Method for truncated cbct to ct generation based on dual-encoder generative adversarial network

Abstract

The application provides a method for generating a truncated CBCT synthetic CT based on a double-encoder generative adversarial network, which comprises the following steps: performing rigid registration on CBCT and pCT, performing cropping, normalization and scaling to form a training sample; constructing an MSCBCT-Net containing two generators and two discriminators, inputting CBCTt and pCT into the first generator to obtain sCT; then inputting sCT and complete CBCT into the second generator to obtain a pseudo CBCT, calculating a cycle consistency loss, and calculating an adversarial loss through the discriminators, combining an identity loss, a structure consistency loss, a gradient edge loss and a histogram matching loss, and training by using LSGAN; directly inputting a CBCTt image to be processed and a corresponding pCT image into the trained MSCBCT-Net to output a synthetic CT, which is used for adaptive radiotherapy dose calculation. The application provides a high-quality, large-view synthetic CT image that can be directly used for dose calculation for adaptive radiotherapy, and improves the accuracy and safety of radiotherapy planning.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network. Background Technology

[0002] Cone-beam computed tomography (CBCT), as the core positioning technology for image-guided radiotherapy (IGRT), still has several key shortcomings in clinical practice, especially in scenarios such as breast cancer that require high-precision adaptive radiotherapy. These limitations seriously affect its application effectiveness.

[0003] Due to the physical limitations of the accelerator gantry, the scanning field of view of CBCT is significantly smaller than that of conventional diagnostic CT. When dealing with larger patients or lateral target areas such as breast cancer, the patient's arm, body contour, and some anatomical structures often exceed the detector's acquisition range, resulting in severe lateral truncation of the projected data. This data loss causes obvious truncation artifacts in the reconstructed images, affecting not only the visualization and identification of the target area and organs at risk, but also directly leading to subsequent image registration misalignment and dose calculation errors, thus threatening the accuracy of radiotherapy.

[0004] To reduce the cumulative radiation dose to patients during fractionated treatments, low-dose or rapid scanning protocols are commonly used in clinical practice. However, this inevitably sacrifices image quality. After reconstruction from projection data with a low signal-to-noise ratio, the images are filled with noise and scattering artifacts, resulting in reconstructed CBCT images riddled with noise and scattering artifacts, extremely low soft tissue contrast, and Hounsfield Unit (HU) values ​​that deviate significantly from the accuracy of real CT. This degradation in image quality makes it difficult to use directly for dose calculation, severely limiting the application value of CBCT in adaptive radiotherapy planning adjustments.

[0005] Existing image enhancement methods have also failed to adequately address the aforementioned issues. Traditional interpolation- or model-based extrapolation techniques typically assume homogeneous tissue outside the field of view, failing to accurately reconstruct heterogeneous structures such as bone, air, or implants, and easily leading to anatomical distortion. While deep learning-based supervised learning methods can improve image appearance quality, they are highly dependent on precisely registered paired CT and CBCT data as labels. Due to the physiological movements, positional changes, and organ deformations of patients between treatment sessions, obtaining an ideally registered dataset is extremely difficult.

[0006] Currently, unsupervised learning methods such as CycleGANs, while reducing reliance on paired data, still suffer from fundamental flaws. Taking the typical approach proposed by Liang et al. in 2019 (see the paper "Generating synthesized computed tomography (CT) from cone-beam computed tomography (CBCT) using CycleGAN for adaptive radiation therapy") as an example, this framework achieves unsupervised image transformation through a bidirectional generator and cycle consistency loss. However, its single-encoder structure struggles to collaboratively model the degradation features of CBCT and the anatomical priors of CT: when processing truncated CBCT, the model tends to generate compensations in missing regions that do not conform to the actual anatomical structure; its ability to suppress low-dose noise is also limited, often leaving strip-like artifacts in the generated synthetic CT. More importantly, the cycle consistency constraint of this framework easily leads to the generator simply replicating the global texture features of the planned CT, failing to effectively retain key information reflecting the specific anatomical state of the day in the CBCT (such as organ displacement, tissue deformation, etc.), the so-called texture duplication phenomenon, thus introducing systematic bias in subsequent dose recalculation.

