A radioactive temporal lobe injury risk prediction model training method, device, equipment and storage medium

Abstract

The application discloses a radioactive temporal lobe injury risk prediction model training method and device, equipment and a storage medium. By fusing CT image samples and dose image samples, and combining clinical information and dose volume information, information omission of artificial feature extraction can be effectively avoided, the prediction accuracy of the model is improved, more reliable results are provided for radioactive temporal lobe injury risk prediction, and thus the current radiotherapy scheme is assisted to be adjusted by doctors, the treatment effect and prognosis effect of patients are improved, and the quality of life is improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of machine learning, and particularly relate to a radioactive temporal lobe injury risk prediction model training method and device, equipment and a storage medium. BACKGROUND

[0002] Nasopharyngeal carcinoma refers to a malignant tumor occurring in the top and lateral wall of the nasopharynx, which is one of the high-incidence malignant tumors in China, with the highest incidence of ear, nose and throat malignant tumors. Radiotherapy is the preferred treatment method for nasopharyngeal carcinoma.

[0003] Radioactive temporal lobe injury is one of the most serious late complications after radical radiotherapy for nasopharyngeal carcinoma, which seriously affects the quality of life of patients. Evaluating the radiotherapy plan can effectively improve the prognosis and quality of life of patients. Doctors can evaluate the radiotherapy plan for nasopharyngeal carcinoma according to the patient's age, tumor stage, dose-volume histogram, and medical image performance, so as to evaluate the prognosis of the patient according to the radiotherapy plan, that is, to estimate the risk of temporal lobe injury, and to adjust the radiotherapy plan according to the risk value, which is beneficial to the treatment and prognosis of the patient.

[0004] The existing radioactive temporal lobe injury risk prediction method mostly relies on artificially designed features, and has low prediction efficiency and low prediction accuracy. SUMMARY

[0005] The present application provides a radioactive temporal lobe injury risk prediction model training method, device, equipment and storage medium, which can effectively avoid information omission of artificial feature extraction, improve the prediction accuracy of the model, provide more reliable results for radioactive temporal lobe injury risk prediction, and thus assist doctors in adjusting the current radiotherapy plan, improve the treatment effect and prognosis effect of patients, and improve the quality of life.

[0006] In a first aspect, the embodiments of the present application provide a radioactive temporal lobe injury risk prediction model training method, comprising:

[0007] obtaining a training sample, the training sample comprising a CT image sample, a dose image sample, a clinical information sample and a dose-volume information sample;

[0008] extracting a first target region image and a second target region image including a temporal lobe from the CT image sample and the dose image sample, respectively;

[0009] inputting the first target region image and the second target region image into a radioactive temporal lobe injury risk prediction model for processing to predict a first risk value of radioactive temporal lobe injury;

[0010] predicting a second risk value of radioactive temporal lobe injury based on the clinical information sample and the dose-volume information sample;

[0011] calculate a risk prediction loss value based on the first risk value and the second risk value;

[0012] update parameters of the radioactive temporal lobe injury risk prediction model based on the risk prediction loss value.

[0013] In a second aspect, the embodiments of the present application further provide a radioactive temporal lobe injury risk prediction device, comprising:

[0014] an acquisition module configured to acquire a CT image, a dose image, clinical information, and dose volume information;

[0015] a first region image extraction module configured to extract a first target region image and a second target region image including a temporal lobe from the CT image and the dose image, respectively;

[0016] a first risk value prediction module configured to input the first target region image and the second target region image into a radioactive temporal lobe injury risk prediction model for processing to predict a first risk value of radioactive temporal lobe injury;

[0017] a second risk value prediction module configured to predict a second risk value of radioactive temporal lobe injury based on the clinical information and the dose volume information;

[0018] a total risk value calculation module configured to calculate a sum of the first risk value and the second risk value to obtain a radioactive temporal lobe injury risk value.

[0019] In a third aspect, the embodiments of the present application further provide a radioactive temporal lobe injury risk prediction model training device, comprising:

[0020] a sample acquisition module configured to acquire training samples, the training samples including CT image samples, dose image samples, clinical information samples, and dose volume information samples;

[0021] a second region image extraction module configured to extract a first target region image and a second target region image including a temporal lobe from the CT image samples and the dose image samples, respectively;

[0022] a first risk value prediction module configured to input the first target region image and the second target region image into a radioactive temporal lobe injury risk prediction model for processing to predict a first risk value of radioactive temporal lobe injury;

[0023] a second risk value prediction module configured to predict a second risk value of radioactive temporal lobe injury based on the clinical information samples and the dose volume information samples;

[0024] The risk prediction loss value calculation module is used to calculate the risk prediction loss value based on the first risk value and the second risk value;

[0025] The parameter update module is used to update the parameters of the radiation-induced temporal lobe injury risk prediction model based on the risk prediction loss value.

[0026] Fourthly, embodiments of the present invention also provide a computer device, comprising:

[0027] One or more processors;

[0028] Storage device for storing one or more programs;

[0029] When the one or more programs are executed by the one or more processors, the one or more processors implement the training method for the radiation-induced temporal lobe injury risk prediction model provided in the first aspect of the present invention.

[0030] Fifthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the training method for the radiation-induced temporal lobe injury risk prediction model provided in the first aspect of the present invention.

[0031] The method for training a radiation-induced temporal lobe injury risk prediction model provided in this invention includes: acquiring training samples, which include CT image samples, dose image samples, clinical information samples, and dose-volume information samples; extracting a first target region image and a second target region image of the temporal lobe from the CT image samples and dose image samples, respectively; inputting the first target region image and the second target region image into the radiation-induced temporal lobe injury risk prediction model for processing; predicting a first risk value for radiation-induced temporal lobe injury; predicting a second risk value for radiation-induced temporal lobe injury based on the clinical information samples and dose-volume information samples; calculating a risk prediction loss value based on the first risk value and the second risk value; and updating the parameters of the radiation-induced temporal lobe injury risk prediction model based on the risk prediction loss value. By fusing CT image samples and dose image samples, and combining clinical information and dose-volume information, the method effectively avoids information omissions caused by manually extracted features, improves the prediction accuracy of the model, provides more reliable results for radiation-induced temporal lobe injury risk prediction, thereby assisting doctors in adjusting current radiotherapy plans, improving patient treatment effects and prognosis, and improving quality of life. Attached Figure Description

[0032] Figure 1A A flowchart of a method for training a radiation-induced temporal lobe injury risk prediction model provided in Embodiment 1 of the present invention;

[0033] Figure 1B A cross-sectional CT image of the head provided in an embodiment of the present invention;

