel radiotherapy; RatioSUV represents the ratio of the mid-term SUV value of a single voxel radiotherapy to the pre-radiotherapy SUV value.

[0010] In some exemplary embodiments, based on the RatioSUV gradient, two major categories of voxels are divided into 14 minor categories of voxels. This includes: for voxels classified as "responsive," their RatioSUV values ​​satisfy 0 < RatioSUV ≤ 0.7. They are then divided into seven gradients based on their RatioSUV values: 0 < RatioSUV ≤ 0.1, 0.1 < RatioSUV ≤ 0.2, 0.2 < RatioSUV ≤ 0.3, 0.3 < RatioSUV ≤ 0.4, 0.4 < RatioSUV ≤ 0.5, 0.5 < RatioSUV ≤ 0.6, and 0.6 < RatioSUV ≤ 0.7. Each voxel is then assigned to its corresponding RatioSUV gradient based on its RatioSUV value, i.e., it is classified as "responsive." Voxels belonging to the "Non-responsive" category are further divided into 7 subcategories, but their labels remain unchanged. For voxels belonging to the "Non-responsive" category, their RatioSUV values ​​satisfy RatioSUV > 0.7. Based on the RatioSUV values, they are divided into seven gradients: 0.7 < RatioSUV ≤ 0.8, 0.8 < RatioSUV ≤ 0.9, 0.9 < RatioSUV ≤ 1.0, 1.0 < RatioSUV ≤ 1.1, 1.1 < RatioSUV ≤ 1.2, 1.2 < RatioSUV ≤ 1.3, and 1.3 < RatioSUV. Based on the RatioSUV values ​​of each voxel, they are assigned to the corresponding RatioSUV gradients. That is, voxels belonging to the "Non-responsive" category are further divided into 7 subcategories, but their labels remain unchanged.

[0011] In some exemplary embodiments, the point cloud is segmented to prepare the dataset, including: for each point cloud segmentation, each point cloud is divided into 14 smaller point clouds according to the category of each voxel in the point cloud, and the smaller point clouds include empty point clouds; for the 14 smaller point clouds segmentation, according to the number of voxels in the point cloud and the characteristics of the model input, the empty point clouds and the minimal point clouds are removed, and the remaining smaller point clouds are divided into multiple point clouds of fixed size according to a fixed number of voxels, thereby completing the preparation of the dataset.

[0012] In some exemplary embodiments, after classifying and labeling the point clouds, and before dividing the training set, validation set, and test set, the method further includes: dividing each patient into three groups based on the proportion of point clouds with the tumor segmentation labeled "1" to the total number of segmented point clouds. The three groups are: low radiotherapy response, intermediate radiotherapy response, and high radiotherapy response. Specifically, in the low radiotherapy response group, the proportion of point clouds with the tumor segmentation labeled "1" to the total number of segmented point clouds is 0 to 0.5; in the intermediate radiotherapy response group, the proportion is 0.5 to 0.8; and in the high radiotherapy response group, the proportion is 0.8 to 1.

[0013] In some exemplary embodiments, constructing a deep learning model based on the Kd-Net network architecture includes: first, constructing the main Kd-Net network architecture; then, constructing the pre-architecture of the Kd-Net network; wherein the main Kd-Net network architecture includes multiple Kd convolutional layers and a fully connected layer; the main Kd-Net network architecture obtains output values ​​through forward propagation, converts the output values ​​into probability values, calculates loss values ​​and related prediction evaluation metrics; finally, based on the pre-architecture of the Kd-Net network, the training set and test set are converted into structured Kd-tree data, and the Kd-tree data is input into the main Kd-Net network architecture for training, adjusting model parameters to achieve model optimization, and obtaining the deep learning model based on the lowest loss value obtained from the input validation set; before inputting the Kd-tree data into the main Kd-Net network architecture for training, the method further includes: converting the Kd-tree data into tensor variables, and transposing the tensor variables as input to the main Kd-Net network architecture.

[0014] In some exemplary embodiments, inputting the tumor dataset to be predicted into the trained deep learning model includes: preprocessing the tumor dataset to be predicted to obtain a preprocessed tumor point cloud dataset; and inputting the preprocessed tumor point cloud dataset into the trained deep learning model to predict the tumor radiotherapy response.

[0015] In some exemplary embodiments, the tumor dataset to be predicted is preprocessed, including: performing feature extraction and data segmentation preprocessing operations on the tumor dataset to be predicted, so that the dimension of the preprocessed tumor point cloud dataset to be predicted is the same as the dimension of the tumor point cloud datasets in the training set, validation set and test set.

