A data processing method, device, apparatus, and storage medium

Abstract

The application provides a data processing method, device and equipment and a storage medium. The method comprises: acquiring a pre-treatment magnetic resonance scan image of a patient, wherein the scan image is a current high-speed low-quality image (HSLQ); inputting the scan image into a deep learning image processing model to obtain a current generated low-speed high-quality image (LSHQ) corresponding to the scan image, wherein the deep learning image processing model is formed by training a cyclic consistent generative adversarial network; registering the current generated LSHQ and a reference image; evaluating a change in a volume and position of a critical organ of the patient based on a registration result, and determining an adaptive plan type for treatment of the patient based on an evaluation result, wherein different adaptive plan types correspond to different plan information. The data processing method can shorten the scan time, improve the quality of the pre-treatment magnetic resonance scan image, and further shorten the total length of the online adaptive radiotherapy process.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a data processing method, apparatus, device and storage medium. Background Technology

[0002] MR-linac is a novel linear accelerator integrated with a magnetic resonance imaging (MRI) system. It can create online adaptive treatment plans based on daily changes in patient anatomy during each fractionation. Unity (Elekta)'s imaging system currently offers 1.5T low-speed high-quality (LSHQ) and high-speed low-quality (HSLQ) MRI scan sequences. The HSLQ sequence has a significantly shorter scan time (approximately 137s) than the LSHQ sequence (approximately 397s); however, the poorer image quality of HSLQ does not meet clinical requirements. The LSHQ sequence can provide high-quality daily MRI images, but its scan time is too long for adaptive radiotherapy. This time-consuming issue is a major challenge in MRI-based adaptive radiotherapy, as longer scan times can affect patient comfort and reduce treatment efficiency.

[0003] One approach to reducing scan time is to use high-speed acquisition sequences to acquire undersampled MRI images, followed by post-processing to improve their quality. Several traditional methods have been proposed to recover undersampled MRI images for high-quality results. These methods, including linear regression, compressed sensing, and random forests, can provide higher quality and contrast MRI images with less scan time. These traditional techniques are capable of generating MRI images with high signal-to-noise ratios and short scan times. However, the difficulty in extracting handcrafted features and the high challenge of parameter tuning limit their application in routine clinical settings.

[0004] Another approach to obtaining high-quality MR images from low-quality MR images is to apply deep learning methods. Some studies have used generative adversarial networks (GANs) to generate high-quality MR images from low-quality input MR images. Furthermore, some studies have reported on using U-net to improve the quality of undersampled MRI images. These studies offer several potential solutions to address the problem of long MRI scan times in routine MR-linac applications. However, the difficulty in collecting paired training data from real-world patients in MR-linac also limits the application of deep learning in the rapid generation of high-quality MR images. Summary of the Invention

[0005] This invention provides a data processing method, apparatus, device, and storage medium that can shorten scanning time, improve the quality of pre-treatment magnetic resonance imaging images, and thus shorten the total duration of online adaptive radiotherapy procedures.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a data processing method, including:

[0007] Acquire the patient's pre-treatment magnetic resonance imaging (MRI) images, which are currently high-speed, low-quality (HSLQ) images.

[0008] The scanned image is input into a deep learning image processing model to obtain a low-speed, high-quality image (LSHQ) corresponding to the scanned image. The deep learning image processing model is formed by training a cycle-consistent generative adversarial network.

[0009] The currently generated LSHQ image is registered with the reference image;

[0010] Based on the registration results, the patient's organ at risk volume and position changes are assessed, and the type of adaptive treatment plan for the patient is determined based on the assessment results. Different types of adaptive treatment plans correspond to different plan information.

[0011] As an optional embodiment, it also includes:

[0012] A model architecture is established based on the aforementioned cycle-consistent generative adversarial network;

[0013] Obtain training data, which includes multiple pairs of historical HSLQ and historical LSHQ corresponding to different patients;

[0014] The model architecture is trained based on the training data to obtain the deep learning image processing model.

[0015] As an optional embodiment, training the model architecture based on the training data includes:

[0016] The model architecture is trained using the training data and based on the five-fold cross-validation method.

