CBCT metal artifact suppression method based on unsupervised preoperative prior repair network

Abstract

The application discloses a CBCT metal artifact suppression method based on an unsupervised preoperative prior repair network and belongs to the technical field of computer image processing. The method comprises the following steps: inputting a preoperative prior CBCT image and a CBCT image to be repaired into a generator after registration, respectively extracting anatomical structure features and non-metal region detail features by using two encoders of a double-flow feature extraction network, fusing the features, and then performing repair reconstruction on the CBCT image by using a decoder; inputting the result into a discriminator to perform authenticity discrimination and metal region discrimination and generate a fused prediction image; taking the prediction image information as a supervision signal to optimize the adversarial loss between the generator and the discriminator and the image similarity loss between the repaired image and the image to be repaired, and performing joint adversarial training on the discriminator and the generator; and processing an intraoperative CBCT image to be repaired by using the trained generator to realize more efficient metal artifact suppression.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method for suppressing metal artifacts in CBCT based on an unsupervised preoperative prior repair network. Background Technology

[0002] Cone-beam computed tomography (CBCT), as an important medical imaging technology, has been widely used in clinical scenarios such as radiotherapy planning, interventional surgical navigation, oral medicine diagnosis, and orthopedic surgical evaluation due to its three-dimensional imaging capabilities, relatively low radiation dose, and flexible scanning modes. However, this technology still faces a significant challenge in practical applications—metal artifacts. When metallic implants (such as dental fillings, orthopedic implants, vascular clips, and surgical clips) are present in the scanning area, severe metal artifacts will occur in the reconstructed image due to the high absorption of X-rays by metallic materials and the resulting beam hardening, scattering, noise interference, and photon starvation. These artifacts mainly manifest as dark bands, bright bars, and radial stripes in the image, which not only significantly reduce the image contrast and signal-to-noise ratio but may also lead to blurred anatomical structures and distorted tissue boundaries, seriously affecting the accuracy of lesion localization and treatment precision, and may even lead to misdiagnosis or treatment deviation.

[0003] Currently, several technical approaches have been developed for suppressing metal artifacts in CBCT images, which can be mainly categorized as follows:

[0004] Projection data-based restoration methods operate primarily within the original projection domain. The core idea is to identify, repair, or replace projection data affected by metal. The specific process typically includes the following steps: First, the projection trajectory of the metal object is located in the reconstructed initial image through thresholding or morphological operations. Then, this information is back-projected into a sinusoidal space to determine the projection areas that are ineffective or distorted due to metal influence. Next, interpolation or extrapolation techniques are used to fill in the missing projection data. Common interpolation methods include linear interpolation, polynomial interpolation, and more complex restoration algorithms based on edge structures or contextual information. Finally, the image is reconstructed using the restored complete projection data. The advantage of this method is that it processes data directly at the data acquisition level, and its theoretical foundation is relatively mature. However, its limitations are also significant: the interpolation process itself is prone to introducing new secondary artifacts, and important anatomical information contained in the projection data may be lost during the restoration process, leading to blurred image details or edge distortion. Furthermore, this method is sensitive to the size and shape of the metal object, and its effectiveness is often poor when dealing with multiple or large metal implants.

[0005] Iterative reconstruction algorithms: These methods construct more accurate physical and statistical models to compensate for various physical effects caused by metals during reconstruction, such as beam hardening, scattering, and noise amplification. Their basic framework is typically based on optimization theory, iteratively updating image estimates and incorporating prior constraints or regularization terms in each iteration to simultaneously achieve artifact suppression and structure preservation. Commonly used objective functions include total variation minimization, weighted least squares criteria, and regularization methods based on sparse representations. Typical implementation techniques include statistical iterative reconstruction algorithms and model-based reconstruction methods, some of which introduce adaptive priors or anatomical constraints to improve performance. While these methods theoretically yield high-quality images and demonstrate strong flexibility in handling complex metal artifacts, their computational complexity is extremely high, requiring multiple iterations to converge, making it difficult to meet the stringent reconstruction speed requirements of clinical practice. This significantly limits their application in real-time or near-real-time scenarios. Furthermore, the performance of iterative algorithms largely depends on parameter tuning, further complicating practical applications.

[0006] Deep learning-based image post-processing methods: In recent years, with the rapid development of artificial intelligence technology, deep learning-based methods for suppressing metal artifacts have gradually become a research hotspot. These methods typically use CBCT images containing artifacts as input, learning a complex nonlinear mapping from artifact-laden images to artifact-free images through an end-to-end convolutional neural network. Depending on the training data, these methods can be further divided into supervised learning and unsupervised or weakly supervised learning. Supervised learning methods rely on a large amount of precisely paired training data (i.e., images containing metal artifacts and their corresponding artifact-free images), and commonly used network structures include U-Net, generative adversarial networks, and their variants. Unsupervised or weakly supervised methods, on the other hand, reduce the dependence on perfectly paired data to some extent by utilizing cycle consistency loss, prior knowledge guidance, or self-supervised strategies. Although deep learning-based methods show significant advantages in artifact suppression and structure preservation, and have fast inference speeds, their performance is still highly dependent on the quality and scale of the training data, and the model decision-making process lacks sufficient interpretability, potentially leading to insufficient generalization ability in clinical applications with large data distribution differences. Furthermore, these methods require significant computational resources for model training and are highly sensitive to the selection of hyperparameters.

[0007] Generally speaking, each of the above methods has its own advantages and limitations. Projection data restoration is relatively simple to implement but prone to information loss; iterative reconstruction has high accuracy but low computational efficiency; deep learning methods have excellent performance but stringent requirements for data quality; and multimodal methods are effective but have high implementation complexity. Therefore, how to develop an efficient, robust, and easily scalable metal artifact suppression algorithm while ensuring image quality and considering the resource and time constraints of clinical scenarios remains a key problem that urgently needs to be solved in the field of CBCT image processing. In particular, achieving high-quality and high-efficiency metal artifact suppression without relying on special hardware equipment or requiring perfectly matched training data has significant clinical significance and research value. Summary of the Invention

[0008] The purpose of this invention is to provide a CBCT metal artifact suppression method based on an unsupervised preoperative prior repair network, which achieves accurate suppression of metal artifacts and complete preservation of anatomical structures under an unsupervised learning framework, improves the clinical applicability of CBCT images, solves the problems of artifact residue and tissue structure distortion in traditional metal artifact suppression algorithms, and reduces the dependence of existing deep learning methods on strictly paired training data.

[0009] To achieve the above objectives, the present invention provides the following technical solution.

