A physics-constrained safe metal artifact reduction inference method

Abstract

The application discloses a physical constraint-based safe metal artifact removal reasoning method, first, a CT image is acquired to establish a data set for a metal artifact removal task, and then the data set is preprocessed; then a PRISM model for metal artifact removal is constructed, the PRISM model is based on a cyclic consistency framework and a double-flow attention generator structure, and is combined with an adversarial discriminator and a composite optimization objective function; then the PRISM model is unsupervisedly trained by using a training set, and network parameters are jointly optimized by the composite optimization objective function; finally, the performance of the PRISM model after training is evaluated by using a test set. The method completes high-fidelity learning of artifact removal and structure preservation under an unsupervised condition, and significantly improves the generalization ability and stability of the model in a real metal artifact scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision and medical image technology, and particularly relates to a safety metal artifact removal reasoning method based on physical constraints. BACKGROUND

[0002] Computed tomography (CT) is an imaging technique widely used in clinical diagnosis, with advantages of high resolution and fast scanning speed. However, when there is a metal implant (such as an orthopedic fixator, a dental prosthesis, an artificial joint, etc.) in the patient's body, the high atomic number of the metal will cause serious attenuation of X-rays, resulting in photon starvation, beam hardening and scattering effects, thus forming obvious stripes, dark bands or local distortions in the reconstructed image, i.e. metal artifacts. Such artifacts not only reduce image quality, but also may mask lesions and interfere with physician interpretation, seriously affecting the accuracy and safety of clinical diagnosis.

[0003] Traditional metal artifact removal methods mostly rely on physical modeling and interpolation reconstruction. For example, the linear interpolation (LI) and normalized metal artifact reduction (NMAR) methods complete the missing ray data in the projection domain to reduce artifacts. However, such methods usually assume that the gray scale distribution in the artifact area is smooth, making it difficult to recover high-frequency details, and are prone to produce new secondary artifacts or edge blurring, resulting in structural distortion.

[0004] In recent years, the rapid development of deep learning technology has provided a new solution for artifact correction. End-to-end models based on convolutional neural networks (CNN) or generative adversarial networks (GAN) can learn the mapping from artifacts to artifact-free on paired training samples, achieving simultaneous recovery of texture and structure. However, it is difficult to obtain paired data for the same patient in the clinic, so researchers mostly use physical simulation to generate synthetic artifact images for training. However, due to the complex physical characteristics of real artifacts (such as different metal types, scanning parameters, and energy spectrum distribution), there is a significant domain gap between synthetic and real data, making the model have insufficient generalization performance in actual scenarios.

[0005] To solve the problem of lack of paired samples, unsupervised generation methods such as CycleGAN and ADN are proposed to achieve domain conversion or feature separation between unpaired data through cycle consistency learning. However, due to the lack of physical constraints, the model is prone to destroy the anatomical structure while removing artifacts, resulting in structural drift or contrast shift; in addition, the instability of adversarial training also introduces problems such as excessive suppression of artifacts or false enhancement of texture, affecting medical safety.

[0006] Therefore, a physical constraint reasoning based safe metal artifact removal method (PRISM) is urgently needed. The method can combine physical consistency modeling and cyclic adversarial learning under unpaired conditions, realize artifact residual separation and structure preservation through physical constraints, align the statistical distribution of real and synthetic artifacts through artifact consistency loss, and combine an adversarial discrimination module (ADM) to strengthen the texture authenticity and detail continuity of the reconstructed image, thereby realizing high-quality artifact removal with structural safety. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the prior art and provide a physical constraint reasoning based safe metal artifact removal method (PRISM). The method realizes automatic identification, explicit separation and accurate reconstruction of metal artifacts in CT images by constructing a cyclic consistency framework and a double-flow attention generator (DAG) structure, combining an adversarial discriminator module (ADM) and a composite optimization objective function. The method completes high-fidelity learning of artifact removal and structure preservation under unsupervised conditions, significantly improving the generalization ability and stability of the model in real metal artifact scenarios.

[0008] To achieve the above purpose, the technical solutions adopted by the present application are as follows:

[0009] A physical constraint based safe metal artifact removal reasoning method comprises the following steps:

[0010] Step 1, obtaining a CT image to establish a dataset for a metal artifact removal task, then preprocessing the dataset and dividing it into a training set and a test set;

[0011] Step 2, constructing a PRISM model for metal artifact removal, the PRISM model is based on a cyclic consistency framework and a double-flow attention generator structure, and combines an adversarial discriminator and a composite optimization objective function;

[0012] Step 3, using the training set to unsupervisedly train the PRISM model, and jointly optimizing network parameters through the composite optimization objective function;

[0013] Step 4, using the test set to evaluate the performance of the trained PRISM model, and verifying the generalization ability and reconstruction stability of the model in real metal artifact scenarios through a combination of pixel-level quantitative indicators and task-driven indicators;

[0014] Step 5, obtaining a real CT image, and using the PRISM model that has completed performance evaluation to perform safe metal artifact removal reasoning.

[0015] Preferably, the preprocessing method comprises metal mask extraction, linear interpolation correction and standardization processing of the original CT image.

