Scattering correction method, cone beam computed tomography (CBCT) system and storage medium

Abstract

The present disclosure relates to a scatter correction method, a cone-beam computed tomography (CBCT) system and a computer storage medium. The method comprises: performing cone-beam CT imaging on an object to be measured to obtain a first set of projection data of the object to be measured, and reconstructing from the first set of projection data to obtain a first volume image containing scatter artifacts; performing fan-beam CT imaging on multiple different heights of the object to be measured to obtain multiple sets of fan-beam projection data of the object to be measured, and reconstructing from the multiple sets of fan-beam projection data to obtain multiple fan-beam slice images respectively, which are used as reference images not containing scatter artifacts; extracting multiple cone-beam slice images from the first volume image; performing image registration on the multiple fan-beam slice images and the multiple cone-beam slice images; and training a deep learning-based scatter correction model using the multiple registered images to obtain a trained scatter correction model.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of image processing, and in particular to a scatter correction method for a cone-beam computed tomography (CBCT) system, a cone-beam computed tomography (CBCT) system and a storage medium. BACKGROUND

[0002] Cone-beam computed tomography (CBCT) is an advanced non-destructive testing technique that uses a cone-beam of X-rays and a two-dimensional area detector to acquire projection data in a single rotation, and reconstructs a volumetric image with three-dimensional information. CBCT has been widely used in industrial product testing due to its ability to provide three-dimensional imaging, axial resolution, and other advantages. Despite the many advantages of CBCT, its image quality is easily affected by X-ray scattering, especially in high-energy imaging systems. Because CBCT uses a large-area detector and a wide cone-beam of X-ray irradiation, it results in a large number of scattered photons when penetrating the measured object. These scattered photons are not the primary photons that directly come from the X-ray source and penetrate the object to reach the detector, and their direction and energy can change. After being received by the detector, they will introduce noise and artifacts, which will seriously degrade the image quality. This image quality degradation caused by scattering is considered one of the fundamental factors that limit the image quality of CBCT, especially in applications that require high image quality. Therefore, developing an effective scatter correction method is crucial for improving the application value of high-energy CBCT.

[0003] To solve the scattering problem in CBCT, a variety of scatter correction methods have been proposed, which can be roughly divided into hardware methods, Monte Carlo (MC) simulation methods, and deep learning-based methods.

[0004] Hardware-based methods reduce the scattered lines reaching the detector by using physical devices in the imaging system. Common hardware includes anti-scatter grids, collimators, etc. Anti-scatter grids also absorb a portion of the primary photons, which may require an increase in radiation dose to maintain the signal-to-noise ratio, and cannot completely eliminate scattering. However, in CBCT, it is very difficult to perform similar strict collimation on the entire cone beam.

[0005] Monte Carlo (MC) simulation method: MC simulation accurately calculates the scatter distribution by tracking the transport process of a large number of photons in the object and detector, and is considered one of the gold standards for scatter estimation. However, the main disadvantage of MC simulation is that the amount of calculation is huge and very time-consuming, making it difficult to meet real-time requirements. In addition, the accuracy of MC simulation is highly dependent on the accurate modeling of the imaging system geometry, X-ray spectrum, and the structure and material composition of the measured object.

[0006] Deep learning based methods: In recent years, deep learning techniques have shown great potential in CBCT scatter correction. These methods usually treat scatter correction as an image translation task, learning the mapping from scatter- containing CBCT images to scatter-free CBCT images. The scatter- containing projection data and the corresponding scatter-free data are usually generated by MC simulation as training pairs. The challenge of this approach lies in the time-consuming MC simulation and the difference between the simulation results and the real physical process, which leads to a very inaccurate learning of the mapping from scatter-containing CBCT images to scatter-free CBCT images, and cannot effectively eliminate the scatter.

[0007] Therefore, it is desirable to have a more accurate, faster and more robust scatter correction method for high-energy CBCT systems and CBCT systems. SUMMARY

[0008] To solve the above technical problems, the embodiments of the present disclosure provide a scatter correction method for a high-energy cone beam CT system, which provides a new and high-quality data source for the training of a deep learning model by using the cone beam and fan beam imaging capabilities of the same high-energy CT system itself to obtain a pair of scatter-containing CBCT slices and approximately scatter-free FBCT slices without moving the object. This scheme of using different modes within the same CT system to obtain high-quality training data with intrinsic registration can overcome the limitations of the prior art in training data, and does not need time-consuming simulation calculations for each correction, thereby realizing more accurate, faster and more robust CBCT scatter correction.

[0009] According to the embodiments of the present disclosure, a scatter correction method for a cone beam computed tomography (CBCT) system is provided, which comprises:

[0010] cone beam CT imaging of the object to be measured by a planar array detector of the CBCT system to obtain a first set of projection data of the object to be measured, and reconstructing according to the first set of projection data to obtain a first volume image containing scatter artifacts;

[0011] fan beam CT imaging of multiple different heights of the object to be measured by a linear array detector of the CBCT system to obtain multiple sets of fan beam projection data of the object to be measured, and respectively reconstructing according to the multiple sets of fan beam projection data to obtain multiple fan beam slice images, which are used as reference images without scatter artifacts;

[0012] According to the height information of the multiple sets of fan beam projection data, multiple cone beam slice images corresponding to the multiple fan beam slice images in the height position are extracted from the first volume image;

[0013] image registration is performed on the plurality of fan slice images and the plurality of cone slice images to obtain a plurality of registered image pairs that are aligned in space;

[0014] a deep learning based scatter correction model is trained using the plurality of registered image pairs, wherein the plurality of cone slice images in the plurality of registered image pairs are used as input of the scatter correction model and the plurality of fan slice images are used as labels of the scatter correction model, to obtain a trained scatter correction model; and

[0015] a volume image obtained by cone CT imaging of the CBCT system or an image slice thereof is input into the trained scatter correction model, and a scatter corrected output image is obtained.

[0016] Preferably, the plurality of fan projection data of the object under test are acquired while keeping the position of the object under test relative to the CBCT system unchanged.