[0010] On one hand, this invention provides a method for suppressing metal artifacts in CBCT based on an unsupervised preoperative prior repair network, comprising:

[0011] Based on the CBCT image to be repaired and a random repair mask, obtain a simulated repair mask;

[0012] Preprocessing is performed on the CBCT image to be repaired and the preoperative a priori CBCT image based on the simulated repair mask to obtain the registered preoperative a priori CBCT image and the CBCT image to be repaired.

[0013] The pre-registered preoperative CBCT image and the CBCT image to be repaired are input into the generator, which includes a dual-stream feature extraction network, a multi-scale feature fusion module, and a decoder. The dual-stream feature extraction network includes a first encoder and a second encoder, which are used to extract anatomical features based on the input preoperative CBCT image and extract non-metallic region detail features based on the input CBCT image to be repaired, respectively. The multi-scale feature fusion module is used to fuse the anatomical features and non-metallic region detail features, and the decoder is used to repair and reconstruct the CBCT image based on the feature fusion result.

[0014] The discriminator is used to determine the authenticity of the CBCT image based on the repaired and reconstructed image and the CBCT image to be repaired. The metal region is predicted based on the simulated repair mask. A prediction map is generated based on the true gender result and the metal region prediction result. Each pixel of the prediction map corresponds to the true prediction value or the metal region prediction value of each local region in the CBCT image.

[0015] The predicted image is used as a supervision signal to optimize the adversarial loss between the generator and the discriminator, and the image similarity loss between the reconstructed image and the CBCT image to be repaired. Joint adversarial training is performed on the discriminator and the generator.

[0016] The trained generator was used to process the intraoperative CBCT images to be repaired, resulting in intraoperative CBCT images with metal artifacts suppressed.

[0017] Optionally, preprocessing is performed on the CBCT image to be repaired and the preoperative a priori CBCT image based on the simulated repair mask, including:

[0018] The CBCT image to be repaired is simulated and repaired according to the simulated repair mask to obtain the CBCT image to be repaired after simulated repair. The anatomical structure of the preoperative CBCT image and the CBCT image to be repaired after simulated repair is aligned by rigid registration or non-rigid registration. Then, the image after anatomical structure alignment is normalized.

[0019] The random repair mask is a binary random repair mask. ;

[0020] The process of obtaining a simulated restoration mask based on the CBCT image to be restored and a random restoration mask includes: extracting a metal mask from the image to be restored using a threshold mask extraction method. ; Obtain binary random repair mask With metal mask The difference set is used to obtain the simulated repair mask. The formula is expressed as: It is used to perform simulated repair operations on clean areas without metal artifacts during the training phase, that is, to fill in the missing areas of the image to be repaired by using the known non-missing area information of the image to be repaired and the area information of the preoperative a priori image.

[0021] The formula for the simulated repair process is expressed as follows: , , The images shown are the CBCT images to be repaired before and after simulated repair processing. This represents the Gaussian filtering and downsampling processing function.

[0022] In the above technical solutions, the preoperative CBCT images are metal-free images. The registered preoperative CBCT images provide important prior information for intraoperative image reconstruction and artifact suppression.

[0023] Optionally, in the dual-stream feature extraction network, the first encoder and the second encoder have the same network topology and different network parameters;

[0024] The identical network topology includes a reflection padding layer and multiple feature extraction units arranged sequentially at different scales, from shallow to deep. Each feature extraction unit includes a convolutional layer, a batch normalization layer, and a ReLU activation function layer. The feature extraction process of the feature extraction unit is expressed as follows:

[0025] ,

[0026] ,

[0027] In the formula, Indicates the first The feature extraction process of each feature extraction unit; The result of processing the input of the first encoder or the input of the second encoder after passing through the reflective filling layer; Indicates the first Convolution operation of each feature extraction unit Represents a non-linear activation function. This indicates a batch normalization operation. This represents the total number of feature extraction units.

[0028] As can be seen from the above expressions, in the first and second encoders, each feature extraction unit performs convolution, batch normalization, and ReLU activation operations based on the input, respectively. Except for the first feature extraction unit, the inputs of other feature extraction units are the outputs of the previous layer's feature extraction units. Through this layer-by-layer stacking structure of feature extraction units, the encoder can progressively extract multi-scale features from low-level texture to high-level semantics. Performing reflection filling before feature extraction maintains the spatial continuity of boundary information during subsequent feature extraction processes.

[0029] Although the two independent encoders share the same topology, they do not share weights. Thanks to the random initialization of independent parameter spaces and the input domain-driven gradient descent strategy, the first encoder can focus on extracting anatomical features from the prior image, while the second encoder can keenly capture the subtle textures of non-metallic regions in the image to be restored. This "isomorphic yet heterogeneous" design reduces the complexity of model design while ensuring the targeted extraction of multi-source information.

[0030] Optionally, the multi-scale feature fusion module includes a feature fusion unit and a feature enhancement unit arranged sequentially. The feature fusion unit is configured in a dual-path configuration corresponding to the first encoder and the second encoder, and each feature fusion unit is configured with a combination of multiple first feature selection fusion modules and first residual blocks in sequence. The feature enhancement unit includes a second feature selection fusion module and multiple cascaded second residual blocks, which are used to enhance the expressive power of the fused features and extract deep semantic information based on the feature fusion results of the feature fusion unit.

[0031] The feature selection and fusion process of the first feature selection and fusion module and the first residual block of each combination is represented as follows:

[0032]

[0033] In the formula, For each feature fusion unit, select the number of layers in the fusion module. and These represent the first encoder and the second encoder, respectively. Anatomical structural features and non-metallic region detail features extracted by each feature extraction unit; This indicates that in the feature fusion unit corresponding to the first encoder, the first... The feature fusion result of the first feature selection fusion module; This indicates that in the feature fusion unit corresponding to the second encoder, the first... The feature fusion result of the first feature selection fusion module; This indicates a concatenation operation along the channel dimension. This represents a 1×1 convolution operation, used to compress the number of channels and achieve a linear combination of features. Indicates the first The convolution operation of the first residual block in the first feature selection fusion module; , These represent the feature fusion units corresponding to the first encoder and the second encoder, respectively. The output of the first residual block in the first feature selection and fusion module, which is also the dual-path feature fusion unit, refines, denoises, and preserves the fidelity of the result of each level of feature selection and fusion through the residual block, so as to better fuse the prior anatomical structure and the detailed texture of the non-metallic region, eliminate metal artifacts and cross-flow interference, and prevent excessive transformation and loss of key features.