[0016] Secondly, this application also provides a tumor radiotherapy response prediction model based on Kd-Net, including a data preprocessing module, a model building module, a model training module, a model evaluation module, and a prediction module connected in sequence. The data preprocessing module is used to extract tumor features and fuse them to form a point cloud, classify and label the point cloud according to the SUV change gradient, and divide it into a training set, a validation set, and a test set. The model building module is used to build a deep learning model according to the Kd-Net network architecture. The model training module is used to train the deep learning model according to the training set and the validation set, adjust the model parameters, and obtain an optimized deep learning model and optimized model parameters. The model evaluation module is used to evaluate the performance of the optimized deep learning model according to the optimized model parameters and the test set to obtain a trained deep learning model. The prediction module is used to input the tumor dataset to be predicted into the trained deep learning model to predict the tumor radiotherapy response.

[0017] The technical solution provided in this application has at least the following advantages:

[0018] This application provides a method and model for predicting tumor radiotherapy response based on Kd-Net. The method includes the following steps: First, tumor features are extracted and fused to form a point cloud. The point cloud is then classified and labeled according to the SUV variation gradient, and divided into training, validation, and test sets. Next, a deep learning model is constructed based on the Kd-Net network architecture. Then, the deep learning model is trained using the training and validation sets, and the model parameters are adjusted to obtain an optimized deep learning model and optimized model parameters. Finally, the optimized deep learning model is evaluated based on the optimized model parameters and the test set to obtain a trained deep learning model. Finally, the tumor dataset to be predicted is input into the trained deep learning model to predict the tumor radiotherapy response.

[0019] This application proposes a method and model for predicting tumor radiotherapy response based on Kd-Net. First, by constructing a Kd-Net-based network architecture and introducing a Kd-tree structure, unstructured point clouds are converted into structured Kd-tree data. This allows the structured data to be input into the neural network, enabling feature engineering to process unstructured point clouds. Second, this application uses point cloud datasets instead of slicing 3D images, avoiding the losses caused by image slicing. Furthermore, point clouds possess arrangement invariance and transformation invariance, meeting the needs of multimodal tumor imaging and the requirement for preprocessing. Finally, this application uses evaluation metrics to compare with traditional network models, verifying the effectiveness, accuracy, and adjustability of the constructed tumor radiotherapy response prediction model. The model provided by this application can predict whether a response will occur after tumor radiotherapy, thereby intelligently assisting doctors in adjusting radiotherapy doses, implementing adaptive and precise radiotherapy decisions for patients, reducing the harm of radiotherapy to patients, and improving doctors' work efficiency and treatment quality. Attached Figure Description

[0020] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0021] Figure 1 A flowchart illustrating a method for predicting tumor radiotherapy response based on Kd-Net, provided in an embodiment of this application;

[0022] Figure 2 A schematic diagram of the data preprocessing flow for a tumor radiotherapy response prediction method based on Kd-Net, provided in an embodiment of this application;

[0023] Figure 3 This is a network architecture diagram of Kd-Net provided in one embodiment of this application;

[0024] Figure 4 A schematic diagram of the structure of a tumor radiotherapy response prediction model based on Kd-Net provided in an embodiment of this application;

[0025] Figures 5A to 5D This is a schematic diagram illustrating the evaluation index results of a Kd-Net-based tumor radiotherapy response prediction model for different patient groups, as provided in an embodiment of this application. Detailed Implementation

[0026] As the background technology indicates, the side effects of radiotherapy vary from person to person and are related to the radiation dose, treatment site, and the patient's individual health condition. Currently, the setting of radiotherapy doses relies primarily on physician experience, and individual differences and accidental factors have a significant impact, making it impossible to guarantee scientific rigor, rationality, and effectiveness. The scientific setting of radiotherapy doses for tumor target areas is of great significance for achieving precision radiotherapy.

[0027] Currently, the setting of radiotherapy doses in tumor radiotherapy mainly relies on physician experience, and individual differences and accidental factors have a significant impact, making it difficult to guarantee scientific accuracy, rationality, and effectiveness. To address the shortcomings of existing technologies, this application provides a method and model for predicting tumor radiotherapy response based on Kd-Net. The method includes the following steps: First, tumor features are extracted and fused to form a point cloud. The point cloud is then classified and labeled according to the SUV variation gradient, and divided into training, validation, and test sets. Next, a deep learning model is constructed based on the Kd-Net network architecture. Then, the deep learning model is trained using the training and validation sets, and the model parameters are adjusted to obtain an optimized deep learning model and optimized model parameters. Next, the performance of the optimized deep learning model is evaluated based on the optimized model parameters and the test set to obtain a trained deep learning model. Finally, the tumor dataset to be predicted is input into the trained deep learning model to predict the tumor radiotherapy response. The purpose of this application is to provide a method and model for predicting tumor radiotherapy response based on Kd-Net. This method and model can predict whether there will be a response after tumor radiotherapy, thereby intelligently assisting doctors in adjusting the radiotherapy dose, implementing adaptive and precise radiotherapy decisions for patients, reducing the harm of radiotherapy to patients, and improving doctors' work efficiency and treatment quality.