[0017] As an optional embodiment, the step of combining the training data and training the model architecture based on five-fold cross-validation includes:

[0018] The training data is folded five times, with the historical HSLQ and historical LSHQ corresponding to the same patient being placed in the same fold.

[0019] In the first iteration, four folds of data are selected to train the model architecture, and the remaining one fold of data is used to test the trained model architecture.

[0020] Repeat the above iterative process until each fold of data has been used as the model architecture after test training.

[0021] As an optional embodiment, it also includes:

[0022] Obtain the standard LSHQ corresponding to the current HSLQ;

[0023] The quality of the currently generated LSHQ, which is generated by the deep learning image processing model, is evaluated based on the standard LSHQ.

[0024] The image quality assessment includes the normalized mean absolute error, structural similarity index, peak signal-to-noise ratio, and edge preservation index of comparing the standard LSHQ with the currently generated LSHQ.

[0025] As an optional embodiment, the registration of the currently generated LSHQ with the reference image includes:

[0026] Obtain the planned CT images for the corresponding patient;

[0027] The currently generated LSHQ images are deformably registered to the planned CT images;

[0028] Calculate the Jacobian determinant value of the image after the currently generated LSHQ is registered to the planned CT image;

[0029] The registration result is determined based on the Jacobian determinant value.

[0030] Another embodiment of the present invention also provides a data processing apparatus, comprising:

[0031] The acquisition module is used to acquire the patient's pre-treatment magnetic resonance imaging (MRI) scan image, which is a current high-speed low-quality (HSLQ) image.

[0032] The input module is used to input the scanned image into a deep learning image processing model to obtain the currently generated low-speed high-quality image LSHQ corresponding to the scanned image;

[0033] The registration module is used to register the current LSHQ image with the reference image;

[0034] The assessment module is used to evaluate the changes in the volume and location of the patient's organs at risk based on the registration results, and to determine the type of adaptive plan to be used for the patient's treatment based on the assessment results. Different types of adaptive plans correspond to different plan information.

[0035] As an optional embodiment, it also includes:

[0036] Build modules are used to establish model architectures based on cycle-consistent generative adversarial networks;

[0037] The acquisition module is used to acquire training data, which includes multiple pairs of historical HSLQ and historical LSHQ corresponding to different patients;

[0038] The training module is used to train the model architecture based on the training data to obtain the deep learning image processing model.

[0039] Another embodiment of the present invention also provides an electronic device, comprising:

[0040] At least one processor; and,

[0041] A memory communicatively connected to the at least one processor; wherein,

[0042] The memory stores instructions that can be executed by the at least one processor to implement the data processing method described above.

[0043] Another embodiment of the present invention provides a computer-readable storage medium that, when the instructions in the storage medium are executed by a processor corresponding to an electronic device, enables the electronic device to implement the data processing method described above.

[0044] As can be seen from the disclosure of the above embodiments, the beneficial effects of the embodiments of the present invention include: by constructing a deep learning image processing model, the model can efficiently and accurately process rapidly acquired low-quality pre-treatment magnetic resonance imaging (MRI) scan images, and generate low-speed, high-quality images required for treatment. Based on these images, the device can directly register them with reference images, and quickly and accurately determine the type of adaptive treatment plan suitable for the patient based on the registration results, thus determining the specific treatment plan. The above process significantly shortens the MRI scan time and is simple to execute, therefore it can not only be widely used in clinical practice, but also significantly improves the execution efficiency of online adaptive radiotherapy procedures.

[0045] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0046] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0047] The accompanying drawings are provided to further illustrate the present application and form part of the specification. They are used together with the embodiments of the present application to explain the application and do not constitute a limitation thereof. In the drawings:

[0048] Figure 1 This is a flowchart of the data processing method in an embodiment of the present invention.

[0049] Figure 2 This is a flowchart illustrating the actual application of the data processing method in this embodiment of the invention.

[0050] Figure 3 This is a flowchart of a data processing method according to another embodiment of the present invention.

[0051] Figure 4 This is a flowchart of a data processing method according to another embodiment of the present invention.

[0052] Figure 5 This is a flowchart of the model architecture training process in an embodiment of the present invention.