[0034] Optionally, in the feature enhancement unit, the processing procedure of the second feature selection and fusion module is represented as follows:

[0035] That is, the second feature selection fusion module is used to fuse the dual-path output features of the feature fusion unit, and then perform residual processing through multiple cascaded second residual blocks;

[0036] The processing procedure for the second residual block is as follows:

[0037]

[0038] In the formula, This indicates the output of the second feature selection fusion module. The number of second residual blocks cascaded in the feature enhancement unit. Indicates the first Convolution operation on the second residual block, Indicates the first The output of the second residual block.

[0039] By combining the above feature fusion and enhancement methods, the anatomical structure information of the prior image and the detailed features of the image to be repaired can be preserved simultaneously. The logic of the dual-path multi-level feature fusion unit ensures the synergistic consistency of multi-source information at multiple scales, while the multi-level residual logic of the feature enhancement unit improves the compactness of the representation through deep feature mining, thereby providing richer feature representations for the subsequent decoder.

[0040] Optionally, the decoder includes multiple image restoration and reconstruction units, each of which includes an upsampling layer, a convolutional layer, and a ReLU activation function layer arranged sequentially, and the multiple image restoration and reconstruction units are arranged sequentially from deep to shallow according to the scale of the convolutional layer;

[0041] The processing procedure of the image restoration and reconstruction unit is expressed as follows:

[0042]

[0043] In the formula, The number of image inpainting and reconstruction units in the decoder. Indicates the first The output of each image restoration and reconstruction unit Indicates the first Convolution operations in convolutional layers of an image restoration and reconstruction unit This indicates an interpolation operation.

[0044] The above technical solutions, through layer-by-layer interpolation followed by convolution, enable the decoder to gradually restore the spatial resolution of the image and effectively suppress metal artifacts while preserving anatomical structures and details.

[0045] Optionally, the discriminator includes an authenticity determination branch and a mask prediction branch;

[0046] The authenticity discrimination branch is used to determine the image to be repaired as real and the repaired area of ​​the reconstructed image as fake, based on the removal of the metal area, to obtain the authenticity prediction map.

[0047] The mask prediction branch is used to perform Gaussian filtering and downsampling on the simulated repair mask to obtain a metal region mask prediction map where the value of each pixel is the probability value of the non-metal region.

[0048] The prediction map output by the discriminator is the result of fusing the realism prediction map and the metal region mask prediction map.

[0049] In the above technical solution, during the training phase, the discriminator uses the repaired image and the image to be repaired as inputs to obtain corresponding prediction maps. Since the image to be repaired is located in a region free of metal artifacts, i.e., the simulated repair region, which is a clean region, it can be used as the learning target for unsupervised training. The discriminator guides the generator to learn the anatomical features of the clean region by classifying the image to be repaired as real and the repaired region of the generated image as fake, based on the removal of the metal region.

[0050] In the final fusion result prediction image, after removing the metal mask... Outside of the identified metallic areas, the pixel values ​​of the areas that retain the original image are used to determine the authenticity confidence level. The pixel values ​​of the metallic areas and the repaired areas are used to determine the non-metallic area confidence level. The closer the non-metallic area confidence level is to 0, the higher the probability that the corresponding area is metallic.

[0051] Optionally, the loss function for the joint adversarial training is expressed as:

[0052] ,

[0053] Indicates the overall loss; This represents the adversarial loss between the generator and the discriminator; This represents the image obtained through restoration and reconstruction. With the CBCT image to be repaired Image similarity loss between them; , These are the weight coefficients for the adversarial loss and the image similarity loss, respectively, used to adjust the weight ratio of the adversarial loss and the similarity loss in the overall objective loss function; where:

[0054] , and These represent the adversarial losses of the generator and discriminator, respectively.

[0055] ,

[0056] ,

[0057] ,

[0058] In the formula, This represents the mathematical expectation operator, used to calculate the average loss under the corresponding data distribution, to ensure the statistical convergence and robustness of the network model; Indicates compliance with the image Distribution The sample takes the expectation, Indicates that the image to be repaired is subject to the specified parameters. Distribution The sample takes the expectation, This represents the calculation of the square of the L2 norm. This represents a pixel-level multiplication operation; The metal mask is extracted from the image to be repaired using a threshold mask extraction method. This is used to constrain the discriminator to perform discrimination operations only in non-metallic regions; This represents the Gaussian filtering and downsampling processing functions; , These represent the preset weight coefficients corresponding to pixel-level L1 loss and structural similarity index, respectively. This represents the calculation of the 1-norm. This represents the function for calculating the structural similarity index.

[0059] In the formula, for those conforming to the image Distribution The sample expectation is used to measure the overall performance of the generator across the entire generated image domain. This loss term is applied only to the simulated inpainting region and is designed to make the generated image as realistic as possible in that region to fool the discriminator.

[0060] In the formula, the first term aims to restore the image obtained through repair and reconstruction. The result is deemed false because the generator only repairs and generates regions simulated during training; therefore, its predicted value is set to... The second item aims to restore the image to be repaired. Non-metallic regions are considered true, and unsupervised training is achieved by using clean regions without artifacts as learning targets.

[0061] Image similarity loss Composed of pixel-level L1 loss and structural similarity index SSIM, it is used to constrain the structural and texture details of the image, so that the non-metallic areas of the restored image are as close as possible to the image to be restored.

[0062] Based on the above, by optimizing the loss function of joint adversarial training to update the generator and discriminator networks, the generator network can learn to use prior information to repair simulated missing regions during the training phase, and then perform metal region repair on newly input intraoperative images during the application phase, effectively suppressing metal artifacts.

[0063] Optionally, the method described in this invention for processing the intraoperative CBCT image to be repaired using a trained generator to obtain an intraoperative CBCT image with metal artifacts suppressed can be as follows: the trained generator is deployed in the CBCT system, and in application, based on the intraoperative CBCT image and its corresponding metal mask, as well as the preoperative prior CBCT image, the generator outputs an intraoperative CBCT image with metal artifacts suppressed.

[0064] As another implementation, the method described in this invention, which uses a trained generator to process the intraoperative CBCT image to be repaired to obtain a metal artifact-suppressed intraoperative CBCT image, includes:

[0065] Intraoperative CBCT images are input into a trained discriminator to obtain the predicted image output by the discriminator.

[0066] Based on the confidence level of the realism or the confidence level of the non-metallic region corresponding to each region in the predicted image, the display correction of artifact regions in the intraoperative CBCT image is performed by weighted calculation.