[0028] The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0029] See Figure 1 This application provides a method for predicting tumor radiotherapy response based on Kd-Net, which includes the following steps:

[0030] Step S101: Extract tumor features and fuse them to form a point cloud. Classify and label the point cloud according to the SUV change gradient, and divide it into training set, validation set and test set.

[0031] Step S102: Construct a deep learning model based on the Kd-Net network architecture.

[0032] Step S103: Based on the training set and validation set, train the deep learning model, adjust the model parameters, and obtain an optimized deep learning model and optimized model parameters.

[0033] Step S104: Based on the optimized model parameters and test set, evaluate the performance of the optimized deep learning model to obtain the trained deep learning model.

[0034] Step S105: Input the tumor dataset to be predicted into the trained deep learning model to predict the tumor radiotherapy response.

[0035] Kd-Net is a deep neural network based on Kd-trees that uses point cloud data in the PyTorch deep learning framework to perform tasks such as model recognition, classification, and regression. It introduces the Kd-tree structure to transform unstructured point clouds into structured Kd-tree data, enabling the input of structured data into the neural network to construct feature engineering for processing unstructured point clouds. Furthermore, Kd-Net uses point cloud data instead of 3D image slices, avoiding the losses associated with image slices. Point clouds also possess arrangement invariance and transformation invariance, meeting the needs of multimodal tumor imaging and the requirements for preprocessing. Based on this, this application proposes a Kd-Net-based method and model for predicting tumor radiotherapy response, capable of predicting whether a response will occur after tumor radiotherapy. This intelligently assists doctors in adjusting radiotherapy doses, enabling adaptive and precise radiotherapy decisions for patients, reducing the harm of radiotherapy to patients, and improving doctors' work efficiency and the quality of diagnosis and treatment.

[0036] Specifically, in step S101, features related to the voxel SUV values ​​during mid-radiotherapy are extracted from the tumor dataset and fused to form a new tumor point cloud dataset. The same tumor point cloud dataset is classified according to the SUV change gradient and labeled, resulting in 14 different point cloud datasets (including empty point clouds) across two main categories. After removing empty and minimal point clouds, each point cloud dataset is further divided into datasets containing a fixed number of voxels as evenly as possible. Then, training, validation, and test sets are created. A leave-one-out cross-validation method is used, grouping patients. In each experiment, one patient is assigned to the test set, and from the remaining patients, one patient is selected from each group to be assigned to the validation set. The remaining patients are then assigned to the training set. However, it is important to note that when dividing the training set, validation set, and test set, multiple point cloud datasets derived from the same tumor point cloud dataset should be included in the training set, validation set, or test set simultaneously. That is, the tumor point cloud dataset should be divided into training set, validation set, and test set based on the patient. This is to prevent the tumor point cloud dataset of the same patient from appearing in the training set, validation set, and test set at the same time, which could cause data leakage and make the evaluation metrics for evaluating the model performance inflated.

[0037] It should be noted that in step S101, the Standard Uptake Value (SUV) is a semi-quantitative indicator commonly used in positron emission tomography (PET) for tumor diagnosis. It refers to the ratio of the radioactivity of the imaging agent taken up by local tissue to the average systemic injected activity. Currently, SUV is widely used for differentiating between benign and malignant tumors, evaluating treatment efficacy, and predicting prognosis.

[0038] In some embodiments, the number of point cloud datasets in the training, validation, and test sets that are actually effective for radiotherapy should be approximately the same as the number of point cloud datasets that are actually ineffective for radiotherapy. To prevent the model from learning only a single type of point cloud dataset, which could lead to inaccurate predictions of tumor radiotherapy responses and overfitting, the number of point cloud datasets in the training, validation, and test sets that are actually effective for radiotherapy and those that are actually ineffective for radiotherapy should be as balanced as possible.

[0039] After dividing the dataset into training, validation, and test sets, a deep learning model is constructed based on the Kd-Net network structure. Then, the deep learning model is trained using the training and validation sets, and the model parameters are adjusted to achieve optimal model performance.

