Image classification method for radioactive temporal lobe injury risk based on multi-time-sequence MRI

Abstract

The invention discloses a radioactive temporal lobe injury risk image classification method based on multi-time-sequence MRI. The method comprises the following steps: acquiring a nasopharynx head and neck multi-sequence nuclear magnetic resonance image and clinical characteristic parameters of a patient; performing time sequence division on the nasopharynx head and neck multi-sequence nuclear magnetic resonance image to obtain a plurality of original sequences; extracting image features of the nasopharynx head and neck multi-sequence nuclear magnetic resonance image in the original sequence; a dynamic MRI radiomics label based on image features is constructed by adopting a long-short time neural network; fusing the dynamic MRI radiomics tag and the clinical characteristic parameters, and establishing a radioactive temporal lobe injury image classification model; and obtaining an image classification result based on the radioactive temporal lobe injury image classification model. According to the method, high-throughput and time-sequence dynamic image space omics information and clinical characteristic indexes are fused, and a more comprehensive radioactive temporal lobe injury image classification model is established by applying an artificial intelligence algorithm.

Description

Technical Field

[0001] The present invention relates to the field of medical-engineering cross-radiomic related technologies, and in particular to an image classification method for radiation temporal lobe injury risk based on multi-time series MRI. Background Art

[0002] Existing prediction methods for radiation-induced temporal lobe damage are mainly based on the prediction of clinical factors such as radiotherapy dose and tumor staging. There are also a small number of studies that predict individual heterogeneity of patients by testing genes, proteins, immune indicators and other indicators. However, genetic immunity and other tests are relatively expensive and invasive, and are difficult to promote and apply in clinical practice.

[0003] As an emerging cross-product of medicine and engineering, radiomics can extract and analyze image features from medical image systems and convert them into high-throughput data that can be mined, providing reference decisions for clinical image classification. Radiomics technology does not require invasive operations and is easy to implement and operate in clinical practice. It has been applied to the prediction of various complications after radiotherapy for nasopharyngeal carcinoma. However, there are few reports on the image classification of radiation-induced temporal lobe damage, or only static image features are used for prediction.

[0004] Long Short Term Memory Neural Network (LSTM) is an improved recurrent neural network that can retain the temporal information of time series during the hidden layer learning process and has good performance in dynamic time series prediction.

[0005] The existing technical methods for image classification of risk of radiation-induced temporal lobe damage in NPC still have the following problems: (1) The prediction method based solely on clinical information is difficult to reflect the spatial biological information of the temporal lobe; (2) The biological characteristics contained in medical images have temporal and spatial heterogeneity, and static imaging genomics research cannot reflect the heterogeneity of tumors and subtle changes in tissues in time series; (3) Genetic immunity and other tests are invasive operations, which are relatively expensive and difficult to promote and apply in clinical practice. Therefore, it is very necessary to provide an image classification method for the risk of radiation-induced temporal lobe damage based on multi-time series MRI. Summary of the invention

[0006] In view of the technical defects or improvement needs in the prior art, the present invention constructs an image classification method for the risk of radiation temporal lobe injury based on multi-time series MRI.

[0007] The present invention provides an image classification method for radiation temporal lobe injury risk based on multi-time series MRI, comprising:

[0008] Obtain the patient's nasopharyngeal head and neck multi-sequence MRI images and clinical characteristic parameters;

[0009] Dividing the multi-sequence nuclear magnetic resonance images of the nasopharynx, head and neck into time series to obtain a number of original sequences;

[0010] Extracting image features of the nasopharyngeal head and neck multi-sequence nuclear magnetic resonance images in the original sequence;

[0011] Using long- and short-time neural networks to construct dynamic MRI radiomics labels based on the image features;

[0012] The dynamic MRI radiomics label and the clinical characteristic parameter are integrated to establish a classification model for radiation-induced temporal lobe injury images;

[0013] An image classification result is obtained based on the radiation temporal lobe injury image classification model.

[0014] Optionally, the process of extracting the image features of the nasopharyngeal head and neck multi-sequence nuclear magnetic resonance images in the original sequence includes:

[0015] Normalizing the nasopharyngeal head and neck multi-sequence nuclear magnetic resonance images to obtain preprocessed images;

[0016] Performing region-of-interest division on the preprocessed image to obtain a region of interest;

[0017] Based on the open source PyRadiomics package, feature extraction is performed on the object of interest to obtain image features.