[0007] Therefore, existing technologies struggle to accurately preserve the true anatomical information of a given CBCT scan while eliminating truncation artifacts and noise interference. There is an urgent need to develop an image generation method that can intelligently fuse prior knowledge of the planned CT scan with the sparse features of the CBCT scan on the current day, in order to construct a full-field, high-quality synthetic CT image that can be used for precise dose calculation. Summary of the Invention

[0008] To address the above issues, this invention utilizes a dual-encoder network within an unsupervised learning framework. It employs truncated low-dose CBCT images and planning CT images as dual inputs to effectively fuse features from both. Furthermore, it introduces anatomical content loss constraints, enabling the generated synthetic CT to achieve CT-grade accuracy in HU values, low noise, and high contrast. Simultaneously, it strictly preserves the true anatomical morphology of the day as seen in the CBCT, while eliminating truncation and scattering artifacts. Ultimately, this provides high-quality, wide-field-of-view synthetic CT images directly usable for dose calculation in adaptive radiotherapy such as for breast cancer, improving the accuracy and safety of radiotherapy planning.

[0009] According to embodiments of the present invention, a method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network is provided.

[0010] In a first aspect of the invention, a method for truncated CBCT synthetic CT generation based on a dual-encoder generative adversarial network is provided. The method includes: Step S01: Perform rigid registration on CBCT and pCT, crop and normalize, and scale to form training samples; Step S02: Construct MSCBCT-Net, which contains two generators and two discriminators. Input CBCTt and pCT into the first generator to obtain sCT; then input sCT and the complete CBCT into the second generator to obtain pseudo CBCT. Calculate the cycle consistency loss and the adversarial loss through the discriminator. Combine identity loss, structural consistency loss, gradient edge loss and histogram matching loss, and train using LSGAN. Step S03: The CBCTt image to be processed and the corresponding pCT image are directly input into the trained MSCBCT-Net output to synthesize a CT for adaptive radiotherapy dose calculation.

[0011] Furthermore, the first generator takes a CBCT image and a pCT image as input, each of which is input into an initial convolutional layer to obtain two feature maps. These two feature maps are then input into four MSE modules to output feature maps. and feature map Feature map and feature map The input CAF module outputs a fused feature map, which is then input into the PFM module and passed through the output convolutional layer to generate the final synthesized CT image.

[0012] Furthermore, the specific processing steps of the MSE module are as follows: After feature mapping is completed by three parallel dilated convolution branches, instance normalization and LeakyReLU activation layers are applied to the output feature map. The dilation rates of the dilated convolutions are 1, 2 and 3, and the convolution kernels are all 3×3. The outputs of each branch are concatenated along the channel dimension, then dimensionality is reduced by 1×1 convolution, and then downsampled by convolution with a stride of 2. After that, instance normalization and LeakyReLU activation layer are performed to output feature maps.

[0013] Furthermore, the specific processing steps of the CAF module are as follows: The query is generated by three independent 1×1 convolutions. ,key ,value ; Calculate bidirectional cross-attention: , , The two attention outputs are added together, and then a gated weight map (Gate) is generated by a 1×1 convolution followed by a sigmoid function. Add to the original feature residual: Output the fused feature map.

[0014] Furthermore, the PFM module comprises four modulation blocks, each of which operates sequentially: After two layers of bilinear upsampling, deformable convolution, and an SE channel attention module, AdaIN fusion is performed. AdaIN fusion adaptively normalizes and fuses the current features with the skip connection features from the corresponding downsampling stage of the MSE module. The fusion formula is as follows: ,in, For the current features, This is a skip connection feature. and These are the channel mean and standard deviation, respectively.

