A CT image motion artifact removal method, system, device and storage medium

Abstract

The application relates to the technical field of medical assistance, and discloses a CT image motion artifact removal method, a system, equipment and a storage medium. The method comprises the following steps: constructing a dual-domain training data set comprising projection domain artifact data, image domain artifact data and artifact-free reference data, and a physical constraint parameter library comprising CT imaging key physical parameters and a physical model; constructing a dual-domain collaborative and physical constraint neural network DCP-Net; training the DCP-Net by using a hybrid loss function in combination with the dual-domain training data set, so as to obtain a target artifact removal model; inputting a CT image to be processed into the target artifact removal model, performing dual-domain collaborative optimization processing on the CT image to be processed by the DCP-Net, and outputting a target repair image. According to the scheme, bidirectional dynamic guidance of projection domain physical correction and image domain structure repair is realized through a dual-domain collaborative interaction module, the generation mechanism of motion artifacts is cracked, and the artifact removal rate is improved.

Description

Technical Field

[0001] This invention belongs to the field of medical auxiliary technology, specifically relating to a method, system, device, and storage medium for removing motion artifacts in CT images. Background Technology

[0002] Computed tomography (CT) technology, with its high spatial resolution, clear anatomical structure rendering capabilities, and rapid imaging characteristics, has become a core supporting technology for critical medical scenarios such as clinical diagnosis, tumor radiotherapy planning, and surgical navigation. During a CT scan, an X-ray source rotates around the human body and emits X-rays. A detector receives the X-ray signals after they penetrate the body and converts them into projection data. Then, through reconstruction algorithms such as filtered back projection (FBP), tomographic images are generated, providing clinicians with information about the internal anatomical structures of the human body.

[0003] However, involuntary movements of patients during clinical scans, such as breathing, heartbeat, limb tremors, and swallowing, are difficult to completely suppress, leading to problems such as streaking artifacts, blurring artifacts, ghosting, or local structural shifts in CT images. These artifacts can severely interfere with the accurate identification of key lesions such as small lung nodules, tumor boundaries, and vascular branches, reducing the accuracy of clinical diagnosis and even causing misdiagnosis, missed diagnosis, or deviation in the localization of the tumor radiotherapy target area, thus affecting the treatment effect.

[0004] Existing CT motion artifact removal techniques are mainly divided into two categories: single-domain processing methods and preliminary dual-domain processing methods. Traditional single-domain processing methods, which only correct artifacts in the projection domain (such as interpolation repair, projection domain filtering, and motion estimation compensation), can initially correct distorted projection data, but lack consideration for prior information about anatomical structures in the image domain. This can easily lead to decreased resolution, loss of detail, or the introduction of secondary artifacts in the reconstructed image. Deep learning methods that only perform repair in the image domain (such as network models based on U-Net and its variants) are mostly data-driven and do not fully integrate the physical mechanisms of CT imaging. They struggle to eliminate motion-induced projection data distortion at its source, are ineffective at suppressing large-scale, irregular motion artifacts, and are prone to losing crucial anatomical details such as organ edges and small nodules due to excessive smoothing. Among existing dual-domain processing methods, a few schemes only employ a serial architecture of projection domain preprocessing + image domain post-repair, which lacks dynamic interaction and bidirectional collaborative optimization between the projection and image domains. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, device, and storage medium for removing motion artifacts in CT images.

[0006] According to one aspect of this application, a method for removing motion artifacts in CT images is disclosed, the method comprising: Acquire clinical CT images without motion artifacts and their corresponding projection data; Motion distortion simulation is performed on the projection data based on rigid body motion model and non-rigid body deformation model to generate projection domain artifact data with motion artifacts. Perform CT image reconstruction on the projection domain artifact data to generate image domain artifact data; Data augmentation operations, including random rotation, scaling, contrast adjustment, and Gaussian noise addition, were performed on the acquired clinical CT images without motion artifacts and the image domain artifact data to generate enhanced CT images without motion artifacts and enhanced image domain artifact data. Based on the projection domain artifact data, the image domain artifact enhancement data, and the motion artifact-free enhanced CT image, the dual-domain training dataset is constructed, wherein the motion artifact-free enhanced CT image is used as the artifact-free reference data in the dual-domain training dataset. A physical constraint parameter library is defined, wherein the physical constraint parameter library includes key physical parameters of CT imaging and a physical model of motion distortion; A dual-domain collaborative and physically constrained neural network (DCP-Net) architecture is constructed, which includes a projection domain physical constraint correction subnetwork, an image domain multi-scale enhancement and insulation subnetwork, and a dual-domain collaborative interaction module connecting the two. A hybrid loss function that integrates physical constraint loss, data fitting loss, and dual-domain collaborative loss is adopted. The dual-domain collaborative and physical constraint neural network DCP-Net architecture is trained with a dual-domain training dataset to obtain the target artifact removal model. The CT motion artifact image to be processed is input into the target artifact removal model to obtain the output target repair image.

