CN-118552765-B - Tumor radiotherapy reaction prediction method and model based on Kd-Net

CN118552765BCN 118552765 BCN118552765 BCN 118552765BCN-118552765-B

Abstract

The embodiment of the application provides a tumor radiotherapy reaction prediction method and model based on Kd-Net, comprising the following steps of extracting tumor characteristics, fusing to form point clouds, classifying and marking the point clouds according to SUV variation gradient, and dividing a training set, a verification set and a test set; the method comprises the steps of constructing a deep learning model based on a Kd-Net network framework, training the deep learning model based on a training set and a verification set, adjusting model parameters, performing performance evaluation on the optimized deep learning model based on the optimized model parameters and a test set to obtain a trained deep learning model, and inputting a tumor data set to be predicted into the trained deep learning model to predict tumor radiotherapy reaction. The prediction method and the model provided by the application can predict whether the tumor reacts after radiotherapy, thereby intelligently assisting a doctor in adjusting the radiotherapy dosage, implementing an adaptive accurate radiotherapy decision on a patient, reducing the harm of radiotherapy to the patient and improving the working efficiency and diagnosis and treatment quality of the doctor.

Inventors

DUAN CHUNYAN
CHEN SHIJUN
Liu Qiantuo

Assignees

同济大学

Dates

Publication Date: 20260508
Application Date: 20240506

Claims (7)