[0016] As preferred, the cycle-consistency framework comprises: a low-quality domain branch, a high-quality domain branch, and an identity reconstruction branch, reversible decoupling and explicit reconstruction of artifact features are achieved through multi-stage mapping.

[0017] As preferred, in step 1, the low-quality branch takes the low-quality image containing metal artifacts and the artifact correction image generated by linear interpolation of the low-quality image as input, obtains the high-quality reconstructed image without artifacts through the generator, and obtains the first artifact information by calculating the difference between the low-quality image and the high-quality reconstructed image.

[0018] The high-quality domain branch adds the first artifact information to the high-quality image without artifacts to generate a synthetic artifact image, and corrects the synthetic artifact image again through the same generator to obtain a cycle-reconstructed high-quality image, and obtains the second artifact information by calculating the difference between the synthetic artifact image and the cycle-reconstructed high-quality image, and reconstructs the cycle-reconstructed low-quality image based on the second artifact information and the high-quality reconstructed image.

[0019] The identity reconstruction branch performs identity mapping on the high-quality image without artifacts, and processes the high-quality image and itself through the generator to obtain an identity reconstructed image, so as to constrain the model not to produce false repair behavior in the artifact-free area.

[0020] As preferred, the dual-flow attention generator in step 2 comprises an image decoding branch and an attention decoding branch, and the image decoding branch and the attention decoding branch share the same set of convolutional encoders in the encoding stage,

[0021] The output features of the convolutional encoders of the image decoding branch are connected through a set of residual connections, and then the artifact area detail features are extracted through the image decoding branch and a plurality of candidate reconstructed images are generated;

[0022] The output features of the convolutional encoders of the attention decoding branch are directly generated through the attention decoding branch to generate a plurality of channel spatial attention maps corresponding to the candidate reconstructed images and a background channel attention map;

[0023] The dual-flow attention generator normalizes the spatial attention map through a Softmax function, element-wise multiplies it with the candidate reconstructed image, sums it up, and then superimposes the element-wise product of the background channel attention map and the prior image to output the final artifact-free image.

[0024] As preferred, the convolutional encoder comprises a 7x7 convolutional layer to map the 3-channel input to 64 feature channels, coupled with instance normalization and ReLU activation; two layers of 3x3 convolutional layers with a stride of 2 to halve the feature map resolution step by step while expanding the channel number to 128 and 256; a number of residual blocks, each of which comprises two layers of 3x3 convolution and instance normalization + ReLU structure, and all of which are filled with reflection before convolution, and the output and input are added through residual connection.

[0025] As preferred, the image decoding branch comprises two layers of deconvolution for upsampling, one layer of 3x3 convolutional layer, each layer of deconvolution adopts convolution transpose with a stride of 2, and is coupled with instance normalization and ReLU activation function to realize the step-by-step compression of the channel number from

[0026] As preferred, the attention decoding branch comprises an independent upsampling module and a prediction head, and the upsampling module comprises two layers of 3x3 deconvolution, normalization and ReLU activation, and one layer of 1x1 convolution.

[0027] As preferred, the adversarial discriminator adopts the PatchGAN structure, and the adversarial discriminators of the PatchGAN structure are configured in the high-quality domain branch and the low-quality domain branch respectively, for evaluating the true and false probability distribution of the reconstructed image and the real image within the local receptive field.

[0028] As preferred, the composite optimization objective function comprises:

[0029] Cycle consistency loss, including low-quality domain cycle consistency loss and high-quality domain cycle consistency loss, which are respectively defined by calculating the L1 norm difference between the low-quality image and the cyclically reconstructed low-quality image, and the L1 norm difference between the high-quality image and the cyclically reconstructed high-quality image;

[0030] Identity loss, which is defined by calculating the L1 norm difference between the identity reconstructed image and the original high-quality image, is used to constrain the output invariance of the model in the artifact-free area;

[0031] Artifact consistency loss, which is defined by calculating the L1 norm difference between the first artifact information and the second artifact information, is used to maintain the consistency of the artifact features separated from the real image and the synthetic image;

[0032] Prior constraint loss, which is defined by calculating the L1 norm difference between the high-quality reconstructed image and the linear interpolation image, is used to guide the generator to avoid abnormal reconstruction behavior in the early stage of training;

[0033] ​​The adversarial loss includes the generator and discriminator loss of the low-quality domain and the high-quality domain, and is used to improve the texture authenticity and detail quality of the generated result.

[0034] Preferably, the model training in step 3 is specifically:

[0035] The deep learning framework is implemented, the batch size is set to 8, the Adam optimization algorithm is used, the initial learning rate is 0.0001, and the learning rate is halved every 20 training periods, and the total training period is 50 periods. During the training process, the cycle consistency loss, the identity loss, the artifact consistency loss, the prior constraint loss and the adversarial loss are calculated at the same time, and the generator and discriminator parameters are jointly optimized with the composite optimization target. The training strategy can realize the optimal balance between artifact removal and structure preservation performance under the condition of no paired samples, ensure the stable convergence of the network and have strong generalization ability.