[0017] Preferably, the method further comprises:

[0018] By replacing different objects under test and repeating the operations of acquiring a volume image, acquiring fan slice images, acquiring cone slice images, and image registration, an extended dataset containing more than the plurality of registered image pairs is accumulated, and the scatter correction model is trained using the extended dataset.

[0019] Preferably, the image registration on the plurality of fan slice images and the plurality of cone slice images comprises:

[0020] the cone slice images are used as fixed images and the fan slice images are used as floating images;

[0021] geometric transformation is performed on the fan slice images to align them with the cone slice images in space.

[0022] Preferably, the method further comprises:

[0023] data augmentation operations are performed on the cone slice images, the data augmentation operations including one or more of random rotation, flipping, cropping, and elastic deformation;

[0024] data augmentation operations are performed on the fan slice images according to the data augmentation operations performed on the cone slice images.

[0025] Preferably, the image registration on the plurality of fan slice images and the plurality of cone slice images employs a rigid registration, an affine registration, or a non-rigid registration algorithm.

[0026] Preferably, the scatter correction model is based on a convolutional neural network U-Net, and training the deep learning-based scatter correction model using the plurality of registered image pairs to obtain a trained scatter correction model comprises:

[0027] inputting the plurality of cone-beam slice images in the plurality of registered image pairs into a convolutional neural network;

[0028] extracting multi-scale features of the plurality of cone-beam slice images step by step via an encoder of the U-Net through a series of convolutional layers and down-sampling operations, the multi-scale features including global features and local features of scatter artifacts;

[0029] implementing fusion of the multi-scale features via a decoder of the U-Net through up-sampling and de-convolution operations while directly passing features at different scales in the encoder to the decoder with skip connections, so as to reconstruct a corrected image without scatter;

[0030] in the training process, using an L1 or L2 loss function to minimize pixel value differences between the corrected image output by the U-Net and the corresponding fan-beam slice image, and optimizing the U-Net parameters through back propagation;

[0031] when a preset convergence criterion, a maximum number of iterations, or a condition of optimal performance on a validation set is met, determining and saving the trained U-Net as a final scatter correction model.

[0032] Preferably, the scatter correction model is based on a generative adversarial network, and training the deep learning-based scatter correction model using the plurality of registered image pairs to obtain a trained scatter correction model comprises:

[0033] inputting the plurality of cone-beam slice images in the plurality of registered image pairs into a generative adversarial network;

[0034] a generator network of the generative adversarial network generates a corrected image without scatter;

[0035] a discriminator network of the generative adversarial network is trained to distinguish between true and false between the corrected image generated by the generator and the corresponding fan-beam slice image;

[0036] using an adversarial loss, the generator is driven to generate images capable of deceiving the discriminator through adversarial learning, while the discriminator is driven to accurately distinguish between real images and generated images;

[0037] using an L1 loss or an L2 loss at the pixel level as a content loss to ensure that the corrected image output by the generator is similar in structure and content to the fan-beam slice image;

[0038] The weight parameters of the generator and the discriminator are optimized by back propagation based on a combination of the adversarial loss and the content loss.

[0039] When a preset convergence criterion, a maximum number of iterations, or a condition of optimal performance on a validation set is met, the trained generator network is determined and saved as a final scatter correction model.

[0040] Preferably, the scatter correction model is based on a Transformer model, and training a deep learning-based scatter correction model using the plurality of registered image pairs to obtain a trained scatter correction model comprises:

[0041] The plurality of cone-beam slice images in the plurality of registered image pairs are processed by blocking, and the blocked image segments are input into the Transformer model as inputs;

[0042] The Transformer model effectively captures long-range dependencies between different blocks in an image via its self-attention mechanism to obtain global and local features of the image, and the long-range dependency features are related to complex global scatter artifacts;

[0043] The captured feature information is remapped and reconstructed into a scatter-free corrected image via a decoding part or a related reconstruction layer of the Transformer model;

[0044] A pixel-level L1 loss or L2 loss function is used to minimize the pixel value difference between the corrected image output by the Transformer model and the corresponding fan-beam slice image, and the weight parameters of the Transformer model are optimized based on the difference by back propagation;

[0045] When a preset convergence criterion, a maximum number of iterations, or a condition of optimal performance on a validation set is met, the trained Transformer model is determined and saved as a final scatter correction model.

[0046] Preferably, the scatter correction model is based on a diffusion model, and training a deep learning-based scatter correction model using the plurality of registered image pairs to obtain a trained scatter correction model comprises:

[0047] A forward process is defined, which gradually adds noise to the fan-beam slice image through a series of time steps;

[0048] A conditional denoising network is trained, which takes a cone-beam slice image containing scatter as a conditional input and receives a noisy image at any time step, and the goal of the conditional denoising network is to predict the noise added to the image at that time step;

[0049] using an L2 loss function to minimize a difference between noise predicted by the conditional denoising network and the actually added noise, and based on the difference to optimize weight parameters of the conditional denoising network by back propagation;

[0050] In inference, the trained conditional denoising network is used to simulate a reverse denoising process to generate a scatter-free corrected image from a noisy image in an iterative manner, referring to the structural information of the scatter- containing cone-beam slice image;

[0051] When a preset convergence criterion, a maximum number of iterations, or a condition of optimal performance on a validation set is met, the trained conditional denoising network is determined and saved as a final scatter correction model.

[0052] Preferably, the linear array detector has a rear collimator to shield scattered lines.