[0034] Figure 1C This is a head cross-sectional dose image provided in an embodiment of the present invention;

[0035] Figure 1D A schematic diagram of a temporal lobe segmentation image provided in an embodiment of the present invention;

[0036] Figure 2A This is a flowchart of a method for training a radiation-induced temporal lobe injury risk prediction model provided in Embodiment 2 of the present invention;

[0037] Figure 2B This is a schematic diagram of cropping the first target region image in an embodiment of the present invention;

[0038] Figure 2C A schematic diagram of the structure of a radiation-induced temporal lobe injury risk prediction model provided in an embodiment of the present invention;

[0039] Figure 3 This is a flowchart of a method for predicting the risk of radiation-induced temporal lobe injury provided in Embodiment 3 of the present invention;

[0040] Figure 4 This is a schematic diagram of a radiation-induced temporal lobe injury risk prediction device provided in Embodiment 4 of the present invention;

[0041] Figure 5 This is a schematic diagram of the structure of a training device for a radiation-induced temporal lobe injury risk prediction model provided in Embodiment 5 of the present invention;

[0042] Figure 6 This is a schematic diagram of the structure of a computer device provided in Embodiment Six of the present invention. Detailed Implementation

[0043] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0044] Example 1

[0045] Figure 1A This is a flowchart of a method for training a radiation-induced temporal lobe injury risk prediction model according to Embodiment 1 of the present invention. This embodiment can be used to improve the prediction efficiency and accuracy of the radiation-induced temporal lobe injury risk prediction model. This method can be executed by the radiation-induced temporal lobe injury risk prediction model training device provided in this embodiment of the invention. This device can be implemented in software and / or hardware, and is typically configured in a computer device, such as... Figure 1A As shown, the method specifically includes the following steps:

[0046] S101. Obtain training samples, which include CT image samples, dose image samples, clinical information samples, and dose-volume information samples.

[0047] Figure 1B The head cross-sectional CT image provided in the embodiment of the present invention, Figure 1C The head transverse dose image provided in this embodiment of the invention requires CT sequence image data to be taken before nasopharyngeal carcinoma patients undergo treatment. Simultaneously, during radiotherapy for nasopharyngeal carcinoma, doctors will determine the prescribed radiotherapy dose for the tumor area, considering factors such as tumor location and disease progression, as well as clinical requirements such as dose limits for surrounding organs at risk. Then, a physical dosimeter sets relevant optimization goals and constraints based on their experience, which are then automatically solved by a Treatment Planning System (TPS) to obtain a dose distribution image that meets the doctor's requirements. This dose distribution is not unique; different optimization goals and constraints result in different dose distributions, manifesting as differences in tumor area dose uniformity, minimum dose, and maximum dose. Both CT images and dose images are three-dimensional sequence images. Although these dose distributions meet the clinical requirements given by the doctor, their quality varies. Theoretically, different quality three-dimensional dose distributions can lead to different cancer prognoses. It should be noted that... Figure 1B and Figure 1C Only cross-sectional CT images and dose images are shown. In embodiments of the present invention, coronal CT images and dose images, as well as sagittal CT images and dose images, may also be included.

[0048] The main influencing factors of radiation-induced temporal lobe injury include clinical factors and dosage factors. Medical imaging can also reflect the extent of radiation-induced temporal lobe lesions to a certain extent. Recent studies have shown that radiation dose is consistently considered a significant risk factor for adverse prognostic factors in radiation-induced temporal lobe injury. Therefore, this embodiment of the invention uses CT images and dose images as training samples to improve the model's predictive accuracy.

[0049] The required patient clinical information includes: gender, age, TNM (Tumor Node Metastasis Classification) stage, and treatment method. Clinical information has a certain influence on radiation-induced temporal lobe injury. This embodiment of the invention aims to incorporate this type of information as training samples into the neural network model, thereby improving the accuracy of the model's predictions.

[0050] The dose-volume information sample consists of dose-volume histogram (DVH) data from the patient's radiotherapy process. DVH information primarily includes: D0.5cc, D1cc, Dmax, Dmin, Dmean, TLV, V5Gy, V10Gy, V15Gy, V20Gy, V25Gy, V30Gy, V35Gy, V40Gy, V45Gy, V50Gy, V55Gy, and V60Gy. Here, V represents volume, D represents dose, and TLV refers to the total volume of the temporal lobe. For example, V10Gy indicates an absorbed dose of 10Gy per unit volume.

[0051] The labels for the training samples are primarily obtained through regular follow-ups after radiotherapy. This invention requires obtaining patient follow-up information as labels for the training samples, including: whether temporal lobe injury (state) occurred, the location of occurrence (left, right, or bilateral), and the corresponding follow-up time t (in months). It is worth noting that in actual studies, we often cannot obtain the exact occurrence time of events for some research subjects, leading to data censoring. Data censoring is mainly caused by two reasons: first, due to time constraints, some individuals may not have experienced the events specified in the study by the study deadline; second, individuals may drop out of the study due to loss to follow-up or other reasons unrelated to the study. Censored data still has value in this invention, but the proportion of censored data should not be too high.

[0052] S102. Extract the first target region image and the second target region image, including the temporal lobe, from the CT image sample and the dose image sample, respectively.

[0053] Figure 1D This is a schematic diagram of a temporal lobe segmentation image provided in an embodiment of the present invention. In this embodiment, the planning system pre-segments the left and right temporal lobes, and after manual verification, obtains the image as shown below. Figure 1D The diagram shows accurate segmented images of the left and right temporal lobes. For example, for the left temporal lobe, a first target region image including the left temporal lobe (segmented image of the left temporal lobe) is extracted from CT image samples, and a second target region image including the left temporal lobe (segmented image of the left temporal lobe) is extracted from dose image samples. For example, the first and second target region images are rectangular images that meet the input requirements of the subsequent radiation-induced temporal lobe injury risk prediction model. The processing for the right temporal lobe is similar to that for the left temporal lobe, and will not be described again here.

[0054] S103. Input the first target region image and the second target region image into the radiation-induced temporal lobe injury risk prediction model for processing, and predict the first risk value of radiation-induced temporal lobe injury.

[0055] In this embodiment of the invention, a first target region image including the temporal lobe from a CT image sample and a second target region image including the temporal lobe from a dose image sample are input into a radiation-induced temporal lobe injury risk prediction model to be trained. Based on the CT image information and dose information, a risk value for radiation-induced temporal lobe injury in the sample is predicted as a first risk value. In this embodiment of the invention, the structure of the radiation-induced temporal lobe injury risk prediction model is not limited.