[0036] Preferably, the performance evaluation in step 4 includes: on the synthetic data set with paired non-artifact gold standard, the peak signal-to-noise ratio, the structural similarity index, the root mean square error and the mean absolute error are used as the pixel-level quantitative indicators.

[0037] On the real clinical data set lacking pixel-level gold standard, the Dice similarity coefficient, the average intersection-over-union and the 95% Hausdorff distance are used as the task-driven indicators, and the downstream task is verified through the pre-trained segmentation network.

[0038] The results of different artifact removal methods are compared through visualization, and the artifact suppression degree, the boundary definition and the tissue continuity are analyzed as the qualitative evaluation indicators.

[0039] Preferably, the CT image obtained in step 1 includes simulation artifact synthesis and real clinical data.

[0040] The present application has the following characteristics and beneficial effects:

[0041] First, the physical constraint reasoning mechanism driven by cycle consistency. The application first combines the cycle consistency framework with the physical constraint artifact modeling, proposes a reversible artifact separation and reconstruction mechanism, and realizes the unified optimization of explicit artifact decoupling and structure preservation. Traditional metal artifact removal methods often rely on paired samples or explicit artifact mask labeling, and it is difficult to achieve stable learning under unsupervised conditions; and the application establishes a three-branch structure of low-quality domain, high-quality domain and identity reconstruction, constructs a closed-loop reasoning system of "artifact generation-artifact removal-artifact regeneration", so that the network can automatically learn the reversible relationship between artifact causes and removal mapping in the absence of paired supervision. Specifically, the low-quality domain branch extracts artifact features A_1, the high-quality domain branch injects it into the artifact-free image to realize artifact regeneration, and the identity branch ensures the structure preservation of the artifact-free image, so as to realize the adaptive balance of semantic decoupling and intensity alignment under the cycle consistency constraint.

[0042] At the same time, the application introduces two innovative constraints of artifact consistency loss and prior constraint loss in the optimization objective: the former explicitly constrains the consistency of artifact residuals in different cycle paths, avoiding artifact residues and over-repair phenomena; the latter uses the linear interpolation result as the intensity prior to guide the generator to maintain a physically reasonable intensity distribution in the initial training stage. This unsupervised cycle reasoning mechanism combined with physical prior makes the model automatically realize artifact identification, explicit separation and high-fidelity reconstruction without artificial annotation, which improves the structure stability and physical interpretability of the image by about 2.22 dB and 0.0138 respectively compared with the suboptimal unsupervised model ADN (PSNR 25.60 dB, SSIM 0.9147).

[0043] Second, the collaborative design of double-flow attention generator (DAG) and composite optimization strategy. To solve the high difference in intensity distribution between artifact area and background area, the application proposes a double-flow attention generator (DAG) structure, which first realizes spatial adaptive artifact repair in the metal artifact removal task. The generator includes image decoding and attention decoding: image decoding generates multiple candidate reconstruction results, and attention decoding generates corresponding spatial weight distribution, both of which are weighted and fused by Softmax to automatically balance artifact suppression and structure fidelity. This design can adaptively adjust the repair intensity according to the artifact density and morphology of different areas, and realize fine-grained artifact removal and texture continuity recovery.

[0044] In addition, the application integrates cycle consistency, identity, artifact consistency, prior constraint and adversarial five losses in a single framework through a composite optimization objective function, realizes the deep coupling of physical interpretability and stable generation adversarial training.

[0045] In summary, the application realizes explicit reversible removal and structural safe reconstruction of metal artifacts under unsupervised conditions through the multi-constraint collaborative optimization strategy of the cycle consistency framework and the double-flow attention generator (DAG), breaks through the limitations of traditional artifact suppression methods in sample dependence, structural distortion and optimization instability, and has significant innovation and application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 FIG. 1 is a schematic diagram of a physical constraint reasoning-based safe metal artifact removal method according to an embodiment of the application.

[0047] Figure 2 For Figure 1 FIG. 3 is a structural schematic diagram of a double-flow attention generator. DETAILED DESCRIPTION

[0048] The application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0049] A safe metal artifact removal reasoning method based on physical constraints is based on a cycle consistency framework and a double-flow attention generator (DAG) structure, and combines an adversarial discrimination module (ADM) and a composite optimization objective function composed of multiple constraint terms to realize automatic identification, explicit separation and accurate reconstruction of metal artifacts in CT images. The method completes high-fidelity learning of artifact removal and structure preservation under unsupervised conditions, significantly improves the generalization ability and stability of the model in real metal artifact scenarios.

[0050] Specifically, as shown in FIG. 1, the method comprises the following steps: Figure 1

[0051] Step 1, acquire CT images to establish a data set for a metal artifact removal task, then pre-process the data set, and divide it into a training set and a test set.

[0052] Specifically, to verify the effectiveness and generalization ability of the PRISM model, in this embodiment, a large-scale simulation data set and two real clinical data sets are used for training and evaluation. Each data set covers multiple artifact types, multiple anatomical regions and multiple scanning conditions, and can fully reflect the complexity of metal artifacts in actual clinical practice.