[0067] Based on the displayed corrected intraoperative CBCT image and its corresponding metal mask, as well as the preoperative prior CBCT image, a generator is used to obtain an intraoperative CBCT image with metal artifacts suppressed.

[0068] The formula for the weighted calculation is expressed as follows: , Intraoperative CBCT images are shown. This indicates that the corrected intraoperative CBCT images are being displayed. This represents the confidence level of the truth value obtained from the truth-based discrimination branch of the discriminator. This represents the confidence level of the non-metallic region obtained from the mask prediction branch of the discriminator.

[0069] The above implementation methods can explicitly correct artifact regions in the input image before generator inference, thereby improving the efficiency of image processing.

[0070] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the CBCT metal artifact suppression method based on dual-domain multi-view weakly supervised segmentation as described in the first aspect.

[0071] Beneficial effects

[0072] Compared with the prior art, the present invention has the following advantages and advancements:

[0073] (1) The present invention is based on the CBCT metal artifact suppression method of unsupervised preoperative prior repair network. The generator is constructed based on the innovative dual-stream network structure. The structural features of the prior image and the detailed features of the image to be repaired are extracted by two independent encoders respectively. Compared with the traditional simple image stitching input method, it can make fuller and more targeted use of complementary information in different source images.

[0074] (2) The present invention introduces a multi-scale feature fusion module into the generator network. By adaptively fusing prior structural features with details to be repaired at different levels of the network, the information interference caused by registration error is effectively overcome, and the accuracy and robustness of prior information guidance are improved.

[0075] (3) The present invention implements a discriminator based on mask prediction. Its dual-function structure, consisting of an authenticity prediction branch and a metal mask prediction branch, can not only supervise the authenticity of the restoration effect at the image level, but also guide the generator to focus on the restoration of artifact regions at the region level by predicting the metal mask, thereby achieving more accurate local image generation and enhancement.

[0076] (4) The present invention adopts an unsupervised learning framework throughout the training process, which does not rely on paired training data with gold standards, reduces the requirements for training sample data collection, and enhances the feasibility and generalization ability of the algorithm in clinical practice. Attached Figure Description

[0077] Figure 1 The diagram shown is a schematic representation of the implementation process of the CBCT metal artifact suppression method based on an unsupervised preoperative prior repair network in an embodiment of the present invention.

[0078] Figure 2 The diagram shown is a schematic representation of the generator's principle framework in an embodiment of the present invention.

[0079] Figure 3 The diagram shows a comparison between the discriminator in this embodiment of the invention and the prior art, wherein the output of the discriminator corresponds from left to right to the traditional discriminator, the patch-based discriminator, and the discriminator of this invention.

[0080] Figure 4 The diagram illustrates the use of prior information contained in the preoperative CBCT image during the repair process in an embodiment of the present invention.

[0081] Figure 5The diagram shows a comparison of the artifact suppression results obtained by the traditional algorithm and the artifact suppression results obtained by the present invention in an embodiment of the present invention. The FDK column represents the image without artifact suppression, the LI column and the TRI column represent the images with artifact suppression using the traditional linear interpolation LI method and the trilinear interpolation TRI method, respectively, and the last column represents the image obtained by the method of the present invention. (a1)-(a3), (b1)-(b3), and (c1)-(c3) represent the images with unsuppressed artifacts, the images with artifact suppression using the linear interpolation LI method, and the images with artifact suppression using the trilinear interpolation TRI method in the three CBCT image restoration scenarios a, b, and c, respectively. Detailed Implementation

[0082] The technical concept of this invention is as follows: First, a preoperative CBCT image and an intraoperative CBCT image to be repaired are acquired after registration. Then, a generator based on a dual-stream feature extraction architecture is used, employing two independent encoders to extract anatomical features from the preoperative image and non-metallic region detail features from the image to be repaired, respectively. Next, a multi-scale feature fusion module achieves deep fusion of features at different levels. Finally, the fused features are input into a decoder for image repair generation, reconstructing a CBCT image that suppresses metal artifacts. During the training phase, this invention uses a discriminator based on mask prediction for unsupervised adversarial training, performing dual discrimination on the repaired image: authenticity judgment and metal region mask prediction. Through a joint optimization process between the generator and the discriminator, using the discriminator's judgment result and the mask prediction result as supervision signals, the generator is guided to maintain non-metallic region details while repairing metallic regions. Finally, the trained network model is deployed to the CBCT system to achieve metal artifact suppression in the intraoperative CBCT image.

[0083] The exemplary solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art.

[0084] Example 1

[0085] This embodiment introduces a CBCT metal artifact suppression method based on unsupervised preoperative prior repair networks, combined with... Figure 1 and Figure 2 The methods include:

[0086] Based on the CBCT image to be repaired and a random repair mask, obtain a simulated repair mask;

[0087] Preprocessing is performed on the CBCT image to be repaired and the preoperative a priori CBCT image based on the simulated repair mask to obtain the registered preoperative a priori CBCT image and the CBCT image to be repaired.

[0088] The pre-registered preoperative CBCT image and the CBCT image to be repaired are input into the generator, which includes a dual-stream feature extraction network, a multi-scale feature fusion module, and a decoder. The dual-stream feature extraction network includes a first encoder and a second encoder, which are used to extract anatomical features based on the input preoperative CBCT image and extract non-metallic region detail features based on the input CBCT image to be repaired, respectively. The multi-scale feature fusion module is used to fuse the anatomical features and non-metallic region detail features, and the decoder is used to repair and reconstruct the CBCT image based on the feature fusion result.

[0089] The discriminator is used to determine the authenticity of the CBCT image based on the repaired and reconstructed image and the CBCT image to be repaired. The metal region is predicted based on the simulated repair mask. A prediction map is generated based on the true gender result and the metal region prediction result. Each pixel of the prediction map corresponds to the true prediction value or the metal region prediction value of each local region in the CBCT image.

[0090] The predicted image is used as a supervision signal to optimize the adversarial loss between the generator and the discriminator, and the image similarity loss between the reconstructed image and the CBCT image to be repaired. Joint adversarial training is performed on the discriminator and the generator.

[0091] The trained generator was used to process the intraoperative CBCT images to be repaired, resulting in CBCT images with metal artifacts suppressed.

[0092] The specific implementation of the method in this embodiment mainly involves the following parts.

[0093] I. Image Preprocessing

[0094] Based on the aforementioned technical concept, during the training phase, the CBCT metal artifact suppression method of this invention uses the CBCT image to be repaired, the simulated repair mask, and the preoperative a priori CBCT image as input data. During the deployment and application phase, it uses the actual intraoperative CBCT image to be repaired, its corresponding metal mask, and the preoperative a priori CBCT image as input data. Figure 4 As shown, preoperative CBCT images provide important prior information, enabling the provision of accurate anatomical information to guide the repair process.

