Model training method, artifact elimination method, device, equipment, medium and product

Abstract

The application relates to a model training method and device, an artifact elimination method and device, a medium and a product. The method comprises the following steps: acquiring training sample data, wherein the training sample data comprises a first sample artifact image and a corresponding artifact-free label image; performing down-sampling on the first sample artifact image to obtain a second sample artifact image, and obtaining a first artifact capture image based on the second sample artifact image and an initial artifact elimination network; the initial artifact elimination network is a two-stage multi-scale artifact elimination network; superimposing the first artifact capture image on the first sample artifact image to obtain a first artifact ablation image corresponding to the first sample artifact image; determining a loss function based on the first artifact ablation image corresponding to the first sample artifact image and the artifact-free label image, and iteratively training the initial artifact elimination network based on the loss function to obtain a target artifact elimination model. The target artifact elimination model can improve the elimination effect of ring-shaped artifacts in a reconstructed image.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a model training method, an artifact removal method, an apparatus, a device, a medium, and a product. Background Technology

[0002] With the development of medical imaging technology, the quality requirements for reconstructed images obtained after scanning by medical imaging equipment are becoming increasingly higher. The higher the image quality, the greater its reference value for clinical applications. Therefore, how to improve the quality of reconstructed images remains a hot research topic and direction.

[0003] Taking a photon-counting detector-based spectral CT device as an example, the uncertainty of the photon-counting detector's response can lead to strip artifacts during scanning, which in turn can cause concentric ring artifacts in the reconstructed image, significantly affecting its image quality and limiting its clinical application.

[0004] Therefore, how to efficiently eliminate ring artifacts in CT reconstructed images remains a pressing technical problem that needs to be solved. Summary of the Invention

[0005] Therefore, it is necessary to provide a model training method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the elimination effect of ring artifacts in CT reconstructed images in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a model training method, including:

[0007] Acquire training sample data, which includes the first sample artifact image and the corresponding artifact-free label image;

[0008] The first sample artifact image is downsampled to obtain the second sample artifact image, and the first artifact capture image is obtained based on the second sample artifact image and the initial artifact removal network; the initial artifact removal network is a two-stage multi-scale artifact removal network.

[0009] The first artifact capture image is superimposed on the first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image;

[0010] Based on the first artifact ablation image and the artifact-free label image corresponding to the first sample artifact image, the loss function is determined, and the initial artifact removal network is iteratively trained based on the loss function to obtain the target artifact removal model.

[0011] In one embodiment, the resolution of the first artifact capture image is the same as the resolution of the first sample artifact image. Based on the second sample artifact image and the initial artifact removal network, the first artifact capture image is obtained, including:

[0012] The second sample artifact image is input into the initial artifact removal network to obtain the second artifact ablation image;

[0013] The second artifact ablation image and the second sample artifact image are differentiated to obtain the second artifact capture image. The artifact features in the second artifact capture image are ring artifact features that are insensitive to resolution changes.

[0014] The first artifact capture image is obtained by upsampling the second artifact capture image.

[0015] In one embodiment, the initial artifact removal network includes a first-stage network and a second-stage network. The first-stage network extracts artifact features based on polar coordinates, and the second-stage network performs artifact ablation based on the Cartesian domain. A second sample artifact image is input into the initial artifact removal network to obtain a second artifact-ablated image, including:

[0016] The second sample artifact image is input into the first-stage network for artifact feature extraction to obtain the third artifact capture image;

[0017] The second sample artifact image and the third artifact capture image are stitched together and then input into the second-stage network for artifact ablation to obtain the second artifact ablation image.

[0018] In one embodiment, the second sample artifact image is input into the first-stage network for artifact feature extraction to obtain a third artifact capture image, including:

[0019] The second sample artifact image is polarized to obtain the sample projection image in polar coordinates.

[0020] The sample projection image is input into the first-stage network for artifact feature extraction, and the first intermediate artifact ablation image in polar coordinates is obtained.

[0021] The first intermediate artifact ablation image is subjected to inverse polarization processing to obtain the second intermediate artifact ablation image in the Cartesian domain.

[0022] The second sample artifact image and the second intermediate artifact ablation image are differentially processed to obtain the third artifact capture image.

[0023] In one embodiment, the first-stage network is an attention-gated artifact capture network, and / or the second-stage network is a noise-reducing attention-gated artifact ablation network.

[0024] In one embodiment, the loss function is a combined loss function, which includes at least two of the following: mean absolute loss, mean squared loss, and structural similarity loss.

[0025] In one embodiment, obtaining training sample data includes:

[0026] Obtain the artifact-free label image and transform it to the projection domain to obtain the artifact-free projection image;

[0027] In the projection domain, multiple strip-shaped artifacts that conform to a normal distribution are randomly generated along the angular direction;

[0028] The Sobel operator with random orientation is used to enhance the edges of each strip artifact, generating an artifact noise image;

[0029] The image with artifact noise is superimposed on the image without artifacts to obtain the image with artifacts. The image with artifacts is then converted to the image domain to obtain the reconstructed image with artifacts.

[0030] Generate a phantom selection mask corresponding to the reconstructed image containing artifacts, and apply Gaussian blur to the phantom selection mask;

[0031] The reconstructed image containing artifacts is weighted and fused with the Gaussian blurred phantom selection mask to obtain the first sample artifact image.

[0032] In one embodiment, a Sobel operator with random orientation is used to enhance the edges of each strip-shaped artifact, generating an artifact noise image, including:

[0033] Multiple strip-shaped artifacts conforming to a normal distribution are randomly generated in the first direction;

[0034] Sobel operator kernel with determined random direction;

[0035] The Sobel operator kernel is convolved with the strip artifacts along the horizontal and vertical directions to obtain the strip artifacts with enhanced edges.

[0036] Generate artifact noise images based on the strip artifacts after edge enhancement.

[0037] Secondly, this application provides a method for artifact removal, including:

[0038] Obtain the initial reconstructed image after image reconstruction; the initial reconstructed image includes ring artifacts;

[0039] The initial reconstructed image is downsampled to obtain a low-resolution reconstructed image.

[0040] The low-resolution reconstructed image is input into the target artifact removal model to obtain the target artifact capture image. The target artifact removal model is obtained by training an initial artifact removal network based on a first sample artifact image and the corresponding unlabeled image. Specifically, the first sample artifact image is downsampled to obtain a second sample artifact image, and a first artifact capture image is obtained based on the second sample artifact image and the initial artifact removal network. The first artifact capture image is superimposed on the first sample artifact image to obtain the corresponding first artifact ablation image. Then, the initial artifact ablation network is trained based on the loss between the first artifact ablation image and the corresponding unlabeled image of the first sample artifact image. The initial artifact removal network is a two-stage multi-scale artifact removal network.

