Deep learning-based CT artifact removal method and system

Abstract

The present invention relates to the technical field of medical images. Disclosed are a deep learning-based CT artifact removal method and system. The method comprises: acquiring a CT image, and separately performing sine transform and wavelet transform processing on the CT image; constructing an image enhancement model, and performing image optimization on the processed image separately by means of the image enhancement model and a random inversion layer which are connected in sequence; coupling the optimized image and the original image and then inputting the coupled image into the image enhancement model for reprocessing; and performing element-wise addition on the reprocessed image and the optimized image to obtain an artifact-removed CT image. In the present invention, wavelet transform is introduced to process the CT image to extract the context and spatial information of the CT image, effectively extracting feature information in an artifact removal process and improving the performance of image enhancement; and a CT image resolution enhancement model based on a VMamba model is established, enhancing the long-term dependencies in network training, effectively recognizing and removing radioactive artifacts, and improving the network training efficiency.

Description

A CT artifact removal method and system based on deep learning TECHNICAL FIELD

[0001] The present application relates to the field of medical imaging technology, more particularly to a CT artifact removal method and system based on deep learning. BACKGROUND

[0002] CT (Computer tomography) technology, as a non-invasive and rapid detection method, has been widely used in the medical field. With the increasing use of metal implants in the medical field, such as dental implants, spinal fixation devices and hip joint replacement, etc., the CT image reconstruction often appears to be a false image with alternating light and dark due to the interference of the implants, which makes it difficult to clearly identify the interface between metal and normal tissue and the surrounding structure, and further affects the evaluation of the surgical effect and the formulation of subsequent treatment decisions by doctors.

[0003] When CT scanning is performed, if there is high-attenuation material (such as metal, high-density bone, etc.) in the scanning range, it will absorb low-energy photons in the X-rays, causing some photons to be unable to reach the detector. This phenomenon causes the energy distribution of the penetrating rays to shift, i.e., the rays are hardened. The formation of artifacts is the result of the combined action of multiple factors, and the main reasons include ray hardening and photon starvation, as well as the interference of factors such as scattering effect, partial volume effect and noise (such as metal artifacts, hardening artifacts, ring artifacts and motion artifacts, etc.). In CT image reconstruction, metal artifacts usually appear as radial or striped brightness abnormalities, which cover the true image information of the surrounding tissue, resulting in errors in the identification and boundary division of the material. In the existing dual-energy CT (Dual-Energy CT) and CBCT (Cone Beam Computed Tomography) image reconstruction tasks, the processing of high-density objects (such as metal implants) will encounter the problem of serious attenuation of X-rays. This attenuation will cause data loss or distortion, thereby forming secondary artifacts.

[0004] The existing methods for solving the artifact removal can be divided into the following: 1. Optimization of CT scanning parameters: CT scanning parameter optimization is a common method for removing artifacts. By increasing the tube voltage, tube current, using a narrow collimator, reducing the scan pitch and increasing the scan layer thickness, etc., the artifacts can be reduced in the data acquisition stage. However, simply optimizing the CT scanning parameters has limited effect on reducing artifacts, and it is difficult to meet the requirements of clinical precise diagnosis. 2. Hardware correction: Hardware correction is to reduce the influence of scattered radiation on the detector by adding specific correction devices to each component of the X-ray imaging system, or adjusting the energy and energy spectrum distribution of the initial radiation. There are the following common methods: ① Collimator correction: This method places a thin plate with high absorption coefficient in front of the detector to absorb scattered radiation. This technology has a significant effect in fan-beam CT, which can eliminate about 89% of scattered radiation, but the effect is not obvious in CBCT. ② Air gap correction: This method increases the distance between the measured object and the detector to correct scattering. Due to the limitation of the projection magnification ratio, this method may cause image blurring. ③ Anti-scattering grid correction: This method sets a grid structure in front of the detector to block scattered radiation. In conditions of high X-ray scattering (e.g., large size of measured object and wide irradiation field), high imaging dose, and the need for low resolution reconstruction, the scattering situation is effectively controlled. However, in low dose and low scattering conditions, such as medical imaging, this method may reduce image quality and increase noise, and the suppression effect of scattered radiation is not good. 3. Software correction: Using the obtained X-ray projection images, according to the method of digital image processing, the scattering distribution map is obtained through the analysis of the image itself and the estimation of the properties of the scanned object for correction. It can be divided into the following two types: ① Monte Carlo simulation: It is a simulation technique based on statistical methods. Through mathematical calculation, the interaction process of multi-energy X-rays with different materials is simulated, and the nonlinear function relationship between multi-energy projection and single-energy projection is obtained, and then the artifact correction is completed. Although it can improve the quality of CBCT images, the calculation complexity is high and the operation time is long. In addition, before performing Monte Carlo simulation, accurate modeling and analysis of the scanned object are required, which is often difficult to achieve in medical applications. ② Deep learning: With the significant progress of computer power, deep learning has rapidly developed in various fields, and its image reconstruction and analysis have been significantly improved. This technology has become a new method for reducing artifacts in CT images caused by implants. The advantage of deep learning is that it can solve the inverse problem in image processing. Using deep learning to reconstruct CT images to improve image quality has become a promising direction with great potential.