[0007] In some embodiments, the construction of the dual-domain collaborative and physically constrained neural network DCP-Net architecture includes: Construct a physical constraint correction subnetwork in the projection domain, with CT projection data containing motion artifacts as input and corrected projection data as output; A multi-scale enhancement and repair subnetwork in the image domain is constructed. Its input is a CT image containing motion artifacts, and its output is a preliminary repaired image. A dual-domain collaborative interaction module is constructed, whose input end is connected to the output end of the projection domain physical constraint correction sub-network and the image domain multi-scale enhancement and restoration sub-network. It is used to fuse and verify the features of the two sub-networks and output the target restored image.

[0008] In some embodiments, the projection domain physical constraint correction subnetwork includes a cascaded adaptive scattering correction module and a motion displacement estimation module, wherein: The adaptive scattering correction module is used to dynamically compensate for scattering interference in the input projection data based on predefined X-ray scattering characteristic parameters. The motion displacement estimation module adopts a convolutional long short-term memory network structure to estimate the motion displacement field between different projection angles during the scanning process, and performs motion alignment correction on the projection data based on the displacement field.

[0009] In some embodiments, the method of using a hybrid loss function that integrates physical constraint loss, data fitting loss, and dual-domain collaborative loss, and training the dual-domain collaborative and physical constraint neural network DCP-Net architecture with a dual-domain training dataset to obtain the target artifact removal model includes: Construct a hybrid loss function, which is composed of a weighted sum of physical constraint loss, perception loss, L1 loss and dual-domain collaborative loss; Using a dual-domain training dataset containing projection domain artifact data, image domain artifact data, and artifact-free reference data, the dual-domain collaborative and physically constrained neural network DCP-Net architecture is trained and optimized based on the hybrid loss function to obtain a target artifact removal model.

[0010] In some embodiments, inputting the CT motion artifact image to be processed into the target artifact removal model to obtain the output target repair image includes: Acquire the CT motion artifact image to be processed; Determine whether the original projection data corresponding to the CT motion artifact image to be processed is simultaneously acquired; if the original projection data to be processed is acquired simultaneously, the CT motion artifact image to be processed and the original projection data are used as input together, so that the target artifact removal model performs physical correction on the projection data, performs structural repair on the CT motion artifact image to be processed, and performs feature fusion and physical consistency verification on the forward projection data of the corrected projection domain data and the structural repair result through the dual-domain collaborative interaction module to generate fused projection domain data, reconstruct the fused projection domain data, and output the target repair image.

[0011] In some embodiments, inputting the CT motion artifact image to be processed into the target artifact removal model to obtain the output target repair image includes: Acquire the CT motion artifact image to be processed; Determine whether the original projection data corresponding to the CT motion artifact image to be processed is acquired simultaneously; When only the CT motion artifact image to be processed is acquired, the built-in forward projection algorithm module generates the corresponding projection data in reverse based on the CT motion artifact image to be processed. The CT motion artifact image to be processed and its reverse-generated projection data are used as inputs, so that the target artifact removal model can physically correct the projection data, perform structural repair on the CT motion artifact image to be processed, and perform feature fusion and physical consistency verification on the forward projection data of the corrected projection domain data and the structural repair result through the dual-domain collaborative interaction module to generate fused projection domain data. The fused projection domain data is then reconstructed to output the target repair image.

[0012] In some embodiments, the CT motion artifact image to be processed and its inversely generated projection data are used as input, so that the target artifact removal model performs physical correction on the projection data, performs structural repair on the CT motion artifact image to be processed, and performs feature fusion and physical consistency verification on the forward projection data of the corrected projection domain data and the structural repair result through a dual-domain collaborative interaction module to generate fused projection domain data, reconstruct the fused projection domain data, and output the target repaired image, including: The CT motion artifact image to be processed and its inversely generated projection data are used as input to enable the target artifact removal model to perform: The projection data is input into the projection domain physical constraint correction subnetwork for physical correction to obtain physical features, which include one or more of motion displacement information, projection intensity distribution and scattering correction parameters. The CT motion artifact image to be processed is input into the image domain multi-scale enhancement and repair subnetwork for initial structural repair to obtain semantic features, which include one or more of anatomical structure classification information and edge gradient information. The dynamic fusion weights of the physical features and the semantic features are determined by using channel attention and spatial attention mechanisms. Based on the dynamic fusion weights, the physical features and semantic features are adaptively fused to generate cross-domain fusion features. The cross-domain fusion features are input into the dual-domain collaborative interaction module to drive the collaborative optimization of projection domain physical correction and image domain structure repair, and output the target fusion features after physical consistency verification. The target fusion features are input into the image domain multi-scale enhancement and insulation subnetwork for target insulation to obtain the target insulated image.