[0015] Furthermore, the second generator structure described in step S02 is asymmetric to the first generator and is used to reverse synthesize pseudo-CBCT from synthetic CT.

[0016] Furthermore, the first discriminator described in step S02 adopts a multi-scale PatchGAN structure containing three parallel discriminative branches, each with the same structure, including: Two 3×3 convolutions with a stride of 2 and three 4×4 convolutions with a stride of 1. Except for the last convolutional layer which uses Sigmoid activation, all other convolutional layers are followed by batch normalization and LeakyReLU. Each branch outputs a feature map, where each pixel corresponds to the probability that a local region in the input image is true; the outputs of the three branches are weighted and averaged to obtain the final discrimination result.

[0017] Furthermore, the structure of the second discriminator described in step S02 is symmetrical to that of the first discriminator, and it is used to distinguish between real CBCT and synthetic CBCT.

[0018] In a second aspect of the invention, an electronic device is provided. The electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the program to implement the method according to the first aspect of the invention.

[0019] In a third aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method according to the first aspect of the invention.

[0020] This invention utilizes a dual-encoder network to effectively fuse features from truncated low-dose CBCT images and planning CT images within an unsupervised learning framework. By introducing anatomical content loss constraints, the generated synthetic CT achieves CT-level accuracy in HU values, low noise, and high contrast, while strictly preserving the true anatomical morphology of the day as seen in the CBCT. Simultaneously, it eliminates truncation and scattering artifacts. Ultimately, this provides high-quality, wide-field-of-view synthetic CT images that can be directly used for dose calculation in adaptive radiotherapy such as for breast cancer, improving the accuracy and safety of radiotherapy planning.

[0021] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0022] The beneficial effects of this invention are: 1. Using truncated low-dose CBCT and localization CT as dual inputs, the complementary information of the two is fused through an unsupervised adversarial framework. While eliminating truncation artifacts and scattering noise, a synthetic CT is generated that retains the real anatomical structure of CBCT and has CT-level HU accuracy. 2. In the MSCBCT-Net network, the Multi-Scale Extraction (MSE) module uses parallel dilated convolution to capture anatomical and textural features under different receptive fields; the Cross-Attention Fusion (CAF) module realizes bidirectional feature interaction and adaptive gating fusion between CBCT and localization CT during the encoding stage, enhancing the representation ability of modal complementary information; the Progressive Feature Modulation (PFM) module gradually upsamples and modulates features through deformable convolution, SE channel attention and AdaIN fusion during the decoding stage to restore the complete anatomical structure. 3. Experimental results on thoracic tumor and breast cancer data show that the method of the present invention exhibits good synthesis accuracy and generalization ability in the task of synthesizing complete CT from truncated CBCT. Attached Figure Description

[0023] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein: Figure 1 A flowchart of a method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to an embodiment of the present invention is shown; Figure 2 A schematic diagram of the MSCBCT-Net network according to an embodiment of the present invention is shown; Figure 3 A generator according to an embodiment of the present invention is shown. Structural diagram; Figure 4A schematic diagram of a multi-scale PatchGAN structure according to an embodiment of the present invention is shown; Figure 5 The qualitative analysis results of chest data according to an embodiment of the present invention are shown; Figure 6 The qualitative analysis results of breast cancer data according to an embodiment of the present invention are shown; Figure 7 A schematic diagram of a device for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to an embodiment of the present invention is shown. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] According to embodiments of the present invention, a method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network is proposed. By using truncated low-dose CBCT images and planning CT images as dual inputs under an unsupervised learning framework, the dual-encoder generative network effectively fuses the features of both images and introduces anatomical content loss constraints. This enables the generated synthetic CT to obtain accurate HU values, low noise, and high contrast at the CT level, while strictly preserving the true anatomical morphology of the day in the CBCT, and eliminating truncation artifacts and scattering artifacts. Ultimately, this provides high-quality, wide-field-of-view synthetic CT images that can be directly used for dose calculation for adaptive radiotherapy such as breast cancer, improving the accuracy and safety of radiotherapy planning.