[0005] Therefore, how to propose a CT artifact removal method and system based on deep learning, reconstruct CT images by using deep learning, solve the CT image artifact problem, reduce the interference and influence of artifacts on images, and reduce the misreading and misjudgment of images in clinical practice are problems that need to be solved by those skilled in the art. SUMMARY

[0006] Therefore, the present application provides a CT artifact removal method and system based on deep learning, which reconstructs CT images by using deep learning, solves the CT image artifact problem, reduces the interference and influence of artifacts on images, and reduces the misreading and misjudgment of images in clinical practice. In order to achieve the above purpose, the present application adopts the following technical solutions:

[0007] A CT artifact removal method based on deep learning, comprising:

[0008] Obtaining a CT image, respectively performing sine transformation and wavelet transformation processing on the CT image;

[0009] Constructing an image enhancement model, and respectively passing the processed image through the image enhancement model and the random back projection layer connected in sequence to optimize the image;

[0010] Coupling the optimized image and the original image and inputting them into the image enhancement model for reprocessing;

[0011] Elementarily adding the reprocessed image and the optimized image to obtain a CT image with artifacts removed.

[0012] Optionally, the obtaining of the CT image comprises collecting medical CT data of a patient, including oral, brain and orthopedic departments, establishing a CT image data set with artifacts, and selecting an image with artifacts as a verification set.

[0013] Optionally, the wavelet transformation processing comprises continuous wavelet transformation and discrete wavelet transformation, and the image signal is convoluted with a wavelet base function.

[0014] Optionally, the construction of the image enhancement model comprises taking a mamba module as a sub-module to build a U-net network structure to obtain a UMamba-Net image enhancement model, and the UMamba-Net image enhancement model comprises three components: an encoder, a decoder and a jump connection.

[0015] Optionally, the mamba module comprises:

[0016] A continuous system is constructed, which maps a one-dimensional input function or sequence to an input through an intermediate hidden state.

[0017] The continuous system is discretized, a time scale parameter is introduced and the state matrix and projection parameters are converted into discrete parameters using a fixed discretization rule;

[0018] After discretization, the calculation is performed by linear recursion or global convolution.

[0019] Optionally, the structure of the UMamba-Net image enhancement model is that a VSS module is coupled with a patch merging layer to obtain a first substructure, and a VSS module is coupled with a patch expanding layer to obtain a second substructure.

[0020] The plurality of first substructures and the plurality of second substructures are symmetrically placed, the plurality of first substructures are sequentially connected from top to bottom, the uppermost layer of the first substructure is taken as input, the plurality of second substructures are sequentially connected from bottom to top, and the uppermost layer of the second substructure is taken as output.

[0021] The lowermost layer of the first substructure is sequentially connected to the lowermost layer of the second substructure through two VSS modules, the output value of the first substructure and the output value of the second substructure of the corresponding lower layer are added element by element as the input value of the corresponding second substructure.