[0056] S104. A second risk value for predicting radiation-induced temporal lobe injury based on clinical information samples and dose-volume information samples.

[0057] For example, in an embodiment of the present invention, clinical information samples and dose-volume information samples can be encoded, vectorized, and then the encoded vectors of the two can be fused. Based on the fused vector, the risk value of radiation-induced temporal lobe injury in the sample can be predicted as a second risk value.

[0058] S105. Calculate the risk prediction loss value based on the first risk value and the second risk value.

[0059] For example, in one embodiment of the invention, the sum of a first risk value and a second risk value can be calculated to obtain a total risk value, and then a predicted risk loss value can be calculated based on the total risk value. In other embodiments of the invention, a first loss value and a second loss value can be calculated respectively based on the first risk value and the second risk value, and then the sum of the first risk value and the second risk value can be calculated to obtain a third loss value, and then a predicted risk loss value can be calculated based on the sum of the first loss value, the second loss value, and the third loss value.

[0060] Since there is no clear risk value for radiation-induced temporal lobe injury as supervisory information for the network, this embodiment of the invention uses the Cox Proportional Hazards Model to calculate the risk prediction loss value. The Cox Proportional Hazards Model, also known as the Cox regression model, can be used to describe the impact of multiple features that do not change over time on the risk value at a given moment.

[0061] S106. Update the parameters of the risk prediction model for radiation-induced temporal lobe injury based on the risk prediction loss value.

[0062] For example, in this embodiment of the invention, the calculated risk prediction loss value is compared with a preset loss value threshold. If the risk prediction loss value is greater than the loss value threshold, the parameters of the radiation-induced temporal lobe injury risk prediction model are updated, and the process returns to the step of obtaining training samples to obtain new training samples. This training is iterated until the risk prediction loss value is less than or equal to the loss value threshold, at which point the model training is considered complete.

[0063] The method for training a radiation-induced temporal lobe injury risk prediction model provided in this invention includes: acquiring training samples, which include CT image samples, dose image samples, clinical information samples, and dose-volume information samples; extracting a first target region image and a second target region image of the temporal lobe from the CT image samples and dose image samples, respectively; inputting the first target region image and the second target region image into the radiation-induced temporal lobe injury risk prediction model for processing; predicting a first risk value for radiation-induced temporal lobe injury; predicting a second risk value for radiation-induced temporal lobe injury based on the clinical information samples and dose-volume information samples; calculating a risk prediction loss value based on the first risk value and the second risk value; and updating the parameters of the radiation-induced temporal lobe injury risk prediction model based on the risk prediction loss value. By fusing CT image samples and dose image samples, and combining clinical information and dose-volume information, the method effectively avoids information omissions caused by manually extracted features, improves the prediction accuracy of the model, provides more reliable results for radiation-induced temporal lobe injury risk prediction, thereby assisting doctors in adjusting current radiotherapy plans, improving patient treatment effects and prognosis, and improving quality of life.

[0064] Example 2

[0065] Figure 2A This is a flowchart of a method for training a radiation-induced temporal lobe injury risk prediction model according to Embodiment 2 of the present invention. This embodiment refines Embodiment 1 above and describes in detail the specific implementation process of each step in the method for training a radiation-induced temporal lobe injury risk prediction model, such as... Figure 2A As shown, the method includes:

[0066] S201. Obtain training samples, which include CT image samples, dose image samples, clinical information samples, and dose-volume information samples.

[0067] As mentioned above, in this embodiment of the invention, CT images, dose images, clinical information, and dose-volume information of nasopharyngeal carcinoma patients are obtained as training samples.

[0068] For example, in a specific embodiment of the present invention, after obtaining the above information and performing data anonymization, the training set, validation set, and test set can be divided according to the proportion of cases. Since the incidence of disease in the left and right temporal lobes is generally different, unilateral temporal lobes are used as separate samples for division. The division is based on the proportion of patients, for example, 6:2:2, that is, the current dataset is divided into 10 parts, with 6 parts as the training set, 2 parts as the validation set, and 2 parts as the test set. When dividing the dataset, it should be ensured that the proportion of positive and negative samples in the training set, validation set, and test set is relatively consistent, and the left and right temporal lobes of the same patient should be in the same dataset (for example, patient A's left temporal lobe should not be in the training set while the right temporal lobe is in the validation set).

[0069] S202. Match the CT image samples and the dose image samples to obtain the first matching image of the CT image samples and the second matching image of the dose image samples.

[0070] Since the acquired CT images and dose images are generally not the same size (different in the x and y directions, but generally the same in the z direction; for example, the CT image size is 512×512×118, and the dose image size is 181×92×118), it is necessary to match the position and size of the CT and dose images. Because the dose image and CT image correspond to the same patient and their actual physical dimensions are consistent, the two images can be matched accordingly. In this embodiment of the invention, CT image samples and dose image samples are matched, resulting in a first matched image of the CT image sample and a second matched image of the dose image sample. Each pixel in both the first and second matched images is matched to its corresponding physical location.

[0071] For example, the matching process between CT image samples and dose image samples is as follows:

[0072] 1. Calculate the common area between CT image samples and dose image samples.

[0073] For example, obtaining the actual physical distance between adjacent pixels in a dose image sample (spacing). A The image is calculated by determining its width (x), height (y), depth (z), and the position of the image's origin in the world coordinate system (i.e., the top-left corner of the image), and then calculating the actual physical width, height, and depth of the dose map (in mm).

[0074] w = x * spacing A _x

[0075] h = y * spacing A _y

[0076] d = z * spacing A _z

[0077] Where w, h, and d represent the actual physical width, height, and depth of the measurement map, respectively, and spacing A _x、spacing A _y、spacing A _z represents the actual physical distance between adjacent pixels in the x, y, and z directions, respectively.

[0078] Therefore, the coordinates of the image's endpoint (i.e., the bottom right corner) in the world coordinate system are:

[0079] p_end_x = position_x + w

[0080] p_end_y = position_y + h

[0081] Where position_x and position_y represent the origin of the image in the world coordinate system.

[0082] Obtain the actual physical distance between adjacent pixels in a CT image sample (spacing) B And the origin of the image in the world coordinate system (i.e., the top left corner of the image). Using the origin of the CT image sample in the world coordinate system as a reference, calculate the common area between the dose image sample and the CT image sample:

[0083] x_min=(position_x-origin_x) / spacing B _x

[0084] y_min=(position_y-prigin_y) / spacing B _y

[0085] x_max=(p_end_x-origin_x) / spacing B _x

[0086] y_max=(p_end_y-origin_y) / spacing B _y

[0087] Where x_min and y_min represent the starting position coordinates in the image coordinates corresponding to the common area of ​​the CT image sample and the dose image sample, respectively, and x_max and y_max represent the ending position coordinates, respectively.