[0026] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0027] Figure 1 This is a schematic flowchart of a method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to an embodiment of the present invention. The method includes: Step S01: Perform rigid registration on CBCT and pCT, crop and normalize, and scale to form training samples; Step S02: Construct MSCBCT-Net, which contains two generators and two discriminators. Input CBCTt and pCT into the first generator to obtain sCT; then input sCT and the complete CBCT into the second generator to obtain pseudo CBCT. Calculate the cycle consistency loss and the adversarial loss through the discriminator. Combine identity loss, structural consistency loss, gradient edge loss and histogram matching loss, and train using LSGAN. Step S03: The CBCTt image to be processed and the corresponding pCT image are directly input into the trained MSCBCT-Net output to synthesize a CT for adaptive radiotherapy dose calculation.

[0028] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0029] To provide a clearer explanation of the above-described method for generating truncated CBCT synthetic CT based on dual-encoder generative adversarial networks, a specific embodiment will be used for illustration below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an improper limitation of the present invention.

[0030] The following example will further illustrate the method of truncated CBCT synthetic CT generation based on dual encoder generative adversarial networks.

[0031] Step S01: Perform rigid registration on CBCT and pCT, crop and normalize, and scale to form training samples.

[0032] Data were collected from 100 patients with thoracic tumors (including lung cancer and esophageal cancer) and 20 patients with breast cancer. For thoracic tumor patients, initial localization pCT, complete CBCT, and relocalization CT were performed. For breast cancer patients, initial localization pCT, truncated CBCT, and relocalization CT were performed. Data from 80 thoracic patients were used for training and validation of the proposed MSCBCT-Net network, and data from the remaining 20 thoracic patients and 20 breast cancer patients were used for testing.

[0033] The Otsu thresholding method was used to separate the background, and the in vitro voxels were assigned a value of -1000 HU. Rigid registration was performed on the initial localization CT (pCT) and CBCT, while deformation registration was performed on the relocalization CT (rCT) and CBCT (CBCT was a fixed image). All images were resampled to a uniform size and voxel size, and intensity values ​​were clipped and normalized to [-1, 1]. For truncated CBCT images of breast cancer, body contours were extracted as a mask to constrain deformation registration to only apply to the truncated region.

[0034] Step S02: Construct MSCBCT-Net, which includes two generators and two discriminators, to process CBCT... t The pCT is input into the first generator to obtain the sCT; then the sCT and the complete CBCT are input into the second generator to obtain the pseudo CBCT. The cycle consistency loss is calculated, and the adversarial loss is calculated through the discriminator. The LSGAN is used for training by combining the identity loss, structural consistency loss, gradient edge loss and histogram matching loss.

[0035] Construct a multi-scale cross-attention bidirectional CT synthesis network (MSCBCT-Net), such as Figure 2 As shown, MSCBCT-Net employs a generative adversarial architecture, containing two generators. , and two discriminators , .

[0036] generator Used to extract low-dose CBCT images (CBCT t High-quality synthetic CT (sCT) is synthesized from planned CT images (pCT). Input is CBCT. t And pCT, the output is sCT.

[0037] generator Structure and Generator Asymmetric, used to reverse synthesize pseudo-CBCT from synthetic CT (or real CT) to support cyclic consistency constraints.

[0038] Discriminator Used to distinguish between real CT and synthetic CT.

[0039] Discriminator Structure and Discriminator Symmetry is used to distinguish between real CBCT and synthetic CBCT.

[0040] The entire network is trained unsupervised through forward loop (CBCT→CT→CBCT) and backward loop (CT→CBCT→CT), without the need for paired CT-CBCT data.

[0041] like Figure 3 As shown, generator It adopts an asymmetric dual-path design, abandons the traditional dual encoder U-Net, and introduces multi-scale feature extraction, cross-attention fusion and progressive feature modulation mechanism.