[0053] According to another embodiment of the present disclosure, a cone-beam computed tomography (CBCT) system is provided, comprising:

[0054] a high-energy X-ray source;

[0055] a planar array detector configured to acquire a first set of projection data of an object to be measured in a cone-beam CT mode;

[0056] a linear array detector configured to acquire a plurality of sets of fan-beam projection data of the object to be measured at different heights in a fan-beam CT mode;

[0057] a control system configured to record respective height information of the sets of fan-beam projection data;

[0058] an image processing unit configured to:

[0059] reconstruct from the first set of projection data to obtain a first volume image containing scatter artifacts;

[0060] reconstruct from the plurality of sets of fan-beam projection data to obtain a plurality of fan-beam slice images, which are used as reference images without scatter artifacts;

[0061] extract a plurality of cone-beam slice images corresponding to the plurality of fan-beam slice images in height position from the first volume image according to the height information of the plurality of sets of fan-beam projection data;

[0062] image registration is performed on the plurality of fan-beam slice images and the plurality of cone-beam slice images to obtain a plurality of registered image pairs aligned in space;

[0063] training a deep learning based scatter correction model using the plurality of registered image pairs, wherein the plurality of cone-beam slice images in the plurality of registered image pairs are as inputs of the scatter correction model and the plurality of fan-beam slice images are as labels of the scatter correction model; and

[0064] inputting a volume image or an image slice thereof obtained by cone-beam CT imaging of the CBCT system into the trained scatter correction model, and obtaining a scatter corrected output image.

[0065] Preferably, the line array detector acquires a plurality of fan-beam projection data of the object to be measured under the condition that the position of the object to be measured relative to the CBCT system is kept unchanged.

[0066] According to another embodiment of the present disclosure, a computer readable storage medium is provided, having stored thereon computer readable instructions which, when executed by a processor, cause the processor to perform the method described above.

[0067] Therefore, according to the precise scatter correction method and system for high-energy CBCT images according to the embodiments of the present disclosure, the inherent scatter suppression capability of fan-beam CT combined with line array detector is utilized, and high-quality training data for deep learning model can be generated without additional hardware. In addition, by utilizing the cone-beam and fan-beam imaging capability of the same high-energy CT system itself, paired CBCT slice containing scatter and FBCT slice approximately without scatter are acquired without movement of the object, providing a large amount of high-fidelity real training data pairs of scatter-containing views and scatter-free views generated by the same imaging system for training of the deep learning model, so that time-consuming simulation calculation is not required for each correction in the training stage. This scheme of utilizing different modes within the same CT system to obtain high-quality training data with inherent registration can overcome the limitations of the prior art in terms of training data, without the need for time-consuming simulation calculation for each correction, thereby realizing more accurate, faster and more robust CBCT scatter correction. BRIEF DESCRIPTION OF DRAWINGS

[0068] The above and other aspects, features, and advantages of certain embodiments of the present disclosure will be more apparent from the following description taken in conjunction with the following drawings, in which:

[0069] Figure 1 is a block diagram illustrating a high-energy CT system capable of cone-beam and fan-beam acquisition according to an embodiment of the present disclosure;

[0070] Figure 2 is a flowchart illustrating a scatter correction method according to an embodiment of the present disclosure;

[0071] Figure 3FIG. 1 is a training flowchart of a convolutional neural network-based scatter correction model according to an embodiment of the present application;

[0072] Figure 4 FIG. 2 is a training flowchart of a generative adversarial network-based scatter correction model according to an embodiment of the present application;

[0073] Figure 5 FIG. 3 is a training flowchart of a Transformer model-based scatter correction model according to an embodiment of the present application;

[0074] Figure 6 FIG. 4 is a training flowchart of a diffusion model-based scatter correction model according to an embodiment of the present application;

[0075] Figure 7 FIG. 5 is a schematic diagram of an image example according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0076] Before undertaking the detailed description below, it can be advantageous to set forth definitions of certain words and phrases used throughout this patent document. The terms "include" and "comprise," as well as derivatives thereof, mean inclusion without limitation. The term "or," is used in the inclusive sense of "and / or" unless it is used in capitalized form, e.g., "OR." The term "control" or "controller" means any device, apparatus, or portion thereof capable of controlling at least one operation. Such a controller can be implemented in hardware or a combination of hardware and software and / or firmware. The functionality associated with any particular controller can be centralized or distributed, whether locally or remotely. The phrase "at least one of," when used with respect to a list of items, means that one or more of the listed items can be used, and that the list is extensible. For example, "at least one of A, B, and C" includes A, B, C, A and B, A and C, B and C, A and B and C, and the like.

[0077] Definitions for other certain words and phrases are provided throughout this patent document. Those of ordinary skill in the art will understand that in many, if not most instances, such definitions apply to prior and future uses of such defined words and phrases.

[0078] Various embodiments of the principles of the present disclosure described in this patent document can be implemented in any of a variety of suitable arrangements. In some cases, the actions described in the specification can be performed in a different order, and still achieve desirable results. Additionally, the process depicted in the figures can not necessarily require the particular order shown, or sequential order, to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous.

[0079] Figure 1 is a block diagram showing a high-energy CT system capable of cone-beam and fan-beam acquisition according to an embodiment of the present application.

[0080] As shown in Figure 1 , a cone-beam computed tomography (CBCT) system 100 according to an embodiment of the present application includes a high-energy X-ray source 101, a two-dimensional (2D) area detector 102, a linear array detector 103, a control system 104, and an image processing unit 105. The CBCT system 100 is designed to perform the scatter correction method according to an embodiment of the present application.

[0081] The hardware structure core of the CBCT system 100 is a high-energy X-ray source 101 and a dual-detector imaging system including the 2D area detector 102 and the linear array detector 103. A key feature of the CBCT system 100 is that it has dual-modality imaging capability: cone-beam CT mode and fan-beam CT mode.

[0082] The term "high-energy" refers to the X-ray energy range used by the system, for example, in the mega-volt (MeV) energy range, which is commonly used in the detection of high-density, large-size metal parts. In the high-energy case, the need for scatter correction is more urgent.

[0083] In the CBCT mode, the high-energy X-ray source 101 generates a cone-shaped X-ray beam, which is received by a 2D area detector 102 after penetrating the object, obtaining an angular projection data. Through the rotation of the gantry, multi-angle projection data is obtained, and then the imaging voxel data is reconstructed by the FDK algorithm.

[0084] In the fan-beam CT (FBCT) mode, the CBCT system 100 uses a linear array detector 103 for data acquisition. The high-energy X-ray source 101 generates a fan-shaped X-ray beam. The linear array detector 103 usually has a narrow effective width. In an embodiment, it is equipped with a high-efficiency back collimator, which can very effectively suppress scattered X-rays reaching the detector. Therefore, the image acquired by the linear array detector 103 is considered to be approximately scatter-free or with very low scatter level. The linear array detector 103 can be moved in the Z-axis and perform multiple scans at multiple fixed positions to acquire multiple tomographic images of the object at different heights.

