CT artifact removal method based on frequency optimization and structure scale enhancement

Abstract

The invention provides a CT (Computed Tomography) artifact removal method based on frequency optimization and structure scale enhancement. The method comprises the following steps: acquiring a CT image containing artifacts to be processed; inputting the CT image containing the artifacts into the trained artifact removal network for processing to obtain a CT image without the artifacts; wherein the artifact removal network is formed by connecting an encoder-decoder structure and a residual error; the encoder comprises at least one encoding level and is used for performing feature extraction and down-sampling on an input image; the decoder comprises at least one decoding level and is used for carrying out feature fusion and up-sampling on features output by the encoder; a frequency domain interaction module, a prior aggregation module and a scale competition module which are connected in sequence are arranged in at least one part of the coding level and the decoding level. By constructing an encoder-decoder network integrated with three core modules, the CT artifact removal effect is collaboratively optimized from three dimensions, and high-precision suppression and high-fidelity reconstruction of various artifacts are jointly realized.

Description

Technical Field

[0001] This invention relates to the fields of computer vision technology and medical image processing technology, and in particular to a method for CT artifact removal based on frequency optimization and structural scale enhancement. Background Technology

[0002] Computed tomography (CT) is a core technology in medical imaging diagnosis, playing a crucial role in clinical disease detection and lesion identification. However, due to various factors such as imperfections in scanning equipment, differences in detector sensitivity, patient movement, and metallic implants, complex degradation phenomena such as ring artifacts, metal artifacts, and motion artifacts often appear in CT images. These artifacts not only reduce image readability but may also obscure subtle lesion information, leading to misdiagnosis or missed diagnosis, posing a serious threat to the accuracy of clinical diagnosis.

[0003] Traditional artifact removal methods are mostly based on physical modeling and reconstruction algorithms, such as filtering correction, projection compensation, and regularized reconstruction. These methods usually rely on strict imaging model assumptions, are sensitive to noise and structural details, and are difficult to maintain robust performance under complex degradation conditions. In recent years, artifact removal methods based on deep learning have become a research hotspot. Among them, convolutional neural networks (CNN) and Transformers have achieved artifact suppression and detail restoration through end-to-end learning. However, these methods also generally have the following shortcomings: (1) spatial domain feature extraction methods are difficult to effectively capture the degradation law of artifacts in the frequency domain, resulting in ring or strip-shaped artifact residues; (2) some networks are prone to losing tissue structure information while removing artifacts, causing blurring or oversmoothing phenomena; (3) when faced with multiple types of artifacts or real complex scanning conditions, the generalization ability of the model is insufficient, and it is difficult to balance robustness and detail fidelity. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the present invention provides a CT artifact removal method based on frequency optimization and structural scale enhancement to solve the technical problems of low accuracy and structural fidelity in the removal of complex artifacts.

[0005] To achieve the above and other related objectives, this invention provides a CT artifact removal method based on frequency optimization and structural scale enhancement, comprising: acquiring a CT image containing artifacts to be processed; inputting the CT image containing artifacts into a trained artifact removal network for processing to obtain a CT image with artifacts removed; wherein the artifact removal network consists of an encoder-decoder structure and a residual connection; the encoder includes at least one encoding layer for feature extraction and downsampling of the input image; the decoder includes at least one decoding layer for feature fusion and upsampling of the features output by the encoder; and in at least a portion of the encoding layer and at least a portion of the decoding layer, there are sequentially connected components: a frequency domain interaction module for converting the input features to the frequency domain for modulation and output; a prior aggregation module for performing sequence modeling on the features output by the frequency domain interaction module to capture global structural dependencies before output; and a scale competition module for performing multi-scale competitive fusion on the features output by the prior aggregation module before output.

[0006] In one embodiment of the present invention, each level of the encoder and each level of the decoder are provided with the frequency domain interaction module, the prior aggregation module and the scale competition module connected in sequence.