[0041] Based on the target artifact capture image, artifact removal is performed on the initial reconstructed image to obtain the target reconstructed image after artifact removal.

[0042] Thirdly, this application also provides a model training apparatus, comprising:

[0043] The sample acquisition module is used to acquire training sample data, which includes the first sample artifact image and the corresponding artifact-free label image.

[0044] The artifact capture module is used to downsample the first sample artifact image to obtain the second sample artifact image, and based on the second sample artifact image and the initial artifact removal network, to obtain the first artifact capture image; the initial artifact removal network is a two-stage multi-scale artifact removal network.

[0045] The artifact removal module is used to overlay the first artifact capture image onto the first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image.

[0046] The model training module is used to determine the loss function based on the first artifact ablation image and the artifact-free label image corresponding to the first sample artifact image, and to iteratively train the initial artifact removal network based on the loss function to obtain the target artifact removal model.

[0047] Fourthly, this application also provides an artifact removal device, comprising:

[0048] The image acquisition module is used to acquire the initial reconstructed image after image reconstruction; the initial reconstructed image includes ring artifacts.

[0049] The image processing module is used to downsample the initial reconstructed image to obtain a low-resolution reconstructed image;

[0050] The artifact capture module is used to input the low-resolution reconstructed image into the target artifact removal model to obtain the target artifact capture image. The target artifact removal model is obtained by training an initial artifact removal network based on a first sample artifact image and the corresponding unlabeled image. Specifically, the first sample artifact image is downsampled to obtain a second sample artifact image, and a first artifact capture image is obtained based on the second sample artifact image and the initial artifact removal network. The first artifact capture image is superimposed on the first sample artifact image to obtain the corresponding first artifact ablation image. Then, the initial artifact ablation network is trained based on the loss between the first artifact ablation image and the corresponding unlabeled image of the first sample artifact image. The initial artifact removal network is a two-stage multi-scale artifact removal network.

[0051] The artifact removal module is used to remove artifacts from the initial reconstructed image based on the target artifact capture image, and obtain the target reconstructed image after artifact removal.

[0052] Fifthly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the model training method in the first aspect and / or the artifact removal method in the second aspect.

[0053] Sixthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the model training method in the first aspect and / or the artifact removal method in the second aspect.

[0054] In a seventh aspect, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the model training method in the first aspect and / or the artifact removal method in the second aspect.

[0055] The aforementioned model training method, artifact removal method, apparatus, computer equipment, storage medium, and computer program product involve the computer equipment acquiring training sample data, which includes a first sample artifact image and a corresponding artifact-free label image. Next, the first sample artifact image is downsampled to obtain a second sample artifact image, and a first artifact capture image is obtained based on the second sample artifact image and an initial artifact removal network. The initial artifact removal network is a two-stage, multi-scale artifact removal network. Further, the first artifact capture image is superimposed on the first sample artifact image to obtain a first artifact ablation image corresponding to the first sample artifact image. Then, based on the first artifact ablation image and the artifact-free label image, a loss function is determined, and the initial artifact removal network is iteratively trained based on the loss function to obtain the target artifact removal model. In other words, the model training method proposed in this application downsamples the high-resolution training samples, i.e., the first sample artifact image, to a low-resolution second sample artifact image that is compatible with the model. Combined with a two-stage multi-scale artifact removal network, artifact removal is performed on the second sample artifact image to capture the first artifact capture image. Based on the first artifact capture image obtained from the low-resolution artifact image, artifacts are removed from the high-resolution first sample artifact image. Then, the initial artifact removal network is trained based on the first artifact ablation image after artifact removal to obtain the final target artifact removal model. That is, this application learns to capture coarse-grained artifact features that still exist after downsampling through a two-stage multi-scale artifact removal network, and superimposes the captured artifact features onto the high-resolution first sample artifact image, which can eliminate ring artifacts while ensuring image resolution. The target artifact removal model trained based on this method can improve the elimination effect of ring artifacts in the reconstructed image. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a diagram illustrating the application environment of the model training method and the artifact removal method in one embodiment.

[0058] Figure 2 This is a flowchart illustrating a model training method in one embodiment;

[0059] Figure 3 This is a flowchart illustrating the model training method in another embodiment;

[0060] Figure 4 This is a schematic diagram of a dual-branch multi-scale fusion noise reduction model in one embodiment;

[0061] Figure 5 This is a schematic diagram of a multi-scale depth reconstruction process in one embodiment;

[0062] Figure 6 This is a flowchart illustrating the model training method in another embodiment;

[0063] Figure 7 This is a schematic diagram of the artifact addition process in one embodiment;

[0064] Figure 8 This is a flowchart illustrating an artifact removal method in one embodiment;

[0065] Figure 9 This is a schematic diagram of the true artifact elimination state in the CT results of one embodiment;

[0066] Figure 10 This is a schematic diagram illustrating performance metric optimization in one embodiment;

[0067] Figure 11 This is a schematic diagram illustrating the water-HAP phantom treatment effect in one embodiment;

[0068] Figure 12 This is a schematic diagram illustrating the differences in CT scans before and after processing in one embodiment;

[0069] Figure 13 This is a schematic diagram illustrating the differences and distribution of CT scans before and after processing in one embodiment.

[0070] Figure 14 This is a structural block diagram of a model training device in one embodiment;

[0071] Figure 15 This is a structural block diagram of an artifact elimination device in one embodiment;

[0072] Figure 16 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0074] Photon-counting spectral CT can simultaneously provide CT data at multiple energy thresholds for multi-energy imaging, resulting in multiple reconstructed images. Based on these reconstructed images, material decomposition can be performed to obtain effective atomic number maps, which can then be used for chemical composition analysis. However, due to the uncertainty in the response of the photon-counting detector, bar artifacts can appear during scanning, leading to concentric ring artifacts in the reconstructed images. This significantly affects image quality and limits its clinical application. In other words, the presence of concentric ring artifacts in the reconstructed images results in inaccuracies in the effective atomic number maps obtained after material decomposition.

[0075] Ring artifacts are generally caused by discontinuous responses between different pixels of a photon counting detector. In related technologies, the following two methods are commonly used to address ring artifacts:

[0076] 1. Remove vertical stripe artifacts from the sinogram image in the projection domain, thus eliminating ring artifacts in the reconstructed image. This method is typically used for severe, dense ring artifacts and relies on the original projection data, requiring a large amount of computational data.

[0077] 2. Performing polar coordinate transformation in the reconstructed image domain will cause artifacts to appear as stripes, making them easier to detect and eliminate. This method does not rely on projection data and has wide applicability, but it may leave residual artifacts or introduce image blurring.

