An agent-based automatic three-dimensional medical image data labeling method and system

Abstract

The application discloses a kind of based on agent automation three-dimensional medical image data labeling method and system, comprising: obtaining original three-dimensional medical image containing artifact, and carrying out replacement physical domain conversion;Quantitative parameters associated with the physical cause of artifact are extracted in replacement physical domain;According to quantitative parameter, determine artifact type and intensity grade, the artifact type includes structure coding artifact and degenerative artifact;When it is structure coding artifact, its structured feature is extracted as enhanced signal injection segmentation network and is labeled;When it is degenerative artifact, it is labeled after artifact repair according to type and intensity grade again.This application can adaptively take differentiated processing according to artifact type, avoid the performance loss caused by single strategy, and through physical parameter accurate quantification artifact intensity, depth fusion of artifact analysis and segmentation task, significantly improve the data labeling accuracy and robustness of artifact-containing image.

Description

Technical Field

[0001] This invention relates to the field of medical image data processing, and in particular to an automated three-dimensional medical image data annotation method and system based on intelligent agents. Background Technology

[0002] Three-dimensional medical images (such as CT and MRI) are often accompanied by various artifacts during acquisition, including metal artifacts, beam hardening artifacts, truncation artifacts, motion artifacts, and geometric bias artifacts. These artifacts reduce image quality, affect subsequent organ segmentation and lesion annotation, and thus limit the training accuracy of medical imaging artificial intelligence models.

[0003] In existing technologies, methods for dealing with artifacts in medical images are mainly divided into two categories: one is artifact correction methods, which use image preprocessing techniques (such as filtering, projection domain interpolation, deconvolution, etc.) to eliminate or suppress artifacts in order to obtain a relatively "clean" image before segmentation and labeling; the other is artifact-robust segmentation methods, which add synthetic artifacts to the training data to enhance the segmentation model's tolerance to artifacts, or design uncertainty estimation modules to label unreliable regions.

[0004] However, the aforementioned prior art has the following technical problems: First, the artifact processing strategy is too simplistic. Existing methods assume all artifacts are harmful noise and apply a uniform correction or robust training strategy to all artifacts, failing to adaptively select different processing methods based on the specific type and severity of the artifacts. This leads to the erroneous removal of potentially usable geometric or phase features from some artifacts that may contain structured information, while the restoration effect of a single strategy is also unsatisfactory for artifacts that severely damage the image structure.

[0005] Second, there is a lack of effective methods for quantifying artifacts. Existing artifact detection methods mostly use image domain statistical features or deep learning black box classification, which makes it difficult to obtain quantifiable parameters that are directly related to the physical causes of artifacts. Therefore, it is impossible to accurately determine the type and intensity level of artifacts, which limits the effectiveness of subsequent differential processing.

[0006] Third, artifacts are disconnected from the segmentation task. Existing methods treat artifact correction or robust training and segmentation annotation as two separate stages. The artifact correction stage does not consider the boundary accuracy requirements of the downstream segmentation task, and the segmentation stage cannot utilize the structured information of the artifacts themselves, resulting in limited overall annotation performance.

[0007] Therefore, there is an urgent need for a three-dimensional medical image data annotation method that can identify artifact types and adaptively apply differentiated processing based on artifact types, in order to overcome the aforementioned shortcomings of existing technologies. Summary of the Invention

[0008] In view of the aforementioned deficiencies of the prior art, the technical problem to be solved by the present invention is to provide an automated three-dimensional medical image data annotation method and system based on intelligent agents, which aims to improve the data annotation accuracy and robustness of images containing artifacts by adopting differentiated processing according to the type of artifact.