[0053] ​In the simulation data part, according to the artifact construction process of InDuDoNet and DICDNet, first, 400 non-artifact CT images are selected from the DeepLesion dataset as the basis data, and 100 metal masks of different sizes and shapes are constructed for simulating metal artifacts. In this embodiment, 90 masks and 200 images are selected for training, and the remaining 10 masks and 200 images are used for testing. The area distribution of the metal region covers multiple levels from large implants to small suture objects. All images are uniformly adjusted to 416x416 pixels, saved in the form of linear attenuation coefficients, ensuring that the simulation data has real physical meaning and provides pixel-level non-artifact gold standard for supervised training.

[0054] In the real clinical data part, the application adopts the SpineWeb dataset, which contains 125 spine CT scans from multiple hospitals. Due to the implantation of metal instruments such as pedicle screws and connecting rods during surgery, there are strong, wide, and complex metal artifacts in the images, which are superimposed with spine structures, fractures, and degenerative lesions, providing important evidence for the robustness of the model under high-difficulty clinical conditions.

[0055] In addition, the application also uses the CTPelvic1K dataset to further evaluate the cross-anatomical generalization ability of the model in the pelvic region. The CLINIC subset contains 103 non-artifact preoperative CT images, which are used as the training basis for high-quality structure learning and downstream segmentation tasks; the CLINIC-metal subset contains 75 postoperative metal-implanted patient images, with significant strip artifacts, stretching artifacts, and bone cortex edge deformation caused by metal, providing real case support for unsupervised training and clinical application testing.

[0056] In the data preprocessing aspect, all data uniformly adopts the ADN preprocessing strategy, first slicing the 3D volume image into 2D images in the axial direction, and selecting the slices containing obvious artifacts for model training; to ensure training stability, all slices are uniformly adjusted to 256x256 pixels, and the gray value is mapped to the [-1,1] interval. To ensure objectivity of evaluation, the application strictly divides the training set and test set according to the patient, and the data division ratio is 8:2, so as to avoid the same patient slices appearing in different data subsets, ensuring the independence and rigor of model performance evaluation.

[0057] Step 2, constructing a PRISM model for metal artifact removal, the PRISM model is based on a cycle consistency framework and a double-flow attention generator structure, and combines an adversarial discriminator and a composite optimization objective function.

[0058] The cycle consistency framework includes three parts: a low-quality domain branch, a high-quality domain branch, and an identity reconstruction branch.

[0059] Specifically, low-quality domain branch: input CT image containing metal artifacts is denoted as , and its linear interpolation image is denoted as . The above input is input into the generator , i.e. double-flow attention generator DAG, to obtain a high-quality reconstructed image after artifact removal:

[0060]

[0061] And by calculating the difference between the low-quality image and the high-quality reconstructed image, first artifact information is obtained:

[0062]

[0063] That is, the input artifact-containing image and the generated artifact-free image are pixel-by-pixel differentiated to depict the artifact component.

[0064] High-quality domain branch: in the high-quality domain branch, the above estimated first artifact information is superimposed on the high-quality artifact-free image to obtain a synthetic artifact image:

[0065]

[0066] The is input into the same generator to obtain a cyclically reconstructed high-quality image

[0067]

[0068] At the same time, the synthetic artifact image is subtracted from the cyclically reconstructed high-quality image to obtain second artifact information :

[0069]

[0070] And based on the second artifact information and the high-quality reconstructed image, a cyclically reconstructed low-quality image is reconstructed.

[0071] Identity reconstruction branch: in order to prevent the generator from introducing pseudo-repair or structural distortion in the clean area that should remain unchanged, an identity reconstruction branch is introduced in the embodiment. The branch directly inputs the high-quality image and itself as the prior input generator:

[0072]

[0073] Through the above operations, the and the original The difference between the two is constrained, which can effectively suppress the excessive repair behavior of the generator in the artifact-free area, and improve the fidelity of the structure and texture.

[0074] Further settings of the embodiment are as shown in the following table: Figure 2 As shown in the table, the generator G in the PRISM model adopts a double-flow attention generator (DAG) structure. The DAG is composed of an image decoding branch and an attention decoding branch, and the image decoding branch and the attention decoding branch share the same set of convolution encoders in the encoding stage, realizing the fine-grained artifact removal mechanism of “multi-candidate reconstruction + spatial attention selection”.

[0075] Specifically, the input is the aforementioned six-channel tensor , wherein: the first channel is the main image channel, and a 3-channel input is formed through channel replication for image flow encoding; the fourth to sixth channels are attention guide inputs for attention flow encoding and prior fusion. Among them , denotes the forward difference artifact guide image between the linear interpolation image and the image containing artifacts, and the area higher than the original image in the interpolation image is retained through the element threshold operation, which is used to explicitly depict the metal artifact affected area.

[0076] Convolution encoder structure: the image flow and the attention flow share the same set of convolution encoders in the encoding stage, which includes: one layer of convolution layer, which maps the 3-channel input to feature channels, and cooperates with instance normalization and ReLU activation; two layers of convolution layers with a step of 2, which gradually reduces the resolution of the feature map by half through downsampling, while expanding the number of channels to and ; a plurality of residual blocks (5 ResNet blocks), each of which includes two layers of convolution + instance normalization + ReLU structure, all of which use reflection padding before convolution, and the output and the input are added through a residual connection to realize stable deep feature modeling.