[0022] Optionally, the VSS module includes: after layer normalization, the input is divided into two branches, in the first branch, the input passes through a linear layer and then an activation function; in the second branch, the input is processed through a linear layer, a depth separable convolution and an activation function, and then is sent to an SS2D module for feature extraction; the extracted features are standardized by layer normalization, and are multiplied element by element with the output of the first branch to combine the two paths, and finally a linear layer is used to mix the features, and the result is combined with the residual connection to form the output of the VSS block.

[0023] Optionally, the activation function is a SiLU activation function.

[0024] Optionally, the SS2D module is a 2D-selective scanning module, which is composed of three parts: a scanning expansion operation, an S6 module and a scanning merging operation, the scanning expansion operation expands the input image along four different directions to form a sequence, the sequence is processed by the S6 module to extract features, the information from each direction is scanned, and the scanning merging operation sums and merges the sequences from different directions, the S6 module is mainly composed of a series of linear transformations and discretization processes, and is used for processing the input feature sequence. The S6 module is a complex component in the Mamba architecture, which is responsible for processing the input feature sequence through a series of linear transformations and discretization processes. It plays a key role in capturing the temporal dynamics of the sequence, which is a key aspect of sequence modeling tasks, including tensor operations and custom discretization methods to handle the complex requirements of sequence data.

[0025] Optionally, a CT artifact removal system based on deep learning comprises:

[0026] A processing module is configured to acquire a CT image, and perform sine transformation and wavelet transformation on the CT image, respectively.

[0027] A model construction optimization module is configured to construct an image enhancement model, and perform image optimization on the processed image by sequentially connecting the image enhancement model and a random back projection layer.

[0028] A reprocessing module is configured to input the optimized image and the original image coupled together into the image enhancement model for reprocessing.

[0029] An artifact removal module is configured to add the reprocessed image and the optimized image element by element to obtain a CT artifact removal image.

[0030] Compared with the prior art, the CT artifact removal method and system based on deep learning have the following beneficial effects:

[0031] The CT artifact removal method based on deep learning comprises the following steps: acquiring a CT image, performing sine transformation and wavelet transformation on the CT image, respectively, constructing an image enhancement model, performing image optimization on the processed image by sequentially connecting the image enhancement model and a random back projection layer, inputting the optimized image and the original image coupled together into the image enhancement model for reprocessing, and adding the reprocessed image and the optimized image element by element to obtain a CT artifact removal image. The method can improve image quality under the same radiation dose, reduce the interference and influence of artifacts on the image, reduce the misreading and misjudgment of the image in the clinic, and improve the performance and efficiency of network training. The method collects medical CT data of patients, establishes a CT picture data set with artifacts, selects part of the artifact pictures for network model verification, introduces wavelet transformation to process the CT image, extracts the context and spatial information of the CT image, effectively extracts feature information in the artifact removal training process, and improves the performance of image enhancement. The method establishes a CT image resolution enhancement model based on a VMamba model, improves the long-term dependence of network training, effectively identifies and removes radioactive artifacts, and improves the network training efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only represent some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0033] Fig. 1 is a flowchart of a CT artifact removal method based on deep learning provided by the present application.

[0034] Fig. 2 is a schematic diagram of the structure of the UMamba-Net image enhancement model provided by the present application.

[0035] Fig. 3 is a schematic diagram of the structure of the VSS module provided by the present application.

[0036] Fig. 4 is a schematic diagram of the comparison between images with and without artifacts provided by the present application. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0038] Deep learning is an important branch of machine learning, which processes complex data patterns by simulating the structure and function of human brain neural networks. The core of deep learning is to use multi-layer neural networks that can learn useful feature representations from large amounts of data without manually designing feature extraction algorithms. Deep learning is to learn the internal laws and representation levels of sample data, and the information obtained in the learning process is very helpful for the interpretation of data such as text, images and sound.

[0039] Deep learning has two major applications in the field of medical imaging. The first application is an intelligent imaging system, including fast reconstruction, noise reduction and super-resolution imaging; the second application is intelligent analysis of medical image big data, including classification, detection, segmentation and registration of medical images. Through deep neural network learning, artificial intelligence assisted medical image processing technology is developed, which can restore image data under high radiation dose conditions under low dose conditions, improve image quality; at the same time, reduce the influence of artifacts on surrounding tissues, and the comparison between images with and without artifacts is shown in Fig. 4.