[0088] 2. Extract common areas from CT image samples as the first matching image.

[0089] For example, using CT image samples as a reference, the starting point (x_min, y_min) and ending point (x_max, y_max) of the aforementioned common region are used as boundaries to crop the CT image samples and obtain the first matching image.

[0090] 3. Linearly sample the dose image sample to the same size as the first matching image to obtain the second matching image.

[0091] The dose image sample is linearly sampled to the same size as the first matching image to obtain the second matching image, so that the first matching image and the second matching image can be completely matched and the physical positions of each pixel point are completely corresponding.

[0092] S203, Locate the center of the temporal lobe.

[0093] As mentioned earlier, the planning and design system pre-segments the left and right temporal lobes, and after manual verification, obtains accurate segmented images of the left and right temporal lobes. The geometric center of the temporal lobe is calculated as the center of the temporal lobe.

[0094] S204. Using the center of the temporal lobe as the center, extract rectangular regions including the temporal lobe from the first matching image and the second matching image respectively as the first target region image and the second target region image.

[0095] Figure 2B This is a schematic diagram of the cropping of the first target region image in an embodiment of the present invention, for example, as shown below. Figure 2B As shown, a square region of fixed length l, centered on the center of the left temporal lobe, is cropped from the first matched image as the first target region image. Similarly, a square region of fixed length l, centered on the center of the left temporal lobe, is cropped from the second matched image as the second target region image. The processing procedure for the right temporal lobe is the same as that for the left temporal lobe, and will not be described again here.

[0096] In this embodiment of the invention, to address the potential issue of insufficient sample size for a single patient, data augmentation transformation operations can be performed on the patient's first target region image and second target image to increase the sample size. Data augmentation transformation methods may include random flipping, random cropping, random scaling, Gaussian blurring, and random erasing, etc., and are not limited to these methods in this embodiment.

[0097] S205. Obtain a binarized image of a local region including the temporal lobe.

[0098] For example, in an embodiment of the present invention, the image of the first target region is binarized, with the element value of the temporal lobe part being 1 and the element value of the remaining part being 0, to obtain a binarized image.

[0099] S206. Perform channel fusion on the first target region image, the second target region image, and the binarized image to obtain a three-channel fused image.

[0100] For example, in this embodiment of the invention, in order to improve computational efficiency, the first target region image and the second target region image can be pre-normalized to normalize the pixel value range of the first target region image and the second target region image to (-1, 1), that is:

[0101]

[0102] Where I represents the pixel value matrix of the current image sequence, I min and I maxThese represent the minimum and maximum values ​​of the current image pixel value matrix, respectively. For example, in this embodiment of the invention, the minimum and maximum values ​​of the pixel value matrix are set to fixed values:

[0103] I min =-200

[0104] I max =200

[0105] Before normalization, I needs to be subjected to a nonlinear transformation, i.e., less than I. min The pixel value becomes -200, which is greater than I. max The pixel value becomes 200.

[0106] Figure 2C This is a schematic diagram of the structure of a radiation-induced temporal lobe injury risk prediction model provided in an embodiment of the present invention, as shown below. Figure 2C As shown, the first target region image, the second target region image, and the binarized image obtained above are used as input to the model. These three images are fused (concatenated) in the channel dimension to form a three-channel fused image. Feature extraction is performed by fully combining CT image samples and dose image samples. Simultaneously, a binarized image of the temporal lobe segmentation is provided as a mask, allowing the network to focus more on image information in the temporal lobe region, thereby extracting more robust and effective features. Since the binarized image (pixel values ​​of 0 and 1) is not normalized, to ensure consistent distribution of the input data, the pixel values ​​of the binarized image are directly subtracted by 0.5, resulting in a pixel range of (-0.5, 0.5) for the binarized image.

[0107] S207. Extract feature maps representing the fused image from the fused image.

[0108] For example, such as Figure 2C As shown, the model uses 3D ResNet as the backbone network. For example, the backbone network is a 3D ResNet34 network, which includes five residually connected convolutional blocks. These five convolutional blocks are connected sequentially, with the output of the previous convolutional block serving as the input to the next. The output features of the five convolutional blocks are C1, C2, C3, C4, and C5, respectively. The resolution of the output features C1, C2, C3, C4, and C5 decreases sequentially, resulting in the high-level semantic features output by the last convolutional block C5, which are used as the feature map of the fused image. To prevent overfitting, in this embodiment, a Dropout layer is added after each residual convolutional block of the 3D ResNet34 for random dropout.

[0109] S208. Extract cross-space dependencies from the feature map to obtain global features.

[0110] In this embodiment of the invention, a non-local spatial attention mechanism is used to extract cross-spatial dependencies in the feature map. Based on these dependencies, each element in the feature map is weighted to enhance the attention score of regions with important information, thus obtaining global features. Specifically, the processing procedure of the non-local spatial attention mechanism is as follows:

[0111] 1. Calculate the spatial correlation between each element in the feature map and all other elements to obtain the correlation matrix.

[0112] In this embodiment of the invention, the spatial correlation of each element in the image features with all other elements is calculated to obtain a correlation matrix. Spatial correlation refers to the degree of closeness of the relationship between different elements in space. The process of calculating the correlation matrix is ​​as follows:

[0113] 1.1 Perform convolution operations in three directions on the feature map to obtain the Q matrix, K matrix and V matrix respectively.

[0114] For example, convolution operations are performed on the feature map in three different directions: width, height, and depth, to obtain the Q matrix, K matrix, and V matrix, respectively.

[0115] 1.2 Multiply the Q matrix and the K matrix to obtain the correlation weight matrix of the Q matrix and the K matrix.

[0116] Calculate the product of the Q matrix and the K matrix to obtain the correlation weight matrix of the Q matrix and the K matrix, which is the degree of correlation between each element of the Q matrix and the K matrix.

[0117] 1.3. Normalize the correlation weight matrix to obtain the correlation weight coefficient matrix of the Q matrix and the K matrix.

[0118] For example, the elements of each row or column of the correlation weight matrix are normalized using the softmax function to obtain a normalized correlation weight coefficient matrix. The element with the highest correlation weight coefficient in a given row or column is most relevant to that row or column.

[0119] 1.4 Multiply the correlation weight coefficient matrix with the V matrix to obtain the correlation matrix, which represents the spatial correlation between each element in the feature map and all other elements.