[0085] The system also includes a control system 104 and a mechanical system 106. The mechanical system 106 is used to support the X-ray source, the gantry, and the detector system. The control system 104 is used to control the relative positions of the X-ray source, the imaging system, and the gantry. The control system 104 is also responsible for accurately recording the Z-axis position information of the linear array detector at each FBCT scan.

[0086] The raw projection data acquired by the detector system is transmitted to the image processing unit 105. The image processing unit 105 is typically a computer equipped with a powerful processor (CPU), graphics processing unit (GPU), and sufficient memory. It is responsible for executing the CT image reconstruction algorithm and the inference application of this invention. The processing results can be displayed via a display device.

[0087] This dual-modal capability, particularly the ability to integrate a linear array detector with high scattering suppression and a planar array detector for rapid volumetric scanning on the same rack, is the physical basis for the ability of this invention to generate high-quality, intrinsically registered training data. The system design ensures that the geometric relationship of the object relative to the imaging system can be precisely controlled and recorded when acquiring data in both modes.

[0088] Below, we will refer to Figure 2 The scattering correction method performed by the CBCT system 100 is described in detail.

[0089] like Figure 2 As shown, the scattering correction method 200 according to an embodiment of this disclosure includes:

[0090] S201: The CBCT system uses an array detector to perform cone-beam CT imaging on the object under test to obtain a first set of projection data of the object under test, and reconstructs a first volumetric image containing scattering artifacts based on the first set of projection data.

[0091] S202: The linear array detector of the CBCT system performs fan-beam CT imaging at multiple different heights of the object under test to obtain multiple sets of fan-beam projection data of the object under test, and reconstructs multiple fan-beam slice images based on the multiple sets of fan-beam projection data. The multiple fan-beam slice images are used as reference images without scattering artifacts.

[0092] S203: Based on the height information of the multiple fan beam projection data, extract multiple cone beam slice images from the first volume image that correspond to the multiple fan beam slice images in height position.

[0093] S204: Perform image registration on the plurality of fan-beam slice images and the plurality of cone-beam slice images to obtain a plurality of registered image pairs aligned in space.

[0094] S205: Use the multiple registered image pairs to train a deep learning-based scattering correction model to obtain a trained scattering correction model, wherein the multiple cone-beam slice images in the multiple registered image pairs are used as inputs to the scattering correction model, and the multiple fan-beam slice images are used as labels for the scattering correction model.

[0095] S206: input the volume image or image slice thereof obtained by the CBCT system into the trained scatter correction model, and obtain the scatter-corrected output image.

[0096] Specifically, the operations performed by the CBCT system 100 include a data acquisition process, a data set preparation process, a model training process, and a model deployment process.

[0097] 1. Data acquisition process

[0098] In the data acquisition stage, the CBCT system 100 performs (1) cone beam CT image acquisition and (2) fan beam CT slice acquisition.

[0099] (1) Cone beam CT image acquisition (cone beam CT mode):

[0100] The object to be measured is placed on the turntable of the CBCT system 100. The turntable rotates the object, the high-energy X-ray source 101 generates a cone-shaped X-ray beam, and after penetrating the object, the first set of cone beam projection data of the object is collected by the area array detector 102.

[0101] These projection data are then reconstructed by the image processing unit 105 to obtain a three-dimensional CBCT image A. The size of the image A can be represented as H1xW1xC, where C represents the number of slices along the Z axis. Due to the wide field of view and large irradiation volume of CBCT, the image A is considered to contain significant X-ray scattering components, resulting in a decrease in image quality.

[0102] This corresponds to the operation in step S201, which performs cone beam CT imaging on the object to be measured by the area array detector 102 of the CBCT system 100 to obtain a first set of projection data of the object to be measured, and reconstructs according to the first set of projection data to obtain a first volume image containing scattering artifacts.

[0103] (2) Fan beam CT image acquisition (cone beam CT mode):

[0104] After completing the CBCT acquisition, the position of the object to be measured on the turntable is kept completely unchanged. The system switches to the FBCT mode and uses the line array detector 103 to perform fan beam CT scanning on the object to be measured at N different Z-axis heights. For each specified height, fan beam projection data at that height is collected, and a two-dimensional fan beam CT slice image is reconstructed. This process is repeated N times to obtain N fan beam CT slice images at different heights, each with a size of H2xW2.

[0105] This corresponds to the operation in step S202, a plurality of different heights of the object to be measured are fan-beam CT imaged by the line array detector 103 of the CBCT system 100 to acquire a plurality of sets of fan-beam projection data of the object to be measured, and a plurality of fan-beam slice images are respectively reconstructed from the plurality of sets of fan-beam projection data, which are used as reference images free of scatter artifacts.

[0106] As previously described, due to the effective suppression of scattered lines by the line array detector 103, the N fan-beam CT slice images are considered to be approximately scatter-free, or their scatter contamination is much lower than that of the CBCT images. Therefore, they can be used as the "gold standard" reference images for subsequent deep learning model training. The number of N should be large enough to cover the representative anatomical regions and different scatter characteristics of the object to be measured.

[0107] 2. Data set preparation process

[0108] After the CBCT volume image A and the N FBCT slice images are acquired, a series of processing steps are needed to build a data set for deep learning model training. The image processing unit 105 is configured to perform this series of processing steps.

[0109] As described above, in the CBCT mode, the image processing unit 105 is configured to reconstruct the first set of projection data of the object acquired by the area array detector 102 to obtain the first volume image A containing scatter artifacts.

[0110] In the FBCT mode, the image processing unit 105 is configured to respectively reconstruct a plurality of fan-beam slice images from a plurality of sets of fan-beam projection data of the object acquired by the line array detector 103. The plurality of fan-beam slice images are used as reference images free of scatter artifacts.

[0111] The image processing unit 105 is configured to perform (1) paired slice extraction, (2) image registration, and (3) label assignment and data set accumulation.