[0007] In one embodiment of the present invention, the frequency domain interaction module processes the input features as follows: normalizes and divides the input features of the frequency domain interaction module into blocks to obtain multiple image blocks; converts each image block to the frequency domain to separate the amplitude spectrum and phase spectrum; performs directional modulation on the amplitude spectrum to obtain the modulated amplitude spectrum; combines the modulated amplitude spectrum and the phase spectrum and converts them back to the spatial domain to obtain the output features of the frequency domain interaction module.

[0008] In one embodiment of the present invention, directional modulation of the amplitude spectrum to obtain a modulated amplitude spectrum includes: performing horizontal and vertical strip pooling on the amplitude spectrum to obtain horizontal features and vertical features, respectively; weighting the horizontal features and the vertical features using a learnable frequency mask; and fusing the weighted horizontal features and the vertical features through a multilayer perceptron to perform cross-frequency interaction to obtain the modulated amplitude spectrum.

[0009] In one embodiment of the present invention, combining the modulated amplitude spectrum and the phase spectrum and converting them back to the spatial domain to obtain the output features of the frequency domain interaction module includes: combining the modulated amplitude spectrum and the phase spectrum and converting them back to the spatial domain to obtain spatial features corresponding to each image block; fusing the spatial features corresponding to all image blocks to obtain converted spatial features; processing and splitting the converted spatial features using depthwise separable convolution to obtain a first sub-feature and a second sub-feature; activating the first sub-feature with the GELU activation function and multiplying it element-wise with the second sub-feature, and then performing a residual concatenation between the multiplication result and the input features of the frequency domain interaction module to obtain the output features of the frequency domain interaction module.

[0010] In one embodiment of the present invention, the prior aggregation module processes the input features as follows: cascaded 3×3 and 1×1 convolutions compress and reorganize the input features of the prior aggregation module into a structure-aligned feature sequence; the feature sequence is modeled through a gated loop mechanism to capture long-range dependencies between elements in the sequence; the modeled feature sequence is reconstructed into spatial features to obtain the output features of the prior aggregation module.

[0011] In one embodiment of the present invention, the scale competition module processes the input features as follows: multiple parallel convolutional branches are used to extract multi-scale features from the input features of the scale competition module to obtain multiple features of different scales; the multiple features of different scales are competitively fused to adaptively emphasize the scale feature with the most information and suppress redundant features, and the fused features are used as the output features of the scale competition module.

[0012] In one embodiment of the present invention, competitive fusion of the multiple features at different scales includes: concatenating the multiple features at different scales; performing global context encoding on the concatenated features to generate a scale importance weight vector; and performing channel-level weighted fusion of the multiple features at different scales based on the scale importance weight vector.

[0013] In one embodiment of the present invention, the training of the artifact removal network is performed by minimizing the total loss function, which is calculated by the following formula: where L total Let L be the total loss function; content For spatial domain content loss, used to calculate the difference between the predicted image and the real image at multiple resolutions; L fft λ is the frequency domain consistency loss, used to calculate the difference between the frequency domain representations of the predicted image and the real image at multiple resolutions; λ is the weighting coefficient that balances the two losses, with a value ranging from 0.1 to 0.5.

[0014] In one embodiment of the present invention, the spatial domain content loss and the frequency domain consistency loss are calculated using the following formulas: In the formula, L total Let L be the total loss function; content For spatial domain content loss, used to calculate the difference between the predicted image and the real image at multiple resolutions; L fft λ is the frequency domain consistency loss, used to calculate the difference between the frequency domain representations of the predicted image and the real image at multiple resolutions; λ is the weighting coefficient that balances the two losses, with a value ranging from 0.1 to 0.5.

[0015] In one embodiment of the present invention, the spatial domain content loss and the frequency domain consistency loss are calculated using the following formulas: , , In the formula, F represents the two-dimensional Fourier transform, and s represents different resolutions.