[0053] Figure 6 This is a schematic diagram of the model architecture in an embodiment of the present invention, where the generator structure corresponds to the first row, the discriminator structure corresponds to the second row, and the residual block structure corresponds to the third row. The number marked under each network block represents the number of channels. Conv: convolutional layer; DeConv: transposed convolutional layer; BN: batch normalization layer; ReLU: rectified linear unit; Tanh: activation function.

[0054] Figure 7 This is a structural block diagram of the data processing device in an embodiment of the present invention. Detailed Implementation

[0055] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but these are not intended to limit the scope of the invention.

[0056] It should be understood that various modifications can be made to the embodiments disclosed herein. Therefore, the following description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this disclosure will be apparent to those skilled in the art.

[0057] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present disclosure and, together with the general description of the disclosure given above and the detailed description of the embodiments given below, serve to explain the principles of the disclosure.

[0058] These and other features of the invention will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0059] It should also be understood that although the invention has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of the invention, which have the features described in the claims and are therefore all within the scope of protection defined herein.

[0060] The above and other aspects, features and advantages of this disclosure will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0061] Specific embodiments of the present disclosure are described thereafter with reference to the accompanying drawings; however, it should be understood that the disclosed embodiments are merely examples of the present disclosure and can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the present disclosure. Therefore, the specific structural and functional details disclosed herein are not intended to be limiting, but merely to serve as the basis and representative basis for the claims to teach those skilled in the art to use the present disclosure in a variety of substantially any suitable detailed structures.

[0062] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in still another embodiment,” all of which may refer to one or more of the same or different embodiments according to this disclosure.

[0063] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0064] Figure 1 This is a flowchart of the data processing method in an embodiment of the present invention, such as... Figure 1 As shown, an embodiment of the present invention provides a data processing method, including:

[0065] S101: Acquire the patient's pre-treatment magnetic resonance imaging (MRI) image, which is the current high-speed, low-quality (HSLQ) image;

[0066] S102: Input the scanned image into the deep learning image processing model to obtain the currently generated low-speed high-quality image LSHQ corresponding to the scanned image. The deep learning image processing model is formed by training a cycle-consistent generative adversarial network.

[0067] S103: Register the currently generated LSHQ image with the reference image;

[0068] S104: Based on the registration results, assess the changes in the volume and location of the patient's organs at risk, and determine the type of adaptive plan for the patient's treatment based on the assessment results. Different types of adaptive plans correspond to different plan information.

[0069] For example, such as Figure 2As shown, patients can undergo pre-treatment MRI scans after the doctor positions them. The device system can obtain the patient's pre-treatment MRI scan image via a network connection, or the doctor can input the scan image into the device system. This pre-treatment MRI scan image is a current high-speed, low-quality (HSLQ) image; that is, it can be obtained quickly, but its quality is low and cannot be directly used, meaning its reference value is low. Once the device system obtains this scan image, it can input it into a deep learning image processing model. This deep learning image processing model is trained based on a recurrent consistency generative adversarial network and is used to process the current HSLQ scan image into a low-speed, high-quality LSHQ scan image. In other words, it transforms a low-quality scan image into a high-quality one. Generating such a high-quality scan image using previous methods would require a significant amount of time, but its high accuracy makes it highly valuable for reference. Once the current LSHQ is obtained based on the deep learning image processing model, it needs to be registered with the reference image. Then, based on the registration results, the changes in the volume and location of the patient's organs at risk are assessed, and the type of adaptive plan for the patient's treatment is determined based on the assessment results. That is, the changes in the volume and location of organs at risk (OARs) are assessed, and the type of adaptive plan is determined based on the assessment results. Then, the adaptive plan is created, transmitted, and used for treatment. Different types of adaptive plans correspond to different plan information.

[0070] Based on the disclosure of the above embodiments, it can be understood that the beneficial effects of this embodiment include: by constructing a deep learning image processing model, this model can efficiently and accurately process rapidly acquired low-quality pre-treatment magnetic resonance imaging (MRI) scan images, and generate low-speed, high-quality images required for treatment. Based on these images, the device can directly register them with reference images, and quickly and accurately determine the type of adaptive treatment plan suitable for the patient based on the registration results, thus determining the specific treatment plan. The above process significantly shortens the MRI scan time and is simple to execute, therefore it can not only be widely applied in clinical practice but also significantly improves the execution efficiency of online adaptive radiotherapy procedures.