[0018] Optionally, the influomics features extracted based on the open source PyRadiomics package include: first-order statistical features, shape-based features, texture features, and high-order statistical features.

[0019] Optionally, the process of constructing a dynamic MRI radiomics label based on the image features using a long-short time neural network includes:

[0020] Standardizing the image features of the original sequence to obtain a training set;

[0021] Inputting the training set into the long-short time neural network optimized based on the Bayesian algorithm for training to obtain the optimal hyperparameter combination;

[0022] Obtaining an optimized long-short time neural network based on the optimal hyperparameter combination;

[0023] Dynamic MRI imaging omics labels are obtained based on the optimized long- and short-time neural networks.

[0024] Optionally, the calculation formula for normalizing the image features of the original sequence is:

[0025]

[0026] In the formula, x scaled represents the standardized eigenvalue, x max Indicates the maximum value of the dynamic sequence feature data, x min Indicates the minimum value of dynamic sequence feature data.

[0027] Optionally, the parameters in the long-short time neural network optimized by the Bayesian algorithm include: the number of neurons, the L2 regularization coefficient, the learning rate and the network structure;

[0028] Wherein, the network structure includes a first structure, a second structure, a third structure and a fourth structure;

[0029] The first structure includes an input layer, a set of combination layers, a fully connected layer, and a regression output layer;

[0030] The second structure includes an input layer, two groups of combination layers, a fully connected layer and a regression output layer;

[0031] The third structure includes an input layer, three groups of combination layers, a fully connected layer and a regression output layer;

[0032] The fourth structure includes an input layer, four groups of combination layers, a fully connected layer and a regression output layer.

[0033] Optionally, the dynamic MRI radiomics labels and the clinical characteristic parameters are fused based on logistic regression analysis to obtain a radiation temporal lobe injury image classification model.

[0034] The present invention has the following technical effects:

[0035] The present invention integrates high-throughput, time-series dynamic image spatial omics information and clinical characteristic indicators, and uses artificial intelligence algorithms to establish a more comprehensive radiation temporal lobe injury image classification model. It can provide early knowledge of the risk of radiation temporal lobe injury in nasopharyngeal carcinoma patients after radiotherapy, which has important clinical significance for the early prevention of radiation brain injury.

[0036] The present invention does not require additional examinations for patients, and can directly convert patients' routine imaging information and clinically related information into quantitative features to construct a prediction model, which is more non-invasive, convenient, fast, economical and feasible. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 A technical roadmap in an embodiment of the present invention;

[0039] Figure 2 LSTM is a long short-term neural network (LSTM) hidden layer neuron structure diagram in an embodiment of the present invention;

[0040] Figure 3 This is a Bayesian optimization LSTM prediction framework diagram based on multiple sequence dynamic nuclear magnetic resonance images in an embodiment of the present invention;

[0041] Figure 4 A predictive nomogram established by combining dynamic and static imaging omics labels with clinical factors in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Embodiment 1

[0044] like Figure 1 As shown, this embodiment provides an image classification method for radiation temporal lobe injury risk based on multi-time series MRI, comprising the following steps:

[0045] Obtain the patient's nasopharyngeal head and neck multi-sequence nuclear magnetic resonance images and clinical characteristic parameters; divide the nasopharyngeal head and neck multi-sequence nuclear magnetic resonance images into time series to obtain several original sequences; the specific implementation process includes:

[0046] 1) Image data acquisition and collection: MRI images are obtained from the medical image storage and communication system (PACS) and stored in the form of DICOM files.

[0047] 2) Clinical characteristics data collection: The patient's age, gender, family cancer history, TNM stage, histopathological subtype, radiotherapy dose, radiotherapy regimen, etc. were obtained from the hospital information system (HIS) or paper medical records and organized in digital form in an Excel table.

[0048] 3) Image preprocessing and region of interest segmentation: The images were preprocessed for normalization, and radiologists manually annotated the temporal lobe voxel masks of each patient in each time series using professional annotation software (3DSlicer).

[0049] Extract the image features of the nasopharyngeal head and neck multi-sequence MRI images in the original sequence; the specific implementation process includes:

[0050] 4) Image feature extraction and screening: After the MRI region of interest (ROI) was delineated, the free open source PyRadiomics V3.0.1 was used to extract image features. The least absolute shrinkage and selection operator (LASSO) method was then used to further select image features.