[0040] Supervised networks, including convolutional neural networks (CNN) and U-Net, are widely used in medical image processing. In addition, unsupervised methods are also introduced to solve the artifacts in the latent space of generative adversarial networks (GAN). Generally, processing the corrupted sinogram can solve the divergent artifacts caused by metal, and the projection data outside the metal artifact is considered as clean data. One of the challenges of sinogram restoration is the secondary artifacts caused by discontinuity after the restoration process. Since each layer of CNN has only a limited receptive field, such as a 3x3 convolution kernel, the deep layer can only partially capture the long-range discontinuity. In order to fully exploit the local information of the surrounding angles and detectors, additional prior information needs to be used for further improvement. For example, metal mask projection and adaptive scale can exploit more information in the corrupted area, thereby obtaining better performance.

[0041] Further, the dual-domain network can solve the artifact removal work. By using two or more rough enhanced images, the dual-domain architecture greatly improves the performance of the single-domain method. However, the fundamental problem of limited receptive field has not been solved in the sinogram and image field. In addition to these two domains, the use of wavelet transform to extract time-frequency localization characteristics can provide high resolution in both time and frequency domains, which can easily provide context and spatial information for global restoration. In the backbone network, since the artifacts present a radioactive state, the UMamba module is used to build the image enhancement network UMamba-Net, which can improve the long-term memory of the network in training, effectively improve the effect of artifact removal, and the network can be trained quickly in the RNN mode, thereby improving the training efficiency of the model artifact removal.

[0042] The embodiment of the application discloses a CT artifact removal method based on deep learning, as shown in FIG. 1, which comprises the following steps:

[0043] Obtaining a CT image, performing sinogram transformation and wavelet transformation on the CT image respectively;

[0044] Constructing an image enhancement model, and performing image optimization on the processed image through the image enhancement model and the random back projection layer connected in sequence respectively;

[0045] Coupling the optimized image and the original image and inputting them into the image enhancement model for reprocessing;

[0046] Elementally adding the reprocessed image and the optimized image to obtain a CT image with artifacts removed.

[0047] Further, the obtaining of the CT image comprises: collecting medical CT data of patients, more than 2000 cases, including oral cavity, neurosurgery and orthopedics, etc., and establishing a CT image data set with artifacts, wherein part of the images with artifacts are selected for verification of the network model.

[0048] Furthermore, the wavelet transform processing includes continuous wavelet transform and discrete wavelet transform, which convolve the image signal with wavelet basis functions.

[0049] Specifically, wavelet transform possesses excellent time-frequency localization properties, providing high resolution simultaneously in both the time and frequency domains. It can accurately locate changes in an image, such as edges and textures. Wavelet transform is implemented by convolving the image signal with wavelet basis functions. The formula for continuous wavelet transform (CWT) is:

[0050] Where x(t) is the original signal, ψ(t) is the mother wavelet function, a and b are the scale and translation parameters, respectively, and * denotes complex conjugate.

[0051] The formula for Discrete Wavelet Transform (DWT) is:

[0052] W ψ [j,k]=∑ n x[n]ψ j,k [n];

[0053] in It is a discrete wavelet transform, where j and k represent the scale and displacement parameters, respectively, and n represents the discrete-time index.

[0054] Wavelet transform is introduced to process CT images to extract CT image context and spatial information, and to effectively extract feature information during artifact removal training, thereby improving the performance of image enhancement (resolution).

[0055] Furthermore, the construction of the image enhancement model includes: using the Mamba module as a sub-module to build a U-net network structure, thereby obtaining the UMamba-Net image enhancement model. The UMamba-Net image enhancement model includes three components: an encoder, a decoder, and hop connections, as shown in Figure 2.

[0056] Furthermore, the Mamba module includes:

[0057] Construct a continuous system through which a one-dimensional input function or sequence is mapped to an input via intermediate hidden states;

[0058] Discretize this continuous system by introducing a time scale parameter and using a fixed discretization rule to convert the state matrix and projection parameters into discrete parameters;

[0059] After discretization, computation is performed using linear recursion or global convolution.