[0040] The loss function used during model training is the Negative Log-Likelihood Loss (NLL Loss). For an input sequence x and a true label y, the negative log-likelihood loss can be calculated using the model-predicted label distribution p(y|x) and the log probability of the true label y. That is, the NLL Loss can be calculated using equation (1):

[0041] NLL(p(y|x),y)=-logp(y|x) (1)

[0042] The model parameters are adjusted based on the NLL loss value after model training convergence to minimize the NLL loss value after convergence during model validation. Finally, the model parameters that minimize the NLL loss value after convergence are selected and imported into the deep learning model, and the model is then validated using a test set.

[0043] In some embodiments, the point cloud is classified and labeled according to the SUV change gradient in step S101, including:

[0044] Step S1011: Based on the SUV change gradient, classify each voxel to obtain two major categories of voxels.

[0045] Step S1012: Based on the RatioSUV gradient, divide the two major voxels into 14 minor voxels.

[0046] Step S1013: Segment the point cloud to complete the preparation of the dataset.

[0047] In step S1011, the voxels are classified to obtain two main categories of voxels, including:

[0048] Step S1011a: Determine whether the SUV value of each voxel before radiotherapy and the SUV value during radiotherapy satisfy the aforementioned correspondence.

[0049] Step S1011b: If yes, the voxel will be classified as “reactive” and labeled “1”.

[0050] Step S1011c: If not, the voxel will be classified as "unresponsive" and labeled "0"; the corresponding relationship is as follows:

[0051]

[0052] RatioSUV≤0.7

[0053] Among them, PreSUV represents the SUV value of a single voxel before radiotherapy, which reflects the degree of tumor lesion through the standard uptake value of voxels; MidSUV represents the SUV value during the mid-term of single voxel radiotherapy; RatioSUV represents the ratio of the mid-term SUV value of a single voxel radiotherapy to the pre-radiotherapy SUV value.

[0054] Specifically, if the SUV value before voxel radiotherapy and the SUV value during the radiotherapy meet the above correspondence, then the tumor radiotherapy targeting that voxel is considered effective; otherwise, the tumor radiotherapy targeting that voxel is considered ineffective.

[0055] In some embodiments, step S1012 divides two major classes of voxels into 14 minor classes of voxels according to the RatioSUV gradient, including:

[0056] Step S1012a: For voxels classified as "responsive", their RatioSUV values ​​satisfy 0 < RatioSUV ≤ 0.7. Based on the RatioSUV values, they are divided into seven gradients, namely 0 < RatioSUV ≤ 0.1, 0.1 < RatioSUV ≤ 0.2, 0.2 < RatioSUV ≤ 0.3, 0.3 < RatioSUV ≤ 0.4, 0.4 < RatioSUV ≤ 0.5, 0.5 < RatioSUV ≤ 0.6, and 0.6 < RatioSUV ≤ 0.7. Based on the RatioSUV value of each voxel, they are assigned to the corresponding RatioSUV gradient, that is, the voxels classified as "responsive" are further divided into 7 subcategories of voxels, but their labels remain unchanged.

[0057] Step S1012b: For voxels classified as "non-responsive", their RatioSUV values ​​satisfy RatioSUV > 0.7. Based on the RatioSUV values, they are divided into seven gradients, namely 0.7 < RatioSUV ≤ 0.8, 0.8 < RatioSUV ≤ 0.9, 0.9 < RatioSUV ≤ 1.0, 1.0 < RatioSUV ≤ 1.1, 1.1 < RatioSUV ≤ 1.2, 1.2 < RatioSUV ≤ 1.3, and 1.3 < RatioSUV. Based on the RatioSUV value of each voxel, they are assigned to the corresponding RatioSUV gradient, that is, the voxels classified as "non-responsive" are further divided into 7 subcategories of voxels, but their labels remain unchanged.

[0058] In some embodiments, the construction of a deep learning model based on the Kd-Net network architecture in step S102 includes:

[0059] Step S1021: Construct the main network architecture of Kd-Net.

[0060] The main architecture of Kd-Net consists of multiple Kd convolutional layers and one fully connected layer. The main architecture of Kd-Net obtains the output value through forward propagation, converts the output value into a probability value, and calculates the loss value and related prediction evaluation indicators.

[0061] Step S1022: Construct the network front architecture of Kd-Net.

[0062] Step S1023: Based on the network pre-architecture of Kd-Net constructed in step S1022, the training set and test set are converted into structured Kd-tree data, and the Kd-tree data is input into the main architecture of Kd-Net network for training. The model parameters are adjusted to achieve the optimal model. Based on the lowest loss value obtained from the input validation set, the deep learning model is obtained.

[0063] Before inputting the Kd-tree data into the main structure of the Kd-Net network for training, the process also includes: converting the Kd-tree data into tensor variables, and transposing the tensor variables as input to the main structure of the Kd-Net network.