[0051] Long- and short-time neural networks are used to construct dynamic MRI radiomics labels based on image features; dynamic MRI radiomics labels and clinical characteristic parameters are integrated to establish a radioactive temporal lobe injury image classification model; image classification results are obtained based on the radioactive temporal lobe injury image classification model. The specific implementation process includes:

[0052] 5) Establishment and prediction of dynamic sequence imaging omics labels: The present invention learns the characteristics of magnetic resonance sequences in different time periods based on the unique memory function of the LSTM model when processing complex time series. At the same time, the Bayesian optimization algorithm is introduced to optimize hyperparameters, and the probability model is used to represent the complex black box function, so that the evaluation process is closer to the actual situation. The sequential optimization method is used to obtain an approximate global optimal solution at a lower evaluation cost, and the best hyperparameter combination is determined for the LSTM model.

[0053] like Figure 2 As shown in the figure, the LSTM module structure in the network produces a memory effect, which is composed of the forget gate ft and the input gate i. t , output gate o t and the state c at the previous moment t-1 The calculation process of the three gates is as follows:

[0054] f t =σ(W f ·[h t-1 , x t ]+b f ) (1)

[0055] i t =σ(W i ·[h t-1 , x t ]+b i ) (2)

[0056] o t =σ(W o .[h t-1 , x t ]+b o ) (3)

[0057] The current input status for:

[0058]

[0059] This can update the state c t :

[0060]

[0061] The output of the LSTM unit is:

[0062] h t =o t .tanh(c t ) (6)

[0063] If time t is the current time, then x t is the input at the current moment, h t Indicates the output at the current moment, h t-1 Indicates the output at the previous moment, c t is the current state of the memory cell, c t -1 is the memory cell state at the previous moment. σ is the sigmoid activation function, f t 、i t and t Represent the outputs of the forget gate, input gate, and output gate respectively. f , W i , W o and W c Represent the weights of the forget gate, input gate, output gate and memory unit respectively. f , b i , b o and b c They represent the bias vectors of the forget gate, input gate, output gate and memory unit respectively.

[0064] Sigmoid and tanh activation functions control the weight of the previous moment's load information transmitted forward within LSTM. When the output value of the activation function is closer to 0, the less information is transmitted forward; when the output value of the activation function is closer to 1, the more information is transmitted forward. Current memory cell state c t is the current input x t , the output h at the previous moment t-1 and the memory state c at the previous momentt-1 Working together, the current output h t is the current input x t 、Current memory cell state c t and the output h at the previous moment t-1 LSTM uses the retained memory information to update weights and biases in real time for automatic training, and uses back propagation to learn parameters. Through repeated iterative parameter updates, the model can more comprehensively learn the nonlinear relationship of the dynamic magnetic resonance sequence.

[0065] However, the parameter adjustment process of the LSTM model relies on rich parameter adjustment experience. Especially when the hyperparameter space is very large, manual configuration of the hyperparameters of the LSTM model is very time-consuming and inefficient. Therefore, optimization algorithms are usually used to optimize hyperparameters. However, the hyperparameter search process is often a high-dimensional non-convex optimization problem of an unknown objective function. If an optimization algorithm with insufficient optimization strength is used, the network will fall into a local optimum, which seriously limits the learning ability of the LSTM model. In this regard, the hyperparameter solution strategy of the Bayesian optimization algorithm can solve the above problems well. The Bayesian method has two obvious characteristics: one is to use prior information to infer the posterior distribution, and the other is to treat the parameters as unknown random variables. Based on the given prior distribution, the Bayesian method uses observed data to update information to obtain the posterior distribution of the parameters.

[0066] The magnetic resonance dynamic prediction model proposed in this invention and the comparison with other methods are shown in the following figure. Figure 3 As shown in the figure, the main process includes three stages: preprocessing, Bayesian optimization LSTM and model evaluation. The min-max normalization method is used for preprocessing the original sequence, and the formula is as follows:

[0067]

[0068] Among them, x scaled represents the standardized eigenvalue, x max Indicates the maximum value of the dynamic sequence feature data, x min Indicates the minimum value of dynamic sequence feature data.