[0042] generator The input is CBCT t Both the image and the pCT image are 512×512×1. They are each passed through an initial convolutional layer (7×7 kernel, stride 1, 64 output channels, followed by batch normalization and LeakyReLU activation) to obtain feature maps of 512×512×64 respectively.

[0043] The two feature maps are input into four MSE modules respectively, and the output feature maps are... and feature map The output dimensions and number of channels are as follows: 256×256×128, 128×128×256, 64×64×512, and 32×32×512, respectively.

[0044] Specifically, the processing steps of the MSE module are as follows: After feature mapping is completed using three parallel dilated convolution branches, instance normalization (IN) and LeakyReLU activation layers are applied to the output feature map. The dilation rates of the dilated convolutions are 1, 2, and 3, and the convolution kernels are all 3×3.

[0045] The outputs of each branch are concatenated along the channel dimension, then dimensionality is reduced by 1×1 convolution, and then downsampled by convolution with a stride of 2. After that, instance normalization and LeakyReLU activation layer are performed to output feature maps.

[0046] This multi-scale design can effectively extract anatomical and textural features from different receptive fields.

[0047] The two feature maps processed by four MSE modules and Input the CAF module, and CAF enables bidirectional feature interaction: The query is generated by three independent 1×1 convolutions. ),key( ),value( ), keeping the number of channels unchanged.

[0048] Calculate bidirectional cross-attention: , , The two attention outputs are added together, and then a gated weight map (Gate) is generated by a 1×1 convolution followed by a sigmoid function.

[0049] Add to the original feature residual: The output fused feature map size remains 32×32×512.

[0050] This module enables the network to adaptively determine how much anatomical structure to retain from CBCT and how much texture information to borrow from pCT, effectively avoiding texture duplication.

[0051] The fused feature map is input to the PFM module and output through the output convolutional layer (3×3 convolution, stride 1, output channel 1, Tanh activation) to generate the final synthesized CT image (512×512×1).

[0052] Specifically, the PFM module contains four modulation blocks, each of which is performed sequentially: Two-layer bilinear upsampling (Bilinear Upsample×2): Gradually restores spatial resolution.

[0053] Deformable Conv (3×3 convolution, stride 1, followed by InstanceNorm and LeakyReLU): Adaptively adjusts the sampling position to better handle truncated boundaries and irregular deformations.

[0054] SE Channel Attention Module: Recalibrates channel weights through compression-excitation operations to enhance useful features.

[0055] AdaIN fusion: Adaptively integrates the current feature with the skip connection features from the corresponding downsampling stage of the MSE module using instance normalization. The fusion formula is as follows: ,in, For the current features, This is a skip connection feature. and These are the channel mean and standard deviation, respectively.

[0056] The output dimensions and number of channels of the four modulation blocks are as follows: 64×64×256, 128×128×128, 256×256×64, and 512×512×32, respectively.

[0057] To improve the ability to distinguish between low-dose noise and truncation artifacts, the discriminator... A multi-scale PatchGAN structure is adopted. For example... Figure 4 As shown, it contains three parallel discriminant branches: First branch L1: Process the original size (512×512) input.

[0058] The second branch L2: processes half of the downsampled (256×256) input.

[0059] The third branch, L3, processes a quarter-downsampled (128×128) input.

[0060] Each branch has the same structure: First, the input is processed by two 3×3 convolutions with a stride of 2, increasing the number of channels from input channel 1 to 64 and then to 128.

[0061] After three more 4×4 convolutions with a stride of 1, the number of channels is 256, 512, and 1 respectively.

[0062] Except for the last convolutional layer which uses Sigmoid activation, all other convolutional layers are followed by batch normalization and LeakyReLU (negative slope 0.2).

[0063] Each branch outputs a feature map, where each pixel corresponds to the probability that a local region in the input image is true.