[0077] Decoder in the image decoding branch: the features processed by the residual block are first recovered in spatial resolution through two layers of deconvolution (up-sampling) modules: each layer of deconvolution uses convolution transpose with a step of 2, and cooperates with instance normalization and ReLU activation function to realize the gradual compression of the number of channels from ; then a layer of convolution layer is used to map the features to channel output, wherein is the number of candidate reconstructed images, which is preferably This outputs nine 3-channel candidate images: The candidate image output uses... Activate to constrain pixel intensity The interval facilitates numerically stable training.

[0078] The decoder in the attention decoding branch: The attention decoding branch shares the output features of the convolutional encoder, but uses independent upsampling and prediction heads: The upsampling part also consists of two layers. It consists of deconvolution + normalization + ReLU, gradually restoring the number of channels to a lower dimension; the last layer is... Convolution maps features to... There are 1 attention channel, corresponding to the attention maps of 9 candidate images and 1 background / prior channel: To achieve normalized allocation of multi-channel attention, the above is performed on a channel-dimension basis. Each attention map is Softmax normalized so that the sum of the attention weights of all channels at the same location is 1, thus forming a complete spatial weighted mask.

[0079] Finally, attention fusion and final output: In the output stage, each attention map is copied to 3 channels, multiplied element-wise with the corresponding candidate reconstructed image, and then superimposed with the prior image. Background channel:

[0080]

[0081] in This represents element-wise multiplication. For prior images, linear interpolation maps are preferred in metal artifact removal tasks. Or its channel copy form.

[0082] Through the above design, DAG can adaptively select different candidate repair results or directly retain the prior between artifacts and background regions, achieving fine-grained artifact suppression and anatomical structure preservation. At the same time, the multi-candidate + attention mechanism provides the network with greater representational power, which is beneficial for handling complex metal artifact patterns.

[0083] A further setting in this embodiment concerns the Adversarial Discriminator (ADM). To improve the realism of the generated results in terms of texture and edge details, this embodiment configures PatchGAN-style adversarial discriminator modules (ADMs) in both the high-quality and low-quality domains:

[0084] PatchGAN discriminator: the discriminator takes a single-channel CT image or a 1-channel feature map composed of CT images as input, and outputs a multi-channel "realness score map" corresponding to the input spatial size, i.e. each local patch corresponds to a true / false prediction value. The discriminator is composed of multiple layers of consecutive 4x4 convolution blocks: the first layer of convolution uses 64 filters with a step size of 2, combined with LeakyReLU activation (negative half-axis slope, for example, take 0.2) for fast extraction of local texture features; in the subsequent convolution blocks, the number of channels doubles with each layer (for example, 128, 256, 512), each layer uses 4x4 convolution + step size 2 or 1, and can optionally use different types of normalization layers such as batch normalization, instance normalization or layer normalization; the last layer uses a 4x4 convolution with 1 channel to output a local discrimination map with a step size of 1 as the basis for the adversarial loss.

[0085] It can be understood that the multi-domain discriminator: the discriminators are set in the high-quality domain and the low-quality domain respectively and , respectively, to judge the difference in statistical distribution between the generated artifact removal results or synthetic artifact results and the corresponding real images.

[0086] In this embodiment, the total loss function of the PRISM model is composed of multiple sub-losses weighted and combined, which is used to balance the artifact removal accuracy, structure preservation and training stability. The total loss can be represented as:

[0087]

[0088] Where:

[0089] Generator and discriminator loss : adopt the form of adversarial loss (such as least squares GAN loss or cross-entropy loss), constrain the overall intensity distribution and local texture distribution of the generated results in the low-quality domain and the high-quality domain to approach the real image distribution, and improve the visual authenticity.

[0090] Low / high domain cycle consistency loss : by constraining the difference between the reconstruction results after "low→high→low" mapping and the original , and the difference between the results after "high→low→high" cycle reconstruction and (adopting norm), ensure that the bidirectional mapping process maintains topological consistency, and avoid serious information loss or mode collapse.

[0091] Identity loss : by measuring with pixel difference, limiting the generator to perform an approximately identity mapping when the input is already a clean image, suppressing excessive modification to artifact-free areas, and protecting anatomical structures and detailed textures.

[0092] artifact consistency loss : for constraining the artifact information estimated from both low-quality and high-quality domains to maintain consistency in spatial distribution and intensity pattern, thereby strengthening the stability and interpretability of artifact separation and reducing ambiguity in artifact estimation.

[0093] prior constraint loss : using a linear interpolation map as a prior, guiding the attention background channel to prefer to preserve or fine-tune the prior image in structurally reliable areas while limiting excessive dependence in severely artifacted areas . This loss can be implemented by weighted constraint on the local difference between the output and .

[0094] The weight coefficients before each loss term can be set or tuned according to task requirements and experimental experience to achieve the optimal balance between artifact suppression degree, structure preservation, and training stability. In this experiment, they were set to 2.0, 2.0, 10.0, 20.0, 10.0, 10.0, and 2.0, respectively.