[0112] (1) Paired slice extraction

[0113] For each FBCT slice acquired, the control system 104 records its exact Z-axis position. According to this recorded height information, in the three-dimensional CBCT volume image A, a corresponding CBCT slice at the same anatomical level position as each FBCT slice is found and extracted along the Z-axis. In this way, the image processing unit 105 obtains N CBCT slices, which correspond one-to-one with the N FBCT slices in the Z-axis position.

[0114] This corresponds to the operation in step S203, the image processing unit 105 can extract, from the first volume image, a plurality of cone-beam slice images corresponding in height position to the plurality of fan-beam slice images.

[0115] (2) Image registration

[0116] Although the CBCT slices and the FBCT slices have been corresponded in Z-axis position, they are not perfectly aligned in the X-Y plane due to the difference in geometric parameters of the two scanning modes. Therefore, accurate two-dimensional image registration is needed for each pair of images. In the registration process, the CBCT slice is taken as the reference image (fixed image), while the FBCT slice is taken as the floating image, which is aligned with the CBCT slice by applying geometric transformation. The selection of keeping the CBCT slice fixed and registering the FBCT slice to it is to ensure that the input of the deep learning model maintains its original, non-interpolated data characteristics as much as possible. After registration, N pairs of spatially accurately aligned images can be obtained.

[0117] This corresponds to the operation in step S204, the image processing unit 105 can perform image registration on the plurality of fan-beam slice images and the plurality of cone-beam slice images to obtain a plurality of registered image pairs that are aligned in space.

[0118] (3) Label assignment and data set accumulation:

[0119] In each image pair, the registered FBCT slice is used as the "gold standard" or "label" image for the deep learning model due to its approximately scatter-free characteristic. While the corresponding CBCT slice is taken as the "input" image containing scatter. In order to train a robust and generalizable deep learning model, a large amount of diverse training data is needed. Therefore, different objects to be tested (e.g., parts of different sizes, shapes, material compositions) can be replaced as needed, and the above steps can be repeated to accumulate a large data set containing hundreds or even thousands of such paired image slices. The diversity of the data set is crucial to ensure that the model can handle scatter artifacts in various situations.

[0120] 3. Model training process

[0121] Using the above prepared data set, a deep learning-based scatter correction model according to an embodiment of the present disclosure can be trained for predicting the corresponding scatter-free slice from the input CBCT slice containing scatter.

[0122] Specifically, in step S205, the image processing unit 105 can use the plurality of registered image pairs to train a deep learning-based scatter correction model, with the plurality of cone-beam slice images in the plurality of registered image pairs as inputs to the scatter correction model and the plurality of fan-beam slice images as labels to the scatter correction model, to obtain a trained scatter correction model.

[0123] In the embodiments of the present disclosure, a plurality of deep learning model architectures can be selected to achieve this goal. The deep learning model architectures can include, but are not limited to, convolutional neural networks (CNNs), generative adversarial networks (GANs), Transformer models, and diffusion models.

[0124] It is flexible to choose which specific model architecture, and a trade-off can be made according to the available computing resources, the size of the dataset, and the specific requirements on the quality of the corrected images. However, the unique high-quality paired training data provided by the present disclosure makes it possible for even relatively simple models to perform well, while also providing a solid foundation for more complex models to fully exploit their potential.

[0125] An optimizer is also needed for deep learning training, and commonly used optimizers include Adam, RMSprop, or stochastic gradient descent (SGD) and its variants.

[0126] The training strategies for deep learning models include:

[0127] Batch Training: The training data is divided into small batches for iterative training.

[0128] Data Augmentation: In order to increase the diversity of training data and improve the generalization ability of the model, data augmentation operations can be performed on the input CBCT slices, such as random rotation, flipping, cropping, elastic deformation, etc. It should be noted that any geometric transformation performed on the CBCT slice must also be applied to the corresponding FBCT label slice to maintain the precise registration relationship between them.

[0129] Learning Rate Scheduling: Dynamically adjust the learning rate during training, for example, use a larger learning rate at the beginning of training, and then gradually reduce it.

[0130] Validation Set: A portion of the training dataset is divided as a validation set, which is used to monitor the performance of the model during training, avoid overfitting, and adjust hyperparameters.

[0131] The evaluation of the deep learning model during training includes:

[0132] During training, the performance of the model should be evaluated periodically on the validation set. Evaluation metrics can include the value of the loss function, as well as image quality objective metrics such as Peak Signal-to-Noise Ratio (PSNR), SSIM, MAE, etc., which are used to compare the corrected image output by the model with the FBCT “ground truth” image.

[0133] In the following, the process of training to obtain a scatter correction model will be described in detail for each specific deep learning model.

[0134] As mentioned before, the CBCT slices in the dataset are used as input to the model, and the corresponding FBCT slices, which are registered, are used as labels for the model.

[0135] <First embodiment>

[0136] In the following, the training process of a convolutional neural network (CNN) based scatter correction model according to embodiments of the present disclosure will be described with reference to Figure 3 The scatter correction model is based on a convolutional neural network (CNN), in particular a CNN with an encoder-decoder structure, such as a U-Net, a variant of U-Net, etc.

[0137] As shown in Figure 3 The training of the convolutional neural network (CNN) based scatter correction model comprises:

[0138] Step S301 : inputting the plurality of cone beam slice images in the plurality of registered image pairs into the convolutional neural network;

[0139] Step S302: stepwise extracting multi-scale features of the plurality of cone beam slice images via the encoder of the U-Net by a series of convolutional layers and down-sampling operations, the multi-scale features including global features and local features of the scatter artifacts;

[0140] Step S303: fusing the multi-scale features via the decoder of the U-Net by up-sampling and de-convolution operations, while directly passing the features at different scales in the encoder to the decoder using skip connections, to reconstruct the scatter-free corrected image;

[0141] Step S304: during the training process, using an L1 or L2 loss function to minimize the pixel value difference between the corrected image output by the U-Net and the corresponding fan beam slice image, and back-propagating to optimize the parameters of the U-Net;

[0142] Step S305: When the preset convergence criterion, the maximum number of iterations, or the optimal performance on the validation set is met, the trained U-Net is determined and saved as the final scatter correction model.