[0013] According to another aspect of this application, a motion artifact removal system for CT images is also disclosed, the system comprising: The data acquisition module is used to acquire clinical CT images without motion artifacts and their corresponding projection data; The projection domain artifact data generation module is used to simulate motion distortion of the projection data based on a rigid body motion model and a non-rigid body deformation model, and generate projection domain artifact data with motion artifacts. The image domain artifact data generation module is used to perform CT image reconstruction operations on the projection domain artifact data to generate image domain artifact data. The image enhancement module is used to perform data enhancement operations, including random rotation, scale scaling, contrast adjustment and Gaussian noise addition, on the acquired clinical CT images without motion artifacts and the image domain artifact data, respectively, to generate enhanced CT images without motion artifacts and image domain artifact enhanced data. A dual-domain training dataset construction module is used to construct a dual-domain training dataset based on the projection domain artifact data, the image domain artifact enhancement data, and the motion artifact-free enhanced CT image, wherein the motion artifact-free enhanced CT image is used as the artifact-free reference data in the dual-domain training dataset. A physical constraint parameter library determination module is used to determine the physical constraint parameter library, wherein the physical constraint parameter library includes key physical parameters of CT imaging and a motion distortion physical model; A dual-domain collaborative and physically constrained neural network architecture construction module is used to construct a dual-domain collaborative and physically constrained neural network DCP-Net architecture. The dual-domain collaborative and physically constrained neural network DCP-Net architecture includes a projection domain physical constraint correction subnetwork, an image domain multi-scale enhancement and repair subnetwork, and a dual-domain collaborative interaction module connecting the two. The target artifact removal model determination module is used to train the dual-domain collaborative and physical constraint neural network DCP-Net architecture using a hybrid loss function that integrates physical constraint loss, data fitting loss and dual-domain collaborative loss, combined with a dual-domain training dataset, to obtain the target artifact removal model. The target restoration image output module is used to input the CT motion artifact image to be processed into the target artifact removal model to obtain the output target restoration image.

[0014] According to another aspect of this application, an electronic device is also disclosed, the electronic device including a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform the various steps of the CT image motion artifact removal method as described above.

[0015] According to another aspect of this application, a computer-readable storage medium is also disclosed, wherein instructions are stored on the computer-readable storage medium, characterized in that, when executed by a processor, the instructions implement the various steps of the CT image motion artifact removal method as described above.

[0016] The present invention includes, but is not limited to, the following beneficial effects: (1) This scheme enables the physical correction of the projection domain to obtain the assistance of image semantic information through cross-domain feature fusion and bidirectional guidance, thereby improving the accuracy of localization and motion distortion correction; at the same time, the restoration of the image domain can also obtain the prior guidance of physical information, so as to better preserve the real anatomical details when removing artifacts, reducing the situation of over-repair or loss of details. On the other hand, through strict physical consistency verification, it provides hard constraints for the output of the entire network, improving the visual clarity of the final generated target restoration image; (2) This scheme establishes a training data foundation that can simultaneously reflect the physical root of artifacts and the diversity of clinical images through the combined construction method of projection domain data simulation and image domain data enhancement. This construction method not only provides a physical simulation basis for the subsequent dual-domain collaborative network to learn and understand the artifact formation mechanism, but also enables the model to adapt in advance to various transformations and noises commonly found in clinical images, thereby significantly improving the accuracy of the restoration. The generalization ability and robustness of the network model in real, complex and paired data scarce clinical application scenarios are enhanced, and the practical value of the artifact removal scheme is ultimately ensured; (3) By constructing DCP-Net, which includes a projection domain physical constraint correction sub-network, an image domain multi-scale enhancement repair sub-network and a dual-domain collaborative interaction module, dual-domain collaborative optimization from the projection domain to the image domain is realized. The projection domain sub-network focuses on motion artifact correction under physical constraints, the image domain sub-network focuses on artifact repair under multi-scale, and the dual-domain collaborative interaction module fuses and verifies the features of the two sub-networks. It uses the physical information of the projection domain to accurately locate and correct the root cause of artifacts, and uses the semantic information of the image domain to optimize the details and naturalness of artifact repair. At the same time, the consistency of the dual-domain processing results is ensured through feature fusion and verification. The final repaired image output achieves a balance between physical law conformity and visual quality, effectively improving the accuracy, robustness and clinical usability of CT image motion artifact removal. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0018] Figure 1 This is a flowchart of a CT image motion artifact removal method according to an embodiment of this application; Figure 2 This is a structural block diagram of a CT image motion artifact removal system according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0019] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] For ease of understanding, the specific process of the embodiments of the present invention will be described below. Figure 1 This is a flowchart of a CT image motion artifact removal method according to an embodiment of this application. (See attached document.) Figure 1 It includes the following steps: S100: Acquire clinical CT images without motion artifacts and their corresponding projection data; S101. Based on the rigid body motion model and the non-rigid body deformation model, the projection data is simulated for motion distortion to generate projection domain artifact data with motion artifacts. S102. Perform CT image reconstruction operation on the projection domain artifact data to generate image domain artifact data; S103. Perform data augmentation operations, including random rotation, scale scaling, contrast adjustment and Gaussian noise addition, on the acquired clinical CT images and image domain artifact data without motion artifacts to generate enhanced CT images and image domain artifact data without motion artifacts. S104. Construct a dual-domain training dataset based on projection domain artifact data, image domain artifact enhancement data, and motion artifact-free enhanced CT images. The motion artifact-free enhanced CT images serve as artifact-free reference data.

[0021] S105. Determine the physical constraint parameter library.