[0095] Through the PRISM model construction process described in the above steps, the present embodiment effectively combines the dual-flow attention generator and the adversarial discriminator module under a unified cycle consistency framework, utilizes the explicit artifact separation and multi-candidate attention fusion mechanism, realizes fine removal of metal artifacts and high-fidelity reconstruction of normal tissue structures, and provides a model basis that can be implemented in engineering and has clear physical and mathematical significance for the training process and inference application in subsequent steps.

[0096] The application realizes deep coupling of physical interpretability and stability of generative adversarial training by integrating cycle consistency, identity, artifact consistency, prior constraint and adversarial five losses in a single framework through composite optimization objective function. In the ablation experiment, the independent contribution and synergistic gain of the double-flow attention module and the key loss term are verified by removing them one by one: first, when the number of image groups in the double-flow attention fusion strategy is reduced or increased, the model artifact removal effect decreases, and the PSNR and SSIM decrease by about 0.93-1.21 dB and 0.0140-0.0257 respectively, indicating the key of the fusion strategy of N=9 in the application. Further, in the loss function ablation, removing each loss function will cause the model effect to decrease, and after removing the artifact consistency loss, the model cannot effectively learn the reversible structure pattern of the artifact residual, and the RMSE and MAE increase significantly (from 0.0245 to 0.0439, from 0.0086 to 0.0170), and obvious residual artifacts and edge blur appear. Finally, after integrating all the above modules, the application achieves the optimal result of PSNR=27.86 dB, SSIM=0.9285, MAE=0.0170 on the SynDeepLesion dataset, obtains a Top-1 rate of 28.75% in the expert subjective score of the SpineWeb clinical set, and realizes the best results of mIoU and Dice of 0.8478 and 0.9158 on CTPelvic1K. The results fully verify the significant advantages and technical value of the application in structure preservation, artifact suppression and clinical application stability.

[0097] Step 3, unsupervised training of the PRISM model using the training set, joint optimization of network parameters through the composite optimization objective function.

[0098] Specifically, the PRISM model is unsupervised trained in the embodiment. Specifically, a training environment is built based on the PyTorch deep learning framework, and a small batch optimization strategy with a batch size of 8 is adopted to improve training stability and reduce memory occupation. The optimization algorithm is selected as the Adam method, the initial learning rate is set to 0.0001, and in the training process, the learning rate is halved every 20 epochs in an exponential decay manner to achieve smooth iterative convergence. The total training period of the model is 5 epochs, which can obtain convergent network parameters under the premise of controllable training time.

[0099] Further, in order to ensure training stability, the weights of all convolution and deconvolution layers in the embodiment are initialized using a Gaussian distribution with a mean of 0 and a standard deviation of 0.02, and the bias term is initialized to 0. This initialization strategy helps to improve the convergence and generation quality in the adversarial training stage.

[0100] Step 4, using the test set to evaluate the performance of the trained PRISM model, by combining pixel-level indicators, task-level structural indicators and visual qualitative analysis, to comprehensively reflect the artifact suppression ability, structure preservation ability and medical usability of the model in real clinical scenarios.

[0101] Specifically, first, in terms of pixel-level evaluation, the embodiment calculates peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), root mean square error (RMSE) and mean absolute error (MAE) and other indicators, which are used to measure the quantitative performance of the model output image in terms of intensity consistency, texture similarity and error control. Such indicators can directly reflect the suppression degree and residual error level of the model in recovering artifact-free images, and are the main numerical indicators for artifact removal effect.

[0102] Secondly, in terms of task-level structure preservation, the artifact-removed image is input into the pre-trained organ segmentation network to calculate Dice coefficient, mean intersection over union (mIoU) and 95% Hausdorff distance (HD95) and other indicators, so as to evaluate the maintenance ability of the model for anatomical boundaries, structure contours and region overlap accuracy. Such indicators can reflect the effectiveness of the artifact-processed image in actual clinical analysis tasks, and are important indirect verification means for metal artifact removal methods.

[0103] Finally, in terms of qualitative analysis, the embodiment visualizes and compares the reconstruction results of different models, and comprehensively judges from the aspects of artifact suppression degree, continuity of bone cortex edge, hierarchical performance of soft tissue structure and preservation of texture details. At the same time, two radiology experts with rich reading experience are invited to subjectively rank the results in a double-blind manner, which is used to further verify the visual quality and medical usability of the method in real clinical environment.

[0104] On the synthetic artifact dataset SynDeepLesion, the proposed PRISM model is compared with a variety of representative metal artifact removal methods, including traditional interpolation methods (LI, NMAR), supervised deep learning methods (IndudoNet+, OSCNet+, FINDNet), and unsupervised generative adversarial methods (ADN, SemiMAR, Calimar-GAN). The results show that the present application achieves the best performance on all key indicators: PSNR reaches 27.86 dB, SSIM is 0.9285, RMSE is 0.0439, and MAE is 0.0170, which is about 2.22 dB and 0.0138 higher than the suboptimal unsupervised model ADN (PSNR 25.60 dB, SSIM 0.9147), and about 4.46 dB higher than SemiMAR and Calimar-GAN on average. This result shows that the PRISM model can significantly enhance the structural consistency and intensity fidelity of image reconstruction without paired training samples.