[0143] It should be noted that for the scatter correction model, the selection of the loss function depends on the model architecture adopted. For a direct regression model similar to U-Net, commonly used loss functions include the mean absolute error (MAE, L1 loss) or the pixel-level mean squared error (MSE, L2 loss). The L1 loss usually produces clearer edges. The L2 loss is more sensitive to outliers.

[0144] <Second Embodiment>

[0145] Below, the training process of a scatter correction model based on a generative adversarial network (GAN) according to an embodiment of the present disclosure will be described with reference to Figure 4

[0146] As shown in Figure 4 , the training of a scatter correction model based on a generative adversarial network (GAN) includes:

[0147] S401: inputting the plurality of cone-beam slice images in the plurality of registered image pairs into a generative adversarial network;

[0148] S402: a generator network of the generative adversarial network generates a scatter-free corrected image;

[0149] S403: a discriminator network of the generative adversarial network is trained to distinguish between true and false between the corrected image generated by the generator and the corresponding fan-beam slice image;

[0150] S404: using an adversarial loss, the generator is driven by adversarial learning to generate images that can deceive the discriminator, while the discriminator is driven to accurately distinguish between real images and generated images;

[0151] S405: using a pixel-level L1 loss or L2 loss as a content loss to ensure that the corrected image output by the generator is similar in structure and content to the fan-beam slice image;

[0152] S406: based on the combination of the adversarial loss and the content loss, back propagation is performed to optimize the weight parameters of the generator and the discriminator;

[0153] S407: when the preset convergence criterion, the maximum number of iterations, or the optimal performance on the validation set is met, the trained generator network is determined and saved as the final scatter correction model. ​

[0154] For GAN models, the loss function is usually composed of two parts: adversarial loss, which is used to drive the generator to produce images that can "fool" the discriminator; and content loss, which is used to ensure that the generated image is similar to the label image in content. The content loss can be L1 or L2 loss.

[0155] <Third Embodiment>

[0156] Below, the training process of the scatter correction model based on the Transformer model according to the third embodiment of the present disclosure will be described with reference to Figure 5

[0157] As shown in Figure 5 , the training of the scatter correction model based on the Transformer model includes:

[0158] S501: block processing is performed on the plurality of cone-beam slice images in the plurality of registered image pairs, and the image segments after the block processing are input as input into the Transformer model;

[0159] S502: via the Transformer model, long-range dependency relationships between different blocks in the image are effectively captured by using the self-attention mechanism thereof, so as to obtain global and local features of the image, and the long-range dependency relationship features are related to complex global scatter artifacts;

[0160] S503: via a decoding part or a related reconstruction layer of the Transformer model, the captured feature information is remapped and reconstructed into a corrected image without scatter;

[0161] S504: a pixel-level L1 loss or L2 loss function is used to minimize the pixel value difference between the corrected image output by the Transformer model and the corresponding fan-beam slice image, and the difference is back-propagated to optimize the weight parameters of the Transformer model;

[0162] S505: when a preset convergence criterion, a maximum number of iterations, or a condition of optimal performance on a validation set is met, the trained Transformer model is determined and saved as a final scatter correction model.

[0163] ​For the Transformer model, other loss terms can also be introduced, such as perceptual loss, which compares the difference between the generated image and the label image in the high-level feature space, helping to improve the visual quality of the generated image. Structural Similarity Index (SSIM) loss or its multi-scale version (MS-SSIM) can also be used to better preserve image structure.

[0164] <Fourth Embodiment>

[0165] Below, the training process of the scatter correction model based on the diffusion model according to the fourth embodiment of the present disclosure will be described with reference to Figure 6

[0166] As shown in Figure 6 , the training of the scatter correction model based on the diffusion model includes:

[0167] S601: Define a forward process that gradually adds noise to the fan-beam slice image through a series of time steps;

[0168] S602: Train a conditional denoising network that takes the scatter-containing cone-beam slice image as a conditional input and receives the noisy image at any time step, the goal of the conditional denoising network is to predict the noise added to the image at that time step;

[0169] S603: Use the L2 loss function to minimize the difference between the noise predicted by the conditional denoising network and the actual added noise, and based on this difference, backpropagate to optimize the weight parameters of the conditional denoising network;

[0170] S604: During inference, the trained conditional denoising network is used to simulate a reverse denoising process to generate a scatter-free corrected image from a noisy image in an iterative manner, referring to the structural information of the scatter-containing cone-beam slice image;

[0171] S605: When the preset convergence criteria, maximum number of iterations, or optimal performance on the validation set are met, determine and save the trained conditional denoising network as the final scatter correction model.

[0172] For the diffusion model, the training goal is usually to predict the noise added to the image during the diffusion process.

[0173] 4. Model deployment process

[0174] Once the deep learning model is successfully trained and validated, it can enter the inference phase for correcting new CBCT images.

[0175] ​Specifically, in step S206, the volume image or image slice thereof obtained by the cone beam CT imaging of the CBCT system is input into the trained scatter correction model, and an output image after scatter correction is obtained.

[0176] The specific process of deployment is as follows:

[0177] (1) For a new object to be measured, a conventional CBCT scan is performed using a high-energy CT system to obtain original projection data.

[0178] (2) The original projection data is reconstructed to obtain a three-dimensional CBCT image containing scatter artifacts.

[0179] (3) The reconstructed CBCT image is processed slice by slice, and each slice is input into the already trained deep learning-based scatter correction model.

[0180] (4) The scatter correction model performs forward propagation calculation on the input CBCT data, and outputs an estimated CBCT image without scatter (or with significantly reduced scatter).

[0181] (5) The output image is the final, scatter-corrected, and image quality improved result. It can be used for subsequent applications such as defect analysis and size measurement.

[0182] Figure 7 An exemplary result is shown, where (a) is the original CBCT slice containing scatter, the corresponding FBCT reference slice, (c) is the FBCT slice after image registration, and (d) is the scatter-corrected CBCT slice obtained after applying the scatter correction model trained according to the present application, which can be seen that the artifacts are reduced and the contrast is improved.