[0016] The beneficial effects of this invention are as follows: This invention proposes a CT artifact removal method based on frequency optimization and structural scale enhancement. This method constructs an encoder-decoder network integrating three core modules: frequency domain interaction, prior aggregation, and scale competition. This collaboratively optimizes the CT artifact removal effect from three dimensions. The frequency domain interaction module directly analyzes and modulates image features in the frequency domain, which can accurately locate and suppress artifacts with specific frequency characteristics (such as ring artifacts), solving the problem that spatial domain methods are insufficient in capturing frequency domain degradation patterns. The prior aggregation module dynamically models the spatial structure of the image as a sequence, effectively capturing long-range dependencies between global pixels, enhancing the model's understanding and preservation of complex tissue structures, and avoiding detail blurring. The scale competition module adaptively fuses multi-scale features, enabling the model to flexibly adjust the receptive field according to the scale characteristics of artifacts, significantly improving the generalization removal ability for artifacts of different types and sizes. The cascaded operation of these three modules jointly achieves high-precision suppression of various artifacts and high-fidelity reconstruction of real anatomical structures under complex degradation conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The accompanying drawings are incorporated in and constitute a part of this specification, illustrating embodiments consistent with this application, and are used together with the description to explain the principles of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0018] Figure 1 A flowchart of an artifact removal method provided in an embodiment of the present invention; Figure 2 This is an architecture diagram of an artifact removal network provided in an embodiment of the present invention; Figure 3 This is an architecture diagram of a frequency domain interaction module provided in an embodiment of the present invention; Figure 4 This is an architecture diagram of a priori aggregation module provided in an embodiment of the present invention; Figure 5 This is an architecture diagram of a scale competition module provided in an embodiment of the present invention. Detailed Implementation

[0019] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other. In addition to the specific methods, equipment, and materials used in the embodiments, based on the knowledge of the prior art and the description of the present invention by those skilled in the art, any prior art methods, equipment, and materials similar to or equivalent to the methods, equipment, and materials in the embodiments of the present invention can be used to implement the present invention.

[0020] It should be understood that the terminology used in the embodiments of this invention is for describing specific implementations and not for limiting the scope of protection of this invention. Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art.

[0021] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In some embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0022] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functions, and operations that may be implemented in the methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0023] Please see Figure 1 and Figure 2 , Figure 1 A CT artifact removal method based on frequency optimization and structural scale enhancement, provided in an embodiment of the present invention, includes: (1) acquiring a CT image containing artifacts to be processed; (2) inputting the CT image containing artifacts into a trained artifact removal network (i.e., Figure 1 The CT image is processed using a frequency domain sensing and structure enhancement network model to obtain artifact-free CT images.

[0024] Figure 2 The diagram shows the architecture of an artifact removal network, which consists of an encoder-decoder structure and a residual connection. The encoder contains at least one encoding layer for feature extraction and downsampling of the input image. The decoder contains at least one decoding layer for feature fusion and upsampling of the features output by the encoder. At least a portion of the encoding layer and at least a portion of the decoding layer include a frequency domain interaction module (FIM), a prior aggregation module (PAM), and a scale competition module (SCM) connected in sequence. Specifically, the frequency domain interaction module converts the input features to the frequency domain for modulation before output; the prior aggregation module performs sequence modeling on the features output by the frequency domain interaction module to capture global structural dependencies before output; and the scale competition module performs multi-scale competitive fusion on the features output by the prior aggregation module before output.

[0025] The encoder-decoder structure employed in this invention is a mature framework for image-to-image translation tasks. Its advantage lies in the fact that the encoder progressively extracts and condenses high-level semantic features of the image through multi-layer downsampling, while the decoder progressively recovers the spatial details of the image through upsampling. On this classic framework, this invention innovatively cascades three dedicated modules—FIM, PAM, and SCM—at key levels. The benefits of this design are: the encoder's downsampling process provides progressively abstracted features for FIM and PAM, facilitating frequency domain analysis and long-range structural modeling; the decoder's upsampling process allows SCM to progressively refine the reconstruction results using multi-scale contextual information. The cascading of the three modules forms a feature enhancement pipeline: FIM first filters out artifact noise in the frequency domain, PAM then enhances the coherence of the organizational structure in the spatial domain, and finally, SCM intelligently fuses multi-scale information to optimize the final output. This structure ensures the logical coherence and effectiveness of the artifact removal process from "frequency purification" to "structural restoration" and then to "multi-scale optimization."