1. The tumor radiotherapy reaction prediction method based on Kd-Net is characterized by comprising the following steps of: Extracting tumor features, fusing to form point clouds, classifying and marking the point clouds according to SUV variation gradients, and dividing a training set, a verification set and a test set; constructing a deep learning model based on a Kd-Net network framework; training the deep learning model based on the training set and the verification set, and adjusting model parameters to obtain an optimized deep learning model and optimized model parameters; Based on the optimized model parameters and the test set, performing performance evaluation on the optimized deep learning model to obtain a trained deep learning model; Inputting a tumor data set to be predicted into the trained deep learning model, and predicting tumor radiotherapy reaction; classifying and labeling the point cloud according to SUV variation gradient, including: Classifying each voxel according to the SUV variation gradient to obtain two major types of voxels; dividing the two major types of voxels into 14 minor types of voxels according to RatioSUV gradients; partitioning the point cloud to finish the preparation of a data set; wherein, the classifying each voxel to obtain two main types of voxels comprises: judging whether the SUV value before radiotherapy of each voxel and the SUV value in the middle of radiotherapy meet the corresponding relation or not; If yes, classifying the voxels into a large class of 'reacted', and attaching a label of '1'; If not, classifying the voxels into a large class of 'no reaction', and adding a label of '0'; The corresponding relation is as follows: wherein PreSUV represents the SUV value before single voxel radiotherapy, the lesion degree of the tumor is reflected by the standard uptake value of the voxels, midSUV represents the SUV value in the middle of single voxel radiotherapy, and RatioSUV represents the ratio of the SUV value in the middle of single voxel radiotherapy to the SUV value before radiotherapy; dividing the two major classes of voxels into 14 minor classes of voxels according to RatioSUV gradients, comprising: For voxels classified as 'reacted' major, the RatioSUV value satisfies 0< RatioSUV less than or equal to 0.7, the values are uniformly divided into seven gradients according to RatioSUV, 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≤0.6,0.6＜RatioSUV≤0.7, is divided into corresponding RatioSUV gradients according to RatioSUV values of all voxels, namely the voxels classified as 'reacted' major are subdivided into 7 minor voxels, but the labels are unchanged; For voxels classified as "non-reactive" major, its RatioSUV value satisfies RatioSUV >0.7, and it is divided into seven gradients according to RatioSUV value, 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≤1.3,1.3＜RatioSUV, divides each voxel RatioSUV value into corresponding RatioSUV gradients, i.e. the voxels classified as "non-reactive" major are subdivided into 7 minor voxels, but their labels are unchanged; The method for constructing the deep learning model based on the Kd-Net network architecture comprises the following steps: Constructing a network main body framework of the Kd-Net, wherein the network main body framework of the Kd-Net comprises a plurality of Kd convolution layers and a full connection layer, acquiring an output value through forward propagation, converting the output value into a probability value, and calculating a loss value and a relevant prediction evaluation index; Constructing a Kd-Net network pre-framework; Converting a training set and a test set into structured Kd-tree data based on a Kd-Net network pre-framework, inputting the Kd-tree data into a Kd-Net network main body framework for training, adjusting model parameters, realizing model optimization, and obtaining a deep learning model according to the lowest loss value obtained by inputting a verification set; Before inputting the Kd-tree data into the Kd-Net network main body framework for training, the method further comprises the following steps: Converting the Kd-tree data into tensor variables, and transposing the tensor variables to be used as input of a Kd-Net network main body framework.
2. The Kd-Net-based tumor radiotherapy response prediction method of claim 1, wherein extracting a fusion of tumor features to form a point cloud comprises: And extracting features related to the SUV value of the voxel in the mid-radiotherapy from the tumor features, and fusing the features to form a three-dimensional point cloud, wherein the features comprise the voxel position, the SUV value of the voxel before radiotherapy and the radiotherapy dosage.
3. The Kd-Net-based tumor radiotherapy response prediction method of claim 1, wherein segmenting the point cloud to complete the preparation of the data set comprises: For the segmentation of each point cloud, according to the category of each voxel in the point cloud, dividing each point cloud into 14 small point clouds, wherein each small point cloud comprises a null point cloud; and for the segmentation of the 14 small point clouds, removing the empty point clouds and the minimum point clouds according to the number of voxels in the point clouds and the input characteristics of the model, and equally dividing the rest small point clouds into a plurality of point clouds with fixed sizes according to the fixed number of voxels, thereby completing the preparation of a data set.
4. The Kd-Net-based tumor radiotherapy response prediction method of claim 1, further comprising, after classifying and labeling the point cloud, and before partitioning the training set, the validation set, and the test set: Dividing each patient into three groups according to the proportion of the point cloud with the label of 1 to the whole point cloud, wherein the three groups are respectively a low radiotherapy reaction, a medium radiotherapy reaction and a high radiotherapy reaction, In the low-radiotherapy reaction group, the proportion of the point cloud with the tumor cut-out label of 1 in the patient group to all the cut-out point clouds is 0-0.5, in the medium-radiotherapy reaction group, the proportion of the point cloud with the tumor cut-out label of 1 in the patient group to all the cut-out point clouds is 0.5-0.8, and in the high-radiotherapy reaction group, the proportion of the point cloud with the tumor cut-out label of 1 in the patient group to all the cut-out point clouds is 0.8-1.
5. The Kd-Net-based tumor radiotherapy response prediction method according to claim 1, characterized by inputting a tumor dataset to be predicted into the trained deep learning model, comprising: preprocessing the tumor data set to be predicted to obtain a preprocessed tumor point cloud data set; And inputting the preprocessed tumor point cloud data set into the trained deep learning model, and predicting tumor radiotherapy reaction.
6. The Kd-Net-based tumor radiotherapy response prediction method according to claim 5, wherein preprocessing the tumor dataset to be predicted comprises: And carrying out feature extraction and data segmentation pretreatment operation on the tumor data set to be predicted, so that the dimension of the pretreated tumor point cloud data set to be predicted is the same as the dimension of the tumor point cloud data set in the training set, the verification set and the test set.
7. A tumor radiotherapy reaction prediction model based on Kd-Net, which is used for realizing the tumor radiotherapy reaction prediction method based on Kd-Net as set forth in any one of claims 1 to 6, and is characterized by comprising a data preprocessing module, a model construction module, a model training module, a model evaluation module and a prediction module which are connected in sequence; the data preprocessing module is used for extracting tumor features, fusing the tumor features to form point clouds, classifying and marking the point clouds according to SUV variation gradients, and dividing a training set, a verification set and a test set; The model construction module is used for constructing a deep learning model according to a Kd-Net network framework; the model training module is used for training the deep learning model according to the training set and the verification set, and adjusting model parameters to obtain an optimized deep learning model and optimized model parameters; the model evaluation module is used for performing performance evaluation on 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 for inputting a tumor data set to be predicted into the trained deep learning model to predict tumor radiotherapy reaction.