[0078] The methods described above are highly dependent on filter design and parameter configuration, requiring proprietary parameter tables for each detector, thus limiting their applicability across different devices. However, with the development of convolutional neural network deep learning models, neural networks are widely used in CT image post-processing to process and restore images. Supervised learning models, in particular, are favored due to their good accuracy and generalization ability. However, current supervised learning methods for eliminating ring artifacts still face the following challenges:

[0079] 1. Artifact Simulation Problem: In supervised learning, artifact simulation is often used to train models by generating a large amount of circular artifact data with different features. A common approach is to project the image domain data into the projection domain and then add noise to the projection domain, such as adding multiple horizontally spaced, uniformly distributed linear noise lines. This is then back-projected back into the image domain to generate circular artifacts. However, the artifacts generated by this method do not accurately reflect reality, and models trained based on this scenario perform poorly in real-world scenarios.

[0080] 2. Model Perception Ability: In Cartesian coordinates (reconstructed image), convolutional network models have difficulty capturing true circular artifact features (incomplete loops with random orientations). However, in polar coordinates (projected domain image), circular features are transformed into band-like features, which are more easily captured by the model. However, the sampling and interpolation involved in the coordinate transformation process often lead to blurred edges and features in the final image.

[0081] Furthermore, since the resulting images from photon counting CT have a high resolution (approximately 2000×2000 pixels), while existing deep learning-processed polar coordinate images are typically 512×512 pixels, the higher resolution places stringent demands on system performance and the model's long-distance perception capabilities.

[0082] Based on this, embodiments of this application provide a model training method and an artifact removal method, which can eliminate ring artifacts in photon-counting CT based on deep learning. Specifically, a ring artifact simulation algorithm is first used to generate training images as sample images. Then, a two-stage multi-scale fusion model is used to process the images to obtain high-quality CT image results. The artifact removal model trained using this method can efficiently and accurately eliminate ring artifacts in CT reconstructed images, improving the elimination effect of ring artifacts and providing high-quality CT reconstructed images for subsequent clinical applications.

[0083] The model training method and artifact removal method provided in this application can be applied to, for example... Figure 1 In the application environment shown, the computer device 101 can run a model training algorithm and / or an artifact removal algorithm to train the artifact removal model and perform post-processing on the reconstructed image in practical applications, i.e., to eliminate ring artifacts in the reconstructed image.

[0084] For example, computer device 101 can be a terminal device or a server. The terminal device can be, but is not limited to, various personal computers, laptops, smartphones, tablets, etc., and the server can be implemented using a standalone server or a server cluster composed of multiple servers.

[0085] In one exemplary embodiment, such as Figure 2 As shown, a model training method is provided, which can be applied to... Figure 1 The following steps, 201 to 204, are used as an example of computer equipment.

[0086] Step 201: Obtain training sample data, which includes the first sample artifact image and the corresponding artifact-free label image.

[0087] The first sample artifact image can be a medical scan image including ring artifacts, which can be a medical reconstructed image obtained by reconstructing the image based on the projection data obtained from the medical scan. Exemplarily, the first sample artifact image can include, but is not limited to, computed tomography (CT) images, cone-beam computed tomography (CBCT) images, positron emission tomography-computed tomography (PET-CT) images, etc. This application embodiment does not specifically limit the image type of the first sample artifact image.

[0088] For example, a computer device can acquire a reconstructed image obtained by a medical scanning device. If the reconstructed image contains annular artifacts, it can be used as a first sample artifact image. If the reconstructed image does not contain annular artifacts, it can be used as an artifact-free labeled image.

[0089] For the first sample artifact image, if there is no corresponding artifact-free label image for the first sample artifact image, artifact removal processing can be performed on the first sample artifact image, and the image after artifact removal can be used as the artifact-free label image corresponding to the first sample artifact image.

[0090] For the obtained artifact-free label image, if there is no corresponding first sample artifact image for the artifact-free label image, artifact overlay processing can be performed on the artifact-free label image, and the image after overlaying the artifacts can be used as the first sample artifact image corresponding to the artifact-free label image.

[0091] For example, the first sample artifact image and the corresponding artifact-free label image can be high-resolution images, such as images with any resolution greater than a preset resolution threshold, where the preset resolution threshold can be the resolution corresponding to the second sample artifact image. It should be noted that in certain special scenarios, the resolution of the first sample artifact image and the corresponding artifact-free label image may also be less than the resolution of the second sample artifact image. In this case, the first sample artifact image can be upsampled to obtain the second sample artifact image.

[0092] Step 202: Downsample the first sample artifact image to obtain a second sample artifact image, and obtain a first artifact capture image based on the second sample artifact image and the initial artifact removal network. The second sample artifact image is a low-resolution image, such as an image with a resolution less than or equal to a preset resolution threshold.

[0093] For example, a high-resolution first sample artifact image can be downsampled to obtain a low-resolution second sample artifact image, which can be adapted to an artifact removal model based on low-resolution image processing.

[0094] Next, the computer device can input the low-resolution second sample artifact image into the initial artifact removal network for artifact removal processing. This initial artifact removal network can be a two-stage, multi-scale artifact removal network. That is, by fusing two processing stages and performing artifact feature analysis at multiple scales, this initial artifact removal network can improve the recognition and removal of ring artifacts. For example, the initial artifact removal network may include a first-stage network and a second-stage network, wherein the first-stage network and / or the second-stage network can employ a multi-scale network structure for artifact feature analysis.

[0095] Furthermore, after processing the low-resolution second sample artifact image through the initial artifact removal network, a first artifact capture image can be obtained. This first artifact capture image is used to characterize the ring artifacts in the first sample artifact image. For example, the first artifact capture image can be a low-resolution artifact image with the same resolution as the second sample artifact image, or it can be a high-resolution artifact image with the same resolution as the first sample artifact image.

[0096] For example, after inputting the second sample artifact image into the initial artifact removal network, a low-resolution artifact ablation image can be obtained, that is, the image after eliminating the ring artifact; then, the second sample artifact image and the artifact ablation image can be differentially calculated and upsampled to the size of the first sample artifact image, thereby obtaining the high-resolution first artifact capture image, that is, the ring artifact feature at the original resolution.

[0097] Step 203: The first artifact capture image is superimposed on the first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image.

[0098] The first artifact ablation image is a high-resolution image with the same resolution as the first sample artifact image.

[0099] In one alternative implementation, the first artifact capture image can be a negative artifact image. For example, by differentiating the artifact ablation image with the second sample artifact image, a first artifact capture image with negative artifacts can be obtained. Based on this, the first artifact capture image with negative artifacts can be superimposed (i.e. summed) onto the first sample artifact image. At this time, the positive artifacts on the first sample artifact image and the negative artifacts on the first artifact capture image cancel each other out, and the first artifact ablation image after artifact removal can be obtained.