[0095] Preprocessing of the input image data is required during the training phase of the network model generator and discriminator, as well as the later model deployment and application phase.

[0096] In some possible embodiments, the preprocessing includes normalizing, resizing, rigid or non-rigid registration of the CBCT image to be repaired and the preoperative a priori CBCT image, and extracting a metal mask, wherein the simulated repair mask is obtained based on the extracted metal mask and a binary random mask.

[0097] As one implementation method, the training phase preprocesses the CBCT image to be repaired and the preoperative prior CBCT image based on a simulated repair mask. This includes: performing simulated repair processing on the CBCT image to be repaired according to the simulated repair mask to obtain the simulated repaired CBCT image; aligning the anatomical structures of the preoperative prior CBCT image and the simulated repaired CBCT image to be repaired through rigid or non-rigid registration; and then performing image normalization processing on the anatomically aligned image. Since the preoperative CBCT image is a metal-free image, registration provides important prior information for intraoperative image reconstruction and artifact suppression.

[0098] During the model deployment and application phase, metal masks can be extracted from the CBCT images to be repaired during surgery. Then, preliminary repair processing can be performed on the CBCT images to be repaired during surgery based on the metal masks, followed by image registration, normalization, and other processing.

[0099] refer to Figure 1 As one implementation method, the metal mask extraction of the CBCT image to be repaired can be performed using a threshold mask extraction method to obtain the metal mask. Then, combine the binary random repair mask. Calculate the binary random repair mask With metal mask The difference set is used to obtain the simulated repair mask. The formula is expressed as: It is used to perform simulated repair operations on clean areas without metal artifacts during the training phase, that is, to fill in the missing areas of the image to be repaired by using the known non-missing area information of the image to be repaired and the area information of the preoperative a priori image.

[0100] The formulas for simulated repair of the CBCT image to be repaired based on the simulated repair mask during the training phase, and the formulas for preliminary repair of the intraoperative CBCT image to be repaired based on the metal mask during the deployment and application phase, are expressed as follows:

[0101]

[0102] in, , These are CBCT images to be repaired during the training phase, before and after simulated repair processing. , The images are intraoperative CBCT images before and after the initial repair based on the metal mask. This represents the Gaussian filtering and downsampling processing function.

[0103] During the training and deployment phases, the model is respectively... and The corresponding preoperative CBCT images are used as input to perform artifact suppression and repair processing.

[0104] II. Generator

[0105] refer to Figure 2 As shown, in one embodiment, the generator includes a dual-stream feature extraction network, a multi-scale feature fusion module, and a decoder; the dual-stream feature extraction network includes a first encoder and a second encoder that are independent of each other, respectively used for: extracting anatomical structural features based on the input preoperative CBCT image, and extracting non-metallic region detail features based on the input CBCT image to be repaired; the multi-scale feature fusion module is used to perform feature fusion on the anatomical structural features and non-metallic region detail features, and the decoder is used to repair and reconstruct the CBCT image based on the feature fusion result.

[0106] 2.1 Two-stream feature extraction network

[0107] In a two-stream feature extraction network, the first encoder and the second encoder have the same network topology but different network parameters;

[0108] The identical network topology includes a reflection padding layer and multiple feature extraction units arranged sequentially at different scales, from shallow to deep. Each feature extraction unit includes a convolutional layer, a batch normalization layer, and a ReLU activation function layer. The feature extraction process of the feature extraction unit is expressed as follows:

[0109] ,

[0110] ,

[0111] In the formula, Indicates the first The feature extraction process of each feature extraction unit; The result of processing the input of the first encoder or the input of the second encoder after passing through the reflective filling layer; Indicates the first Convolution operation of each feature extraction unit Represents a non-linear activation function. This indicates a batch normalization operation. This represents the total number of feature extraction units.

[0112] As can be seen above, in the first and second encoders, each feature extraction unit performs convolution, batch normalization, and ReLU activation operations based on the input. Except for the first feature extraction unit, the inputs of other feature extraction units are the outputs of the previous layer's feature extraction units. Through this layer-by-layer stacking structure of feature extraction units, the encoder can progressively extract multi-scale features from low-level texture to high-level semantics. Performing reflection filling before feature extraction maintains the spatial continuity of boundary information during subsequent feature extraction processes.

[0113] Although the first and second encoders are structurally isomorphic, they do not share weights. However, thanks to the random initialization of independent parameter spaces and the gradient descent strategy driven by the input domain, the first encoder can focus on extracting anatomical features from the prior image, while the second encoder can keenly capture the subtle textures of non-metallic regions in the image to be restored. This "isomorphic yet heterogeneous" design reduces the complexity of model design while ensuring the targeted extraction of multi-source information.

[0114] 2.2 Multi-scale feature fusion module

[0115] The multi-scale feature fusion module includes a feature fusion unit and a feature enhancement unit arranged sequentially. The feature fusion unit is configured in a dual-path configuration corresponding to the first encoder and the second encoder, and each feature fusion unit is configured with a combination of multiple first feature selection fusion modules and first residual blocks in sequence. The feature enhancement unit includes a second feature selection fusion module and multiple cascaded second residual blocks, which are used to enhance the expressive power of the fused features and extract deep semantic information based on the feature fusion results of the feature fusion unit.

[0116] The feature selection and fusion process of the first feature selection and fusion module and the first residual block of each combination is represented as follows:

[0117]

[0118] In the formula, For each feature fusion unit, select the number of layers in the fusion module. and These represent the first encoder and the second encoder, respectively. Anatomical structural features and non-metallic region detail features extracted by each feature extraction unit; This indicates that in the feature fusion unit corresponding to the first encoder, the first... The feature fusion result of the first feature selection fusion module; This indicates that in the feature fusion unit corresponding to the second encoder, the first... The feature fusion result of the first feature selection fusion module; This indicates a concatenation operation along the channel dimension. This represents a 1×1 convolution operation, used to compress the number of channels and achieve a linear combination of features. Indicates the first The convolution operation of the first residual block in the first feature selection fusion module; , These represent the feature fusion units corresponding to the first encoder and the second encoder, respectively. The output of the first residual block in the first feature selection and fusion module, which is also the dual-path feature fusion unit, refines, denoises, and preserves the fidelity of the result of each level of feature selection and fusion through the residual block, so as to better fuse the prior anatomical structure and the detailed texture of the non-metallic region, eliminate metal artifacts and cross-flow interference, and prevent excessive transformation and loss of key features.