[0105] At the same time, the box plot and violin plot analysis results of the indicators further show that the PRISM model has the highest median and smallest variance in the distribution of PSNR and SSIM, showing stable structural fidelity; the error distribution is the most concentrated in RMSE and MAE, verifying the robustness of the model under different metal sizes and multi-scale artifacts. In contrast, SemiMAR and Calimar-GAN have a larger fluctuation in the distribution, indicating that they are more sensitive to changes in metal intensity, while PRISM can maintain stable artifact suppression under the guidance of physical constraints.

[0106] Comprehensive evaluation shows that the PRISM model of the present application achieves the best results in key indicators such as PSNR, SSIM, mIoU, and HD95. On the synthetic artifact dataset SynDeepLesion, the present application is about 2.22 dB and 0.0138 higher than the unsupervised method ADN (PSNR 25.60 dB, SSIM 0.9147), and more than 4 dB higher than SemiMAR and Calimar-GAN on average, indicating that the model still has excellent intensity recovery and structural consistency performance without paired samples. Further distribution analysis shows that the PRISM model has the highest median and smallest variance in PSNR and SSIM, and the most concentrated error in RMSE and MAE, indicating that it has high stability under different metal sizes, multi-scale artifacts, and different scanning conditions.

[0107] Step 5, obtain real CT images, and use the PRISM model that has completed performance evaluation to perform safe metal artifact removal inference.

[0108] On the real clinical data SpineWeb, a double-blind expert evaluation method is used for subjective quality ranking. The statistical results show that the PRISM model is selected in the "best visual reconstruction result" with the highest proportion of 28.75%, which is significantly better than ADN, OSCNet+, and other comparison methods. The results show that PRISM can significantly weaken the artifacts while fully preserving the bone cortex, pedicle and soft tissue details, and the image level is clear and the boundary transition is natural. Compared with the over-smoothing or residual artifact phenomenon commonly existing in other methods, the present application realizes the synchronous optimization of texture details and structure fidelity.

[0109] On the Clinic-metal subset of CTPelvic1K, a downstream structure segmentation task is further introduced as an indirect performance verification indicator. Through the transfer evaluation of the U-Net network pre-trained on the artifact-free CLINIC subset, the PRISM model achieves the best results on mIoU (0.8478) and Dice (0.9158), with an improvement of 0.0435 and 0.0078 over OSCNet+, and an improvement of 00.0197 and 0.0296 over ADN. Although its HD95 (4.30 px) is slightly higher than that of OSCNet+ (3.49 px), the overall structure consistency and region overlap accuracy are better, which verifies the anatomical fidelity and downstream task compatibility of the model after artifact suppression.

[0110] In addition, in terms of computing efficiency, the lightweight structure of the single generator of the present application significantly reduces the computing overhead. The parameter amount is only 8.42 M, which is reduced by about 70% compared with ADN, and the inference time of a single CT slice is 0.217 s, which is increased by about 49.3% compared with Calimar-GAN. This performance enables the model to be directly embedded in a clinical image processing system to realize real-time metal artifact removal and safe image reconstruction.

[0111] In summary, the method of the present application comprehensively surpasses the comparison models in terms of key indicators such as PSNR, SSIM, mIoU and Dice under unsupervised conditions, and has high precision, high stability and low computing cost, fully embodying its practical value and promotion potential in the actual clinical metal artifact suppression task.

[0112] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and do not limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A reasoning method for safe metal artifact removal based on physical constraints, characterized in that, Includes the following steps: Step 1: Obtain CT images to create a dataset for the metal artifact removal task, then preprocess the dataset and divide it into training and test sets; Step 2: Construct the PRISM model for metal artifact removal. The PRISM model is based on a cycle consistency framework and a two-stream attention generator structure, combined with an adversarial discriminator and a composite optimization objective function. The cycle consistency framework includes a low-quality domain branch, a high-quality domain branch, and an identity reconstruction branch, which achieves reversible decoupling and explicit reconstruction of artifact features through multi-stage mapping. The low-quality branch takes a low-quality image containing metal artifacts and an artifact-corrected image generated by linear interpolation as input, obtains a high-quality reconstructed image after artifact removal through a generator, and obtains the first artifact information by calculating the difference between the low-quality image and the high-quality reconstructed image. In the high-quality domain branch, the first artifact information is added to the high-quality artifact-free image to generate a synthetic artifact image. The synthetic artifact image is then corrected again using the same generator to obtain a cyclically reconstructed high-quality image. At the same time, the second artifact information is obtained by calculating the difference between the synthetic artifact image and the cyclically reconstructed high-quality image. Based on the second artifact information and the high-quality reconstructed image, a cyclically reconstructed low-quality image is reconstructed. The identity reconstruction branch performs identity mapping on high-quality, artifact-free images. It processes the high-quality image and itself through a generator to obtain an identity reconstructed image, thereby constraining the model to avoid erroneous repair behavior in artifact-free regions. The dual-stream attention generator includes an image decoding branch and an attention decoding branch, which share the same set of convolutional encoders during the encoding phase. The output features of the convolutional encoder in the image decoding branch are passed through a set of residual connections and then used by the image decoding branch to extract the detailed features of the artifact region and generate multiple sets of candidate reconstructed images. The output features of the convolutional encoder in the attention decoding branch are directly used to generate a multi-channel spatial attention map and a background channel attention map corresponding to the candidate reconstructed image. Step 3: Use the training set to perform unsupervised training on the PRISM model, and jointly tune the network parameters using the composite optimization objective function; Step 4: Use the test set to evaluate the performance of the trained PRISM model by combining pixel-level metrics, task-level structural metrics, and visual qualitative metrics. The task-level structural metrics are obtained by inputting the artifact-removed image into the pre-trained organ segmentation network to calculate the Dice coefficient, average intersection-over-union ratio, and 95% Hausdorff distance to evaluate the model's ability to maintain the accuracy of anatomical boundaries, structural contours, and regional overlap. Step 5: Obtain real CT images and use the PRISM model, which has completed performance evaluation, to perform safe metal artifact removal inference.