[0100] In another alternative implementation, the first artifact capture image can also be a positive artifact image. For example, by differentiating the second sample artifact image with the artifact ablation image, a first artifact capture image with positive artifacts can be obtained. Based on this, the first sample artifact image can be differentiated from the first artifact capture image with positive artifacts. At this time, after differentiating the positive artifacts on the first sample artifact image with the positive artifacts on the first artifact capture image, the first artifact ablation image after artifact removal can be obtained.

[0101] Furthermore, it should be noted that when the resolution of the first artifact capture image is the same as the resolution of the first sample artifact image, the first artifact capture image can be directly superimposed on the first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image; when the resolution of the first artifact capture image is the same as the resolution of the second sample artifact image, the first artifact capture image can be upsampled first to obtain an artifact capture image with the same resolution as the first sample artifact image, and then the artifact capture image can be superimposed on the first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image.

[0102] Step 204: Based on the first artifact ablation image and the artifact-free label image corresponding to the first sample artifact image, determine the loss function, and iteratively train the initial artifact removal network based on the loss function to obtain the target artifact removal model.

[0103] The first artifact ablation image is the artifact-free image corresponding to the first sample artifact image obtained after artifact removal by the initial artifact removal network, also known as the artifact-free output image. By comparing the artifact-free output image output by the network with the artifact-free label image, and calculating the network's loss function, the initial artifact removal network is iteratively trained using this loss function until a preset iteration stopping condition is met, thus obtaining the trained target artifact removal model. The preset iteration stopping condition includes, but is not limited to, reaching a preset number of iterations or reaching a preset iteration standard. The preset iteration standard may include the value of the loss function being less than a preset loss threshold or the artifact removal effect reaching a preset removal effect.

[0104] For example, the loss function can be a single loss function, such as any one of the mean absolute error loss function, mean squared error loss function, structural similarity loss function, gradient loss function, and perceptual loss function.

[0105] For example, the loss function can also be a combined loss function, which includes at least two of the following loss functions: mean absolute loss, mean squared loss, structural similarity loss, gradient loss, and perceptual loss. In an exemplary embodiment, the combined loss function can be a weighted sum of mean absolute loss, mean squared loss, and structural similarity loss, which can be expressed by the formula:

[0106] (1)

[0107] in,

[0108] (2)

[0109] (3)

[0110] (4)

[0111] For the average absolute loss, For mean square loss, SSIM is the structural similarity loss. SSIM is a structural metric that primarily considers three key features of an image: luminance, contrast, and structure, as detailed below:

[0112] (5)

[0113] in, By comparing the mean brightness of the input image, we can obtain: This can be represented as:

[0114] (6)

[0115] (7)

[0116] This is obtained by comparing the grayscale standard deviation of the input image, and can be expressed as:

[0117] (8)

[0118] (9)

[0119] By comparing the normalized input images using correlation coefficients, it can be represented as:

[0120] (10)

[0121] (11)

[0122] The weights in formula (1) , and The algorithm will be dynamically adjusted during training, focusing on MSE or MAE loss in the early stages and gradually increasing the weight of SSIM in the middle and later stages of training. The control strategy is as follows:

[0123] (12)

[0124] (13)

[0125] (14)

[0126] Where epoch represents the current epoch, and Total epoch represents the total number of epochs. for The initial weights, for The initial weights are then determined. In each epoch, the loss function calculates the loss value and updates the network parameters using stochastic gradient descent.

[0127] For example, after the model training is completed, the model performance can be evaluated according to the structural similarity index, peak signal-to-noise ratio and mean square error, etc. When the model performance meets the preset evaluation requirements, the initial artifact removal network after training can be used as the target artifact removal model.

[0128] In the above model training method, the computer device acquires training sample data, which includes a first sample artifact image and a corresponding artifact-free label image. Then, the first sample artifact image is downsampled to obtain a second sample artifact image. Based on the second sample artifact image and an initial artifact removal network, a first artifact capture image is obtained. The initial artifact removal network is a two-stage, multi-scale artifact removal network. Further, the first artifact capture image is superimposed on the first sample artifact image to obtain a first artifact ablation image corresponding to the first sample artifact image. Then, based on the first artifact ablation image and the artifact-free label image, a loss function is determined, and the initial artifact removal network is iteratively trained based on the loss function to obtain the target artifact removal model. In other words, the model training method proposed in this application downsamples the high-resolution training samples, i.e., the first sample artifact image, to a low-resolution second sample artifact image that is compatible with the model. Combined with a two-stage multi-scale artifact removal network, artifact removal is performed on the second sample artifact image to capture the first artifact capture image. Based on the first artifact capture image obtained from the low-resolution artifact image, artifacts are removed from the high-resolution first sample artifact image. Then, the initial artifact removal network is trained based on the first artifact ablation image after artifact removal to obtain the final target artifact removal model. That is, this application learns to capture coarse-grained artifact features that still exist after downsampling through a two-stage multi-scale artifact removal network, and superimposes the captured artifact features onto the high-resolution first sample artifact image, which can eliminate ring artifacts while ensuring image resolution. The target artifact removal model trained based on this method can improve the elimination effect of ring artifacts in the reconstructed image.

[0129] In an exemplary embodiment, the resolution of the first artifact capture image is consistent with the resolution of the first sample artifact image; based on this, as... Figure 3 As shown, step 202 above, "obtaining the first artifact capture image based on the second sample artifact image and the initial artifact removal network," may include steps 301 to 303. Wherein:

[0130] Step 301: Input the second sample artifact image into the initial artifact removal network to obtain the second artifact ablation image.

[0131] The resolution of the second artifact ablation image is consistent with that of the second sample artifact image. In this example, a two-stage, multi-scale initial artifact removal network is used to identify and remove the ring artifact features in the low-resolution second sample artifact image, thereby obtaining the artifact-removed second artifact ablation image.

[0132] In one exemplary embodiment, such as Figure 4 As shown, a network structure for an initial artifact removal network is provided. This initial artifact removal network may include a first-stage network and a second-stage network. The first-stage network extracts artifact features based on the polar coordinate domain, corresponding to the first stage being the polar domain artifact removal stage. The second-stage network performs artifact ablation based on the Cartesian domain, corresponding to the second stage being the attention-based denoising stage. In an optional implementation, the first-stage network may be an attention-gated artifact capture network, and / or the second-stage network may be a denoising attention-gated artifact ablation network.

[0133] Based on the above network structure, when processing the second sample artifact image through the initial artifact removal network, the second sample artifact image can first be input into the first-stage network to extract artifact features in the polar coordinate domain, i.e., the projection domain, to obtain a third artifact capture image (such as a low-resolution artifact prior image); wherein the resolution of the third artifact capture image is consistent with the resolution of the second sample artifact image; then, the second sample artifact image and the third artifact capture image are concatenated and stitched together before being input into the second-stage network to perform artifact ablation in the Cartesian domain, i.e., the image domain, to obtain a second artifact ablation image. That is, the first-stage network can be a multi-scale network, wherein the multi-scale can include polar coordinate scale and Cartesian coordinate scale, i.e., projection domain scale and image domain scale.