[0060] Further, the structure of the UMamba-Net image enhancement model is: a VSS module is coupled with a patch merging layer to obtain a first substructure, and a VSS module is coupled with a patch expansion layer to obtain a second substructure;

[0061] The plurality of first substructures and the plurality of second substructures are symmetrically placed, the plurality of first substructures are sequentially connected from top to bottom, the uppermost layer of the first substructure is taken as input, the plurality of second substructures are sequentially connected from bottom to top, and the uppermost layer of the second substructure is taken as output;

[0062] The lowermost layer of the first substructure is sequentially connected to the lowermost layer of the second substructure through two VSS modules, the output value of the first substructure and the output value of the second substructure at the lower layer of the corresponding second substructure are added element by element as the input value of the corresponding second substructure.

[0063] In specific embodiments, the UMamba-Net image enhancement model specifically comprises:

[0064] Modern SSMs (such as Mamba) not only establish long-range dependencies, but also exhibit linear complexity with respect to input size. Mamba is taken as a sub-module to build a UMamba-Net image enhancement model with a U-net network structure, and the UMamba-Net mainly consists of three main parts: an encoder, a decoder and a skip connection.

[0065] Step one: in modern SSM-based models, i.e., structured state space sequence model (S4) and Mamba, a continuous system is relied on, which maps a one-dimensional input function or sequence (denoted as x(t)∈R) to an input y(t)∈R through an intermediate hidden state h(t)∈R N The above process is represented as a linear ordinary differential equation (ODE): h'(t)=Ah(t)+Bx(t); y(t)=Ch(t);

[0066] Wherein, A∈R N×N represents a state matrix, B∈R N×1 , C∈R N×1 respectively represent projection parameters.

[0067] Step two: S4 and Mamba discretize this continuous system to make it more suitable for deep learning scenarios.

[0068] A time scale parameter Δ is introduced, and A and B are converted into discrete parameters and Zero-order hold (ZHO) is taken as the discretization rule, which is defined as follows:

[0069] where I is an identity matrix of the same dimension as matrix A.

[0070] Step three: After discretization, the SSM-based model can be calculated in two ways: linear recurrence or global convolution, as follows:

[0071] where, represents a structured convolution kernel, and L represents the length of the input sequence x.

[0072] The CT image resolution enhancement model based on the VMamba model is established, the long-term dependence of network training is improved, the radioactive artifacts are effectively identified and removed, and the network training efficiency is improved.

[0073] Further, the VSS module includes: the VSS module is derived from the structure of the VMamaba, and is a core module of the VM-UNet. After layer normalization, the input is divided into two branches. In the first branch, the input passes through a linear layer and then an activation function; in the second branch, the input is processed through a linear layer, a depth separable convolution and an activation function, and then is sent to the SS2D module for feature extraction; the extracted features are standardized by layer normalization, and are multiplied element by element with the output of the first branch to combine the two paths, and finally a linear layer is used to mix the features, and the result is combined with the residual connection to form the output of the VSS block, as shown in FIG. 3.

[0074] Further, the activation function is a SiLU activation function.

[0075] Further, the SS2D module is a 2D-selective scanning module, which is composed of three parts: a scanning expansion operation, an S6 module and a scanning merging operation. The scanning expansion operation expands the input image into a sequence along four different directions (from top left to bottom right, from bottom right to top left, from top right to bottom left, and from bottom left to top right), the sequence is processed by the S6 module to extract features, and the information from each direction is thoroughly scanned to capture diverse features. The information from each direction is scanned, and the scanning merging operation sums and merges the sequences from different directions.

[0076] In the specific embodiment, a CT artifact removal system based on deep learning includes:

[0077] The processing module is configured to acquire the CT image, and perform sine transformation and wavelet transformation on the CT image, respectively.

[0078] The model construction optimization module is configured to construct an image enhancement model, and the processed image is input into the image enhancement model and a random back projection layer connected in sequence for image optimization.

[0079] The reprocessing module is configured to input the optimized image and the original image into the image enhancement model for reprocessing after coupling.