[0071] Furthermore, such as Figure 3 As shown, the method in this embodiment further includes:

[0072] S105: Establishing a model architecture based on generative adversarial networks with cycle consistency;

[0073] S106: Obtain training data, which includes multiple pairs of historical HSLQ and historical LSHQ corresponding to different patients;

[0074] S107: Train the model architecture based on the training data to obtain a deep learning image processing model.

[0075] For example, Figure 5 The LSHQ image generated by the model is denoted as synLSHQ, such as... Figure 5 and Figure 6 As shown, this embodiment uses a CycleGAN (Cycle Consistent Generative Adversarial Network) consisting of two discriminators (D) and two generators (G) to learn the mapping between HSLQ and LSHQ images, i.e., for training. During model testing and usage, HSLQ images are used as input, and LSHQ images are synthesized by the CycleGAN's generators AB.

[0076] Specifically, in this embodiment, the generator includes several downsampling and upsampling operators, as well as nine residual blocks. 3×3 convolutional layers and transposed convolutional layers are used for downsampling and upsampling, respectively. The 7×7 convolutional layers with reflection padding have a kernel count of 0, while the other convolutional layers have a kernel count of 1. The hyperbolic tangent function (Tanh) is used as the activation function before generating the output data. The residual blocks are formed using skip connections. The discriminator is a 70×70 patched GAN network.

[0077] Furthermore, the loss function in the network in this embodiment is as follows:

[0078] The generator GAB network is used to translate data pattern B in the training data into A as close as possible, and the discriminator DB is used to distinguish the synthesized data pattern B from the actual data B. This constitutes the adversarial loss.

[0079]

[0080] Then, the synthesized data style B is translated into cyclic data A by GAB, where cycle-consistent loss is used to preserve the image structure of A. The following formulas represent the two cycles involved in this embodiment:

[0081]

[0082]

[0083] Based on the above-mentioned cyclic and adversarial losses, the complete loss function can be obtained as follows:

[0084]

[0085] Once the model architecture is determined using the methods described above, training data can be obtained and input into the model architecture for training. It is worth noting that the training data in this embodiment consists of multiple pairs of historical HSLQ and historical LSHQ data corresponding to different patients; that is, training data composed of real patient scan data. This high degree of realism allows the trained model to clearly understand the patient's actual scan information based on the patient's HSLQ, and to calculate and adjust accordingly, thereby generating an LSHQ that reflects the patient's actual physical condition. In the actual training process, the training data can be 1.5T MRI image data from real patients using MR-linac. In this embodiment, 30 pairs of images (layers: n = 340 × 30) were evaluated. These 30 pairs of images include 30 sets each of HSLQ and LSHQ images. Each set of images underwent preprocessing to adapt to the requirements of the deep learning network used in this embodiment, such as denoising and normalization, before being input into the model architecture for training.

[0086] Furthermore, in this embodiment, the model architecture trained based on training data includes:

[0087] S201: Combine training data and train the model architecture based on the five-fold cross-validation method.

[0088] Of course, in practical applications, other regression algorithms can also be used to train the model architecture, and it is not necessarily limited to the five-fold cross-validation method. In this embodiment, the five-fold cross-validation method is used because of the algorithm's optimization attribute and its ability to adapt to scenarios with limited training data.

[0089] Specifically, such as Figure 4 As shown, when combining training data and training the model architecture based on five-fold cross-validation, the following steps are included:

[0090] S202: The training data is folded five times, with historical HSLQ and historical LSHQ corresponding to the same patient being placed into the same fold.

[0091] S203: In the first iteration, four folds of data are selected for training the model architecture, and the remaining one fold of data is used to test the trained model architecture.

[0092] S204: Repeat the above iterative process until each fold of data has been used as the tested and trained model architecture.