[0134] For example, continue to refer to Figure 4 As shown, the network structure of the first-stage network may include a polarization layer, an artifact capture layer (i.e., a network containing a noise reduction step and an attention module), an anti-polarization layer, and a difference layer. Based on this, the process of extracting features from the second sample artifact image through the first-stage network to obtain the third artifact capture image may include: inputting the second sample artifact image (such as an 8~16 bit integer PNG format DICOM image, i.e., a low-resolution original image with artifacts) into the polarization layer to polarize the second sample artifact image and obtain the sample projection image in the polar coordinate domain; Next, the sample projection image is input to the artifact capture layer for artifact feature extraction, resulting in a first intermediate artifact ablation image in the polar coordinate domain. Then, the first intermediate artifact ablation image is input to the inverse polarization layer to perform inverse polarization processing, resulting in a second intermediate artifact ablation image in the Cartesian domain. Finally, the second sample artifact image and the second intermediate artifact ablation image are input to the difference layer to perform difference processing, resulting in a third artifact capture image (such as a low-resolution artifact prior image).

[0135] Step 302: Difference is performed on the second artifact ablation image and the second sample artifact image to obtain the second artifact capture image.

[0136] The artifact feature in the second artifact capture image is the ring artifact feature, which is insensitive to resolution changes. It should be noted that for high-resolution CT images, tissue detail information is lost during downsampling, and this process is irreversible. That is, after upsampling the downsampled CT image to obtain a high-resolution CT image, the lost tissue detail information will not be present in the upsampled CT image; that is, the upsampled CT image differs in detail from the original CT image, resulting in a decrease in image quality. However, for ring artifacts, resolution is irrelevant. At extremely low resolutions, the features of ring artifacts are easily distinguishable; for example, the ring artifacts extracted at 256×256 and 1024×1024 sizes are identical. Therefore, the high-resolution image can be downsampled and compressed first, then the artifacts can be extracted by the model and restored to the original size. This can then be superimposed on the original image for artifact removal. In this way, the original image does not lose detail (because there is no compression), and the artifacts are eliminated, thus improving the artifact removal effect while ensuring image quality.

[0137] For example, a computer device can perform a difference operation on the second artifact ablation image and the second sample artifact image to obtain a second artifact capture image representing negative artifacts. Alternatively, the computer device can perform a difference operation on the second sample artifact image and the second artifact ablation image to obtain a second artifact capture image representing positive artifacts.

[0138] Step 303: Upsample the second artifact capture image to obtain the first artifact capture image.

[0139] The second artifact capture image is an artifact image with the same resolution as the second sample artifact image, i.e., a low-resolution artifact image. Based on this, the low-resolution second artifact capture image can be upsampled to obtain a high-resolution first artifact capture image with the same resolution as the first sample artifact image.

[0140] Combination Figure 2 and Figure 3 The example provides a complete model training process; refer to [reference]. Figure 5As shown, the high-resolution first sample artifact image is first downsampled to obtain a low-resolution second sample artifact image. Then, the low-resolution second sample artifact image is input into a two-stage, multi-scale initial artifact removal network to obtain a low-resolution second artifact ablation image. The low-resolution second artifact ablation image and the second sample artifact image are then differiated to obtain a low-resolution second artifact capture image. Next, the low-resolution second artifact capture image is upsampled to obtain a high-resolution first artifact capture image. This high-resolution first artifact capture image is then superimposed on the high-resolution first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image. Finally, based on the first artifact ablation image corresponding to the high-resolution first sample artifact image and the artifact-free label image, the network's loss function is determined. Based on this loss function, the initial artifact removal network adapted to the low resolution is iteratively trained to obtain the target artifact removal model.

[0141] The core concept of the multi-scale perception-based depth image reconstruction algorithm proposed in this application is: a deep learning model that integrates polar coordinates and Cartesian domains captures the coarse-grained artifact features that still exist after downsampling, then upsamples the captured feature results to the original size, and then superimposes them onto the high-resolution initial image, thereby eliminating ring artifacts while ensuring image resolution.

[0142] Assuming the high-resolution first sample artifact image is 2000×2000 pixels, after downsampling, a low-resolution second sample artifact image of 768×768 pixels is obtained, and this image is input into a dual-stage multi-scope artifact removal network model (DSMS). Figure 4 As shown), this model mainly consists of two stages:

[0143] The first stage involves feature extraction based on the polar coordinate domain. An attention-gated Unet is used to process the circular artifact features that are easier to capture in polar coordinates. The captured results (the third artifact capture image) and the input tensor (the second sample artifact image) are then concatenated and output to the next stage.

[0144] Second stage: Cartesian domain-based artifact ablation: A denoising attention-gated Unet network is used to process features at both scales simultaneously and output the final loop removal result.

[0145] The core model of this second-stage model can be AG-Unet, which is a noise-reducing attention-gated U-shaped network. It adopts a spatial / channel attention mechanism based on CBAM (Convolutional Block Attention Module) and adds rich attention gating to the skip connection skip layer to improve the model's perception ability.

[0146] In this embodiment, a low-resolution image processing artifact removal model can be used to process low-resolution artifact images to capture coarse-grained artifact features. Then, these coarse-grained artifact features are upsampled from low resolution to high resolution, and artifact ablation is performed at high resolution, outputting a high-resolution artifact-ablated image. This eliminates ring artifacts in the image while ensuring high-resolution image quality, comprehensively improving both image quality and artifact removal effectiveness. Furthermore, the two-stage artifact removal model efficiently balances artifact removal effectiveness and image feature preservation because the second stage, AG-Unet, considers both the perception results from the polar coordinate domain feature capture stage and the original image input, and uses an attention mechanism to repair only existing artifact features without introducing new artifacts. During model training, the model directly accepts artifact images and corresponding unlabeled images for training, and iterative training under the guidance of a combined dynamic loss function improves the model's effectiveness in eliminating ring artifacts.

[0147] The model trained using this method achieves end-to-end artifact removal at low resolution. Further difference operations are performed on the artifact-ablated image and the artifact image at low resolution to obtain artifact features. These features exhibit the same shape properties (coarse-grained) at both high and low resolutions. Therefore, the artifacts obtained at low resolution can be upsampled to their original size and added to the original high-resolution image to achieve high-quality artifact ablation while preserving the fine-grained organ features of the original image.

[0148] In one exemplary embodiment, a method for obtaining training sample data is provided, such as... Figure 6 As shown, step 201 above may include steps 601 to 603. Wherein:

[0149] Step 601: Obtain the artifact-free label image and convert it to the projection domain to obtain the artifact-free projection image.