[0009] To achieve the above objectives, the first aspect of this invention discloses an automated three-dimensional medical image data annotation method based on intelligent agents, the method comprising: Step S1: Obtain the original three-dimensional medical image containing artifacts, and perform a substitution physical domain transformation on the original three-dimensional medical image; wherein, the substitution physical domain includes one or more of the following: projection domain sine curve, frequency domain Fourier spectrum, multi-energy channel virtual monoenergy spectrum, and image geometry domain; Step S2: Extract at least one quantization parameter associated with the physical origin of the artifact in the alternative physical domain; wherein the quantization parameter includes at least one of the following: metal trajectory discontinuity angle for characterizing metal artifacts, dark band penetration energy index for characterizing beam hardening artifacts, ringing phase angle for characterizing truncated ringing artifacts, fuzzy kernel notch frequency vector for characterizing motion artifacts, and truncated profile asymmetry coefficient for characterizing geometric offset artifacts. Step S3: Determine the artifact type and intensity level corresponding to the artifacts in the original three-dimensional medical image according to the quantization parameters; wherein, the artifact type includes structural coding artifacts and degradation artifacts, the structural coding artifacts include at least one of bundle hardening dark band artifacts and truncated ringing artifacts, and the degradation artifacts include at least one of metal artifacts, motion artifacts and geometric offset artifacts. Step S4: In response to the artifact type being the structure-encoded artifact, the structured features of the artifact in the original 3D medical image are extracted according to the artifact type and the intensity level, and the structured features are injected into the 3D segmentation network as an enhancement signal to enhance the segmentation network's ability to identify target boundaries, thereby annotating the original 3D medical image; In response to the artifact type being the degradation artifact, the original 3D medical image is repaired according to the artifact type and the intensity level, and then annotated.

[0010] Optionally, in step S4, the structured features of artifacts in the original three-dimensional medical image are extracted based on the artifact type and the intensity level, including: Based on the intensity level, the sampling density or feature dimension of the feature extraction is adaptively adjusted; and, When the structure encoding artifact is a bundle hardening dark band artifact, the dark band region is detected, the dark band region is refined in three dimensions to obtain the skeleton, the signed distance from each voxel in the dark band to the skeleton is calculated to generate a distance field feature channel, the normalized gradient direction of each voxel in the dark band is calculated and forced to point to the adjacent high-density region to generate a gradient direction field feature channel. When the structure encoding artifact is a truncated ringing artifact, strong edges in the image are detected, an intensity profile is extracted in the direction perpendicular to the edge, local extreme points are detected, the local phase angle, amplitude and local oscillation frequency of the ringing are calculated, and phase angle feature channel, amplitude feature channel and frequency feature channel are generated. When the intensity level exceeds a preset threshold, the number of sampling points of the skeleton is increased or the local extreme points are extracted at multiple scales.

[0011] Optionally, in step S4, injecting the structured features as an enhancement signal into the 3D segmentation network includes: The structured features are encoded into at least one enhanced feature channel, which is then concatenated with the original three-dimensional medical image in the channel dimension to form an enhanced input. The enhanced input is fed into a 3D segmentation network, which has a learnable gating module after at least one scale layer of its encoder. The gating module is used to concatenate the feature map of the encoder at the current scale with the downsampled enhanced feature channel, generate gating weights through convolution and activation functions, and multiply the gating weights element-wise with the encoder feature map to adaptively adjust the contribution of the structured features to the segmentation result.

[0012] Optionally, in step S3, determining the artifact type and intensity level based on the quantization parameters includes: The quantization parameters are input into a classifier based on a physical rule base, which contains a deterministic mapping relationship between parameter value ranges and artifact types. When the discontinuity angle of the metal trajectory is greater than the first angle threshold, it is determined to be a metal artifact, and the intensity level is determined according to the ratio of the angle to the maximum angle value. When the penetration energy index of the dark band is greater than the first energy index threshold and the truncation profile asymmetry coefficient is less than the first symmetry threshold, it is determined to be a beam-hardened dark band artifact, and the intensity level is determined according to the magnitude of the energy index. When the local variance of the ringing phase angle is less than the second variance threshold, it is determined to be a truncated ringing artifact, and the intensity level is determined according to the amplitude of the phase angle. When the notch depth detected by the fuzzy kernel notch frequency vector in the frequency domain exceeds the first depth threshold, it is determined to be a motion artifact, and the intensity level is determined according to the notch depth and frequency distribution range. When the asymmetry coefficient of the truncated contour is greater than the second symmetry threshold, it is determined to be a geometric offset artifact, and the intensity level is determined according to the value of the asymmetry coefficient.