[0093] For example, this training and subsequent testing were both performed on an Nvidia GeForce RTX 3090 GPU, but the same training and testing could also be performed on other processing devices. Furthermore, during training, the training data needs to be divided into five folds for cross-validation. Data from the same patient must be selected into the same fold. In the first data iteration, the first four folds of data can be used to train the model, and the remaining fold can be used to test the model. This process is repeated until every fold of data in the five folds is registered as part of the test set.

[0094] Optionally, the data processing method in this embodiment further includes:

[0095] S301: Obtain the standard LSHQ corresponding to the current HSLQ;

[0096] S302: Perform quality assessment on the currently generated LSHQ based on the standard LSHQ generated by the deep learning image processing model;

[0097] S303: Image quality assessment includes the normalized mean absolute error, structural similarity index, peak signal-to-noise ratio, and edge preservation index of the comparison between the standard LSHQ and the currently generated LSHQ.

[0098] Specifically, the LSHQ image corresponding to the current HSLQ, i.e., the standard LSHQ, can be set as the reference image to evaluate the image quality of the current HSLQ and the synLSHQ generated by the deep learning image processing model. The method of obtaining this labeled LSHQ is not unique; for example, it can be formed based on long-time scanning methods used in existing technologies. Specific metrics for image quality evaluation may include, but are not limited to, normalized mean absolute error (nMAE), structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and edge preservation index (EKI). Lower nMAE, higher SSIM, higher PSNR, and higher EKI values ​​all indicate better image quality. Specific formulas involved include:

[0099]

[0100] Where I1 and I2 represent the composite image and the reference image with a total of N pixels.

[0101] SSIM represents the similarity between a reference image and a synthetic image (i.e., an image generated by a model). A higher SSIM value indicates less image distortion. SSIM is defined as:

[0102]

[0103] m x and These are the mean and variance of x, respectively; m v and These are the mean and variance of y, respectively; s xv Let x be the covariance of x and y; c1 = (k1L) 2 k1 = 0.01, c2 = (k2L) 2 k2 = 0.03, L is a value of 2 B The range of pixel values ​​of -1.

[0104] PSNR describes the ratio of the maximum possible power of a signal to the power of destructive noise that affects its representation accuracy. PSNR is usually expressed in logarithmic decibels (dBs). A higher PSNR value indicates better image quality. The relevant formulas include:

[0105]

[0106] MAX I 2 This represents the maximum possible pixel value of the image, and m and n describe the size of the image.

[0107] The Edge Preservation Index (EKI) is used to assess edge preservation capability, and the calculation formula is as follows:

[0108]

[0109] g i and f i Representing the reference image and the composite image respectively, Δg i and Δf i respectively via g i and f i The filtering and mean are μ g and μ f Obtained from the Qualcomm Laplace filter.

[0110] In practical applications, some or all of the above indicators can be selected for quality assessment.

[0111] Furthermore, the currently generated LSHQ image is registered with the reference image, including:

[0112] S401: Obtain the planned CT images for the corresponding patient;

[0113] S402: Deformably register the currently generated LSHQ images to the planned CT images;

[0114] S403: Calculate the Jacobian determinant of the image after the currently generated LSHQ is registered to the planned CT image;

[0115] S404: Determine the registration result based on the Jacobian determinant value.

[0116] Continue to combine Figure 2 As shown, during online adaptive radiotherapy in clinical practice, the synLSHQ (an LSHQ image generated by the model) and the patient's original HSLQ image, i.e., the current HALQ mentioned above, can be deformably registered to the patient's planned CT images. Jacobi determinant value (JDV) and geometric indices are two methods recommended by the AAPM for assessing registration accuracy. In practice, any one or both of these indices can be combined for quality assessment. A mean IDV value closer to 1 indicates better registration accuracy. Geometric indices include dice similarity coefficient (DSC) and mean consistency distance (MDA); higher DSC and lower MDA indicate better image registration accuracy. Furthermore, during quality assessment, experienced clinicians can delineate the clinical target volume (CTV) and organ at risk (OAR) on the planned CT, deformed HSLQ, and synLSHQ images, respectively. The consistency of the corresponding regions of interest on the planned CT and synLSHQ images also reflects the accuracy of registration. All of the above methods can be used. In this embodiment, the Jacobian determinant value is preferred for matching accuracy evaluation, as the calculation process is simpler and faster.