[0119] 2.3 Feature Enhancement Unit

[0120] In the feature enhancement unit, the processing procedure of the second feature selection and fusion module is represented as follows:

[0121] That is, the second feature selection fusion module is used to fuse the dual-path output features of the feature fusion unit, and then perform residual processing through multiple cascaded second residual blocks;

[0122] The processing procedure for the second residual block is as follows:

[0123]

[0124] In the formula, This indicates the output of the second feature selection fusion module. The number of second residual blocks cascaded in the feature enhancement unit. Indicates the first Convolution operation on the second residual block, Indicates the first The output of the second residual block.

[0125] By combining the above feature fusion and enhancement methods, the anatomical structure information of the prior image and the detailed features of the image to be repaired can be preserved simultaneously. The logic of the dual-path multi-level feature fusion unit ensures the synergistic consistency of multi-source information at multiple scales, while the multi-level residual logic of the feature enhancement unit improves the compactness of the representation through deep feature mining, thereby providing richer feature representations for the subsequent decoder.

[0126] 2.4 Decoder

[0127] The decoder includes multiple image inpainting and reconstruction units. Each image inpainting and reconstruction unit includes an upsampling layer, a convolutional layer, and a ReLU activation function layer arranged sequentially. The multiple image inpainting and reconstruction units are arranged sequentially from deep to shallow according to the scale of the convolutional layer.

[0128] The processing procedure of the image restoration and reconstruction unit is expressed as follows:

[0129]

[0130] In the formula, The number of image inpainting and reconstruction units in the decoder. Indicates the first The output of each image restoration and reconstruction unit Indicates the first Convolution operations in convolutional layers of an image restoration and reconstruction unit This indicates an interpolation operation.

[0131] By using layer-by-layer interpolation followed by convolution, the decoder can gradually restore the spatial resolution of the image, effectively suppress metal artifacts while preserving anatomical structures and details, and output a CBCT-restored image with metal artifacts suppressed.

[0132] III. Discriminator

[0133] refer to Figure 3 This invention employs a discriminator based on mask prediction to determine the authenticity of the restored image and predict its mask. This discriminator is used for joint adversarial training with the generator to optimize the network parameters of both the generator and the discriminator. Specifically:

[0134] The discriminator includes a authenticity determination branch and a mask prediction branch;

[0135] The authenticity discrimination branch is used to determine the image to be repaired as real and the repaired area of ​​the reconstructed image as fake, based on the removal of the metal area, to obtain the authenticity prediction map.

[0136] The mask prediction branch is used to perform Gaussian filtering and downsampling on the simulated repair mask to obtain a metal region mask prediction map where the value of each pixel is the probability value of the non-metal region.

[0137] The prediction map output by the discriminator is the result of fusing the realism prediction map and the metal region mask prediction map.

[0138] During the training phase, the discriminator uses the restored image and the image to be restored as inputs to obtain corresponding prediction maps. Since the image to be restored is located in a clean region free of metal artifacts (i.e., the simulated restoration region), it can serve as the learning target for unsupervised training. The discriminator guides the generator to learn the anatomical features of the clean region by classifying the image to be restored as real and the restored region of the generated image as fake, based on the removal of the metal region.

[0139] In the final fusion result prediction image, after removing the metal mask... Outside of the identified metallic areas, the pixel values ​​of the areas that retain the original image are used to determine the authenticity confidence level. The pixel values ​​of the metallic areas and the repaired areas are used to determine the non-metallic area confidence level. The closer the non-metallic area confidence level is to 0, the higher the probability that the corresponding area is metallic.

[0140] Refer again Figure 3 Compared to traditional discriminators, the discriminator of this invention has a dual-function output structure, which can not only determine the authenticity of an image but also predict the mask of metallic regions. By synchronously performing adversarial training with the generator network during training, and using the repaired image and its corresponding simulated repair mask as guidance, the parameters of the discriminator are optimized.

[0141] The discriminator of this invention can also be used to perform display correction on intraoperative CBCT images to be repaired during the model deployment and application phase, thereby improving image processing efficiency. Specifically:

[0142] Intraoperative CBCT images are input into a trained discriminator to obtain the predicted image output by the discriminator.

[0143] Based on the confidence level of the realism or the confidence level of the non-metallic region corresponding to each region in the predicted image, the display correction of artifact regions in the intraoperative CBCT image is performed by weighted calculation.

[0144] Based on the displayed corrected intraoperative CBCT image and its corresponding metal mask, as well as the preoperative prior CBCT image, a generator is used to obtain an intraoperative CBCT image with metal artifacts suppressed.

[0145] The formula for the weighted calculation is expressed as follows: , Intraoperative CBCT images are shown. This indicates that the corrected intraoperative CBCT images are being displayed. This represents the confidence level of the truth value obtained from the truth-based discrimination branch of the discriminator. This represents the confidence level of the non-metallic region obtained from the mask prediction branch of the discriminator. Accordingly, based on the corrected intraoperative CBCT image and its corresponding metallic mask, the preliminary repair formula is expressed as: Then, the preoperative CBCT images and the intraoperative CBCT images after preliminary repair were compared. Input generator.

[0146] IV. Joint Combat Training

[0147] The joint adversarial training process of the present invention uses the discrimination result of the discriminator and the mask prediction result as supervision signals to perform joint adversarial training with the generator.

[0148] The optimization objective functions during joint training include: adversarial loss between the generator and the discriminator, and image similarity loss between the restored image and the real, artifact-free image.

[0149] The loss function for joint adversarial training is expressed as:

[0150] ,

[0151] Indicates the overall loss; This represents the adversarial loss between the generator and the discriminator; This represents the image obtained through restoration and reconstruction. With the CBCT image to be repaired Image similarity loss between them; , These are the weight coefficients for the adversarial loss and the image similarity loss, respectively, used to adjust the weight ratio of the adversarial loss and the similarity loss in the overall objective loss function; where:

[0152] , and These represent the adversarial losses of the generator and discriminator, respectively.

[0153] ,

[0154] ,

[0155] ,

[0156] In the formula, This represents the mathematical expectation operator, used to calculate the average loss under the corresponding data distribution, to ensure the statistical convergence and robustness of the network model; Indicates compliance with the image Distribution The sample takes the expectation, Indicates that the image to be repaired is subject to the specified parameters. Distribution The sample takes the expectation, This represents the calculation of the square of the L2 norm. This represents a pixel-level multiplication operation; The metal mask is extracted from the image to be repaired using a threshold mask extraction method. This is used to constrain the discriminator to perform discrimination operations only in non-metallic regions; This represents the Gaussian filtering and downsampling processing functions; , These represent the preset weight coefficients corresponding to pixel-level L1 loss and structural similarity index, respectively. This represents the calculation of the 1-norm. This represents the function for calculating the structural similarity index.