[0026] In a specific embodiment of the present invention, each level of the encoder and each level of the decoder are provided with a frequency domain interaction module, a priori aggregation module and a scale competition module connected in sequence. Figure 2 The diagram illustrates this structure, with FIM, PAM, and SCM modules at each level of the encoder and decoder. The advantage of this integrated design lies in achieving multi-level, multi-granularity feature collaborative optimization. In shallow layers, the network primarily processes detailed textures and edges. At this stage, FIM effectively removes high-frequency artifacts, PAM establishes local structural relationships, and SCM integrates subtle scale variations. As the network deepens, feature maps become more abstract, representing higher-level semantic information. At this point, FIM can handle a wider range of frequency domain patterns, PAM can model more global anatomical structural dependencies (such as organ shape), and SCM can decide on larger-scale feature selection. This comprehensive optimization, from fine-to-coarse and then back to fine-to-fine, allows the network to continuously suppress artifacts and enhance structure throughout the entire feature extraction and reconstruction path, thereby achieving optimal overall performance.

[0027] Please see Figure 3 In a specific embodiment of the present invention, the frequency domain interaction module processes the input features according to steps 3.1 to 3.4.

[0028] Step 3.1: Normalize and patch the input features of the frequency domain interaction module to obtain multiple image patches. Normalization and patching are performed primarily based on two considerations: First, normalization (such as LayerNorm) can stabilize the distribution of input features, accelerate network training convergence, and improve the model's robustness to CT images of different contrasts; second, patching divides the entire feature map into multiple local image patches. This firstly reduces the data scale of each Fourier Transform (FFT) calculation, making it easier to process high-resolution images; more importantly, artifacts in CT images (especially ring-shaped and metallic artifacts) often exhibit typical frequency domain characteristics in local regions. Patching allows the FIM module to focus on the independent frequency domain characteristics of each local region, achieving targeted analysis and suppression of local artifacts and avoiding feature confusion that may be caused by global FFT.

[0029] Step 3.2: Transform each image patch to the frequency domain to separate the amplitude and phase spectra. Specifically, this is done using a two-dimensional Fast Fourier Transform (FFT), and the process can be represented as follows: In the formula, X i For the i-th image patch, M and e jθ These are the separated amplitude and phase spectra, respectively. This transformation allows the model to precisely manipulate the amplitude and phase components, preserving the integrity of the organization by retaining the original phase spectrum, and focusing the modulation emphasis on the amplitude spectrum components of the coding direction.

[0030] Step 3.3: Perform directional modulation on the amplitude spectrum to obtain the modulated amplitude spectrum.

[0031] In a specific embodiment of the present invention, step 3.3 includes steps 3.3.1 to 3.3.3.

[0032] Step 3.3.1: Perform horizontal and vertical strip pooling on the amplitude spectrum to obtain the horizontal feature F. h and vertical feature F v It can be expressed by the formula: .

[0033] To enhance frequency representation, strip pooling mechanisms are employed in both the vertical and horizontal directions, enabling modules to aggregate directional frequency responses. Strip pooling is a variant of the ordinary pooling operation; horizontal strip pooling (StripPool) is a specific example. h Pooling is performed within a very narrow horizontal strip window (e.g., a window size of H×1), thus aggregating statistical information from an entire row of pixels. It is particularly sensitive to frequency variations in the horizontal direction (corresponding to vertical edges or horizontal stripes). Similarly, vertical strip pooling...v The sampling is performed within a vertical strip window (e.g., 1×W) and is sensitive to frequency variations in the vertical direction (corresponding to horizontal edges or vertical stripes). In CT images, many artifacts (such as detector ring artifacts, which locally approximate concentric rings with varying tangential directions) manifest as stripes or edges in specific directions. Through strip pooling in these two orthogonal directions, the FIM module can effectively separate and capture the directional frequency energy distribution closely related to artifacts, providing a precise target for subsequent learnable mask-weighted suppression.