[0064] The outputs of the three branches are weighted and averaged to obtain the final discrimination result. This multi-scale design allows the discriminator to simultaneously focus on global structure and local texture details, more effectively guiding the generator to synthesize high-quality CT images. In this embodiment, the weighting coefficients for the three branches are 0.5, 0.3, and 0.2, respectively.

[0065] Step S03: Process the CBCT scan to be processed t The image and the corresponding pCT image are directly input into the trained MSCBCT-Net output to synthesize a CT, which is used for adaptive radiotherapy dose calculation.

[0066] To verify the effectiveness of MSCBCT-Net in synthetic CT (sCT), this embodiment compares MSCBCT-Net with two advanced generative adversarial networks (UnetGAN and ResGAN). All models were trained on the same complete chest CBCT dataset and then applied to both the chest test set and the breast cancer CBCT test set (for breast cancer data, due to the lack of a complete CBCT dataset, the pre-trained model was used directly for generalization testing). The three methods share the same network architecture, differing only in generator design. Quantitative evaluation metrics include mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM), and paired t-tests (significance level α=0.05) were used for statistical comparison.

[0067] Table 1. Comparison of quantitative indicators of different methods for the thoracic tumor test set.

[0068] Tables 1 and 2 present the quantitative results for the chest and breast test sets, respectively. The data are expressed as mean ± standard deviation. As shown in Table 1, all sCT methods significantly reduced MAE and improved PSNR and SSIM compared to the original CBCT, indicating a significant improvement in image quality. Among them, MSCBCT-Net achieved the best results, with an MAE of 21.03 ± 3.26 HU, significantly lower than UnetGAN (23.06 ± 3.90 HU) and ResGAN (22.47 ± 3.57 HU). Paired t-tests showed that MSCBCT-Net differed significantly from the other two methods in both MAE and PSNR (p < 0.05); in SSIM, it differed significantly from UnetGAN (p < 0.05), but not significantly from ResGAN (p = 0.369).

[0069] Table 2 Comparison of quantitative indicators of different methods in breast cancer testing sets (without original CBCT baseline)

[0070] After generalization to breast data, the indicators of each method in Table 2 decreased slightly, but MSCBCT-Net still maintained the best performance, with an MAE of 25.04±5.89 HU, a PSNR of 29.40±1.58 dB, and an SSIM of 93.81±2.28%, which were significantly better than UnetGAN and ResGAN in all indicators (p<0.05).

[0071] In terms of qualitative analysis, the results of generating chest data are as follows: Figure 5 As shown, rigidly registered CBCT and planned CT (pCT) exhibit structural differences in liver, heart, and body contours. Compared to CBCT, the sCT generated by all three methods effectively reduces scattering and noise artifacts, while restoring textures similar to CT and preserving the anatomical structures of CBCT, with natural transitions in truncated regions. Among them, MSCBCT-Net removes scattering artifacts more thoroughly than UnetGAN and ResGAN, and can more reasonably reference the texture information of pCT, achieving a more realistic synthetic effect while preserving details of the heart and lungs. For breast cancer data ( Figure 6 All methods can eliminate artifacts and preserve anatomical structures, but UnetGAN and ResGAN produce obvious dark stripes at the truncation boundary, reducing the realism of the synthesized image; while MSCBCT-Net shows stronger generalization ability, with smooth and realistic transitions in its truncation region, and reasonably integrates the texture features of pCT while preserving the CBCT structure.

[0072] Both qualitative and quantitative results show that MSCBCT-Net consistently outperforms the two baseline methods, UnetGAN and ResGAN, in synthetic CT tasks using chest and breast data, generating higher-quality synthetic CT images. This verifies the effectiveness and superior generalization performance of our method in artifact elimination and structural fidelity.