[0069] Then, a training set is constructed, and the LSTM with the optimal hyperparameter combination is obtained through the Bayesian optimization algorithm, so as to further learn the parameters of LSTM. Among them, the number of neurons represents the number of neurons in each layer of the network, which has a decisive influence on the fitting ability and training speed of LSTM; L2 regularization is based on the attenuation of weights by the L2 method, and the size of its coefficient can determine the penalty degree of the loss function; the learning rate controls the update of network parameters and plays a key role in the learning rate and convergence of the network; the number of hidden layers determines the depth of the network and the learning ability of LSTM. The hyperparameters of Bayesian optimization and their search range are shown in Table 1. In addition, the search range of the network structure includes four structures. The first structure consists of an input layer, a group of combination layers (consisting of unidirectional or bidirectional LSTM and dropout layers), a fully connected layer, and a regression output layer; the second structure consists of an input layer, two groups of combination layers, a fully connected layer, and a regression output layer; the third structure consists of an input layer, three groups of combination layers, a fully connected layer, and a regression output layer; the fourth structure consists of an input layer, four groups of combination layers, a fully connected layer, and a regression output layer.

[0070] Table 1

[0071]

[0072] For other network parameters of the Bayesian optimization LSTM prediction model, the input step size of LSTM is set to 24, the Epoch is set to 400, the mini-batch is set to 32, the Dropout ratio is set to 0.5, the parameter training algorithm uses the adaptive moment estimation algorithm (Adam), and the maximum number of iterations of the Bayesian optimization algorithm is set to 60. Through the pre-training process, the combination of network parameters is set to the above initial values, and network training is very easy to complete. During the network training process, the back propagation algorithm and the optimization algorithm are used to adjust the network parameters by minimizing the error between the current predicted value and the actual value of the network. In the model evaluation stage, in order to verify the superiority of the Bayesian optimization strategy prediction model, the effect of the predetermined prediction framework model is compared with the prediction effect of the simple static prediction and the following models. The comparison model is as follows:

[0073] (1) LSTM model: This method is the basic model of this prediction framework and is a special structure of the recurrent neural network (RNN). It overcomes the limitation of RNN that gradient vanishing easily when processing long sequences. It can process sequence data of any length and learn the long-term dependency of sequence data.

[0074] (2) Bidirectional Long-Short Term Memory (Bi-LSTM) model: This method uses a forward and reverse order strategy to extract data features, retaining the importance and complementarity of forward and backward information for time series prediction, and obtaining richer features to enhance the learning ability of the network.

[0075] (3) Convolutional LSTM model (CNN-LSTM): This method is an effective fusion of convolutional neural network (CNN) and LSTM neural network. It fully considers the significant advantages of convolutional neural network in data feature extraction and integrates the strong memory ability of LSTM neural network in sequence data. It can well extract the characteristic relationship of variables in high-dimensional space and become a more advanced time series prediction method.

[0076] (4) Bayesian Optimization LSTM Model (BO-LSTM): Based on the proposed prediction framework, this model eliminates the process of extracting data features by the CEEMD algorithm, directly substitutes the original outpatient volume data after sequence reconstruction into the LSTM model, and uses the Bayesian optimization algorithm to optimize the hyperparameters of the LSTM model based on this data.

[0077] 6) Establishment and verification of fusion model: Based on dynamic MRI imaging genomics labels, combined with clinical parameters such as age, gender, degree of pathological differentiation, and radiation dose parameters, logistic regression analysis was used to analyze the factors related to radiation-induced temporal lobe injury. The variables were then filtered step by step to simplify the prediction model. Finally, an optimal nomogram model combining imaging genomics characteristics and clinical indicators was drawn. According to the corresponding scales of the nomograms of each indicator, qualitative and quantitative predictions of radiation-induced temporal lobe injury in patients were made to obtain the probability of temporal lobe injury.

[0078] As a preferred solution of this embodiment, the solution is further explained in conjunction with the following specific implementation process.

[0079] 1. MRI Image Acquisition

[0080] The magnetic resonance images of patients were obtained from the hospital's medical image storage and transmission system (Picture Archiving and Communication System, PACS). The image data included the multi-sequence nuclear magnetic resonance images of the nasopharynx, head and neck of patients in regular follow-up within 2 weeks before treatment, 3 months after treatment, 6 months after treatment, 1 year, 2 years, 3 years, 5 years after treatment, etc. The collected images included the three sequences of T1 sequence (T1WI), T2 sequence (T2WI) and T1 enhanced sequence (CET1) at the above time. The image data format is Digital Imaging and Communications in Medicine (DICOM) format.