[0064] Specifically, this application converts the training and test sets into structured Kd-tree data through the constructed Kd-Net network pre-architecture, further converts the Kd-tree data into tensor variables, transposes the tensor variables as input to the main Kd-Net network architecture, and finally inputs the transposed tensor variables into the constructed Kd-Net network architecture for training, adjusts the model parameters to achieve model optimization, and obtains the deep learning model based on the lowest loss value obtained from the input validation set.

[0065] In some embodiments, step S105, inputting the tumor dataset to be predicted into the trained deep learning model, includes:

[0066] Step S1051: Preprocess the tumor dataset to be predicted to obtain the preprocessed tumor point cloud dataset.

[0067] Step S1052: Input the preprocessed tumor point cloud dataset into the trained deep learning model to predict the tumor radiotherapy response.

[0068] In some embodiments, the preprocessing of the tumor dataset to be predicted in step S1051 includes: performing feature extraction and data segmentation preprocessing operations on the tumor dataset to be predicted, so that the dimension of the preprocessed tumor point cloud dataset to be predicted is the same as the dimension of the tumor point cloud datasets in the training set, validation set and test set.

[0069] Specifically, the preprocessing of the tumor dataset to be predicted is the same as that for the training, validation, and test sets. Feature extraction and data segmentation are performed on the tumor dataset to be predicted, ensuring that the dimensionality of the preprocessed tumor point cloud dataset is the same as that of the tumor point cloud datasets in the training, validation, and test sets. The preprocessed tumor point cloud dataset is then input into a trained deep learning model to obtain prediction results, which can then predict whether each tumor will respond to radiotherapy.

[0070] See Figure 2This application provides a schematic diagram of the data preprocessing process for a tumor radiotherapy response prediction method based on Kd-Net, including the following steps: First, extract tumor features and fuse them to form a point cloud. Specifically, extract features related to the voxel SUV value during radiotherapy from the tumor features, namely voxel position (IndSparse), pre-radiotherapy voxel SUV value (PrePETsparse), and radiotherapy dose (DoseSparse). Then, classify and label the point cloud according to the SUV change gradient; first, divide each voxel into two major categories, namely "no response" (RatioPETsparse > 0.7) and "response" (RatioPETsparse ≤ 0.7), and add labels "0" and "1" respectively; then, according to the RatioSUV gradient, divide the two major categories of voxels into 14 subcategories of voxels; finally, further segment the point cloud (each subcategory consists of 128 points forming a point cloud), remove empty point clouds and minimal point clouds, and then divide each point cloud dataset as evenly as possible into point cloud datasets containing a fixed number of voxels, such as... Figure 2 As shown, point cloud datasets with more than 128 points and point cloud datasets with less than 128 points are created. From the point cloud datasets with more than 128 points, 128 points are randomly selected to form a single point cloud. Similarly, from the point cloud datasets with less than 128 points, empty points are padded to bring the total number of points to 128, ensuring each point cloud contains 128 points. This completes the preparation of the tumor point cloud dataset. Finally, the tumor point cloud dataset is divided into training, validation, and test sets, based on the patient. It should be noted that the number of points in each point cloud within the tumor point cloud dataset should be determined based on the specific circumstances.

[0071] Specifically, features related to the voxel SUV value during radiotherapy are extracted from tumor features, including voxel location, pre-radiotherapy voxel SUV value, and radiotherapy dose, and these features are fused to form a three-dimensional point cloud.

[0072] In some embodiments, dividing the training set, validation set, and test set further includes: segmenting the point cloud to complete the preparation of the dataset.

[0073] Specifically, for the segmentation of each point cloud, each point cloud is divided into 14 smaller point clouds (including empty point clouds) according to the category of each voxel in the point cloud; for the segmentation of the 14 smaller point clouds (including empty point clouds), based on the number of voxels in the point cloud and the characteristics of the model input, empty point clouds and minimal point clouds are removed, and the remaining smaller point clouds are divided into multiple point clouds of fixed size as evenly as possible according to a fixed number of voxels, thus completing the preparation of the dataset.

[0074] See Figure 3This application provides a network architecture diagram of Kd-Net, including the main network architecture and the pre-network architecture of Kd-Net. The main network architecture of Kd-Net includes multiple Kd convolutional layers and a fully connected layer. Forward propagation yields output values, which are then converted into probability values ​​to calculate loss values ​​and related prediction evaluation metrics. The pre-network architecture of Kd-Net includes converting the training set, validation set, and test set into structured Kd-tree data, further converting the Kd-tree data into tensor variables, and then transposing these tensor variables as input to the main network architecture of Kd-Net. The transposed tensor variables are input into the constructed main network architecture of Kd-Net for training. Model parameters are adjusted to achieve model optimization. Based on the lowest loss value obtained from the input validation set, the best-trained deep learning model is selected.