[0150] Step 602: In the projection domain, randomly generate multiple strip-shaped artifacts that conform to a normal distribution along the angular direction.

[0151] Step 603: The Sobel operator with random orientation is used to enhance the edges of each strip artifact to generate an artifact noise image.

[0152] Step 604: Superimpose the artifact noise image onto the artifact-free projection image to obtain the artifact projection image.

[0153] Step 605: Convert the artifact projection image to the image domain to obtain the reconstructed image containing the artifact.

[0154] Step 606: Generate a phantom selection mask corresponding to the reconstructed image containing artifacts, and perform Gaussian blur processing on the phantom selection mask.

[0155] Step 607: The reconstructed image containing artifacts is weighted and fused with the Gaussian blurred phantom selection mask to obtain the first sample artifact image.

[0156] refer to Figure 7 As shown, this illustrates the process of generating a first sample artifact image based on an artifact-free label image. First, the artifact-free label image is transformed to the projection domain, i.e., orthographically projected to obtain an artifact-free projection image. Simultaneously, multiple normally distributed strip artifacts are randomly generated along the angular direction, and the edges of each strip artifact are enhanced using a Sobel operator with random directions, generating an artifact noise image. Next, the artifact noise image is superimposed on the artifact-free projection image to obtain an artifact projection image, which is then back-projected to obtain a reconstructed image containing artifacts in the image domain. Then, a phantom selection mask corresponding to the artifact-containing reconstructed image is generated, and Gaussian blurring is applied to the phantom selection mask. Finally, the artifact-containing reconstructed image and the Gaussian-blurred phantom selection mask are weighted and fused to obtain the first sample artifact image corresponding to the artifact-free label image.

[0157] For example, the process of using a Sobel operator with random directions to enhance the edges of each strip artifact and generate an artifact noise image may further include: determining a Sobel operator kernel with random directions; performing convolution operations on the strip artifacts along the horizontal and vertical directions respectively using the Sobel operator kernel to obtain the edge-enhanced strip artifacts; and generating an artifact noise image based on the edge-enhanced strip artifacts.

[0158] In one alternative implementation, multiple horizontally distributed bar-shaped artifacts are first randomly generated. Then, for each bar-shaped artifact, a Sobel kernel with a random orientation is used to perform a vertical convolution operation, thereby enhancing the vertical edges and generating the final artifact-noise image. For example, a computer device can use a Sobel operator with a random orientation to process each horizontal bar to enhance the upper and lower edges of each bar. Based on this, the generated artifact-noise image can be superimposed on the artifact-free projection image to obtain the artifact-projection image. The artifact-projection image is then reconstructed in the Cartesian domain to obtain the first sample artifact image corresponding to the artifact-free label image.

[0159] In this embodiment, by generating incomplete annular artifacts in the projection domain, it is possible to generate annular artifacts that are more consistent with the real situation of photon counting CT in the projection domain. This allows the training samples to better match the annular artifacts in the reconstructed CT images of the real scene, improving the accuracy and realism of the model training samples. In turn, this improves the accuracy and reliability of model training, thereby enhancing the training model's ability to eliminate annular artifacts.

[0160] In one exemplary embodiment, such as Figure 8 As shown, an artifact removal method is provided, which is applied to... Figure 1 The following steps are used as an example of computer equipment, including steps 801 to 804. Wherein:

[0161] Step 801: Obtain the initial reconstructed image after image reconstruction; the initial reconstructed image includes ring artifacts.

[0162] For example, annular artifacts may include circular artifacts or arc-shaped artifacts.

[0163] Step 802: Downsample the initial reconstructed image to obtain a low-resolution reconstructed image.

[0164] Step 803: Input the low-resolution reconstructed image into the target artifact removal model to obtain the target artifact capture image. The target artifact removal model is obtained by training the initial artifact removal network based on the first sample artifact image and the corresponding unlabeled image. Specifically, the first sample artifact image is downsampled to obtain the second sample artifact image, and the first artifact capture image is obtained based on the second sample artifact image and the initial artifact removal network. The first artifact capture image is superimposed on the first sample artifact image to obtain the corresponding first artifact ablation image. Then, the initial artifact ablation network is trained based on the loss between the first artifact ablation image and the unlabeled image corresponding to the first sample artifact image. The initial artifact removal network is a two-stage multi-scale artifact removal network.

[0165] Step 804: Based on the target artifact capture image, perform artifact removal on the initial reconstructed image to obtain the target reconstructed image after artifact removal.

[0166] By employing the model training methods described in the above embodiments, a target artifact elimination model for efficiently eliminating ring artifacts can be trained. In the application stage, based on this target artifact elimination model, precise elimination of ring artifacts in high-resolution reconstructed images obtained from medical scanning equipment can be achieved. Specifically, the initial reconstructed image including ring artifacts is first downsampled to a low-resolution reconstructed image. Then, the low-resolution reconstructed image is input into the target artifact elimination model to obtain a target artifact capture image. Here, the target artifact capture image can be a high-resolution artifact image, that is, by differentiating the low-resolution artifact elimination image output by the target artifact elimination model from the low-resolution reconstructed image to obtain the low-resolution artifact capture image, and then upsampling the low-resolution artifact capture image to obtain the high-resolution target artifact capture image.

[0167] Then, the high-resolution target artifact capture image is superimposed on the initial reconstructed image to eliminate the ring artifacts in the initial reconstructed image, resulting in the target reconstructed image after artifact removal.

[0168] The training process of the target artifact removal model can be referred to the relevant descriptions of the model training methods provided in the above embodiments, and will not be repeated here.

[0169] This artifact removal method employs a two-stage, multi-scale artifact removal model to learn and capture coarse-grained artifact features that still exist after downsampling. These captured artifact features are then superimposed onto the high-resolution initial reconstructed image, enabling the elimination of ring artifacts while maintaining image resolution. This targeted artifact removal model improves the removal effect of ring artifacts in the reconstructed image, enhances the quality of the reconstructed image, and thus facilitates subsequent clinical applications, thereby increasing its clinical value.

[0170] The feasibility and effectiveness of the above method are verified through simulation experiments below. Taking photon-counting CT images as an example, the effect of photon-counting CT in eliminating material artifacts is as follows: Figure 9 As shown, it can be seen that in the image with artifact removal, the ring artifacts present in the original image are well eliminated.

[0171] The artifact removal performance metrics of the model are as follows: Figure 10As shown in Table 1, the testing method involved adding artifacts to the original image, followed by using a model to remove the artifacts. The results with and without artifacts were then compared to the original image, and the differences in various metrics were calculated. It can be seen that, in addition to a significant improvement in the SSIM structural similarity index, the signal-to-noise ratio of the output results was also significantly improved.