[0034] Step 3.3.2: Use a learnable frequency mask to weight the horizontal and vertical features respectively, expressed by the formula: , In the formula, W h and W v A learnable frequency mask is used. In this step, a learnable frequency mask is introduced to adaptively adjust the preserved directional energy and suppress artifact frequency components.

[0035] Step 3.3.3: The weighted horizontal and vertical features are fused using a multilayer perceptron to perform cross-frequency interaction, resulting in the modulated amplitude spectrum, which can be expressed by the formula: .

[0036] In this step, pooled features are input into a lightweight multilayer perceptron (MLP) to simulate cross-frequency dependencies, thereby enabling adaptive interaction between different frequency bands.

[0037] Step 3.4: Combine the modulated amplitude spectrum and phase spectrum and convert them back to the spatial domain to obtain the output characteristics of the frequency domain interactive module.

[0038] In a specific embodiment of the present invention, step 3.4 includes steps 3.4.1 to 3.4.4.

[0039] Step 3.4.1: Combine the modulated amplitude spectrum and phase spectrum and transform them back into the spatial domain to obtain the spatial features corresponding to each image patch. The transformation here uses inverse Fourier transform processing.

[0040] Step 3.4.2: Fuse the spatial features corresponding to all image blocks to obtain the transformed spatial features. The fusion operation in step 3.4.2 is a necessary inverse process of the block processing (step 3.1). In step 3.1, the input features are divided into multiple non-overlapping image blocks for independent frequency domain processing. After steps 3.2 to 3.4.1, each image block generates its corresponding spatial features, which have undergone frequency domain modulation and inverse transformation. At this point, these processed image blocks correspond one-to-one with the original partitioning in spatial location. The "fusion" operation reassembles these processed image blocks into a complete, spatially continuous feature map according to the original spatial arrangement order. This ensures that the output of the FIM module is still a feature tensor with a complete spatial structure, which can be seamlessly input into subsequent network layers (such as the PAM module) for further processing.

[0041] Step 3.4.3: Use depthwise separable convolution to process and split the transformed spatial features to obtain the first sub-feature h1 and the second sub-feature h2.

[0042] Step 3.4.4: After activating the first sub-feature with the GELU activation function, multiply it element-wise with the second sub-feature, and then perform residual connection between the multiplication result and the input feature of the frequency domain interaction module to obtain the output feature of the frequency domain interaction module.

[0043] The frequency domain interaction module achieves precise surgical manipulation of CT image features in the frequency domain through a process of "block-FFT-modulation-inverse FFT". Its core innovation lies in directional modulation: it utilizes strip pooling to capture the directional frequency features of artifacts, adaptively enhances useful frequency bands and suppresses artifact bands through a learnable mask, and then achieves cross-band information interaction via MLP. After the inverse transform, spatial features are refined and fused through depthwise separable convolution and gated residual connections (GELU activation + dot product). This module effectively separates and removes artifacts with obvious features in the frequency domain (such as ring and stripe artifacts), while preserving the tissue structure information of the original image to the greatest extent by retaining phase and residual connections, providing a "clean" and "fidelity-preserving" feature foundation for subsequent processing.

[0044] Please see Figure 4 In a specific embodiment of the present invention, the prior aggregation module processes the input features according to the following steps 4.1 to 4.3.

[0045] Step 4.1: The cascaded 3×3 and 1×1 convolutions compress and reorganize the input features of the prior aggregation module into a structure-aligned feature sequence. This step can be expressed by the formula: , .

[0046] Step 4.2: Model the feature sequence using a gated loop mechanism to capture long-range dependencies between elements in the sequence. Compared to statically extracting prior information from isolated regions, this invention treats the feature map as a dynamically evolving spatial trajectory. By reshaping spatial features into a sequence, directional transitions can be modeled step by step, thereby promoting dynamic structural evolution and enhancing the ability to extract structural priors. The specific process is as follows: , , , , In the formula, and Used to control information selection and suppress artifacts, x t h represents the current feature obtained through convolution. t-1 This represents the evolutionary characteristics of the previous iteration. First, the reset gate r is applied. t Adjust the historical memory and then compare it with the current input x. t New feature candidates are generated through fusion. Subsequently, the gating variable z... t The update mechanism determines whether to retain the existing structure or introduce new feature representations. Ultimately, h t This framework enables the continuous evolution of features during the iteration process. It treats the two-dimensional spatial layout as a dynamic sequence {x}. t Instead of a static mesh, each time step t corresponds to a spatial step size. The feature state is adaptively updated based on the accumulated structural context and the current spatial content.