[0013] Optionally, in step S4, when the artifact type is a degenerative artifact, artifact repair is performed according to the artifact type and intensity level, including: When the artifact type is a metal artifact, different intensity projection domain interpolation repairs are selected according to the intensity level corresponding to the discontinuity angle of the metal trajectory: linear interpolation is used when the intensity level is lower than the first metal threshold, spline interpolation is used when the intensity level is between the first metal threshold and the second metal threshold, and dual-energy material decomposition correction is used when the intensity level is higher than the second metal threshold. When the artifact type is motion artifact, non-blind deconvolution repair with different iteration numbers is selected according to the intensity level corresponding to the blur kernel notch frequency vector: the lower the intensity level, the fewer the deconvolution iterations; the higher the intensity level, the more the deconvolution iterations, supplemented by multi-frame registration. When the artifact type is a geometric offset artifact, symmetrical continuation reconstruction with different continuation ranges is selected according to the intensity level corresponding to the truncation profile asymmetry coefficient: the higher the intensity level, the larger the continuation range.

[0014] The second aspect of this invention discloses an automated three-dimensional medical image data annotation system based on intelligent agents, the system comprising: a physical domain conversion module, a parameter extraction module, an artifact type and level determination module, and an annotation module; The physical domain conversion module is used to acquire the original three-dimensional medical image containing artifacts and perform a substitution physical domain conversion on the original three-dimensional medical image; wherein, the substitution physical domain includes one or more of the following: projection domain sine curve, frequency domain Fourier spectrum, multi-energy channel virtual monoenergy spectrum, and image geometry domain; The parameter extraction module is used to extract at least one quantization parameter associated with the physical cause of artifacts in the alternative physical domain; wherein the quantization parameter includes at least one of the following: a metal trajectory discontinuity angle for characterizing metal artifacts, a dark band penetration energy index for characterizing beam hardening artifacts, a ringing phase angle for characterizing truncated ringing artifacts, a fuzzy kernel notch frequency vector for characterizing motion artifacts, and a truncated profile asymmetry coefficient for characterizing geometric offset artifacts. The artifact type and intensity level determination module is used to determine the artifact type and intensity level corresponding to the artifacts in the original three-dimensional medical image according to the quantization parameters; wherein, the artifact type includes structural coding artifacts and degradation artifacts, the structural coding artifacts include at least one of bundle hardening dark band artifacts and truncated ringing artifacts, and the degradation artifacts include at least one of metallic artifacts, motion artifacts and geometric offset artifacts. The annotation module is configured to, in response to the artifact type being the structure-encoded artifact, extract the structured features of the artifact in the original 3D medical image according to the artifact type and the intensity level, and inject the structured features as an enhancement signal into the 3D segmentation network to enhance the segmentation network's ability to recognize target boundaries, thereby annotating the original 3D medical image; and in response to the artifact type being the degradation artifact, perform artifact repair on the original 3D medical image according to the artifact type and the intensity level, and then annotate it.

[0015] Optionally, the annotation module extracts the structured features of artifacts in the original 3D medical image based on the artifact type and the intensity level, specifically including: Based on the intensity level, the sampling density or feature dimension of the feature extraction is adaptively adjusted; and, When the structure encoding artifact is a bundle hardening dark band artifact, the dark band region is detected, the dark band region is refined in three dimensions to obtain the skeleton, the signed distance from each voxel in the dark band to the skeleton is calculated to generate a distance field feature channel, the normalized gradient direction of each voxel in the dark band is calculated and forced to point to the adjacent high-density region to generate a gradient direction field feature channel. When the structure encoding artifact is a truncated ringing artifact, strong edges in the image are detected, an intensity profile is extracted in the direction perpendicular to the edge, local extreme points are detected, the local phase angle, amplitude and local oscillation frequency of the ringing are calculated, and phase angle feature channel, amplitude feature channel and frequency feature channel are generated. When the intensity level exceeds a preset threshold, the number of sampling points of the skeleton is increased or the local extreme points are extracted at multiple scales.