[0075] See Figure 4 This application also provides a tumor radiotherapy response prediction model based on Kd-Net, including a data preprocessing module 101, a model building module 102, a model training module 103, a model evaluation module 104, and a prediction module 105 connected in sequence. The data preprocessing module 101 extracts tumor features and fuses them to form a point cloud, classifies and labels the point cloud according to the SUV change gradient, and divides it into training, validation, and test sets. The model building module 102 constructs a deep learning model based on the Kd-Net network architecture. The model training module 103 trains the deep learning model using the training and validation sets, adjusts the model parameters, and obtains an optimized deep learning model and optimized model parameters. The model evaluation module 104 evaluates the performance of the optimized deep learning model using the optimized model parameters and the test set, obtaining a trained deep learning model. The prediction module 105 inputs the tumor dataset to be predicted into the trained deep learning model to predict the tumor radiotherapy response, i.e., whether a response will occur after radiotherapy.

[0076] The following describes in detail the tumor radiotherapy response prediction model based on Kd-Net provided in this application with reference to a specific embodiment.

[0077] First, the tumor radiotherapy response prediction model based on Kd-Net proposed in this application embodiment is implemented using Python 3.9 on a Windows 11 system.

[0078] Furthermore, it should be noted that the embodiments of this application analyze and evaluate the radiotherapy response to non-small cell lung cancer.

[0079] First, tumor features are extracted and fused to form point clouds through the data preprocessing module. The point clouds are then classified and labeled according to the SUV change gradient. The same tumor point cloud dataset is divided into 14 different point cloud datasets (including empty point clouds) in two major categories after being classified and labeled according to the SUV change gradient. After removing empty point clouds and minimal point clouds, each point cloud dataset is divided as evenly as possible into point cloud datasets containing a fixed number of voxels.

[0080] Then, the preprocessed tumor dataset is divided into training, validation, and test sets. Leave-one-out cross-validation is used, where patients are grouped, and in each experiment, one patient from each group is assigned to the test set. From the remaining patients, one patient from each group is selected for the validation set, and the remaining patients are assigned to the training set. However, it's crucial that multiple point cloud datasets from the same tumor dataset are simultaneously included in the training, validation, and test sets. This means dividing the tumor point cloud dataset by patient to prevent data leakage and inflated evaluation metrics caused by the same patient's tumor data appearing in all three sets. Furthermore, the number of point cloud datasets in the training, validation, and test sets that were actually effective for radiotherapy should be roughly equal to the number of those that were ineffective. This prevents the model from learning only a single type of point cloud dataset, leading to inaccurate predictions of tumor radiotherapy responses and overfitting.

[0081] Next, based on the Kd-Net network architecture, a deep learning model is constructed using the model building module. The training, validation, and test sets are converted into structured Kd-tree data, which is then further converted into tensor variables. These tensor variables are transposed and used as input to the main Kd-Net network architecture. The transposed tensor variables are then input into the constructed Kd-Net network architecture and trained using the training module. Model parameters are adjusted to achieve optimal model performance. Based on the lowest loss value obtained from the validation set, the best-trained deep learning model is selected.

[0082] Then, based on the test set, the performance of the deep learning model was evaluated using the evaluation module. Evaluation metrics included accuracy, area under the receiver operating characteristic curve (AUC), positive predictive value (PPV), negative predictive value (NPV), sensitivity, and specificity. For all patients, the model achieved an AUC of 0.66 on the test set; for patients with high radiotherapy response, the model achieved an AUC of 0.79; for patients with moderate radiotherapy response, the model achieved an AUC of 0.55; and for patients with low radiotherapy response, the model achieved an AUC of 0.68 (see Figure 5). For all patients, the model achieved an AUC of 0.66, accuracy of 0.69, PPV of 0.69, NPV of 0.70, sensitivity of 0.83, and specificity of 0.52 on the test set (see Table 1). Compared to patients with moderate and low radiotherapy response, patients with high radiotherapy response showed a higher AUC (0.79), which is superior to the traditional voxel prediction classifier based on variogram-weighted GLM (AUC = 0.72). Furthermore, for all patients, the model achieved a Sensitivity of 0.83 on the test set, indicating that the model can more effectively predict "responsive" voxels, and the proportion of "responsive" voxels is even higher in patients with high radiotherapy response. Therefore, considering all these indicators, the model demonstrates better predictive performance for patients with high radiotherapy response.

[0083] Table 1 shows the evaluation index results of the Kd-Net-based tumor radiotherapy response prediction model for all patients in one embodiment.