[0022] Understandably, clinical CT images free of motion artifacts can be manually selected. The image data covers respiratory motion in various types of organ tumors (including major organs in the thoracic and upper abdominal regions such as the liver, lungs, and kidneys). All data is stored in the DICOM standard format to ensure compatibility with clinical imaging systems. Motion distortion simulation is performed on the projection data of the manually selected motion artifact-free clinical CT images using rigid body motion and non-rigid body deformation models, generating projection domain artifact data with different motion intensities (displacement range 0-20 pixels) and artifact types. Next, filtered backprojection reconstruction is performed on the aforementioned projection domain artifact data to obtain image domain artifact data. Simultaneously, motion artifact-free CT images from sources independent of the artifact-laden CT images are selected as reference data. To enhance data diversity and simulate image differences in real clinical scans, the same enhancement operations were performed on both the acquired motion artifact-free clinical CT images and image domain artifact data. These operations included random rotation (±15°), scaling (0.8–1.2x), adaptive contrast adjustment (±20%), and Gaussian noise addition (signal-to-noise ratio 30–50 dB), generating enhanced motion artifact-free CT images and image domain artifact-enhanced images. The resulting dual-domain training dataset contained three types of data: projection domain artifact data, image domain artifact-enhanced data, and motion artifact-free CT images. The motion artifact-free CT images were used as artifact-free reference data in the dual-domain training dataset. Based on the physical principles of CT imaging, key physical parameters were extracted and stored, specifically including: X-ray attenuation coefficient (human soft tissue: 0.15–0.2 cm). - ¹, Bones: 0.5-0.8cm - ¹, Air: 0.001cm - ¹) Scanning geometric parameters (detector size, scanning angle range, source-to-detector distance, source-to-rotation center distance), filtered back projection (FBP) reconstruction parameters (filter kernel type, reconstruction step size), and X-ray scattering characteristic parameters (scattering kernel width, scattering intensity coefficient); construct an X-ray projection physical model based on Beer-Lambert Law, as well as a motion distortion physical model describing the mapping relationship between motion displacement and projection data distortion, forming a standardized physical constraint parameter library to provide physical mechanism support for the network.

[0023] S106. Construct the DCP-Net architecture for dual-domain collaborative and physically constrained neural networks.

[0024] The dual-domain collaborative and physically constrained neural network DCP-Net architecture includes a projection domain physical constraint correction subnetwork, an image domain multi-scale enhancement and insulation subnetwork, and a dual-domain collaborative interaction module connecting the two.

[0025] Specifically, the construction of the dual-domain collaborative and physically constrained neural network DCP-Net architecture includes: Construct a physical constraint correction subnetwork in the projection domain, with CT projection data containing motion artifacts as input and corrected projection data as output; A multi-scale enhancement and repair subnetwork in the image domain is constructed. Its input is a CT image containing motion artifacts, and its output is a preliminary repaired image. A dual-domain collaborative interaction module is constructed, whose input is connected to the output of the projection domain physical constraint correction sub-network and the image domain multi-scale enhancement and repair sub-network. This module is used to fuse and verify the features of the two sub-networks and output the target repair image. Specifically, the CT motion artifact image to be processed and its inversely generated projection data are used as inputs, so that the target artifact removal model performs physical correction on the projection data and structural repair on the CT motion artifact image to be processed. The dual-domain collaborative interaction module performs feature fusion and physical consistency verification on the forward projection data of the corrected projection domain data and the structural repair result to generate fused projection domain data. The fused projection domain data is then reconstructed to output the target repair image.

[0026] Understandably, the dual-domain collaborative and physically constrained neural network DCP-Net architecture adopts an integrated architecture of projection domain correction branch - image domain repair branch - dual-domain collaborative interaction module. The core modules of the above integrated architecture include a projection domain physical constraint correction subnetwork, an image domain multi-scale enhancement repair subnetwork, and a dual-domain collaborative interaction module. Each module embeds corresponding CT imaging physical constraints to achieve bidirectional collaborative optimization of projection domain physical correction and image domain structural repair. The specific design is as follows: 1) Projection Domain Physical Constraint Correction Subnetwork: Input the original motion artifact projection data, embed the X-ray projection physical model and motion distortion correction constraints to achieve accurate correction of the projection data, and output the corrected projection data; this subnetwork is composed of an adaptive scattering correction module and a motion displacement estimation module connected in series. 2) Image Domain Multi-Scale Enhancement and Repair Sub-Network: The input is a CT image with motion artifacts. Prior constraints on anatomical structures in the image domain are embedded to achieve artifact suppression and preservation of key anatomical details, outputting a preliminary repaired image. This sub-network adopts an improved hybrid architecture of multi-scale Transformer and CNN (Convolutional Neural Network), consisting of a series of multi-scale feature divide-and-conquer units, hierarchical multi-scale Transformer encoding units, multi-scale CNN detail enhancement units, and dynamic feature fusion units. 3) Dual-domain collaborative interaction module: As the core link connecting the projection domain correction branch and the image domain restoration branch, it realizes dynamic interaction and physical consistency verification of dual-domain features, and outputs the final restored image; this module consists of a priori guiding bridge unit and a physical consistency verification unit connected in parallel: a. Adaptive Scattering Correction Module: Based on the X-ray scattering characteristic parameters in the physical constraint parameter library, the scattering kernel width (0.5-2.0mm) is adjusted through a dynamic convolution kernel generation mechanism to compensate for scattering interference in the projection data, correct the superposition distortion of scattering signals caused by motion, and ensure the physical accuracy of the projection data; at the same time, a projection intensity threshold constraint based on the attenuation characteristics of human tissue X-rays is introduced to remove abnormal noise points in the projection data. b. Motion displacement estimation module: It adopts a 3-layer convolutional LSTM (ConvLSTM, Convolutional Long Short-Term Memory) network structure, combined with projection geometric constraints such as scanning angle and detector position, to accurately estimate the motion displacement field between adjacent projection angles; it uses an inverse deformation sampling algorithm to align the distorted projection data, correct the spatial misalignment of the projection data, and output the preliminarily corrected projection data.