[0016] Optionally, the annotation module injects the structured features as enhancement signals into the 3D segmentation network, specifically including: The structured features are encoded into at least one enhanced feature channel, which is then concatenated with the original three-dimensional medical image in the channel dimension to form an enhanced input. The enhanced input is fed into a 3D segmentation network, which has a learnable gating module after at least one scale layer of its encoder. The gating module is used to concatenate the feature map of the encoder at the current scale with the downsampled enhanced feature channel, generate gating weights through convolution and activation functions, and multiply the gating weights element-wise with the encoder feature map to adaptively adjust the contribution of the structured features to the segmentation result.

[0017] Optionally, the artifact type level determination module is specifically used for: The quantization parameters are input into a classifier based on a physical rule base, which contains a deterministic mapping relationship between parameter value ranges and artifact types. When the discontinuity angle of the metal trajectory is greater than the first angle threshold, it is determined to be a metal artifact, and the intensity level is determined according to the ratio of the angle to the maximum angle value. When the penetration energy index of the dark band is greater than the first energy index threshold and the truncation profile asymmetry coefficient is less than the first symmetry threshold, it is determined to be a beam-hardened dark band artifact, and the intensity level is determined according to the magnitude of the energy index. When the local variance of the ringing phase angle is less than the second variance threshold, it is determined to be a truncated ringing artifact, and the intensity level is determined according to the amplitude of the phase angle. When the notch depth detected by the fuzzy kernel notch frequency vector in the frequency domain exceeds the first depth threshold, it is determined to be a motion artifact, and the intensity level is determined according to the notch depth and frequency distribution range. When the asymmetry coefficient of the truncated contour is greater than the second symmetry threshold, it is determined to be a geometric offset artifact, and the intensity level is determined according to the value of the asymmetry coefficient.

[0018] Optionally, when the artifact type is a degenerative artifact, the annotation module performs artifact repair based on the artifact type and intensity level, specifically including: When the artifact type is a metal artifact, different intensity projection domain interpolation repairs are selected according to the intensity level corresponding to the discontinuity angle of the metal trajectory: linear interpolation is used when the intensity level is lower than the first metal threshold, spline interpolation is used when the intensity level is between the first metal threshold and the second metal threshold, and dual-energy material decomposition correction is used when the intensity level is higher than the second metal threshold. When the artifact type is motion artifact, non-blind deconvolution repair with different iteration numbers is selected according to the intensity level corresponding to the blur kernel notch frequency vector: the lower the intensity level, the fewer the deconvolution iterations; the higher the intensity level, the more the deconvolution iterations, supplemented by multi-frame registration. When the artifact type is a geometric offset artifact, symmetrical continuation reconstruction with different continuation ranges is selected according to the intensity level corresponding to the truncation profile asymmetry coefficient: the higher the intensity level, the larger the continuation range.