[0084]

[0085] Finally, the preprocessed tumor dataset to be predicted is input into the trained deep learning model. The trained deep learning model provides the prediction results through the prediction module, which can predict whether there will be a response after tumor radiotherapy.

[0086] Based on the above technical solutions, this application provides a method and model for predicting tumor radiotherapy response based on Kd-Net. The method includes the following steps: First, extracting tumor features and fusing them to form a point cloud; classifying and labeling the point cloud according to the SUV change gradient; and dividing it into training set, validation set, and test set. Then, constructing a deep learning model based on the Kd-Net network architecture. Next, training the deep learning model based on the training set and validation set, adjusting the model parameters, and obtaining an optimized deep learning model and optimized model parameters. Then, evaluating the performance of the optimized deep learning model based on the optimized model parameters and the test set to obtain a trained deep learning model. Finally, inputting the tumor dataset to be predicted into the trained deep learning model to predict the tumor radiotherapy response.

[0087] This application proposes a method and model for predicting tumor radiotherapy response based on Kd-Net. First, by constructing a Kd-Net-based network architecture and introducing a Kd-tree structure, unstructured point clouds are converted into structured Kd-tree data. This allows the structured data to be input into the neural network, enabling feature engineering to process unstructured point clouds. Second, this application uses point cloud datasets instead of slicing 3D images, avoiding the losses caused by image slicing. Furthermore, point clouds possess arrangement invariance and transformation invariance, meeting the needs of multimodal tumor imaging and the requirement for preprocessing. Finally, this application uses evaluation metrics to compare with traditional network models, verifying the effectiveness, accuracy, and adjustability of the constructed tumor radiotherapy response prediction model. The model provided by this application can predict whether a response will occur after tumor radiotherapy, thereby intelligently assisting doctors in adjusting radiotherapy doses, implementing adaptive and precise radiotherapy decisions for patients, reducing the harm of radiotherapy to patients, and improving doctors' work efficiency and treatment quality.

[0088] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A method for predicting tumor radiotherapy response based on Kd-Net, characterized in that, Includes the following steps: Tumor features are extracted and fused to form a point cloud. The point cloud is then classified and labeled according to the SUV variation gradient, and divided into training set, validation set and test set. Deep learning models are built based on the Kd-Net network architecture. Based on the training set and the validation set, the deep learning model is trained, the model parameters are adjusted, and an optimized deep learning model and optimized model parameters are obtained. Based on the optimized model parameters and the test set, the performance of the optimized deep learning model is evaluated to obtain the trained deep learning model. The tumor dataset to be predicted is input into the trained deep learning model to predict the tumor radiotherapy response; The point cloud is classified and labeled according to the SUV variation gradient, including: Based on the SUV variation gradient, each voxel is classified into two major categories. Based on the RatioSUV gradient, the two major categories of voxels are divided into 14 subcategories of voxels; Segment the point cloud to prepare the dataset; The classification of voxels yields two main categories, including: Determine whether the SUV values ​​of each voxel before radiotherapy and the SUV values ​​during radiotherapy correspond to each other; If so, classify the voxel into the "responsive" category and add the label "1"; If not, the voxel is classified as "unresponsive" and labeled "0". The correspondence is as follows: Among them, PreSUV represents the SUV value of a single voxel before radiotherapy, which reflects the degree of tumor lesion through the standard uptake value of voxels; MidSUV represents the SUV value during the mid-stage of single voxel radiotherapy; RatioSUV represents the ratio of the mid-stage SUV value of a single voxel radiotherapy to the SUV value before radiotherapy. Based on the RatioSUV gradient, the two major voxels are divided into 14 subclasses, including: For voxels classified as "responsive", their RatioSUV values ​​satisfy 0 < RatioSUV ≤ 0.

7. Based on the RatioSUV values, they are divided into seven gradients: 0 < RatioSUV ≤ 0.1, 0.1 < RatioSUV ≤ 0.2, 0.2 < RatioSUV ≤ 0.3, 0.3 < RatioSUV ≤ 0.4, 0.4 < RatioSUV ≤ 0.5, 0.5 < RatioSUV ≤ 0.6, and 0.6 < RatioSUV ≤ 0.

7. Based on the RatioSUV values ​​of each voxel, they are assigned to the corresponding RatioSUV gradients. That is, voxels classified as "responsive" are further divided into 7 subcategories of voxels, but their labels remain unchanged. For voxels classified as "non-responsive", their RatioSUV values ​​satisfy RatioSUV > 0.