[0183] At least one embodiment according to the present disclosure can be a computer readable storage medium. The computer readable storage medium has computer instructions stored thereon, which, when executed by a processor, perform one or more steps of the various methods described above and additional aspects thereof.

[0184] Exemplarily, the non-transitory computer readable storage medium can be any combination of one or more computer readable storage media, for example, one computer readable storage medium contains program code for executing the various methods described above.

[0185] Exemplarily, when the program code is read by a computer, the computer can execute the program code stored in the computer storage medium to perform one or more steps to implement, for example, the various methods described above and additional aspects thereof according to at least one embodiment of the present disclosure.

[0186] Exemplarily, the computer-readable storage medium can include a memory card of a smart phone, a storage component of a tablet computer, a hard disk of a personal computer, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disc read-only memory (CD-ROM), a flash memory, and other non-transitory readable storage medium or any combination thereof.

[0187] The method and system for precise scatter correction of high-energy CBCT images according to the embodiments of the present disclosure utilize the inherent scatter suppression capability of fan-beam CT combined with a line array detector, and can generate high-quality training data for a deep learning model without additional hardware. In addition, by utilizing the cone-beam and fan-beam imaging capabilities of the same high-energy CT system, paired CBCT slices with scatter and FBCT slices with approximate no scatter are obtained without movement of the object, providing a large amount of high-fidelity real training data pairs of scatter-containing views and scatter-free views generated by the same imaging system for training of the deep learning model, so that time-consuming simulation calculations are not required for each correction in the training stage. This scheme of utilizing different modes within the same CT system to obtain high-quality training data with inherent registration can overcome the limitations of the prior art in terms of training data, and does not require time-consuming simulation calculations for each correction, thereby realizing more accurate, faster and more robust CBCT scatter correction.

[0188] The method for precise scatter correction of high-energy CBCT images according to the embodiments of the present disclosure proposes a novel deep learning-based solution to the problem of image quality degradation caused by X-ray scatter that is prevalent in high-energy CBCT images. The core is to utilize the same high-energy CT system to first obtain CBCT volume data containing scatter and multiple FBCT slices with approximate no scatter corresponding to the anatomical position under the condition of keeping the object to be measured stationary. Through accurate registration, the CBCT slice and the corresponding FBCT slice form a training pair for training a deep learning model. After training is completed, the model can be used to correct newly acquired CBCT images, effectively removing scatter artifacts and improving the quantitative accuracy and visual quality of the images. The present invention overcomes the limitations of the prior art in terms of training data acquisition by generating high-quality, real and perfectly registered training data.

[0189] It should be noted that the flowcharts and block diagrams in the drawings are illustrations of the possible architectures, functional processes, and operations of systems, methods, and computer program products in accordance with various embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or combinations of hardware and software.

[0190] In general, the various example embodiments of the application can be implemented in hardware or special-purpose circuits, software, firmware, logic, or any combination thereof. Some aspects of the application can be implemented in hardware, while other aspects can be implemented in firmware or software which can be executed by a controller, microprocessor or other computing device, although the application is not limited thereto. While various aspects of the application can be illustrated and described as block diagrams, flow charts, or using some other pictorial representation, it is well understood that these forms of representation are used by those skilled in the art to facilitate discussion of the concepts involved.

[0191] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0192] The foregoing is a summary of the present application and is not to be considered as limiting its scope. While several example embodiments of the application have been described, it will be apparent to those of ordinary skill in the art that many modifications are possible without departing from the novel teachings and advantages of the application. The embodiments chosen and described are meant to be illustrative only and are not intended to limit the scope of the application. It is to be understood that the above description is intended to be illustrative and not restrictive. Many embodiments will be apparent to those of skill in the art upon reviewing the above description. The scope of the application should, therefore, be determined not with reference to the above description, but should instead be determined with reference to the appended claims, along with their full scope of equivalents.

Claims

1. A scattering correction method for a cone-beam computed tomography (CBCT) system, characterized in that, include: The CBCT system uses an array detector to perform cone-beam CT imaging on the object under test to obtain a first set of projection data of the object under test, and reconstructs a first volumetric image containing scattering artifacts based on the first set of projection data. The CBCT system uses a linear array detector to perform fan-beam CT imaging at multiple different heights of the object under test to obtain multiple sets of fan-beam projection data of the object under test. Multiple fan-beam slice images are then reconstructed based on the multiple sets of fan-beam projection data. These multiple fan-beam slice images are used as reference images without scattering artifacts. Based on the height information of the multiple fan beam projection data, extract multiple cone beam slice images from the first volume image that correspond to the multiple fan beam slice images in height position; Image registration is performed on the plurality of fan-beam slice images and the plurality of cone-beam slice images to obtain multiple registered image pairs aligned in space; The multiple registered image pairs are used to train a deep learning-based scattering correction model to obtain a trained scattering correction model, wherein the multiple cone-beam slice images in the multiple registered image pairs are used as inputs to the scattering correction model, and the multiple fan-beam slice images are used as labels for the scattering correction model. as well as The volumetric image or its image slice obtained by cone-beam CT imaging of the CBCT system is input into the trained scattering correction model to obtain the scattering-corrected output image.

2. The method according to claim 1, characterized in that, While keeping the position of the object under test relative to the CBCT system unchanged, acquire multiple sector projection data of the object under test.

3. The method according to claim 2, characterized in that, Also includes: By changing different test objects and repeatedly performing the operations of acquiring volume images, acquiring fan-beam slice images, acquiring cone-beam slice images, and image registration, an extended dataset containing more than the number of registered image pairs is accumulated, and the scattering correction model is trained using the extended dataset.

4. The method according to claim 1, characterized in that, Image registration of the plurality of fan-beam slice images and the plurality of cone-beam slice images includes: The cone-beam slice image is used as a fixed image, and the fan-beam slice image is used as a floating image; The fan-beam slice image is geometrically transformed to be spatially aligned with the cone-beam slice image.

5. The method according to claim 4, characterized in that, Also includes: Perform data augmentation operations on the cone-beam slice image, including one or more of random rotation, flipping, cropping, and elastic deformation; Based on the data augmentation operation performed on the cone-beam slice image, a data augmentation operation is performed on the fan-beam slice image.