[0120] Multiplying the correlation weight coefficient matrix by the V matrix yields a correlation matrix representing the spatial correlation between each element in the feature map and all other elements. In this embodiment of the invention, to enable subsequent multiplication with the feature map, the channels of the correlation matrix need to be expanded. Typically, a one-dimensional convolution operation is performed on the correlation matrix to expand it to the same size as the feature map.

[0121] 2. Calculate the product of the correlation matrix and the feature map to obtain the global features.

[0122] Calculating the product of the correlation matrix and the feature map, and applying the spatial attention mechanism to all elements of the feature map, essentially means that each output element is a weighted average of all other elements, ultimately yielding the global features.

[0123] S209, First risk value for predicting radiation-induced temporal lobe injury based on global features.

[0124] like Figure 2C As shown, global features are input into a global average pooling layer (Average Pool 3D) for global average pooling processing to achieve feature dimensionality reduction and obtain dimensionality-reduced features. After the global average pooling layer, two fully connected layers (FC2 and FC Risk1) are concatenated. The risk value of the sample having radiation-induced temporal lobe damage is obtained through mapping by the two fully connected layers and is used as the first risk value Risk1.

[0125] Furthermore, in this embodiment of the invention, after the global average pooling layer, two fully connected layers (FC1 and FCClassification) are connected in series, and the binary classification result of whether the sample has radiation-induced temporal lobe damage is obtained through mapping by the two fully connected layers.

[0126] S210, a second risk value for predicting radiation-induced temporal lobe injury based on clinical information samples and dose-volume information samples.

[0127] For example, such as Figure 2C As shown, clinical information samples and dose-volume information samples are input into fully connected layers FC3 and FC4 for encoding to obtain a first feature and a second feature, respectively. Next, the first and second features are fused to obtain a fused feature. Then, a second risk value for radiation-induced temporal lobe injury is predicted based on the fused feature. For example, the fused feature is mapped through a fully connected layer (FC Risk2) to obtain the risk value of radiation-induced temporal lobe injury in the sample, which serves as the second risk value, Risk2.

[0128] S211. Calculate the risk prediction loss value based on the first risk value and the second risk value.

[0129] For example, in this embodiment of the invention, the sum of the first risk value and the second risk value is calculated to obtain the total risk value. Then, based on the proportional hazards regression model, the first loss value, the second loss value, and the third loss value corresponding to the first risk value, the second risk value, and the total risk value are calculated respectively. Then, the first loss value, the second loss value, and the third loss value are assigned corresponding weights and added together to obtain the risk prediction loss value. Specifically, the calculation formula for the risk prediction loss value loss1 is as follows:

[0130] loss1 = λ1·cox_loss risk1 +λ2·cox_loss risk2 +λ3·cox_loss risk

[0131] Among them, cox_loss risk1 The first loss value is cox_loss risk2 The second loss value is cox_loss risk Let λ1, λ2, and λ3 be the weights corresponding to the first, second, and third loss values, respectively. The first, second, and third loss values ​​can be calculated using the following formula:

[0132]

[0133] Where N is the total number of samples in the training set, and j is the number of samples in the training set that have experienced radiation-induced temporal lobe injury (i.e., the label δ). j =1), R j t represents the risk value (which can be the first risk value, the second risk value, or the total risk value) predicted by the model for sample j. j Let be the survival time of sample j, and be t. i The training time set is based on samples at risk (i.e., the survival time of sample i is greater than the survival time of sample j: t). i >t j ).

[0134] S212. Calculate the classification loss value based on the classification results.

[0135] In this embodiment of the invention, a classification branch is designed for classification supervision constraints. For example, based on the classification result, the cross-entropy loss is calculated as the classification loss. classification .

[0136] S213. Update the parameters of the radiation-induced temporal lobe injury risk prediction model based on the risk prediction loss value.

[0137] As described above, the model in this embodiment of the invention includes one classification branch and two risk prediction branches (one based on image data, and the other based on clinical information and DHV information). In this embodiment, considering that the risk values ​​predicted by the two risk prediction branches should be consistent, a self-supervised consistency constraint is designed. Specifically, the consistency loss value between the first risk value and the second risk value is calculated as the consistency constraint between them. The formula for calculating the consistency loss value is as follows:

[0138] loss2 = |σ(risk1) - σ(risk2)|

[0139] Where σ represents the sigmoid activation function.

[0140] Furthermore, to prevent overfitting and improve the generalization performance of the model, this embodiment of the invention also designs a regularization constraint term "Regular" for the parameters of the entire model. The regularization constraint term "Regular" is the L1 regularization and L2 regularization of the model parameters.

[0141] Therefore, the total loss value of the model is loss all for:

[0142] loss all =loss1+λ4loss classification +λ5loss2+λ6Regular

[0143] Right now:

[0144]

[0145] Where λ1 to λ6 represent the weights of different losses.

[0146] After obtaining the total loss value of the model, the parameters of the radiation-induced temporal lobe injury risk prediction model are updated based on the total loss value. For example, the calculated total loss value is compared with a preset loss value threshold. If the total loss value is greater than the loss value threshold, the parameters of the radiation-induced temporal lobe injury risk prediction model are updated, and the process returns to the step of obtaining training samples to acquire new training samples. This training is iterated until the total loss value is less than or equal to the loss value threshold, at which point the model training is considered complete.

[0147] The radiation-induced temporal lobe injury risk prediction model training method provided in this invention, by fusing CT image samples and dose image samples, and combining clinical information and dose-volume information, effectively avoids information omissions in manually extracted features, improves the model's prediction accuracy, and provides more reliable results for radiation-induced temporal lobe injury risk prediction. This assists physicians in adjusting current radiotherapy plans, improving patient treatment outcomes and prognosis, and enhancing quality of life. Furthermore, a multi-branch network model and a corresponding loss function are designed to effectively improve the model's prediction accuracy for the radiation-induced temporal lobe injury risk value prediction problem.

[0148] Example 3

[0149] Figure 3 The flowchart of a method for predicting the risk of radiation-induced temporal lobe injury provided in Embodiment 3 of the present invention is as follows: Figure 3 As shown, the method includes:

[0150] S301. Acquire CT images, dose images, clinical information, and dose-volume information.

[0151] S302. Extract the first target region image and the second target region image, including the temporal lobe, from the CT image and the dose image, respectively.

[0152] S303. Input the first target region image and the second target region image into the radiation-induced temporal lobe injury risk prediction model for processing, and predict the first risk value of radiation-induced temporal lobe injury.