[0117] This embodiment improves the image quality of high-speed imaging in MRI-guided radiotherapy by using deep learning methods. This saves 59.8% of the total generation time during image scanning, further improving efficiency and reducing the acquisition time of high-quality MRI radiotherapy (HSLQ) images. Furthermore, compared to HSLQ images, the model-generated HSLQ images show significant improvements in nMAE (57%), SSIM (4.6%), PSNR (28.1%), and EKI (5.6%), ensuring the image quality of HSLQ images. In addition, the generated HSLQ images also show improved registration accuracy, superior average JDV (2.3%), and better geometric indices, making them highly valuable for clinical applications.

[0118] like Figure 7 As shown, another embodiment of the present invention also provides a data processing apparatus, including:

[0119] The acquisition module is used to acquire the patient's pre-treatment magnetic resonance imaging (MRI) scan image, which is a current high-speed low-quality (HSLQ) image.

[0120] The input module is used to input the scanned image into a deep learning image processing model to obtain the currently generated low-speed high-quality image LSHQ corresponding to the scanned image;

[0121] The registration module is used to register the current LSHQ image with the reference image;

[0122] The assessment module is used to evaluate the changes in the volume and location of the patient's organs at risk based on the registration results, and to determine the type of adaptive plan to be used for the patient's treatment based on the assessment results. Different types of adaptive plans correspond to different plan information.

[0123] As an optional embodiment, the apparatus in this embodiment further includes:

[0124] Build modules are used to establish model architectures based on cycle-consistent generative adversarial networks;

[0125] The first acquisition module is used to acquire training data, which includes multiple pairs of historical HSLQ and historical LSHQ corresponding to different patients;

[0126] The training module is used to train the model architecture based on the training data to obtain the deep learning image processing model.

[0127] As an optional embodiment, training the model architecture based on the training data includes:

[0128] The model architecture is trained using the training data and based on the five-fold cross-validation method.

[0129] As an optional embodiment, the step of combining the training data and training the model architecture based on five-fold cross-validation includes:

[0130] The training data is folded five times, with the historical HSLQ and historical LSHQ corresponding to the same patient being placed in the same fold.

[0131] In the first iteration, four folds of data are selected to train the model architecture, and the remaining one fold of data is used to test the trained model architecture.

[0132] Repeat the above iterative process until each fold of data has been used as the model architecture after test training.

[0133] As an optional embodiment, the apparatus in this embodiment further includes:

[0134] The second obtaining module obtains the standard LSHQ corresponding to the current HSLQ;

[0135] The processing module is used to perform a quality assessment on the current LSHQ generated by the deep learning image processing model according to the standard LSHQ.

[0136] The image quality assessment includes the normalized mean absolute error, structural similarity index, peak signal-to-noise ratio, and edge preservation index of the comparison between the standard LSHQ and the current LSHQ.

[0137] As an optional embodiment, the registration of the current LSHQ with the reference image includes:

[0138] Obtain the planned CT images for the corresponding patient;

[0139] The current LSHQ is deformably registered to the planned CT image;

[0140] Calculate the Jacobian determinant value of the image after the current LSHQ is registered to the planned CT image;

[0141] The registration result is determined based on the Jacobian determinant value.

[0142] Another embodiment of the present invention also provides an electronic device, comprising:

[0143] One or more processors;

[0144] Memory, configured to store one or more programs;

[0145] When the one or more programs are executed by the one or more processors, the one or more processors perform the above processing method.

[0146] An embodiment of the present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the processing method described above. It should be understood that the various solutions in this embodiment have the corresponding technical effects in the above-described method embodiments, and will not be repeated here.

[0147] This invention also provides a computer program product, which is tangibly stored on a computer-readable medium and includes computer-readable instructions. When executed, these computer-executable instructions cause at least one processor to perform processing methods as described in the embodiments above. It should be understood that the various solutions in this embodiment have the corresponding technical effects in the above-described method embodiments, and will not be repeated here.