[0027] c. Multi-scale feature divide-and-conquer unit: The input image is decomposed into feature maps of four scales (scale factors are 1, 1 / 2, 1 / 4, and 1 / 8) by four convolutional layers of different lengths. These feature maps are matched with fine stripe artifacts (1-2 pixels), medium-scale blur artifacts (3-5 pixels), large-scale displacement artifacts (6-10 pixels), and ultra-large-scale distortion artifacts (>10 pixels) respectively, so as to achieve targeted feature extraction of artifacts of different scales. d. Hierarchical multi-scale Transformer encoding unit: Employs a multi-scale window self-attention mechanism (MS-WMSA) to adaptively adjust the window size (8×8, 16×16, 32×32, 64×64) for feature maps of different scales, accurately capturing the long-range spatial correlation of artifacts; embeds rotationally symmetric stripe prior constraints, and learns the directional characteristics of artifacts through a direction-shared convolutional representation mechanism to achieve accurate separation of artifact features from normal tissue features; e. Multi-scale CNN detail enhancement unit: It includes a multi-scale dilated convolution module and an edge-aware residual block. The multi-scale dilated convolution module has three parallel convolution branches, and the dilation rate is set to 1, 2 and 4 according to the artifact scale of the branches, which expands the receptive field without reducing the feature map resolution and achieves efficient suppression of artifacts at different scales. The edge-aware residual block extracts the image edge gradient information based on the second derivative filtering layer, and the edge features are weighted and fused with the original features through residual connections to enhance the preservation of key anatomical details such as organ edges and small nodules. f. Dynamic Feature Fusion Unit: Based on the anatomical structure priority mask, the fusion weights of Transformer global features and CNN local features are dynamically allocated. The CNN local detail features are preferentially preserved for key anatomical regions such as tumor target area and small nodules, while the Transformer global artifact suppression capability is enhanced for background regions, and the preliminary repaired image is output.

[0028] g. Prior guidance bridge unit: A dual-stream feature fusion mechanism is adopted to perform cross-domain correlation analysis on the physical features corrected in the projection domain (including motion displacement information, projection intensity distribution, and scattering correction parameters) and the semantic features extracted in the image domain (including anatomical structure classification information and edge gradient information); the dynamic weights of the dual-domain features are calculated through a channel-space dual attention mechanism to achieve adaptive fusion of the dual-domain features; the fused features obtained after adaptive fusion are used to guide the image domain restoration branch to accurately locate artifact regions, and are also used to adjust the displacement estimation accuracy of the projection domain correction branch, realizing bidirectional dual-domain guided optimization; h. Physical Consistency Verification Unit: Based on CT imaging physical constraints such as X-ray attenuation law and projection-image mapping relationship, a physical consistency loss function is constructed to verify whether the image corresponding to the fused features meets physical laws; specifically including: ① Projection consistency verification: The image domain repaired image is reverse-generated into projection data through the projection physical model, and the difference is calculated with the projection data after projection domain correction; ② Attenuation coefficient consistency verification: The attenuation coefficient of each tissue region in the repaired image is verified to conform to the human tissue attenuation characteristics in the physical constraint parameter library; if physical inconsistencies exist, a correction gradient is generated and backpropagated to the projection domain correction branch and the image domain repair branch to achieve iterative optimization of network parameters and ensure the physical rationality of the output results.

[0029] S107. A hybrid loss function that integrates physical constraint loss, data fitting loss, and dual-domain collaborative loss is adopted. The dual-domain collaborative and physical constraint neural network DCP-Net architecture is trained with dual-domain training dataset to obtain the target artifact removal model.

[0030] Specifically, step S107 includes: A hybrid loss function is constructed, which consists of a weighted sum of physical constraint loss, perceptual loss, L1 loss, and dual-domain collaborative loss. Using a dual-domain training dataset containing projection domain artifact data, image domain artifact enhancement data, and motion-free artifact-enhanced CT images, the dual-domain collaborative and physical constraint neural network DCP-Net architecture is trained and optimized based on the hybrid loss function to obtain the target artifact removal model.

[0031] Specifically: A hybrid loss function is constructed: physical constraint loss, data fitting loss, and dual-domain collaborative loss are integrated to ensure that network training conforms to the physical laws of CT imaging and accurately fits the distribution of artifact-free images. The specific expression is as follows: Loss = α × Loss1 + β × Loss2 + γ × Loss3 + δ × Loss4, where α, β, γ, and δ are loss weights (α = 0.3, β = 0.4, γ = 0.2, δ = 0.1). The definitions of each loss term are as follows: a. Physical constraint loss (Loss1): includes projection consistency loss and attenuation coefficient consistency loss; the projection consistency loss uses the mean squared error (MSE) loss to calculate and repair the difference between the reverse-generated projection data and the corrected projection data of the image; the attenuation coefficient consistency loss uses the L1 loss to calculate and repair the difference between the image attenuation coefficient and the standard tissue attenuation coefficient in the physical constraint parameter library. b. Perceptual loss (Loss2): Based on the convolutional features of the 5th layer of the pre-trained VGG-19 network, the feature distance between the restored image and the artifact-free reference image is calculated to ensure the structural similarity and visual consistency of the restored image. c. L1 loss (Loss3): Calculates the pixel-level difference between the restored image and the artifact-free reference image to improve the detail restoration of the image; d. Dual-domain collaborative loss (Loss4): Cosine similarity loss is used to calculate the cross-domain correlation between the physical features of the projection domain and the semantic features of the image domain, thereby enhancing the dual-domain collaborative optimization effect.