[0081] 2. Image preprocessing and image region of interest segmentation

[0082] The three-dimensional pixels of each collected magnetic resonance image were resampled to the same interval and the DICOM image matrix was normalized. The image preprocessing process was implemented by PyRadiomics (version 3.0.1). The region of interest (ROI) of the image was segmented manually. A specialist with more than 10 years of work experience used the open source 3D Slicer image computing platform software to manually draw the contour of the middle and lower poles of the temporal lobe layer by layer on the axial images of each sequence. The drawing range included the middle and lower pole level of the temporal lobe, the temporal lobe range from the cerebral peduncle level to the level of temporal lobe disappearance. Each side of the temporal lobe was individually outlined and named and saved. After the ROI was drawn, a specialist with more than 20 years of work experience checked the accuracy of the outline.

[0083] 3. Image feature extraction

[0084] After the ROI was delineated, the open source PyRadiomics package was used to extract features. The specific features were subdivided into the following categories: first order statistic, shape-based, texture, and high order features. A total of 2961 radiomic features were extracted from multiple MRI sequences before radiotherapy (987 features were extracted from each sequence), and 2861, 2940, 2784, 2835, 2891, and 2982 radiomic features were extracted three months, six months, one year, two years, three years, and five years after radiotherapy, respectively.

[0085] 4. Establishment of dynamic imaging omics labels and prediction models

[0086] The magnetic resonance dynamic prediction model proposed in this invention and the comparison with other methods are as follows: Figure 3 As shown in Figure 1, the main process includes three stages: preprocessing, Bayesian optimization LSTM and model evaluation. The min-max normalization method is used to preprocess the original sequence.

[0087] Then, a training set is constructed, and the LSTM with the optimal hyperparameter combination is obtained through the Bayesian optimization algorithm, so as to further learn the parameters of LSTM. Among them, the number of neurons represents the number of neurons in each layer of the network, which has a decisive influence on the fitting ability and training speed of LSTM; L2 regularization is based on the attenuation of weights by the L2 method, and the size of its coefficient can determine the penalty degree of the loss function; the learning rate controls the update of network parameters and plays a key role in the learning rate and convergence of the network; the number of hidden layers determines the depth of the network and the learning ability of LSTM. In addition, the search range of the network structure includes four structures. The first structure consists of an input layer, a group of combination layers (consisting of unidirectional or bidirectional LSTM and dropout layers), a fully connected layer, and a regression output layer; the second structure consists of an input layer, two groups of combination layers, a fully connected layer, and a regression output layer; the third structure consists of an input layer, three groups of combination layers, a fully connected layer, and a regression output layer; the fourth structure consists of an input layer, four groups of combination layers, a fully connected layer, and a regression output layer.

[0088] For other network parameters of the Bayesian optimization LSTM prediction model, the input step size of LSTM is set to 24, the Epoch is set to 400, the mini-batch is set to 32, the Dropout ratio is set to 0.5, the parameter training algorithm uses the adaptive moment estimation algorithm (Adam), and the maximum number of iterations of the Bayesian optimization algorithm is set to 60. Through the pre-training process, the combination of network parameters is set to the above initial values, and network training is very easy to complete. In the process of network training, the back propagation algorithm and the optimization algorithm are used to adjust the network parameters by minimizing the error between the current prediction value and the actual value of the network. In order to verify the superiority of the prediction performance of the proposed prediction framework, the prediction effects of the simple static non-sequential prediction model (N-LSTM), the Bayesian LSTM model (BO-LSTM), the convolutional LSTM model (CNN-LSTM), the bidirectional LSTM model (Bi-LSTM) and the LSTM model are compared, and the results are as follows:

[0089] Table 2

[0090]

[0091]