[0157] In the formula, for those conforming to the image Distribution The sample expectation is used to measure the overall performance of the generator across the entire generated image domain. This loss term is applied only to the simulated inpainting region and is designed to make the generated image as realistic as possible in that region to fool the discriminator.

[0158] In the formula, the first term aims to restore the image obtained through repair and reconstruction. The result is deemed false because the generator only repairs and generates regions simulated during training; therefore, its predicted value is set to... The second item aims to restore the image to be repaired. Non-metallic regions are considered true, and unsupervised training is achieved by using clean regions without artifacts as learning targets.

[0159] Image similarity loss It is composed of pixel-level L1 loss and the structural similarity index SSIM. L1 loss is used to constrain the structure and texture details of an image, making the non-metallic regions of the reconstructed image as close as possible to the original image. It is a structural similarity index used to preserve the structural and texture details of an image.

[0160] Based on the above, by optimizing the loss function of joint adversarial training to update the generator and discriminator networks, the generator network can learn to use prior information to repair simulated missing regions during the training phase, and then perform metal region repair on newly input intraoperative images during the application phase, effectively suppressing metal artifacts.

[0161] V. Model Deployment and Application

[0162] After the model is trained, the generator and discriminator are deployed in the CBCT system, which can then perform image restoration based on the newly input CBCT image to be restored during the operation and the prior CBCT image before the operation.

[0163] The present invention describes using a trained generator to process the intraoperative CBCT image to be repaired, obtaining a CBCT image with metal artifacts suppressed. This can be achieved by using the generator to output a metal artifact-suppressed intraoperative CBCT image based on the intraoperative CBCT image and its corresponding metal mask, as well as a preoperative a priori CBCT image. That is, according to the formula: First, perform preliminary image restoration, then... and preoperative CBCT images Registration and normalization operations are performed, and the data is input into the generator. The repaired intraoperative CBCT image is obtained through feedforward network inference.

[0164] Alternatively, a CBCT image with metal artifact suppression can be obtained by combining a discriminator and a generator. Specifically, this involves: inputting the intraoperative CBCT image into a trained discriminator to obtain a predicted image output by the discriminator; and, based on the realism confidence score or non-metallic region confidence score corresponding to each region in the predicted image, performing a weighted calculation to correct the display of artifact regions in the intraoperative CBCT image. Then, preliminary repair is performed based on the corrected intraoperative CBCT images and their corresponding metal masks: Then to and preoperative CBCT images Registration and normalization operations are performed, and the data is input into the generator to obtain intraoperative CBCT images with suppressed metal artifacts.

[0165] To verify the practical effectiveness of the method of this invention, a set of clinically acquired CBCT datasets containing metal implants was used to verify the performance improvement of this invention compared to traditional metal artifact suppression algorithms. A comparative experiment was conducted using CBCT data from the head and neck and pelvic regions as examples. Figure 5 The paper presents comparative images of artifact suppression effects achieved using traditional metal artifact suppression algorithms LI and TRI in three CBCT image restoration scenarios (a, b, and c), compared to those achieved using the method proposed in this invention. The comparison results demonstrate that the method provided by this invention effectively suppresses metal artifacts while better preserving the anatomical details and tissue texture information of the original image, significantly improving the visual quality and clinical usability of the image, thus proving the effectiveness and advancement of this invention.

[0166] Example 2

[0167] This embodiment describes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the CBCT metal artifact suppression method based on an unsupervised preoperative prior repair network as described in Embodiment 1.

[0168] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for suppressing metal artifacts in CBCT based on an unsupervised preoperative prior repair network, characterized in that, include: Based on the CBCT image to be repaired and a random repair mask, obtain a simulated repair mask; Preprocessing is performed on the CBCT image to be repaired and the preoperative a priori CBCT image based on the simulated repair mask to obtain the registered preoperative a priori CBCT image and the CBCT image to be repaired. The registered preoperative CBCT image and the CBCT image to be repaired are input into the generator, which includes a dual-stream feature extraction network, a multi-scale feature fusion module, and a decoder. The dual-stream feature extraction network includes a first encoder and a second encoder that are independent of each other, and are used to extract anatomical features based on the input preoperative CBCT image and extract non-metallic region detail features based on the input CBCT image to be repaired. The multi-scale feature fusion module is used to perform feature fusion on the anatomical structural features and non-metallic region detail features, and the decoder is used to repair and reconstruct CBCT images based on the feature fusion results; The discriminator is used to determine the authenticity of the CBCT image based on the repaired and reconstructed image and the CBCT image to be repaired. The metal region is predicted based on the simulated repair mask. A prediction map is generated based on the true gender result and the metal region prediction result. Each pixel of the prediction map corresponds to the true prediction value or the metal region prediction value of each local region in the CBCT image. The predicted image is used as a supervision signal to optimize the adversarial loss between the generator and the discriminator, and the image similarity loss between the reconstructed image and the CBCT image to be repaired. Joint adversarial training is performed on the discriminator and the generator. The trained generator was used to process the intraoperative CBCT images to be repaired, resulting in intraoperative CBCT images with metal artifacts suppressed.

2. The CBCT metal artifact suppression method based on unsupervised preoperative prior repair network according to claim 1, characterized in that, Preprocessing is performed on the CBCT image to be repaired and the preoperative a priori CBCT image based on a simulated repair mask, including: The CBCT image to be repaired is simulated and repaired according to the simulated repair mask to obtain the simulated repaired CBCT image. The anatomical structure of the preoperative CBCT image and the simulated repaired CBCT image is aligned by rigid registration or non-rigid registration. Then, the image after anatomical structure alignment is normalized. The random repair mask is a binary random repair mask. ; The process of obtaining a simulated restoration mask based on the CBCT image to be restored and a random restoration mask includes: extracting a metal mask from the image to be restored using a threshold mask extraction method. ; Obtain binary random repair mask With metal mask The difference set is used to obtain the simulated repair mask. ; The formula for the simulated repair process is expressed as follows: , , The images shown are the CBCT images to be repaired before and after simulated repair processing. This represents the Gaussian filtering and downsampling processing function.