2. The reasoning method for safe metal artifact removal based on physical constraints according to claim 1, characterized in that, In step 1, the preprocessing method includes metal mask extraction, linear interpolation correction, and standardization of the original CT image.

3. The reasoning method for removing secure metal artifacts based on physical constraints according to claim 1, characterized in that, The dual-stream attention generator normalizes the spatial attention map using the Softmax function, sums it by element-wise multiplication with the candidate reconstructed image, and then superimposes the element-wise product of the background channel attention map and the prior image to output the final artifact-free image.

4. The reasoning method for removing secure metal artifacts based on physical constraints according to claim 3, characterized in that, The convolutional encoder includes a 7×7 convolutional layer that maps the 3-channel input to 64 feature channels, and incorporates instance normalization and ReLU activation; two consecutive 3×3 convolutional layers with a stride of 2, which halve the feature map resolution step by step through downsampling, while expanding the number of channels to 128 and 256; and several residual blocks, each containing two 3×3 convolutional layers and an instance normalization + ReLU structure, with reflection padding applied before each convolution. The output and input are added together through a residual concatenation.

5. The reasoning method for removing secure metal artifacts based on physical constraints according to claim 4, characterized in that, The image decoding branch includes two deconvolution layers for upsampling and one convolution layer. Each deconvolution layer uses convolution transpose with a stride of 2, and is combined with instance normalization and ReLU activation function to achieve progressive compression of the number of channels.

6. The reasoning method for removing secure metal artifacts based on physical constraints according to claim 5, characterized in that, The attention decoding branch includes an independent upsampling module and a prediction head. The upsampling module includes two 3×3 deconvolution layers, normalization and ReLU activation, and one 1×1 convolution layer.

7. The reasoning method for removing secure metal artifacts based on physical constraints according to claim 1, characterized in that, The adversarial discriminator adopts the PatchGAN structure, with adversarial discriminators of the PatchGAN structure configured in the high-quality domain branch and the low-quality domain branch, respectively, to evaluate the authenticity probability distribution of the reconstructed image and the real image within the local receptive field.

8. The reasoning method for removing secure metal artifacts based on physical constraints according to claim 7, characterized in that, The composite optimization objective function includes: Cyclic consistency loss, including low-quality domain cyclic consistency loss and high-quality domain cyclic consistency loss, is defined by calculating the L1 norm difference between the low-quality image and the cyclically reconstructed low-quality image, and between the high-quality image and the cyclically reconstructed high-quality image, respectively. The identity loss, defined by calculating the L1 norm difference between the identity reconstructed image and the original high-quality image, is used to constrain the output invariance of the model in artifact-free regions. The artifact consistency loss is defined by calculating the L1 norm difference between the first artifact information and the second artifact information, and is used to maintain the consistency of artifact features separated from the real image and the synthetic image. The prior constraint loss, defined by calculating the L1 norm difference between the high-quality reconstructed image and the linearly interpolated image, is used to guide the generator to avoid abnormal reconstruction behavior in the early stages of training. Adversarial losses, including generator and discriminator losses for low-quality and high-quality domains, are used to improve the texture realism and detail quality of the generated results.

9. The reasoning method for removing secure metal artifacts based on physical constraints according to claim 1, characterized in that, The model training in step 3 specifically involves: This was implemented using a deep learning framework with a batch size of 8. The Adam optimization algorithm was used, with an initial learning rate of 0.0001, which was halved every 20 training epochs, for a total of 50 training epochs.

10. The reasoning method for safe metal artifact removal based on physical constraints according to claim 1, characterized in that, The performance evaluation in step 4 includes: On a synthetic dataset with a gold standard for paired artifact-free images, peak signal-to-noise ratio, structural similarity index, root mean square error, and mean absolute error are used as pixel-level quantitative indicators. On real clinical datasets lacking pixel-level gold standards, Dice similarity coefficient, mean crossover ratio, and 95% Hausdorff distance were used as task-driven metrics, and downstream task validation was performed using a pre-trained segmentation network.

11. The reasoning method for removing secure metal artifacts based on physical constraints according to claim 1, characterized in that, The CT images obtained in step 1 include simulated artifact synthesis and real clinical data.

Images