[0032] Network Training: The constructed training dataset is divided into training and validation sets in an 8:2 ratio. The AdamW adaptive optimizer is used, and the dual-domain collaborative and physically constrained neural network DCP-Net architecture is initialized with parameters from the physical constraint parameter library. During training, the initial learning rate is set to 0.0001, and the learning rate is dynamically adjusted using a cosine annealing strategy. The physical consistency verification unit provides real-time feedback on the satisfaction of physical constraints, and the loss weights are dynamically fine-tuned. When the peak signal-to-noise ratio (PSNR) of the validation set is ≥39dB, the structural similarity (SSIM) is ≥0.975, and the physical consistency error is ≤2%, training is stopped, the network parameters are saved, and the optimal artifact removal model is obtained.

[0033] S108. Input the CT motion artifact data to be processed into the target artifact removal model to obtain the output target repair image.

[0034] In some embodiments, step S108, inputting the CT motion artifact data to be processed into the target artifact removal model to obtain the output target restoration image includes: Acquire the CT motion artifact image to be processed; Determine whether the original projection data corresponding to the CT motion artifact image to be processed is acquired simultaneously; When the original projection data to be processed is acquired at the same time, the CT motion artifact image to be processed and the original projection data are used as input together; When only the CT motion artifact image to be processed is acquired, the corresponding projection data is generated in reverse based on the CT motion artifact image to be processed through the built-in physical model-based projection generation module. The CT motion artifact image to be processed and its inversely generated projection data are used as inputs, so that the target artifact removal model can physically correct the projection data, perform structural repair on the CT motion artifact image to be processed, and perform feature fusion and physical consistency verification on the correction and repair results through a dual-domain collaborative interaction module, and output the target repaired image.

[0035] Specifically: Data input: Obtain the CT motion artifact image to be processed and the corresponding raw projection data; if only the CT motion artifact image to be processed is obtained, generate the corresponding projection data through the network's built-in physical model-based forward projection algorithm.

[0036] Dual-domain collaborative artifact removal: The projection data is input into the projection domain physical constraint correction sub-network, and after adaptive scattering correction and motion displacement estimation, the corrected projection data is obtained; the artifact image is input into the image domain multi-scale enhancement and restoration sub-network, and after multi-scale feature extraction and detail enhancement, the preliminary restored image is obtained; the dual-domain collaborative interaction module completes the dynamic fusion of dual-domain features and physical consistency verification, and the final artifact-removed image is output after reconstruction.

[0037] Output results: The final output images are stored in DICOM format, which can be directly connected to clinical imaging systems such as medical image archiving and communication systems, supporting doctors to view, diagnose, and plan treatment.

[0038] In some embodiments, the CT motion artifact image to be processed and its inversely generated projection data are used as input, so that the target artifact removal model performs physical correction on the projection data, performs structural repair on the CT motion artifact image to be processed, and performs feature fusion and physical consistency verification on the correction and repair results through a dual-domain collaborative interaction module. The output target repaired image specifically includes: The CT motion artifact image to be processed and its inversely generated projection data are used as input to enable the target artifact removal model to perform: The projection data is input into the projection domain physical constraint correction subnetwork for physical correction to obtain physical features. The physical features include one or more of the following: motion displacement information, projection intensity distribution, and scattering correction parameters. The CT motion artifact image to be processed is input into the image domain multi-scale enhancement and repair subnetwork for initial structural repair, and semantic features are obtained. The semantic features include one or more of the following: anatomical structure classification information and edge gradient information. The dynamic fusion weights of physical and semantic features are determined by channel attention and spatial attention mechanisms. Based on dynamic fusion weights, physical features and semantic features are adaptively fused to generate cross-domain fusion features. The cross-domain fusion features are input into the dual-domain collaborative interaction module to drive the collaborative optimization of projection domain physical correction and image domain structure repair, and output the target fusion features after physical consistency verification. The target fusion features are reconstructed to obtain the target restoration image.