[0047] Step 4.3: Reconstruct the modeled feature sequence into spatial features to obtain the output features of the prior aggregation module.

[0048] Please see Figure 5 In a specific embodiment of the present invention, the scale competition module processes the input features according to the following steps 5.1 to 5.2.

[0049] Step 5.1: Use multiple parallel convolutional branches to extract multi-scale features from the input features of the scale competition module, obtaining features at multiple different scales. This step can be expressed by the formula: , In the formula, s represents different scales. Parallel depthwise separable convolutions are used, and responses in different spatial ranges can be extracted as the kernel size increases. Each F s Capture artifact patterns within a specific spatial range.

[0050] Step 5.2: Competitively fuse features at multiple different scales to adaptively emphasize the scale feature with the most information and suppress redundant features, and use the fused feature as the output feature of the scale competition module.

[0051] In a specific embodiment of the present invention, competitive fusion of features at multiple different scales is performed, including: (1) splicing features at multiple different scales; (2) performing global context encoding on the spliced ​​features to generate a scale importance weight vector A; and (3) performing channel-level weighted fusion of features at multiple different scales based on the scale importance weight vector.

[0052] In this embodiment, a cross-scale competitive modulation mechanism is introduced instead of uniformly aggregating them to adaptively suppress redundant responses and emphasize the most informative components. For this purpose, multi-scale features are concatenated in parallel, and global contextual information is encoded through a lightweight attention mechanism. The above process can be expressed by the following formula: , , , In the formula, The scale selection vector reflects the relative importance of each branch; This represents the channel recalibration vector based on scale-competitive projection. The cross-scale competitive allocation mechanism introduces softmax-driven cross-scale branch sparse competition, followed by learnable projection to achieve fine-grained channel modulation. This design is beneficial for cross-scale competition, retaining the most significant artifact components and suppressing redundant interference. The final output is obtained by applying channel-weighted calibration, where ⊙ represents channel-weighted multiplication, ensuring that features at different scales are dynamically adjusted according to their importance in artifact removal.

[0053] It should be noted that the steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they contain the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0054] It should be noted that in practical applications, the original DICOM format CT images may vary in size. To adapt to the fixed input size (e.g., 256×256 pixels) of the artifact removal network described in this invention, necessary preprocessing is required for CT images containing artifacts, such as scaling or cropping them to the input size required by the network. Accordingly, the size of the artifact-removed image output by the network is consistent with the network input size. Therefore, post-processing may be necessary before outputting the final result, such as scaling the network output image back to the size of the original CT image for comparison with the original image or for clinical diagnosis.

[0055] It should be noted that the artifact removal network needs to be trained. The training process generally includes: (1) constructing a dataset; (2) constructing an artifact removal network; and (3) training the artifact removal network using the dataset.

[0056] In a specific embodiment of the present invention, the dataset can be constructed by adding artifacts to normal CT images. The specific steps include: (1) acquiring DICOM images of different tissues in Stagell-Colorectal-CT and Deep-Lesion, adjusting the window width and window level using ImageJ software, and generating PNG images; (2) dividing the images into three parts, adding different degrees of annular artifacts to the first part, adding different degrees of blur kernels to the second part, and adding metal masks of different shapes to the third part, thus obtaining annular artifact, motion artifact, and metal artifact datasets respectively. In this dataset, the image with added artifacts is used as the input to the model, and the image without added artifacts is used as the real image, constituting a sample. In addition, the number of samples in the dataset can be increased through dataset augmentation techniques.