[0019] The beneficial effects of this invention are as follows: 1. This invention accurately identifies artifact types and their intensity levels by extracting quantitative parameters related to the physical causes of artifacts. For structure-coded artifacts such as bundle-hardened dark band artifacts and truncated ringing artifacts, this invention extracts their structured features as enhancement signals injected into the segmentation network, actively utilizing the geometric or phase information contained in the artifacts to improve boundary recognition accuracy; for degenerative artifacts such as metal artifacts, motion artifacts, and geometric offset artifacts, corresponding repair strategies are selected according to the intensity level. This differentiated processing avoids the problem of loss of beneficial information or insufficient repair caused by a single processing strategy in the prior art. 2. This invention converts the image to alternative physical domains such as projection domain sine wave, frequency domain Fourier spectrum, multi-energy channel virtual monoenergetic spectrum, and image geometric domain, extracting parameters such as the discontinuity angle of the metal trajectory, the dark band penetration energy index, the ringing phase angle, the fuzzy kernel notch frequency vector, and the truncated contour asymmetry coefficient. These parameters directly map the physical causes of artifacts, enabling precise quantification of artifact types and intensity levels. This provides a reliable basis for differentiated processing and overcomes the coarseness of existing technologies that rely on image domain statistical features or black-box classification. 3. For structure-encoded artifacts, this invention directly injects extracted structured features such as distance field, gradient direction field, and ringing phase into the 3D segmentation network. A learnable gating module adaptively adjusts the contribution weights of these features, allowing the segmentation network to actively utilize boundary information within the artifacts. For degenerative artifacts, the repaired image is directly input into the segmentation network, and the choice of repair strategy depends on the needs of the downstream segmentation task. This deep coupling mechanism significantly improves overall annotation accuracy compared to the independent processes of artifact correction and segmentation in existing technologies.

[0020] In summary, this invention can adaptively switch processing strategies based on the type of artifact in 3D medical images containing artifacts, effectively utilize the structured information of beneficial artifacts, and simultaneously repair harmful artifacts in a targeted manner, overcoming the shortcomings of traditional methods that uniformly correct all artifacts or ignore artifact information. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an automated three-dimensional medical image data annotation method based on an intelligent agent, according to a specific embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an automated three-dimensional medical image data annotation system based on an intelligent agent, provided in a specific embodiment of the present invention; Figure 3 This is a data annotation diagram provided in a specific embodiment of the present invention. Detailed Implementation

[0022] This invention discloses an automated three-dimensional medical image data annotation method and system based on intelligent agents. Those skilled in the art can refer to the content of this document and appropriately modify the technical details for implementation. It should be particularly noted that all similar substitutions and modifications are obvious to those skilled in the art and are considered to be included in this invention. The methods and applications of this invention have been described through preferred embodiments. Those skilled in the art can obviously modify or appropriately change and combine the methods and applications described herein without departing from the content, spirit, and scope of this invention to implement and apply the technology of this invention.

[0023] The applicant's research revealed that not all artifacts are detrimental to segmentation tasks. For example, fasciculation artifacts in CT images spatially conform precisely to the outer edges of high-density tissues such as bone, with the inner boundary of the fasciculation band naturally aligned geometrically with the actual bone cortex boundary. Similarly, truncated ringing artifacts (Gibbs artifacts) in MRI images have an analytical mapping relationship between their ringing phase and the sub-pixel position of the actual edge. These two types of artifacts actually encode the structured information of the actual anatomical boundaries and can be considered "structure-encoded artifacts." Conversely, metal artifacts, motion artifacts, and geometric offset artifacts severely disrupt image structure and are classified as "degenerative artifacts."

[0024] Therefore, embodiments of the present invention provide an automated three-dimensional medical image data annotation method based on intelligent agents, such as... Figure 1 As shown, the method includes: Step S1: Obtain the original 3D medical image containing artifacts, and perform a substitution physical domain transformation on the original 3D medical image.

[0025] The alternative physical domains include one or more of the following: projection domain sine curve, frequency domain Fourier spectrum, multi-energy channel virtual monoenergy spectrum, and image geometry domain.