[0039] Furthermore, Figure 2 This is a structural block diagram of the CT image motion artifact removal system according to an embodiment of this application, as shown below. Figure 2 As shown, the system includes: The data acquisition module is used to acquire clinical CT images without motion artifacts and their corresponding projection data; The projection domain artifact data generation module is used to simulate motion distortion of the projection data based on a rigid body motion model and a non-rigid body deformation model, and generate projection domain artifact data with motion artifacts. The image domain artifact data generation module is used to perform CT image reconstruction operations on the projection domain artifact data to generate image domain artifact data. The image enhancement module is used to perform data enhancement operations, including random rotation, scale scaling, contrast adjustment and Gaussian noise addition, on the acquired clinical CT images without motion artifacts and the image domain artifact data, respectively, to generate enhanced CT images without motion artifacts and image domain artifact enhanced data. A dual-domain training dataset construction module is used to construct the dual-domain training dataset based on the projection domain artifact data, the image domain artifact enhancement data, and the motion artifact-free enhanced CT image. The dual-domain training dataset includes projection domain artifact data, image domain artifact data, and artifact-free reference data, with the motion artifact-free enhanced CT image serving as the artifact-free reference data. The physical constraint parameter library determination module is used to determine the physical constraint parameter library, which includes key physical parameters of CT imaging and physical models of motion distortion. A dual-domain collaborative and physically constrained neural network architecture building module is used to construct the dual-domain collaborative and physically constrained neural network DCP-Net architecture. The dual-domain collaborative and physically constrained neural network DCP-Net architecture includes a projection domain physical constraint correction subnetwork, an image domain multi-scale enhancement and repair subnetwork, and a dual-domain collaborative interaction module connecting the two. The target artifact removal model determination module is used to train the dual-domain collaborative and physical constraint neural network DCP-Net architecture using a hybrid loss function that integrates physical constraint loss, data fitting loss and dual-domain collaborative loss, combined with a dual-domain training dataset, to obtain the target artifact removal model. The target restoration image output module is used to input the CT motion artifact image to be processed into the target artifact removal model to obtain the output target restoration image.

[0040] The application of the relevant modules of the system in this example can be found in the above introduction to the principles of the method, and will not be repeated here.

[0041] above Figure 2 The CT image motion artifact removal system in this embodiment of the invention is described in detail from the perspective of modular functional entities. The electronic equipment in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0042] Figure 3 This is a schematic diagram of the structure of an electronic device 300 provided in an embodiment of the present invention. The electronic device 300 can vary significantly due to differences in configuration or performance, and may include one or more central processing units (CPUs) 310 and memory 320, and one or more storage media 330 for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the electronic device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the electronic device 300.

[0043] Electronic device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated electronic device structure does not constitute a limitation on electronic devices and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0044] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of any of the above-described CT image motion artifact removal methods.

[0045] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0046] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for removing motion artifacts in CT images, characterized in that, The method includes: Acquire clinical CT images without motion artifacts and their corresponding projection data; Motion distortion simulation is performed on the projection data based on rigid body motion model and non-rigid body deformation model to generate projection domain artifact data with motion artifacts. Perform CT image reconstruction on the projection domain artifact data to generate image domain artifact data; Data augmentation operations, including random rotation, scaling, contrast adjustment, and Gaussian noise addition, were performed on the acquired clinical CT images without motion artifacts and the image domain artifact data to generate enhanced CT images without motion artifacts and enhanced image domain artifact data. Based on the projection domain artifact data, the image domain artifact enhancement data, and the motion artifact-free enhanced CT image, a dual-domain training dataset is constructed, wherein the motion artifact-free enhanced CT image is used as the artifact-free reference data in the dual-domain training dataset. A physical constraint parameter library is defined, wherein the physical constraint parameter library includes key physical parameters of CT imaging and a physical model of motion distortion; A dual-domain collaborative and physically constrained neural network (DCP-Net) architecture is constructed, which includes a projection domain physical constraint correction subnetwork, an image domain multi-scale enhancement and insulation subnetwork, and a dual-domain collaborative interaction module connecting the two. Construct a hybrid loss function, which is composed of a weighted sum of physical constraint loss, perception loss, L1 loss and dual-domain collaborative loss; Using a dual-domain training dataset containing projection domain artifact data, image domain artifact data, and artifact-free reference data, the dual-domain collaborative and physically constrained neural network DCP-Net architecture is trained and optimized based on the hybrid loss function to obtain a target artifact removal model. The CT motion artifact image to be processed is input into the target artifact removal model to obtain the output target repair image.

2. The method for removing motion artifacts in CT images according to claim 1, characterized in that, The construction of the dual-domain collaborative and physically constrained neural network DCP-Net architecture includes: Construct a physical constraint correction subnetwork in the projection domain, with CT projection data containing motion artifacts as input and corrected projection data as output; A multi-scale enhancement and repair subnetwork in the image domain is constructed. Its input is a CT image containing motion artifacts, and its output is a preliminary repaired image. A dual-domain collaborative interaction module is constructed, whose input end is connected to the output end of the projection domain physical constraint correction sub-network and the image domain multi-scale enhancement and restoration sub-network. It is used to fuse and verify the features of the two sub-networks and output the target restored image.

3. The method for removing motion artifacts in CT images according to claim 2, characterized in that, The projection domain physical constraint correction subnetwork includes a cascaded adaptive scattering correction module and a motion displacement estimation module, wherein: The adaptive scattering correction module is used to dynamically compensate for scattering interference in the input projection data based on predefined X-ray scattering characteristic parameters. The motion displacement estimation module adopts a convolutional long short-term memory network structure to estimate the motion displacement field between different projection angles during the scanning process, and performs motion alignment correction on the projection data based on the displacement field.