3. The CBCT metal artifact suppression method based on unsupervised preoperative prior repair network according to claim 1, characterized in that, In the dual-stream feature extraction network, the first encoder and the second encoder have the same network topology but different network parameters. The identical network topology includes a reflection padding layer and multiple feature extraction units arranged sequentially at different scales, from shallow to deep. Each feature extraction unit includes a convolutional layer, a batch normalization layer, and a ReLU activation function layer. The feature extraction process of the feature extraction unit is expressed as follows: ， ， In the formula, Indicates the first The feature extraction process of each feature extraction unit; The result of processing the input of the first encoder or the input of the second encoder after passing through the reflective filling layer; Indicates the first Convolution operation of each feature extraction unit Represents a non-linear activation function. This indicates a batch normalization operation. This represents the total number of feature extraction units.

4. The CBCT metal artifact suppression method based on unsupervised preoperative prior repair network according to claim 3, characterized in that, The multi-scale feature fusion module includes a feature fusion unit and a feature enhancement unit arranged in sequence. The feature fusion unit is configured in a dual-path configuration corresponding to the first encoder and the second encoder, and each feature fusion unit is configured in a combination of multiple first feature selection fusion modules and first residual blocks in sequence. The feature enhancement unit includes a second feature selection fusion module and multiple cascaded second residual blocks, which are used to enhance the expressive power of the fused features and extract deep semantic information based on the feature fusion results of the feature fusion unit. The feature selection and fusion process of the first feature selection and fusion module and the first residual block of each combination is represented as follows: In the formula, For each feature fusion unit, select the number of layers in the fusion module. and These represent the first encoder and the second encoder, respectively. Anatomical structural features and non-metallic region detail features extracted by each feature extraction unit; This indicates that in the feature fusion unit corresponding to the first encoder, the first... The feature fusion result of the first feature selection fusion module; This indicates that in the feature fusion unit corresponding to the second encoder, the first... The feature fusion result of the first feature selection fusion module; This indicates a concatenation operation along the channel dimension. This represents a 1×1 convolution operation. Indicates the first The convolution operation of the first residual block in the first feature selection fusion module; , These represent the feature fusion units corresponding to the first encoder and the second encoder, respectively. The output of the first residual block in the first feature selection fusion module.

5. The CBCT metal artifact suppression method based on unsupervised preoperative prior repair network according to claim 4, characterized in that, In the feature enhancement unit, the processing procedure of the second feature selection and fusion module is as follows: ； The processing procedure for the second residual block is as follows: In the formula, This indicates the output of the second feature selection fusion module. The number of second residual blocks cascaded in the feature enhancement unit. Indicates the first Convolution operation on the second residual block, Indicates the first The output of the second residual block.

6. The CBCT metal artifact suppression method based on unsupervised preoperative prior repair network according to claim 5, characterized in that, The decoder includes multiple image restoration and reconstruction units. Each image restoration and reconstruction unit includes an upsampling layer, a convolutional layer, and a ReLU activation function layer arranged sequentially. The multiple image restoration and reconstruction units are arranged sequentially from deep to shallow according to the scale of the convolutional layer. The processing procedure of the image restoration and reconstruction unit is expressed as follows: In the formula, The number of image inpainting and reconstruction units in the decoder. Indicates the first The output of each image restoration and reconstruction unit Indicates the first Convolution operations in convolutional layers of an image restoration and reconstruction unit This indicates an interpolation operation.

7. The CBCT metal artifact suppression method based on unsupervised preoperative prior repair network according to claim 2, characterized in that, The discriminator includes an authenticity determination branch and a mask prediction branch; The authenticity discrimination branch is used to determine the image to be repaired as real and the repaired area of ​​the reconstructed image as fake, based on the removal of the metal area, to obtain the authenticity prediction map. The mask prediction branch is used to perform Gaussian filtering and downsampling on the simulated repair mask to obtain a metal region mask prediction map where the value of each pixel is the probability value of the non-metal region. The prediction map output by the discriminator is the result of fusing the realism prediction map and the metal region mask prediction map.

8. The CBCT metal artifact suppression method based on unsupervised preoperative prior repair network according to claim 7, characterized in that, The loss function for the joint adversarial training is expressed as: ， Indicates overall loss; This represents the adversarial loss between the generator and the discriminator; This represents the image obtained through restoration and reconstruction. With the CBCT image to be repaired Image similarity loss between them; , These are the weight coefficients for the adversarial loss and the image similarity loss, respectively; where: , and These represent the adversarial losses of the generator and discriminator, respectively. ， ， ， In the formula, This represents the mathematical expectation operator, used to calculate the average loss under a given data distribution; Indicates compliance with the image Distribution The sample takes the expectation, Indicates that the image to be repaired is subject to the specified parameters. Distribution The sample takes the expectation, This represents the calculation of the square of the L2 norm. This represents a pixel-level multiplication operation; The threshold mask extraction method is used to extract the metal mask from the image to be repaired. This is used to constrain the discriminator to perform discrimination operations only in non-metallic regions; This represents the Gaussian filtering and downsampling processing functions; , These represent the preset weight coefficients corresponding to pixel-level L1 loss and structural similarity index, respectively. This represents the calculation of the 1-norm. This represents the function for calculating the structural similarity index.

9. The CBCT metal artifact suppression method based on unsupervised preoperative prior repair network according to claim 7, characterized in that, The process of using a trained generator to process the intraoperative CBCT image to be repaired to obtain a metal artifact-suppressed intraoperative CBCT image includes: Based on intraoperative CBCT images and their corresponding metal masks, as well as preoperative prior CBCT images, a generator is used to output intraoperative CBCT images with metal artifacts suppressed. or, Intraoperative CBCT images are input into a trained discriminator to obtain the predicted image output by the discriminator. Based on the confidence level of the realism or the confidence level of the non-metallic region corresponding to each region in the predicted image, the display correction of artifact regions in the intraoperative CBCT image is performed by weighted calculation. Based on the displayed corrected intraoperative CBCT image and its corresponding metal mask, as well as the preoperative prior CBCT image, a generator is used to obtain an intraoperative CBCT image with metal artifacts suppressed. The formula for the weighted calculation is expressed as follows: , Intraoperative CBCT images are shown. This indicates that the corrected intraoperative CBCT images are being displayed. This represents the confidence level of the truth value obtained from the truth-based discrimination branch of the discriminator. This represents the confidence level of the non-metallic region obtained from the mask prediction branch of the discriminator.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the CBCT metal artifact suppression method based on an unsupervised preoperative prior repair network as described in any one of claims 1 to 9.

Images