[0080] The artifact removal module is configured to obtain an artifact removal CT image by element addition of the reprocessed image and the optimized image.

[0081] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other.

[0082] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based CT artifact removal method, characterized in that, The application relates to a method for removing artifacts in CT images. The method comprises the following steps: acquiring a CT image, performing a sine transform and a wavelet transform on the CT image respectively; constructing an image enhancement model, and performing image optimization on the processed image through an image enhancement model and a random back projection layer connected in sequence; inputting the optimized image and the original image into the image enhancement model for reprocessing after coupling; performing element addition on the reprocessed image and the optimized image to obtain a CT image with artifacts removed. 2.The CT artifact removal method based on deep learning according to claim 1, characterized in that, The method comprises the following steps: collecting medical CT data of a patient, including oral cavity, brain and orthopedics, establishing a CT image data set with artifacts, and selecting an image with artifacts as a verification set. 3.The CT artifact removal method based on deep learning according to claim 1, characterized in that, The wavelet transform processing comprises continuous wavelet transform and discrete wavelet transform, and the image signal is convolved with a wavelet base function. 4.The CT artifact removal method based on deep learning according to claim 1, characterized in that, The method comprises the following steps: taking a mamba module as a sub-module to build a U-net network structure to obtain a UMamba-Net image enhancement model, and the UMamba-Net image enhancement model comprises three components: an encoder, a decoder and a skip connection. 5.The CT artifact removal method based on deep learning according to claim 4, characterized in that, The mamba module comprises the following steps: constructing a continuous system, mapping a one-dimensional input function or sequence to an input through an intermediate hidden state; discretizing the continuous system, introducing a time scale parameter and converting a state matrix and a projection parameter into discrete parameters using a fixed discretization rule; after discretization, performing calculation through linear recursion or global convolution. The structure of the UMamba-Net image enhancement model is as follows: coupling a VSS module with a Patch merging layer to obtain a first substructure, and coupling the VSS module with a Patch expansion layer to obtain a second substructure; 6.The CT artifact removal method based on deep learning according to claim 4, characterized in that, symmetrically arranging a plurality of first substructures and a plurality of second substructures, sequentially connecting the first substructures from top to bottom, taking the uppermost first substructure as input, sequentially connecting the second substructures from bottom to top, and taking the uppermost second substructure as output; sequentially connecting the lowermost first substructure to the lowermost second substructure through two VSS modules, performing element addition on the output value of the first substructure and the output value of the second substructure in the corresponding second substructure to obtain the input value of the corresponding second substructure. After layer normalization, the input is divided into two branches, in the first branch, the input passes through a linear layer and then an activation function; in the second branch, the input is processed through a linear layer, a depth separable convolution and an activation function, and then is sent to a SS2D module for feature extraction; the extracted features are standardized through layer normalization, and are multiplied with the output of the first branch element by element to combine the two paths, finally, a linear layer is used to mix the features, and the result is combined with the residual connection to form the output of the VSS block.

7. The CT artifact removal method based on deep learning according to claim 6, characterized in that, The activation function is a SiLU activation function. 8.The CT artifact removal method based on deep learning according to claim 7, characterized in that, ​ 9.The CT artifact removal method based on deep learning according to claim 7, characterized in that, The SS2D module is a 2D-selective scanning module, which is composed of three parts: a scanning expansion operation, an S6 module and a scanning merging operation, the scanning expansion operation expands the input image to constitute a sequence in four different directions, the sequence is processed by the S6 module to extract features, the information from each direction is scanned, and the scanning merging operation sums and merges the sequences from different directions. 10.A CT artifact removal system based on deep learning, characterized in that, It comprises: a processing module for acquiring a CT image, and performing sine transformation and wavelet transformation processing on the CT image respectively; a model construction optimization module for constructing an image enhancement model, and performing image optimization on the processed image through the image enhancement model and a random back projection layer connected in sequence respectively; a reprocessing module for coupling the optimized image and the original image and inputting them into the image enhancement model for reprocessing; an artifact removal module for performing element addition on the reprocessed image and the optimized image to obtain a CT image with artifacts removed.

Images