[0057] In a specific embodiment of the present invention, the artifact removal network is trained by minimizing the total loss function, which is calculated using the following formula: , In the formula, L total L is the total loss function; content For spatial domain content loss, used to calculate the difference between the predicted image and the real image at multiple resolutions; L fft λ is the frequency domain consistency loss, used to calculate the difference between the frequency domain representations of the predicted image and the real image at multiple resolutions; λ is the weighting coefficient that balances the two losses, with a value ranging from 0.1 to 0.5.

[0058] Using the total loss function in model training offers significant advantages. Spatial domain content loss directly constrains the output image to approximate a clean, real image in terms of pixel values, forming the basis for ensuring global structural restoration and detail preservation. Frequency domain consistency loss, on the other hand, constrains the output image to have a consistent spectrum with the real image in the frequency domain, specifically penalizing abnormal frequency components caused by artifacts, and is highly effective in removing artifacts with obvious features in the frequency domain. Combining the two with appropriate weights (λ=0.1~0.5) achieves a perfect balance between spatial detail preservation and frequency domain artifact suppression, thereby training a model with superior performance.

[0059] In a specific embodiment of the present invention, the spatial domain content loss and the frequency domain consistency loss are calculated using the following formulas: , , In the formula, F represents the two-dimensional Fourier transform, and s represents different resolutions. The multi-scale loss calculation method (i.e., calculating L at three resolutions: s=1, 0.5, and 0.25) is used. content and L fft This further enhances the training effect. Calculating the loss at high resolution (s=1) ensures the network can recover the finest anatomical details; calculating the loss at medium to low resolution (s=0.5, 0.25) forces the network to focus on and reconstruct the correct overall contours, large organ morphologies, and medium-scale tissue structures. This multi-scale pyramid-style supervision strategy ensures the network produces accurate outputs not only at the pixel level but also at different semantic levels, effectively avoiding the problem of reconstructed images appearing reasonable locally but with distorted overall structure (or vice versa), greatly improving the global consistency and visual realism of the reconstruction results.

[0060] The effectiveness of this invention will be verified through experiments below.

[0061] Experimental conditions: The experiment used a training set with a batch size of 8 pixels and input image patch sizes of 256×256 pixels. Each patch was randomly horizontally flipped with a probability of 0.5 for data augmentation. The optimizer used was Adam, with an initial learning rate of... Then, a cosine annealing scheduler is used to gradually reduce the learning rate to A total of 200 training cycles were conducted. All experiments were implemented using the PyTorch framework and an NVIDIA GeForce RTX4090 GPU.

[0062] Evaluation metrics: This invention uses two widely used evaluation metrics, namely Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), to evaluate the performance of the artifact removal network.

[0063] A systematic comparative experiment was conducted on the Ring artifact, Motion artifact, and Metal artifact datasets, comparing the results with six other image restoration models (Uformer, NERD, MFDNet, MPRNet, ELFormer, and FSNet). The test results are shown in Table 1.

[0064] Table 1: Comparison of model results on the ring, motion, and metal artifact datasets.

[0065] As can be seen from the table above, the artifact removal network in this invention performs best in all experimental results, demonstrating its superiority in CT artifact removal tasks.

[0066] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A CT artifact removal method based on frequency optimization and structure scale enhancement, characterized in that, The method comprises the following steps: obtaining a CT image containing artifacts to be processed; inputting the CT image containing artifacts into a trained artifact removal network for processing to obtain a CT image after removing artifacts; wherein the artifact removal network is composed of an encoder-decoder structure and a residual connection; the encoder comprises at least one encoding level for feature extraction and down-sampling of the input image; the decoder comprises at least one decoding level for feature fusion and up-sampling of the features output by the encoder; in at least part of the encoding level and at least part of the decoding level, there are sequentially connected: a frequency domain interaction module for converting the input features to the frequency domain for modulation and output; a prior aggregation module for sequence modeling of the features output by the frequency domain interaction module to capture global structure dependencies and output; a scale competition module for multi-scale competitive fusion of the features output by the prior aggregation module and output.

2. The method for CT artifact removal based on frequency optimization and structure scale enhancement according to claim 1, characterized in that, In each level of the encoder and each level of the decoder, the frequency domain interaction module, the prior aggregation module and the scale competition module are sequentially connected.