[0092] 5. Establishment and validation of radiomics labels and prediction models

[0093] The time series MRI features of each patient extracted in step 3 were screened. We used the least absolute shrinkage and selection operator (LASSO) method for feature selection, and used a 10-fold cross-validation method to compare the features selected in each cross-validation run, and screened the best feature subset in the last cross-validation result. Subsequently, 7 machine learning classifiers were used to generate the radiomics label (Rad_Score). These 7 machine learning methods included K-NearestNeighbor (KNN), Logistic Regression (Log), Naive Bayesian (NB), Decision Tree (DT), Stochastic Gradient Descent (SGD), Gradient Boosting Decision Tree (GBT), and Support Vector Machine (SVM). In order to avoid overfitting and improve the stability of the model, this study used a 10-fold cross-validation method to build a classification model based on the selected features. Based on the recall rate and cross-validation score of the model, the most stable classification model is obtained, and the decision function of this classification model is used to generate radiomics labels. So far, each patient's magnetic resonance imaging feature each time has obtained a radiomics label score, and the radiomics features of the dynamic sequence can also be combined to obtain a new dynamic imaging label score. Then, in the R language, each patient's age, gender, degree of pathological differentiation, radiation dose parameters and other clinical parameters, as well as dynamic and static multi-time series radiomics labels, are subjected to logistic regression analysis to screen factors related to the target results and their correlations, and draw a nomogram. From Figure 4 It can be seen that the actual scores of different variables correspond to the coordinate points of the top scale, from which the specific values ​​of clinical factors, static MRI imaging genomics labels, dynamic MRI imaging genomics labels, and dynamic MRI imaging genomics labels + clinical factors can be read, and the sum of the scores of each variable is the total score below the scale. The corresponding total score can be used to derive the corresponding risk probability, which is used to predict the risk of radiation-induced temporal lobe damage in patients with nasopharyngeal carcinoma after radiotherapy. The results of the nomogram established with the above four parameters show that dynamic MRI imaging genomics labels + clinical factors have the largest weight in the prediction model.

[0094] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An image classification method for radiation-induced temporal lobe injury risk based on multi-time series MRI, characterized in that: The following steps are involved: Obtain the patient's nasopharyngeal head and neck multi-sequence MRI images and clinical characteristic parameters; Dividing the multi-sequence nuclear magnetic resonance images of the nasopharynx, head and neck into time series to obtain a number of original sequences; Extracting image features of the nasopharyngeal head and neck multi-sequence nuclear magnetic resonance images in the original sequence; Using long- and short-time neural networks to construct dynamic MRI radiomics labels based on the image features; The dynamic MRI radiomics label and the clinical characteristic parameter are integrated to establish a classification model for radiation-induced temporal lobe injury images; An image classification result is obtained based on the radiation temporal lobe injury image classification model.

2. The image classification method for radiation-induced temporal lobe injury risk based on multi-time series MRI according to claim 1, characterized in that: The process of extracting the image features of the nasopharyngeal head and neck multi-sequence nuclear magnetic resonance images in the original sequence includes: Normalizing the nasopharyngeal head and neck multi-sequence nuclear magnetic resonance images to obtain preprocessed images; Performing region-of-interest division on the preprocessed image to obtain a region of interest; Based on the open source PyRadiomics package, feature extraction is performed on the object of interest to obtain image features.

3. The image classification method for radiation-induced temporal lobe injury risk based on multi-time series MRI according to claim 2, characterized in that: The influomics features extracted based on the open source PyRadiomics package include: first-order statistical features, shape-based features, texture features, and high-order statistical features.

4. The image classification method for radiation-induced temporal lobe injury risk based on multi-time series MRI according to claim 2, characterized in that: The process of constructing a dynamic MRI radiomics label based on the image features using a long-short time neural network includes: Standardizing the image features of the original sequence to obtain a training set; Inputting the training set into the long-short time neural network optimized based on the Bayesian algorithm for training to obtain the optimal hyperparameter combination; Obtaining an optimized long-short time neural network based on the optimal hyperparameter combination; Dynamic MRI imaging omics labels are obtained based on the optimized long- and short-time neural networks.

5. The image classification method for radiation temporal lobe injury risk based on multi-time series MRI according to claim 4, characterized in that: The calculation formula for standardizing the image features of the original sequence is: In the formula, x scaled represents the standardized eigenvalue, x max Indicates the maximum value of the dynamic sequence feature data, x min Indicates the minimum value of dynamic sequence feature data.

6. The image classification method for radiation-induced temporal lobe injury risk based on multi-time series MRI according to claim 4, characterized in that: The parameters in the long-short time neural network optimized by the Bayesian algorithm include: the number of neurons, the L2 regularization coefficient, the learning rate and the network structure; Wherein, the network structure includes a first structure, a second structure, a third structure and a fourth structure; The first structure includes an input layer, a set of combination layers, a fully connected layer, and a regression output layer; The second structure includes an input layer, two groups of combination layers, a fully connected layer and a regression output layer; The third structure includes an input layer, three groups of combination layers, a fully connected layer and a regression output layer; The fourth structure includes an input layer, four groups of combination layers, a fully connected layer and a regression output layer.

7. The image classification method for radiation temporal lobe injury risk based on multi-time series MRI according to claim 6, characterized in that: The dynamic MRI radiomics labels and the clinical characteristic parameters were fused based on logistic regression analysis to obtain a radiation temporal lobe injury image classification model.

Images