6. The method according to claim 1, characterized in that, Image registration of the plurality of fan-beam slice images and the plurality of cone-beam slice images is performed using rigid registration, affine registration, or non-rigid registration algorithms.

7. The method according to claim 1, characterized in that, The scattering correction model is based on the convolutional neural network U-Net. Using the multiple registered images to train a deep learning-based scattering correction model to obtain a trained scattering correction model includes: The multiple cone-beam slice images from the multiple registered image pairs are input into a convolutional neural network; The encoder of the U-Net extracts multi-scale features of the multiple cone-beam slice images step by step through a series of convolutional layers and downsampling operations. The multi-scale features include global and local features of scattering artifacts. The U-Net decoder uses upsampling and deconvolution operations, and skip connections to directly pass features of different scales from the encoder to the decoder, thereby achieving the fusion of multi-scale features and reconstructing a scatter-free corrected image. During training, L1 or L2 loss functions are used to minimize the pixel value difference between the corrected image output by the U-Net and the corresponding fan slice image, and the U-Net parameters are optimized through backpropagation. When the preset convergence criteria, maximum number of iterations, or optimal performance on the validation set are met, the trained U-Net is determined and saved as the final scattering correction model.

8. The method according to claim 1, characterized in that, The scattering correction model is based on generative adversarial networks. Using the multiple registered images to train a deep learning-based scattering correction model to obtain a trained scattering correction model includes: The multiple cone-beam slice images from the multiple registered image pairs are input into a generative adversarial network; The generator network of the generative adversarial network generates a scatter-free corrected image. The discriminator network of the generative adversarial network is trained to distinguish between the genuine and fake corrected images generated by the generator and the corresponding fan-shaped slice images; Using adversarial loss, the generator is driven by adversarial learning to generate images that can deceive the discriminator, while simultaneously driving the discriminator to accurately distinguish between real and generated images; Pixel-level L1 or L2 loss is used as content loss to ensure that the corrected image output by the generator is similar to the fan-shaped slice image in structure and content. Based on the combination of the adversarial loss and the content loss, backpropagation is performed to optimize the weight parameters of the generator and the discriminator; When the preset convergence criteria, maximum number of iterations, or optimal performance on the validation set are met, the trained generator network is determined and saved as the final scattering correction model.

9. The method according to claim 1, characterized in that, The scattering correction model is based on the Transformer model. Using the multiple registered images to train a deep learning-based scattering correction model to obtain a trained scattering correction model includes: The multiple cone-beam slice images in the multiple registered image pairs are divided into blocks, and the block-shaped image fragments are used as input to the Transformer model. The Transformer model effectively captures long-range dependencies between different blocks in an image using its self-attention mechanism to obtain global and local features of the image. These long-range dependency features are related to complex global scattering artifacts. The captured feature information is remapped and reconstructed into a scatter-free corrected image via the decoding part or related reconstruction layer of the Transformer model; Pixel-level L1 or L2 loss functions are used to minimize the pixel value difference between the corrected image output by the Transformer model and the corresponding fan slice image, and backpropagation is performed based on this difference to optimize the weight parameters of the Transformer model. When the preset convergence criteria, maximum number of iterations, or optimal performance on the validation set are met, the trained Transformer model is determined and saved as the final scattering correction model.

10. The method according to claim 1, characterized in that, The scattering correction model is based on the diffusion model. Using the multiple registered images to train a deep learning-based scattering correction model to obtain a trained scattering correction model includes: Define a forward process that gradually adds noise to the fan-slice image through a series of time steps; Train a conditional denoising network that takes a cone-beam slice image containing scattering as a conditional input and receives a noisy image at any time step. The goal of the conditional denoising network is to predict the noise that will be added to the image at that time step. The L2 loss function is used to minimize the difference between the noise predicted by the conditional denoising network and the actual noise added, and backpropagation is performed based on this difference to optimize the weight parameters of the conditional denoising network. During inference, the trained conditional denoising network is used to simulate an inverse denoising process, starting from the noisy image in an iterative manner and referring to the structural information of the cone-beam slice image containing scattering to generate a scatter-free corrected image. When the preset convergence criteria, maximum number of iterations, or optimal performance on the validation set are met, the trained conditional denoising network is determined and saved as the final scattering correction model.

11. The method according to claim 1, characterized in that, The linear array detector has a back collimator to shield scattered radiation.

12. A cone-beam computed tomography (CBCT) system, characterized in that, include: A high-energy X-ray source; A single area array detector, configured to acquire the first set of projection data of the object under test in cone-beam CT mode; A linear array detector, configured to acquire multiple sets of fan-beam projection data of the object under test at multiple different heights in fan-beam CT mode; The control system is configured to record the respective height information of the multiple sets of fan beam projection data; Image processing unit, configured for: A first volumetric image containing scattering artifacts is reconstructed based on the first set of projection data. Multiple fan beam slice images are reconstructed based on the multiple sets of fan beam projection data, and the multiple fan beam slice images are used as reference images without scattering artifacts; Based on the height information of the multiple fan beam projection data, extract multiple cone beam slice images from the first volume image that correspond to the multiple fan beam slice images in height position; Image registration is performed on the plurality of fan-beam slice images and the plurality of cone-beam slice images to obtain multiple registered image pairs aligned in space; The multiple registered image pairs are used to train a deep learning-based scattering correction model to obtain a trained scattering correction model, wherein the multiple cone-beam slice images in the multiple registered image pairs are used as inputs to the scattering correction model, and the multiple fan-beam slice images are used as labels for the scattering correction model. as well as The volumetric image or its image slice obtained by cone-beam CT imaging of the CBCT system is input into the trained scattering correction model to obtain the scattering-corrected output image.

13. The CBCT system as described in claim 12, characterized in that, While keeping the position of the object under test relative to the CBCT system unchanged, the linear array detector acquires multiple sector projection data of the object under test.

14. A computer-readable storage medium, characterized in that, It stores computer-readable instructions thereon, which, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 11.

Images