3. The method for CT artifact removal based on frequency optimization and structure scale enhancement according to claim 1, characterized in that, The frequency domain interaction module processes the input features as follows: normalizing and blocking the input features of the frequency domain interaction module to obtain a plurality of image blocks; convert each image block to the frequency domain to separate the amplitude spectrum and the phase spectrum; directionally modulating the amplitude spectrum to obtain a modulated amplitude spectrum; combining the modulated amplitude spectrum and the phase spectrum and converting back to the spatial domain to obtain the output features of the frequency domain interaction module.

4. The method for CT artifact removal based on frequency optimization and structure scale enhancement according to claim 3, characterized in that, Directionally modulating the amplitude spectrum to obtain a modulated amplitude spectrum comprises: performing horizontal and vertical band pooling on the amplitude spectrum to obtain horizontal features and vertical features, respectively; using a learnable frequency mask to weight the horizontal features and the vertical features, respectively; performing cross-frequency interaction by fusing the weighted horizontal features and the vertical features through a multilayer perceptron to obtain the modulated amplitude spectrum.

5. The method for CT artifact removal based on frequency optimization and structure scale enhancement according to claim 3, characterized in that, Combining the modulated amplitude spectrum and the phase spectrum and converting back to the spatial domain to obtain the output features of the frequency domain interaction module comprises: combining the modulated amplitude spectrum and the phase spectrum and converting back to the spatial domain to obtain spatial features corresponding to each image block; fusing the spatial features corresponding to all image blocks to obtain converted spatial features; processing and splitting the converted spatial features using depth separable convolution to obtain first and second sub-features; activating the first sub-features through a GELU activation function, element-wise multiplying the first sub-features and the second sub-features, and then performing residual connection between the multiplication result and the input features of the frequency domain interaction module to obtain the output features of the frequency domain interaction module.

6. The method for CT artifact removal based on frequency optimization and structure scale enhancement according to claim 1, characterized in that, The prior aggregation module processes the input features as follows: a cascaded 3x3 and 1x1 convolution compresses and reorganizes the input features of the prior aggregation module into a structured feature sequence; a gating cycle mechanism is used to model the feature sequence to capture long-range dependencies between elements in the sequence. The reconstructed spatial features are obtained by reconstructing the modeled feature sequence into spatial features, and the output features of the prior aggregation module are obtained.

7. The method for CT artifact removal based on frequency optimization and structure scale enhancement according to claim 1, characterized in that, The scale competition module processes the input features as follows: A plurality of parallel convolution branches are used to perform multi-scale feature extraction on the input features of the scale competition module to obtain features of multiple different scales; Competitive fusion is performed on the features of the multiple different scales to adaptively emphasize the scale features with the largest amount of information and suppress redundant features, and the fused features are taken as the output features of the scale competition module.

8. The CT artifact removal method based on frequency optimization and structural scale enhancement according to claim 7, characterized in that, The competitive fusion of the features of the multiple different scales includes: The features of the multiple different scales are spliced; Global context encoding is performed on the spliced features to generate a scale importance weight vector; Based on the scale importance weight vector, channel-level weighted fusion is performed on the features of the multiple different scales.

9. The method for CT artifact removal based on frequency optimization and structure scale enhancement according to claim 1, characterized in that, The training of the artifact removal network is performed by minimizing a total loss function, which is calculated by the following formula: where L total is the total loss function; L content is a spatial domain content loss that computes the difference between the predicted image and the real image at multiple resolutions; L fft is a frequency domain consistency loss that computes the difference between the frequency domain representation of the predicted image and the real image at multiple resolutions; λ is a weight coefficient for balancing the two losses, and its value is 0.1 to 0.

5.

10. The method of claim 9, wherein the frequency-optimized and structure-scale-enhanced CT artifact removal is based on a frequency-optimized and structure-scale-enhanced CT artifact removal method, wherein the method comprises: The spatial domain content loss and the frequency domain consistency loss are calculated by the following formulas, respectively: ​ ， ， In the formula, F represents a two-dimensional Fourier transform, and s represents different resolutions.

Images