[0153] S304, a second risk value for predicting radiation-induced temporal lobe injury based on clinical and dose-volume information.

[0154] S305. Calculate the sum of the first risk value and the second risk value to obtain the risk value of radiation-induced temporal lobe injury.

[0155] Specifically, the specific processes of each step in the above model application are similar to the data processing methods of the corresponding steps in the model training process, and will not be repeated here in the embodiments of the present invention.

[0156] The radiation-induced temporal lobe injury risk prediction method provided in this invention, by fusing CT images and dose images, and combining clinical information and dose-volume information, can effectively avoid information omissions in manually extracted features, improve the prediction accuracy of the model, and provide more reliable results for radiation-induced temporal lobe injury risk prediction. This assists doctors in adjusting current radiotherapy plans, improving patients' treatment effects and prognoses, and enhancing their quality of life.

[0157] Example 4

[0158] Figure 4This is a schematic diagram of a radiation-induced temporal lobe injury risk prediction device provided in Embodiment 4 of the present invention, as shown below. Figure 4 As shown, the device includes:

[0159] The acquisition module 401 is used to acquire CT images, dose images, clinical information, and dose-volume information.

[0160] The first region image extraction module 402 is used to extract a first target region image including the temporal lobe and a second target region image from the CT image and the dose image, respectively.

[0161] The first risk value prediction module 403 is used to input the first target region image and the second target region image into the radiation-induced temporal lobe injury risk prediction model for processing, and to predict the first risk value of radiation-induced temporal lobe injury.

[0162] The second risk value prediction module 404 is used to predict a second risk value of radiation-induced temporal lobe injury based on the clinical information and the dose-volume information.

[0163] The total risk value calculation module 405 is used to calculate the sum of the first risk value and the second risk value to obtain the risk value of radiation-induced temporal lobe injury.

[0164] The above-mentioned radiation-induced temporal lobe injury risk prediction device can execute the radiation-induced temporal lobe injury risk prediction method provided in the above embodiments of the present invention, and has the corresponding functional modules and beneficial effects of executing the radiation-induced temporal lobe injury risk prediction method.

[0165] Example 5

[0166] Figure 5 This is a schematic diagram of the structure of a training device for a radiation-induced temporal lobe injury risk prediction model provided in Embodiment 5 of the present invention, as shown below. Figure 5 As shown, the device includes:

[0167] The sample acquisition module 501 is used to acquire training samples, which include CT image samples, dose image samples, clinical information samples, and dose-volume information samples.

[0168] The second region image extraction module 502 is used to extract a first target region image including the temporal lobe and a second target region image from the CT image sample and the dose image sample, respectively.

[0169] The first risk value prediction module 503 is used to input the first target region image and the second target region image into the radiation-induced temporal lobe injury risk prediction model for processing, and to predict the first risk value of radiation-induced temporal lobe injury.

[0170] The second risk value prediction module 504 is used to predict a second risk value of radiation-induced temporal lobe injury based on the clinical information sample and the dose-volume information sample.

[0171] The risk prediction loss value calculation module 505 is used to calculate the risk prediction loss value based on the first risk value and the second risk value;

[0172] The parameter update module 506 is used to update the parameters of the radiation-induced temporal lobe injury risk prediction model based on the risk prediction loss value.

[0173] In some embodiments of the present invention, the second region image extraction module 502 includes:

[0174] The matching submodule is used to match the CT image sample and the dose image sample to obtain a first matching image of the CT image sample and a second matching image of the dose image sample, wherein each pixel in the first matching image and the second matching image is matched with its corresponding physical location;

[0175] The central positioning submodule is used to locate the center of the temporal lobe;

[0176] The target region image cropping submodule is used to crop a rectangular region including a preset range of the temporal lobe from the first matching image and the second matching image, with the center of the temporal lobe as the center, as the first target region image and the second target region image, respectively.

[0177] In some embodiments of the present invention, the matching submodule includes:

[0178] A common area calculation unit is used to calculate the common area between the CT image sample and the dose image sample;

[0179] The first matching image cropping unit is used to crop the common region from the CT image sample as the first matching image;

[0180] The second matching image cropping unit is used to linearly sample the dose image sample to the same size as the first matching image to obtain the second matching image.

[0181] In some embodiments of the present invention, the first risk value prediction module 503 further includes:

[0182] The binarized image acquisition submodule is used to acquire a binarized image including a local region of the temporal lobe;

[0183] The first risk value prediction submodule is used to input the first target region image, the second target region image and the binarized image into the radiation-induced temporal lobe injury risk prediction model for processing, and to predict the first risk value of radiation-induced temporal lobe injury.

[0184] In some embodiments of the present invention, the first risk value prediction submodule includes:

[0185] The channel fusion unit is used to perform channel fusion on the first target region image, the second target region image, and the binarized image to obtain a three-channel fused image;

[0186] The feature extraction unit is used to extract feature maps representing the fused image from the fused image;

[0187] The relation extraction unit is used to extract cross-space dependencies from the feature map to obtain global features;

[0188] The first risk value prediction unit is used to predict the first risk value of radiation-induced temporal lobe injury based on the global features.

[0189] In some embodiments of the present invention, the relationship extraction unit includes:

[0190] The correlation calculation subunit is used to calculate the spatial correlation between each element in the feature map and all other elements to obtain the correlation matrix.

[0191] The product sub-unit is used to calculate the product of the correlation matrix and the feature map to obtain the global features.

[0192] In some embodiments of the present invention, the correlation calculation subunit includes:

[0193] The convolution component is used to perform convolution operations on the feature map in three directions to obtain the Q matrix, K matrix and V matrix respectively;

[0194] The first matrix multiplication component is used to multiply the Q matrix and the K matrix to obtain the correlation weight matrix of the Q matrix and the K matrix;

[0195] A normalization component is used to normalize the correlation weight matrix to obtain the correlation weight coefficient matrix of the Q matrix and the K matrix.

[0196] The second matrix multiplication component is used to multiply the correlation weight coefficient matrix with the V matrix to obtain a correlation matrix representing the spatial correlation between each element in the feature map and all other elements.

[0197] In some embodiments of the present invention, the second risk value prediction module 504 includes:

[0198] The encoding submodule is used to encode the clinical information sample and the dose-volume information sample respectively to obtain the first feature and the second feature;

[0199] The feature fusion submodule is used to perform channel fusion on the first feature and the second feature to obtain a fused feature;

[0200] The second risk value prediction submodule is used to predict the second risk value of radiation-induced temporal lobe injury based on the fusion features.