[0148] It should be noted that the computer storage medium of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. Computer-readable media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access storage media (RAM), read-only storage media (ROM), erasable programmable read-only storage media (EPROM or flash memory), optical fibers, portable compact disk read-only storage media (CD-ROM), optical storage media, magnetic storage media, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program configured for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, antenna, optical fiber, RF, etc., or any suitable combination thereof.

[0149] It should be understood that although this application is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0150] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its spirit and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A data processing method, characterized in that, include: Acquire the patient's pre-treatment magnetic resonance imaging (MRI) images, which are currently high-speed, low-quality (HSLQ) images. The scanned image is input into a deep learning image processing model to obtain a low-speed, high-quality image (LSHQ) corresponding to the scanned image. The deep learning image processing model is formed by training a cycle-consistent generative adversarial network. Registering the currently generated LSHQ image with the reference image includes: Obtain the planned CT images for the corresponding patient; The currently generated LSHQ image is deformably registered to the planned CT image; Calculate the Jacobian determinant value of the image after the currently generated LSHQ is registered to the planned CT image; The registration result is determined based on the Jacobian determinant value; Based on the registration results, the patient's organ at risk volume and position changes are assessed, and the type of adaptive plan for the patient's treatment is determined based on the assessment results. Different types of adaptive plans correspond to different plan information. A model architecture is established based on the aforementioned cycle-consistent generative adversarial network; Obtain training data, which includes multiple pairs of historical HSLQ and historical LSHQ corresponding to different patients; The model architecture is trained based on the training data to obtain the deep learning image processing model; wherein, training the model architecture includes: The model architecture is trained using the training data and based on the five-fold cross-validation method. The training data is folded five times, with the historical HSLQ and historical LSHQ corresponding to the same patient being placed in the same fold. In the first iteration, four folds of data are selected to train the model architecture, and the remaining one fold of data is used to test the trained model architecture. Repeat the above iterative process until each fold of data has been used as the model architecture after test training.

2. The method according to claim 1, characterized in that, Also includes: Obtain the standard LSHQ corresponding to the current HSLQ; The quality of the currently generated LSHQ, which is generated by the deep learning image processing model, is evaluated based on the standard LSHQ. The image quality assessment includes at least one of the following: normalized mean absolute error, structural similarity index, peak signal-to-noise ratio, and edge preservation index, comparing the standard LSHQ and the currently generated LSHQ.

3. A data processing apparatus, characterized in that, include: The acquisition module is used to acquire the patient's pre-treatment magnetic resonance imaging (MRI) scan image, which is a current high-speed low-quality (HSLQ) image. The input module is used to input the scanned image into a deep learning image processing model to obtain the currently generated low-speed high-quality image LSHQ corresponding to the scanned image; Build modules are used to establish model architectures based on cycle-consistent generative adversarial networks; The registration module is used to register the currently generated LSHQ image with the reference image, including: Obtain the planned CT images for the corresponding patient; The currently generated LSHQ image is deformably registered to the planned CT image; Calculate the Jacobian determinant value of the image after the currently generated LSHQ is registered to the planned CT image; The registration result is determined based on the Jacobian determinant value; The assessment module is used to assess changes in the volume and location of the patient's organs at risk based on the registration results, and to determine the type of adaptive treatment plan for the patient based on the assessment results. Different types of adaptive treatment plans correspond to different plan information. A model architecture is established based on the aforementioned cycle-consistent generative adversarial network; The acquisition module is used to acquire training data, which includes multiple pairs of historical HSLQ and historical LSHQ corresponding to different patients; The training module, used to train the model architecture based on the training data to obtain the deep learning image processing model, includes: The model architecture is trained using the training data and based on the five-fold cross-validation method. The training data is folded five times, with the historical HSLQ and historical LSHQ corresponding to the same patient being placed in the same fold. In the first iteration, four folds of data are selected to train the model architecture, and the remaining one fold of data is used to test the trained model architecture. Repeat the above iterative process until each fold of data has been used as the model architecture after test training.

4. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to implement the data processing method as described in claim 1 or 2.

5. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor corresponding to the electronic device, the electronic device is able to implement the data processing method as described in claim 1 or 2.

Images