4. The method for removing motion artifacts in CT images according to claim 1, characterized in that, The process of inputting the CT motion artifact image to be processed into the target artifact removal model to obtain the output target restoration image includes: Acquire the CT motion artifact image to be processed; Determine whether the original projection data corresponding to the CT motion artifact image to be processed is acquired simultaneously; When the original projection data is acquired simultaneously, the CT motion artifact image to be processed and the original projection data are used as input together, so that the target artifact removal model can physically correct the projection data, perform structural repair on the CT motion artifact image to be processed, and perform feature fusion and physical consistency verification on the forward projection data of the corrected projection domain data and the structural repair result through the dual-domain collaborative interaction module to generate fused projection domain data, reconstruct the fused projection domain data, and output the target repair image.

5. The method for removing motion artifacts in CT images according to claim 1, characterized in that, The process of inputting the CT motion artifact image to be processed into the target artifact removal model to obtain the output target restoration image includes: Acquire the CT motion artifact image to be processed; Determine whether the original projection data corresponding to the CT motion artifact image to be processed is acquired simultaneously; When only the CT motion artifact image to be processed is acquired, the built-in forward projection algorithm module generates the corresponding projection data in reverse based on the CT motion artifact image to be processed. The CT motion artifact image to be processed and its reverse-generated projection data are used as inputs, so that the target artifact removal model can physically correct the projection data, perform structural repair on the CT motion artifact image to be processed, and perform feature fusion and physical consistency verification on the forward projection data of the corrected projection domain data and the structural repair result through the dual-domain collaborative interaction module to generate fused projection domain data. The fused projection domain data is then reconstructed to output the target repair image.

6. The method for removing motion artifacts in CT images according to claim 5, characterized in that, The process involves using the CT motion artifact image to be processed and its inversely generated projection data as input, enabling the target artifact removal model to physically correct the projection data, perform structural repair on the CT motion artifact image to be processed, and perform feature fusion and physical consistency verification on the forward projection data of the corrected projection domain data and the structural repair result through a dual-domain collaborative interaction module to generate fused projection domain data. The fused projection domain data is then reconstructed, and the output target repaired image includes: The CT motion artifact image to be processed and its inversely generated projection data are used as input to enable the target artifact removal model to perform: The projection data is input into the projection domain physical constraint correction subnetwork for physical correction to obtain physical features, which include one or more of motion displacement information, projection intensity distribution and scattering correction parameters. The CT motion artifact image to be processed is input into the image domain multi-scale enhancement and repair subnetwork for initial structural repair to obtain semantic features, which include one or more of anatomical structure classification information and edge gradient information. The dynamic fusion weights of the physical features and the semantic features are determined by using channel attention and spatial attention mechanisms. Based on the dynamic fusion weights, the physical features and semantic features are adaptively fused to generate cross-domain fusion features. The cross-domain fusion features are input into the dual-domain collaborative interaction module to drive the collaborative optimization of projection domain physical correction and image domain structure repair, and output the target fusion features after physical consistency verification. The target fusion features are input into the image domain multi-scale enhancement and insulation subnetwork for target insulation to obtain the target insulated image.

7. A CT image motion artifact removal system, characterized in that, The system includes: The data acquisition module is used to acquire clinical CT images without motion artifacts and their corresponding projection data; The projection domain artifact data generation module is used to simulate motion distortion of the projection data based on a rigid body motion model and a non-rigid body deformation model, and generate projection domain artifact data with motion artifacts. The image domain artifact data generation module is used to perform CT image reconstruction operations on the projection domain artifact data to generate image domain artifact data. The image enhancement module is used to perform data enhancement operations, including random rotation, scale scaling, contrast adjustment and Gaussian noise addition, on the acquired clinical CT images without motion artifacts and the image domain artifact data, respectively, to generate enhanced CT images without motion artifacts and image domain artifact enhanced data. A dual-domain training dataset construction module is used to construct a dual-domain training dataset based on the projection domain artifact data, the image domain artifact enhancement data, and the motion artifact-free enhanced CT image, wherein the motion artifact-free enhanced CT image is used as the artifact-free reference data in the dual-domain training dataset. A physical constraint parameter library determination module is used to determine the physical constraint parameter library, wherein the physical constraint parameter library includes key physical parameters of CT imaging and a motion distortion physical model; A dual-domain collaborative and physically constrained neural network architecture construction module is used to construct a dual-domain collaborative and physically constrained neural network DCP-Net architecture. The dual-domain collaborative and physically constrained neural network DCP-Net architecture includes a projection domain physical constraint correction subnetwork, an image domain multi-scale enhancement and repair subnetwork, and a dual-domain collaborative interaction module connecting the two. The hybrid loss function construction module is used to construct a hybrid loss function, which is composed of a weighted sum of physical constraint loss, perception loss, L1 loss and dual-domain collaborative loss. The target artifact removal model determination module uses a dual-domain training dataset containing projection domain artifact data, image domain artifact data, and artifact-free reference data to train and optimize the dual-domain collaborative and physically constrained neural network DCP-Net architecture based on the hybrid loss function to obtain the target artifact removal model. The target restoration image output module is used to input the CT motion artifact image to be processed into the target artifact removal model to obtain the output target restoration image.

8. An electronic device, characterized in that, The electronic device includes a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform the steps of the CT image motion artifact removal method as described in any one of claims 1-6.

9. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the CT image motion artifact removal method as described in any one of claims 1-6.

Images