[0172] Table 1

[0173]

[0174] The crucial role of photon-counting CT is energy spectrum decomposition, which demands extremely high accuracy in the CT values ​​obtained after material decomposition. When using bone or teeth as the target, a water-hydroxyapatite (HAP) phantom is typically used for calibration. However, artifacts during the calibration process can severely impact the accuracy of the final material decomposition. Therefore, high demands are placed on the model's processing capabilities and the final output results. Figure 11 and Figure 12 As shown, the model can handle artifacts in HAP images quite well. Statistical analysis shows that ( Figure 13 The mean difference between the processed CT mean and the true CT value was 2.61 HU, and the variance was 0.608 HU, both of which are at a low level.

[0175] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0176] Based on the same inventive concept, this application also provides a model training apparatus for implementing the model training method described above, and an artifact removal apparatus for implementing the artifact removal method. The solution provided by this apparatus is similar to the solution described in the above method. Therefore, the specific limitations of one or more model training apparatus embodiments and artifact removal apparatus embodiments provided below can be found in the limitations of the model training method and artifact removal method described above, and will not be repeated here.

[0177] In one exemplary embodiment, such as Figure 14As shown, a model training device is provided, including: a sample acquisition module 1401, an artifact capture module 1402, an artifact removal module 1403, and a model training module 1404, wherein:

[0178] The sample acquisition module 1401 is used to acquire training sample data, which includes the first sample artifact image and the corresponding artifact-free label image.

[0179] The artifact capture module 1402 is used to downsample the first sample artifact image to obtain the second sample artifact image, and to obtain the first artifact capture image based on the second sample artifact image and the initial artifact removal network; the initial artifact removal network is a two-stage multi-scale artifact removal network.

[0180] The artifact removal module 1403 is used to superimpose the first artifact capture image onto the first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image.

[0181] The model training module 1404 is used to determine the loss function based on the first artifact ablation image and the artifact-free label image corresponding to the first sample artifact image, and to iteratively train the initial artifact removal network based on the loss function to obtain the target artifact removal model.

[0182] In one embodiment, the resolution of the first artifact capture image is the same as the resolution of the first sample artifact image. The artifact capture module 1402 includes:

[0183] The artifact removal unit is used to input the second sample artifact image into the initial artifact removal network to obtain the second artifact ablation image;

[0184] The difference unit is used to perform difference on the second artifact ablation image and the second sample artifact image to obtain the second artifact capture image. The artifact features in the second artifact capture image are ring artifact features that are insensitive to resolution changes.

[0185] The upsampling unit is used to upsample the second artifact capture image to obtain the first artifact capture image.

[0186] In one embodiment, the initial artifact removal network includes a first-stage network and a second-stage network. The first-stage network extracts artifact features based on the polar coordinate domain, and the second-stage network performs artifact ablation based on the Cartesian domain. The artifact removal unit is specifically used to input the second sample artifact image into the first-stage network for artifact feature extraction to obtain a third artifact capture image. The second sample artifact image and the third artifact capture image are then stitched together and input into the second-stage network for artifact ablation to obtain a second artifact ablation image.

[0187] In one embodiment, the artifact removal unit is specifically used to polarize the second sample artifact image to obtain a sample projection image in the polar coordinate domain; input the sample projection image into the first-stage network for artifact feature extraction to obtain a first intermediate artifact ablation image in the polar coordinate domain; perform inverse polarization processing on the first intermediate artifact ablation image to obtain a second intermediate artifact ablation image in the Cartesian domain; and perform differential processing on the second sample artifact image and the second intermediate artifact ablation image to obtain a third artifact capture image.

[0188] In one embodiment, the first-stage network is an attention-gated artifact capture network, and / or the second-stage network is a noise-reducing attention-gated artifact ablation network.

[0189] In one embodiment, the loss function is a combined loss function, which includes at least two of the following: mean absolute loss, mean squared loss, and structural similarity loss.

[0190] In one embodiment, the sample acquisition module 1401 includes:

[0191] The image acquisition unit is used to acquire an artifact-free label image and convert the artifact-free label image to the projection domain to obtain an artifact-free projection image.

[0192] The noise generation unit is used to generate a noise image with artifacts based on preset artifact features; the preset artifact features include strip-shaped artifacts that conform to a normal distribution.

[0193] The sample acquisition unit is used to superimpose the artifact noise image onto the artifact-free projection image to obtain the artifact projection image, and convert the artifact projection image to the image domain to obtain the first sample artifact image corresponding to the artifact-free label image.

[0194] In one embodiment, the noise generation unit is specifically used to randomly generate multiple strip-shaped artifacts conforming to a normal distribution in a first direction; to perform edge enhancement on each strip-shaped artifact in a second direction to generate an artifact noise image; the second direction and the first direction are two directions that are perpendicular to each other.

[0195] In one exemplary embodiment, such as Figure 15 As shown, an artifact removal device is provided, including: an image acquisition module 1501, a downsampling module 1502, an artifact capture module 1503, and an artifact removal module 1504, wherein:

[0196] The image acquisition module 1501 is used to acquire the initial reconstructed image after image reconstruction; the initial reconstructed image includes ring artifacts.

[0197] Image processing module 1502 is used to downsample the initial reconstructed image to obtain a low-resolution reconstructed image;

[0198] The artifact capture module 1503 is used to input the low-resolution reconstructed image into the target artifact removal model to obtain the target artifact capture image. The target artifact removal model is obtained by training an initial artifact removal network based on a first sample artifact image and the corresponding unlabeled image. Specifically, the first sample artifact image is downsampled to obtain a second sample artifact image, and a first artifact capture image is obtained based on the second sample artifact image and the initial artifact removal network. The first artifact capture image is superimposed on the first sample artifact image to obtain the corresponding first artifact ablation image. Then, the initial artifact ablation network is trained based on the loss between the first artifact ablation image and the corresponding unlabeled image of the first sample artifact image. The initial artifact removal network is a two-stage multi-scale artifact removal network.

[0199] The artifact removal module 1504 is used to remove artifacts from the initial reconstructed image based on the target artifact capture image, so as to obtain the target reconstructed image after artifact removal.

[0200] Each module in the aforementioned model training device and artifact removal device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0201] In one exemplary embodiment, a computer device is provided, taking a server as an example, and its internal structure diagram can be as follows: Figure 16 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores training sample data and the trained target artifact removal model. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a model training method and / or an artifact removal method.

[0202] Those skilled in the art will understand that Figure 16The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0203] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the model training method and / or artifact removal method in any of the above embodiments.

[0204] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the model training method and / or artifact removal method in any of the above embodiments.

[0205] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the model training method and / or artifact removal method in any of the above embodiments.