Description

Tumor radiotherapy reaction prediction method and model based on Kd-Net Technical Field The application relates to the technical field of medical deep learning, in particular to a tumor radiotherapy reaction prediction method and model based on Kd-Net. Background Radiotherapy is one of three treatment means of malignant tumors, belongs to local area treatment, and can be used for radically curing or relieving local primary tumors or metastasis. Side effects of radiotherapy vary from person to person and are related to the radiation dose, location, and individual health of the patient at the time of treatment. The most common side effects of radiotherapy comprise skin injury of irradiation field, nausea and vomiting, fever, peripheral blood image drop, hypodynamia and the like, and besides, the side effects of dry mouth and throat, radiation pneumonitis, radiation esophagitis, alopecia and the like can be caused due to different irradiation parts. Either high or low radiation doses may have a small impact on the patient. However, the setting of radiotherapy dosage at present mainly depends on the experience of doctors, the individual difference and accidental factors have great influence, and scientificity, rationality and effectiveness cannot be ensured. Therefore, scientific setting of the radiotherapy dosage of the tumor target area has great significance for realizing accurate radiotherapy. Disclosure of Invention The embodiment of the application provides a Kd-Net-based tumor radiotherapy reaction prediction method and model, which can predict whether the tumor radiotherapy reaction is performed or not, thereby intelligently assisting a doctor in adjusting radiotherapy dosage, implementing adaptive accurate radiotherapy decision on a patient, reducing the harm of radiotherapy to the patient and improving the working efficiency and diagnosis and treatment quality of the doctor. In order to solve the technical problems, in a first aspect, the embodiment of the application provides a tumor radiotherapy reaction prediction method based on Kd-Net, which comprises the following steps of firstly, extracting tumor characteristics to fuse and form point clouds, classifying and marking the point clouds according to SUV variation gradients, and dividing a training set, a verification set and a test set; then constructing a deep learning model based on a Kd-Net network framework, training the deep learning model based on a training set and a verification set, adjusting model parameters to obtain an optimized deep learning model and optimized model parameters, evaluating the performance of the optimized deep learning model based on the optimized model parameters and a test set to obtain a trained deep learning model, and finally inputting a tumor data set to be predicted into the trained deep learning model to predict tumor radiotherapy reaction. In some exemplary embodiments, extracting a fusion of tumor features to form a point cloud includes extracting features related to mid-radiotherapy voxel SUV values from tumor features, fusing the features to form a three-dimensional point cloud, the features including voxel locations, pre-radiotherapy voxel SUV values, and radiotherapy doses. In some exemplary embodiments, classifying and labeling the point cloud according to SUV variation gradients comprises classifying each voxel according to SUV variation gradients to obtain two major types of voxels, classifying the two major types of voxels into 14 minor types of voxels according to RatioSUV gradients, segmenting the point cloud to complete preparation of a data set, wherein classifying each voxel to obtain two major types of voxels comprises judging whether SUV values before radiotherapy of each voxel and SUV values in middle of radiotherapy satisfy a corresponding relation, if yes, classifying the voxels into 'major types with response', and attaching a label '1', if no, classifying the voxels into 'major types without response', and attaching a label '0', wherein the corresponding relation is as follows: RatioSUV≤0.7 Wherein PreSUV represents the SUV value before single voxel radiotherapy, the lesion degree of the tumor is reflected by the standard uptake value of the voxels, midSUV represents the SUV value in the middle of single voxel radiotherapy, and RatioSUV represents the ratio of the SUV value in the middle of single voxel radiotherapy to the SUV value before radiotherapy. In some exemplary embodiments, the two major classes of voxels are divided into 14 minor classes of voxels according to RatioSUV gradients, including voxels classified as "reactive" major classes whose RatioSUV value satisfies 0< RatioSUV≤0.7, which are equally divided into seven gradients according to RatioSUV values, i.e., 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≤0.6,0.6＜RatioSUV≤0.7, classifies each voxel according to RatioSUV values into corresponding RatioSUV gradients, i.e., voxe