[0201] In some embodiments of the present invention, the risk prediction loss value calculation module 505 includes:

[0202] The total risk value calculation submodule is used to calculate the sum of the first risk value and the second risk value to obtain the total risk value;

[0203] The loss value calculation submodule is used to calculate the first loss value, the second loss value, and the third loss value corresponding to the first risk value, the second risk value, and the total risk value, respectively, based on the proportional risk regression model;

[0204] The risk prediction loss value calculation submodule is used to assign corresponding weights to the first loss value, the second loss value and the third loss value, and add them together to obtain the risk prediction loss value.

[0205] In some embodiments of the present invention, the radiation-induced temporal lobe injury risk prediction model also outputs a binary classification result indicating whether the CT image sample has radiation-induced temporal lobe injury, and the device further includes:

[0206] The classification loss calculation module is used to calculate the classification loss value based on the classification result.

[0207] In some embodiments of the present invention, the parameter update module 506 includes:

[0208] The consistency loss calculation submodule is used to calculate the consistency loss value between the first risk value and the second risk value;

[0209] The regularization constraint term calculation submodule is used to calculate the regularization constraint term of the radiation-induced temporal lobe injury risk prediction model;

[0210] The parameter update submodule is used to update the parameters of the radiation-induced temporal lobe injury risk prediction model based on the risk prediction loss value, the classification loss value, the consistency loss value, and the regularization constraint term.

[0211] The above-mentioned radiation-induced temporal lobe injury risk prediction model training device can execute the radiation-induced temporal lobe injury risk prediction model training method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of executing the radiation-induced temporal lobe injury risk prediction model training method.

[0212] Example 6

[0213] Embodiment 6 of the present invention provides a computer device,Figure 6 This is a schematic diagram of the structure of a computer device provided in Embodiment Six of the present invention, as shown below. Figure 6 As shown, the computer device includes:

[0214] Processor 601, memory 602, communication module 603, input device 604, and output device 605; the number of processors 601 in the computer device can be one or more. Figure 6 Taking a processor 601 as an example; the processor 601, memory 602, communication module 603, input device 604, and output device 605 in a computer device can be connected via a bus or other means. Figure 6 Taking a bus connection as an example, the aforementioned processor 601, memory 602, communication module 603, input device 604, and output device 605 can be integrated into a computer device.

[0215] The memory 602, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the modules corresponding to the radiation-induced temporal lobe injury risk prediction model training method or the radiation-induced temporal lobe injury risk prediction method in the above embodiments. The processor 601 executes various functional applications and data processing of the computer device by running the software programs, instructions, and modules stored in the memory 602, thereby implementing the aforementioned radiation-induced temporal lobe injury risk prediction model training method or radiation-induced temporal lobe injury risk prediction method.

[0216] The memory 602 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on the use of the microcomputer. Furthermore, the memory 602 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 602 may further include memory remotely located relative to the processor 601, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0217] The communication module 603 is used to establish a connection with external devices (such as smart terminals) and to realize data interaction with external devices. The input device 604 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the computer device.

[0218] The computer device provided in this embodiment can execute the training method for the radiation-induced temporal lobe injury risk prediction model or the radiation-induced temporal lobe injury risk prediction method provided in any of the above embodiments of the present invention, and has corresponding functions and beneficial effects.

[0219] Example 7

[0220] Embodiment 7 of the present invention provides a storage medium containing computer-executable instructions, on which a computer program is stored. When the program is executed by a processor, it implements the training method for the radiation-induced temporal lobe injury risk prediction model or the radiation-induced temporal lobe injury risk prediction method provided in any of the above embodiments of the present invention.

[0221] Training methods for radiation-induced temporal lobe injury risk prediction models include:

[0222] Acquire training samples, which include CT image samples, dose image samples, clinical information samples, and dose-volume information samples;

[0223] The first target region image, including the temporal lobe, and the second target region image are extracted from the CT image sample and the dose image sample, respectively.

[0224] The first target region image and the second target region image are input into the radiation-induced temporal lobe injury risk prediction model for processing to predict the first risk value of radiation-induced temporal lobe injury.

[0225] A second risk value for radiation-induced temporal lobe injury is predicted based on the clinical information sample and the dose-volume information sample.

[0226] Calculate the predicted risk loss value based on the first risk value and the second risk value;

[0227] The parameters of the radiation-induced temporal lobe injury risk prediction model are updated based on the risk prediction loss value.

[0228] Methods for predicting the risk of radiation-induced temporal lobe injury include:

[0229] Acquire CT images, dose images, clinical information, and dose-volume information;

[0230] The first target region image, including the temporal lobe, and the second target region image are extracted from the CT image and the dose image, respectively.

[0231] The first target region image and the second target region image are input into the radiation-induced temporal lobe injury risk prediction model for processing to predict the first risk value of radiation-induced temporal lobe injury.

[0232] A second risk value for radiation-induced temporal lobe injury is predicted based on the clinical information and the dose-volume information;

[0233] The sum of the first risk value and the second risk value is calculated to obtain the risk value of radiation-induced temporal lobe injury.

[0234] It should be noted that the embodiments of the apparatus, computer equipment, and storage media are basically similar to the method embodiments, so the descriptions are relatively simple. For relevant details, please refer to the descriptions of the method embodiments.

[0235] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a robot, personal computer, server, or network device, etc.) to execute the radiation-induced temporal lobe injury risk prediction model training method or radiation-induced temporal lobe injury risk prediction method described in any embodiment of the present invention.

[0236] It is worth noting that the various modules, sub-modules, units, sub-units, and components included in the above-mentioned device are only divided according to functional logic, but are not limited to the above division, as long as they can achieve the corresponding functions; in addition, the specific names of each functional module are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0237] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0238] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0239] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for training a radiation-induced temporal lobe injury risk prediction model, characterized in that, include: Acquire training samples, which include CT image samples, dose image samples, clinical information samples, and dose-volume information samples; The first target region image, including the temporal lobe, and the second target region image are extracted from the CT image sample and the dose image sample, respectively. The first target region image and the second target region image are input into the radiation-induced temporal lobe injury risk prediction model for processing to predict the first risk value of radiation-induced temporal lobe injury. A second risk value for radiation-induced temporal lobe injury is predicted based on the clinical information sample and the dose-volume information sample. Calculate the predicted risk loss value based on the first risk value and the second risk value; The parameters of the radiation-induced temporal lobe injury risk prediction model are updated based on the predicted risk loss value. The first target region image and the second target region image are input into a radiation-induced temporal lobe injury risk prediction model for processing to predict a first risk value for radiation-induced temporal lobe injury. The process also includes: Obtain a binarized image including a local region of the temporal lobe; The first target region image, the second target region image, and the binarized image are input into the radiation-induced temporal lobe injury risk prediction model for processing to predict the first risk value of radiation-induced temporal lobe injury. The second risk value for radiation-induced temporal lobe injury is predicted based on the clinical information sample and the dose-volume information sample, including: The clinical information sample and the dose-volume information sample are encoded respectively to obtain a first feature and a second feature; Channel fusion is performed on the first feature and the second feature to obtain the fused feature; A second risk value for radiation-induced temporal lobe injury is predicted based on the fusion features.