[0206] It should be noted that the data involved in this application (including but not limited to data used for analysis, data stored, data displayed, etc.) are all information and data that have been fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0207] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0208] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0209] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A model training method, characterized in that, The method includes: Acquire training sample data, which includes a first sample artifact image and a corresponding artifact-free label image; The first sample artifact image is downsampled to obtain the second sample artifact image, and a first artifact capture image is obtained based on the second sample artifact image and the initial artifact removal network; the initial artifact removal network is a two-stage multi-scale artifact removal network. The first artifact capture image is superimposed on the first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image; Based on the first artifact ablation image corresponding to the first sample artifact image and the artifact-free label image, a loss function is determined, and the initial artifact removal network is iteratively trained based on the loss function to obtain the target artifact removal model.

2. The method according to claim 1, characterized in that, The resolution of the first artifact capture image is the same as the resolution of the first sample artifact image. The step of obtaining the first artifact capture image based on the second sample artifact image and the initial artifact removal network includes: The second sample artifact image is input into the initial artifact removal network to obtain the second artifact ablation image; The second artifact ablation image and the second sample artifact image are differentially analyzed to obtain the second artifact capture image. The artifact features in the second artifact capture image are ring artifact features that are insensitive to resolution changes. The first artifact capture image is obtained by upsampling the second artifact capture image.

3. The method according to claim 2, characterized in that, The initial artifact removal network includes a first-stage network and a second-stage network. The first-stage network extracts artifact features based on the polar coordinate domain, and the second-stage network performs artifact ablation based on the Cartesian domain. The step of inputting the second sample artifact image into the initial artifact removal network to obtain the second artifact ablation image includes: The second sample artifact image is input into the first stage network for artifact feature extraction to obtain the third artifact capture image; The second sample artifact image and the third artifact capture image are stitched together and then input into the second stage network for artifact ablation to obtain the second artifact ablation image.

4. The method according to claim 3, characterized in that, The step of inputting the second sample artifact image into the first-stage network for artifact feature extraction to obtain the third artifact capture image includes: The second sample artifact image is polarized to obtain the sample projection image in polar coordinates. The sample projection image is input into the first stage network for artifact feature extraction to obtain the first intermediate artifact ablation image in polar coordinates. The first intermediate artifact ablation image is subjected to inverse polarization processing to obtain the second intermediate artifact ablation image in the Cartesian domain. The second sample artifact image and the second intermediate artifact ablation image are differentially processed to obtain the third artifact capture image.

5. The method according to claim 3, characterized in that, The first stage network is an attention-gated artifact capture network, and / or the second stage network is a noise-reducing attention-gated artifact ablation network.

6. The method according to any one of claims 1-5, characterized in that, The loss function is a combined loss function, which includes at least two of the following: mean absolute loss, mean squared loss, and structural similarity loss.

7. The method according to any one of claims 1-5, characterized in that, The acquisition of training sample data includes: The artifact-free label image is obtained and converted to the projection domain to obtain the artifact-free projection image; In the projection domain, multiple strip-shaped artifacts conforming to a normal distribution are randomly generated along the angular direction; The Sobel operator with random orientation is used to enhance the edges of each strip artifact, generating an artifact noise image; The artifact noise image is superimposed on the artifact-free projection image to obtain the artifact projection image; The artifact projection image is reconstructed into the image domain to obtain a reconstructed image containing artifacts; Generate a phantom selection mask corresponding to the reconstructed image containing artifacts, and apply Gaussian blur to the phantom selection mask; The reconstructed image containing artifacts is weighted and fused with the Gaussian blurred phantom selection mask to obtain the first sample artifact image.

8. The method according to claim 7, characterized in that, The step of using a Sobel operator with random directions to enhance the edges of each strip-shaped artifact to generate an artifact noise image includes: Sobel operator kernel with determined random direction; The Sobel operator kernel is convolved with the strip artifact along the horizontal and vertical directions to obtain the strip artifact with enhanced edges. The artifact noise image is generated based on the edge-enhanced strip artifacts.

9. A method for artifact removal, characterized in that, The method includes: Obtain the initial reconstructed image after image reconstruction; the initial reconstructed image includes annular artifacts; The initial reconstructed image is downsampled to obtain a low-resolution reconstructed image; The low-resolution reconstructed image is input into the target artifact removal model to obtain the target artifact capture image. The target artifact removal model is obtained by training an initial artifact removal network based on a first sample artifact image and the corresponding unlabeled image. Specifically, the first sample artifact image is downsampled to obtain a second sample artifact image, and a first artifact capture image is obtained based on the second sample artifact image and the initial artifact removal network. The first artifact capture image is superimposed on the first sample artifact image to obtain a corresponding first artifact ablation image. Then, the initial artifact ablation network is trained based on the loss between the first artifact ablation image and the unlabeled image corresponding to the first sample artifact image. The initial artifact removal network is a two-stage multi-scale artifact removal network. Based on the target artifact capture image, the initial reconstructed image is subjected to artifact removal to obtain the target reconstructed image after artifact removal.

10. A model training device, characterized in that, The device includes: The sample acquisition module is used to acquire training sample data, which includes a first sample artifact image and a corresponding artifact-free label image. The artifact capture module is used to downsample the first sample artifact image to obtain a second sample artifact image, and to obtain a first artifact capture image based on the second sample artifact image and an initial artifact removal network; the initial artifact removal network is a two-stage multi-scale artifact removal network. The artifact removal module is used to overlay the first artifact capture image onto the first sample artifact image to obtain the first artifact ablation image corresponding to the first sample artifact image. The model training module is used to determine a loss function based on the first artifact ablation image corresponding to the first sample artifact image and the artifact-free label image, and to iteratively train the initial artifact removal network based on the loss function to obtain the target artifact removal model.

11. An artifact elimination device, characterized in that, The device includes: The image acquisition module is used to acquire the initial reconstructed image after image reconstruction; the initial reconstructed image includes annular artifacts; The image processing module is used to downsample the initial reconstructed image to obtain a low-resolution reconstructed image; An artifact capture module is used to input the low-resolution reconstructed image into a target artifact removal model to obtain a target artifact capture image. The target artifact removal model is obtained by training an initial artifact removal network based on a first sample artifact image and a corresponding unlabeled image. Specifically, the first sample artifact image is downsampled to obtain a second sample artifact image, and a first artifact capture image is obtained based on the second sample artifact image and the initial artifact removal network. The first artifact capture image is then superimposed on the first sample artifact image to obtain a corresponding first artifact ablation image. Furthermore, the initial artifact ablation network is trained based on the loss between the first artifact ablation image and the corresponding unlabeled image of the first sample artifact image. The initial artifact removal network is a two-stage, multi-scale artifact removal network. The artifact removal module is used to remove artifacts from the initial reconstructed image based on the target artifact capture image, so as to obtain the target reconstructed image after artifact removal.

12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images