7. Based on the RatioSUV values, they are divided into seven gradients: 0.7 < RatioSUV ≤ 0.8, 0.8 < RatioSUV ≤ 0.9, 0.9 < RatioSUV ≤ 1.0, 1.0 < RatioSUV ≤ 1.1, 1.1 < RatioSUV ≤ 1.2, 1.2 < RatioSUV ≤ 1.3, and 1.3 < RatioSUV. Based on the RatioSUV values ​​of each voxel, they are assigned to the corresponding RatioSUV gradient, that is, voxels classified as "non-responsive" are further divided into 7 subcategories of voxels, but their labels remain unchanged. Based on the Kd-Net network architecture, a deep learning model is constructed, including: The network architecture of Kd-Net is constructed, which includes multiple Kd convolutional layers and one fully connected layer. The network architecture of Kd-Net obtains output values ​​through forward propagation, converts the output values ​​into probability values, and calculates loss values ​​and related prediction evaluation indicators. Building the network front-end architecture of Kd-Net; Based on the Kd-Net network pre-architecture, the training and test sets are converted into structured Kd-tree data, and the Kd-tree data is input into the Kd-Net network main architecture for training. The model parameters are adjusted to achieve the optimal model. The deep learning model is obtained based on the lowest loss value obtained from the input validation set. Before inputting the Kd-tree data into the main Kd-Net network architecture for training, the following steps are also included: The Kd-tree data is converted into tensor variables, and the tensor variables are transposed and used as input to the main architecture of the Kd-Net network.

2. The tumor radiotherapy response prediction method based on Kd-Net according to claim 1, characterized in that, Tumor features are extracted and fused to form a point cloud, including: Features related to the mid-radiotherapy voxel SUV value are extracted from tumor features and fused to form a three-dimensional point cloud; the features include voxel location, pre-radiotherapy voxel SUV value, and radiotherapy dose.

3. The tumor radiotherapy response prediction method based on Kd-Net according to claim 1, characterized in that, Segment the point cloud to prepare the dataset, including: For each point cloud segmentation, each point cloud is divided into 14 smaller point clouds according to the category to which each voxel belongs, and the smaller point clouds include empty point clouds; For the segmentation of the 14 small point clouds, based on the number of voxels in the point cloud and the characteristics of the model input, empty point clouds and minimal point clouds are removed, and the remaining small point clouds are divided into multiple point clouds of fixed size according to a fixed number of voxels, thereby completing the preparation of the dataset.

4. The tumor radiotherapy response prediction method based on Kd-Net according to claim 1, characterized in that, After classifying and labeling the point cloud, and before dividing it into training, validation, and test sets, the process also includes: Based on the proportion of point clouds labeled "1" segmented from the tumor to the total number of segmented point clouds, each patient was divided into three groups: low-grade radiation response, moderate-grade radiation response, and high-grade radiation response. In the low radiotherapy response group, the proportion of tumor point clouds labeled "1" in the total number of segmented point clouds was 0–0.5%; in the moderate radiotherapy response group, the proportion of tumor point clouds labeled "1" in the total number of segmented point clouds was 0.5–0.8%; and in the high radiotherapy response group, the proportion of tumor point clouds labeled "1" in the total number of segmented point clouds was 0.8–1%.

5. The tumor radiotherapy response prediction method based on Kd-Net according to claim 1, characterized in that, The tumor dataset to be predicted is input into the trained deep learning model, including: The tumor dataset to be predicted is preprocessed to obtain a preprocessed tumor point cloud dataset. The preprocessed tumor point cloud dataset is input into the trained deep learning model to predict the tumor radiotherapy response.

6. The tumor radiotherapy response prediction method based on Kd-Net according to claim 5, characterized in that, The tumor dataset to be predicted is preprocessed, including: The tumor dataset to be predicted is preprocessed by feature extraction and data segmentation to make the dimension of the preprocessed tumor point cloud dataset to be predicted the same as the dimension of the tumor point cloud datasets in the training set, validation set and test set.

7. A tumor radiotherapy response prediction model based on Kd-Net, wherein the model is used to implement the tumor radiotherapy response prediction method based on Kd-Net as described in any one of claims 1 to 6, characterized in that, It includes a data preprocessing module, a model building module, a model training module, a model evaluation module, and a prediction module, which are connected in sequence. The data preprocessing module is used to extract tumor features and fuse them to form a point cloud. The point cloud is then classified and labeled according to the SUV change gradient, and divided into training set, validation set and test set. The model building module is used to build a deep learning model based on the network architecture of Kd-Net; The model training module is used to train the deep learning model based on the training set and the validation set, adjust the model parameters, and obtain an optimized deep learning model and optimized model parameters. The model evaluation module is used to evaluate the performance of the optimized deep learning model based on the optimized model parameters and the test set, so as to obtain the trained deep learning model. The prediction module is used to input the tumor dataset to be predicted into the trained deep learning model to predict the tumor radiotherapy response.

Images