2. The training method for the radiation-induced temporal lobe injury risk prediction model according to claim 1, characterized in that, Extracting a first target region image and a second target region image, including the temporal lobe, from the CT image samples and the dose image samples, respectively, includes: The CT image sample and the dose image sample are matched to obtain a first matching image of the CT image sample and a second matching image of the dose image sample, wherein each pixel in the first matching image and the second matching image is matched with its corresponding physical location; Locate the center of the temporal lobe; Centered on the center of the temporal lobe, rectangular regions including the temporal lobe within a preset range are extracted from the first matching image and the second matching image, respectively, as the first target region image and the second target region image.

3. The training method for the radiation-induced temporal lobe injury risk prediction model according to claim 2, characterized in that, Matching the CT image samples and the dose image samples to obtain a first matching image of the CT image samples and a second matching image of the dose image samples includes: Calculate the common region between the CT image sample and the dose image sample; The common region is extracted from the CT image sample as the first matching image; The dose image sample is linearly sampled to the same size as the first matching image to obtain the second matching image.

4. The training method for the radiation-induced temporal lobe injury risk prediction model according to claim 1, characterized in that, The first target region image, the second target region image, and the binarized image are input into a radiation-induced temporal lobe injury risk prediction model for processing to predict a first risk value for radiation-induced temporal lobe injury, including: Channel fusion is performed on the first target region image, the second target region image, and the binarized image to obtain a three-channel fused image; Extract feature maps representing the fused image from the fused image; The global features are obtained by extracting cross-space dependencies from the feature map. The first risk value for radiation-induced temporal lobe injury is predicted based on the global features.

5. The method for training a radiation-induced temporal lobe injury risk prediction model according to claim 4, characterized in that, From the feature map, cross-space dependencies are extracted to obtain global features, including: Calculate the spatial correlation between each element in the feature map and all other elements to obtain the correlation matrix; The global features are obtained by multiplying the correlation matrix and the feature map.

6. The training method for the radiation-induced temporal lobe injury risk prediction model according to claim 5, characterized in that, Calculate the spatial correlation between each element in the feature map and all other elements to obtain a correlation matrix, including: The feature map is convolved in three directions to obtain the Q matrix, K matrix and V matrix respectively; Multiply the Q matrix and the K matrix to obtain the correlation weight matrix between the Q matrix and the K matrix; The correlation weight matrix is ​​normalized to obtain the correlation weight coefficient matrix of the Q matrix and the K matrix; Multiplying the correlation weight coefficient matrix with the V matrix yields a correlation matrix representing the spatial correlation between each element in the feature map and all other elements.

7. The method for training a radiation-induced temporal lobe injury risk prediction model according to any one of claims 1-3 and 4-6, characterized in that, The risk prediction loss value is calculated based on the first risk value and the second risk value, including: Calculate the sum of the first risk value and the second risk value to obtain the total risk value; Based on the proportional hazards regression model, the first loss value, the second loss value, and the third loss value corresponding to the first risk value, the second risk value, and the total risk value are calculated respectively. The first loss value, the second loss value, and the third loss value are assigned corresponding weights and added together to obtain the risk prediction loss value.

8. The method for training a radiation-induced temporal lobe injury risk prediction model according to any one of claims 1-3 and 4-6, characterized in that, The radiation-induced temporal lobe injury risk prediction model also outputs a binary classification result indicating whether the CT image sample contains radiation-induced temporal lobe injury. The method further includes: The classification loss value is calculated based on the classification results.

9. The method for training a radiation-induced temporal lobe injury risk prediction model according to claim 8, characterized in that, The parameters of the radiation-induced temporal lobe injury risk prediction model are updated based on the risk prediction loss value, including: Calculate the consistency loss value between the first risk value and the second risk value; Calculate the regularization constraint term of the radiation-induced temporal lobe injury risk prediction model; The parameters of the radiation-induced temporal lobe injury risk prediction model are updated based on the risk prediction loss value, the classification loss value, the consistency loss value, and the regularization constraint term.

10. A device for predicting the risk of radiation-induced temporal lobe injury, characterized in that, The radiation-induced temporal lobe injury risk prediction model applied to any one of claims 1-9 includes: The acquisition module is used to acquire CT images, dose images, clinical information, and dose-volume information; The first region image extraction module is used to extract a first target region image including the temporal lobe and a second target region image from the CT image and the dose image, respectively. The first risk value prediction module is used to input the first target region image and the second target region image into the radiation-induced temporal lobe injury risk prediction model for processing, and to predict the first risk value of radiation-induced temporal lobe injury. The second risk value prediction module is used to predict a second risk value of radiation-induced temporal lobe injury based on the clinical information and the dose-volume information. The total risk value calculation module is used to calculate the sum of the first risk value and the second risk value to obtain the risk value of radiation-induced temporal lobe injury.

11. A training device for a radiation-induced temporal lobe injury risk prediction model, characterized in that, A method for training a radiation-induced temporal lobe injury risk prediction model according to any one of claims 1-9, comprising: The sample acquisition module is used to acquire training samples, which include CT image samples, dose image samples, clinical information samples, and dose-volume information samples. The second region image extraction module is used to extract a first target region image including the temporal lobe and a second target region image from the CT image sample and the dose image sample, respectively. The first risk value prediction module is used to input the first target region image and the second target region image into the radiation-induced temporal lobe injury risk prediction model for processing, and to predict the first risk value of radiation-induced temporal lobe injury. The second risk value prediction module is used to predict a second risk value of radiation-induced temporal lobe injury based on the clinical information sample and the dose-volume information sample. The risk prediction loss value calculation module is used to calculate the risk prediction loss value based on the first risk value and the second risk value; The parameter update module is used to update the parameters of the radiation-induced temporal lobe injury risk prediction model based on the risk prediction loss value.

12. A computer device, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the training method for the radiation-induced temporal lobe injury risk prediction model as described in any one of claims 1-9.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the method for training a radiation-induced temporal lobe injury risk prediction model as described in any one of claims 1-9.

Images