[0073] like Figure 7 As shown, the device includes a central processing unit (CPU), which can perform various appropriate actions and processes based on computer program instructions stored in read-only memory (ROM) or loaded from storage units into random access memory (RAM). The RAM can also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0074] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0075] The processing unit executes the various methods and processes described above, such as method steps S01 to S03. For example, in some embodiments, method steps S01 to S03 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of method steps S01 to S03 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute method steps S01 to S03 by any other suitable means (e.g., by means of firmware).

[0076] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0077] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0078] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0079] Furthermore, although the operations are described in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0080] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network, characterized in that, The method includes: Step S01: Perform rigid registration on CBCT and pCT, crop and normalize, and scale to form training samples; Step S02: Construct MSCBCT-Net, which contains two generators and two discriminators. Input CBCTt and pCT into the first generator to obtain sCT; then input sCT and the complete CBCT into the second generator to obtain pseudo CBCT. Calculate the cycle consistency loss and the adversarial loss through the discriminator. Combine identity loss, structural consistency loss, gradient edge loss and histogram matching loss, and train using LSGAN. Step S03: The CBCTt image to be processed and the corresponding pCT image are directly input into the trained MSCBCT-Net output to synthesize a CT for adaptive radiotherapy dose calculation.

2. The method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to claim 1, characterized in that, The first generator takes a CBCT image and a pCT image as inputs. Each image is input into an initial convolutional layer to obtain two feature maps. These two feature maps are then input into four MSE modules to output feature maps. and feature map Feature map and feature map The input CAF module outputs a fused feature map, which is then input into the PFM module and passed through the output convolutional layer to generate the final synthesized CT image.

3. The method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to claim 2, characterized in that, The specific processing steps of the MSE module are as follows: After feature mapping is completed by three parallel dilated convolution branches, instance normalization and LeakyReLU activation layers are applied to the output feature map. The dilation rates of the dilated convolutions are 1, 2 and 3, and the convolution kernels are all 3×3. The outputs of each branch are concatenated along the channel dimension, then dimensionality is reduced by 1×1 convolution, and then downsampled by convolution with a stride of 2. After that, instance normalization and LeakyReLU activation layer are performed to output feature maps.

4. The method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to claim 2, characterized in that, The specific processing steps of the CAF module are as follows: The query is generated by three independent 1×1 convolutions. ,key ,value ; Calculate bidirectional cross-attention: ， ， The two attention outputs are added together, and then a gated weight map (Gate) is generated by a 1×1 convolution followed by a sigmoid function. Add to the original feature residual: Output the fused feature map.

5. The method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to claim 2, characterized in that, The PFM module contains four modulation blocks, each of which operates sequentially: After two layers of bilinear upsampling, deformable convolution, and an SE channel attention module, AdaIN fusion is performed. AdaIN fusion adaptively normalizes and fuses the current features with the skip connection features from the corresponding downsampling stage of the MSE module. The fusion formula is as follows: ,in, For the current features, This is a skip connection feature. and These are the channel mean and standard deviation, respectively.

6. The method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to claim 1, characterized in that, The second generator structure described in step S02 is asymmetric to the first generator and is used to reverse synthesize pseudo-CBCT from synthetic CT.

7. The method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to claim 1, characterized in that, The first discriminator described in step S02 adopts a multi-scale PatchGAN structure containing three parallel discriminative branches, each with the same structure, including: Two 3×3 convolutions with a stride of 2 and three 4×4 convolutions with a stride of 1. Except for the last convolutional layer which uses Sigmoid activation, all other convolutional layers are followed by batch normalization and LeakyReLU. Each branch outputs a feature map, where each pixel corresponds to the probability that a local region in the input image is true; the outputs of the three branches are weighted and averaged to obtain the final discrimination result.

8. The method for generating truncated CBCT synthetic CT based on a dual-encoder generative adversarial network according to claim 1, characterized in that, The structure of the second discriminator described in step S02 is symmetrical to that of the first discriminator, and it is used to distinguish between real CBCT and synthetic CBCT.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images