[0026] It should be noted that in step S1, the original three-dimensional medical image containing artifacts is first acquired, such as a DICOM format image sequence from a CT or MRI machine. To facilitate the subsequent extraction of parameters related to the physical causes of artifacts, the original three-dimensional medical image needs to undergo a substitution physical domain transformation. The substitution physical domain includes one or more of the following: projection domain sine wave, frequency domain Fourier spectrum, multi-energy channel virtual single-energy spectrum, and image geometry domain. Specifically, for CT images, the image can be transformed from the spatial domain to the projection domain sine wave using Radon transform to analyze the trajectory discontinuity of metal artifacts in the sine wave; alternatively, the image can be transformed to the frequency domain using Fourier transform to detect frequency domain notches caused by motion artifacts; for CT scans with dual-energy or multi-energy spectral data, virtual single-energy spectra at different energy levels can be reconstructed to analyze the energy dependence of beam hardening artifacts; furthermore, the asymmetry of the contour can be directly analyzed in the image geometry domain to identify geometric offset artifacts. This transformation step provides a physically meaningful domain space for the subsequent extraction of quantization parameters.

[0027] Step S2: Extract at least one quantization parameter in the alternative physical domain that is associated with the physical cause of the artifact.

[0028] The quantization parameters include at least one of the following: the metal trajectory discontinuity angle for characterizing metal artifacts, the dark band penetration energy index for characterizing beam hardening artifacts, the ringing phase angle for characterizing truncated ringing artifacts, the fuzzy kernel notch frequency vector for characterizing motion artifacts, and the truncated profile asymmetry coefficient for characterizing geometric offset artifacts.

[0029] It should be noted that in step S2, at least one quantization parameter related to the physical cause of the artifact is extracted in the alternative physical domain. The quantization parameters include: a metal trajectory discontinuity angle for characterizing metal artifacts (the ratio of the angular span of the metal projection trajectory to the signal discontinuity gradient measured in the projection domain sine wave); a dark band penetration energy index for characterizing beam hardening artifacts (the second derivative or curvature feature of the curve is extracted by fitting a nonlinear curve of the voxel value of the dark band region as a function of energy in a multi-energy channel virtual monoenergy spectrum); a ringing phase angle for characterizing truncated ringing artifacts (the zero-crossing phase shift of the oscillating fringes near high-contrast edges is measured in the frequency domain Fourier spectrum or spatial domain); a blur kernel notch frequency vector for characterizing motion artifacts (the approximate vector of the motion blur kernel is reconstructed by detecting the position and direction of the energy notch in the frequency domain Fourier spectrum); and a truncated contour asymmetry coefficient for characterizing geometric offset artifacts (the degree of asymmetry of the curvature of the four quadrant edges is calculated in the image geometry domain). By extracting these quantization parameters that directly map to physical causes, we can accurately distinguish artifact types and quantify their intensity, providing a reliable basis for subsequent adaptive processing.

[0030] Step S3: Determine the artifact type and intensity level corresponding to the artifacts in the original three-dimensional medical image based on the quantization parameters.

[0031] The artifact types include structure-encoded artifacts and degradation artifacts. Structure-encoded artifacts include at least one of bundle-hardened dark band artifacts and truncated ringing artifacts. Degradation artifacts include at least one of metallic artifacts, motion artifacts, and geometric bias artifacts.

[0032] It should be noted that in step S3, the artifact type and intensity level corresponding to the artifacts in the original 3D medical image are determined based on the quantization parameters. Specifically, one or more extracted quantization parameters are input into a classifier based on a physical rule base, and the artifact type is determined by preset thresholds and logical rules. For example, when the discontinuity angle of the metal trajectory is greater than the first angle threshold, it is determined to be a metal artifact, and the intensity level is determined according to the ratio of the angle to the maximum value; when the penetration energy index of the dark band is greater than the first energy index threshold and the truncated contour asymmetry coefficient is less than the first symmetry threshold, it is determined to be a beam hardening dark band artifact, and the intensity level is determined according to the energy index value; when the local variance of the ringing phase angle is less than the second variance threshold, it is determined to be a truncated ringing artifact, and the intensity level is determined according to the phase angle amplitude; when the blur kernel notch depth exceeds the first depth threshold, it is determined to be a motion artifact; when the truncated contour asymmetry coefficient is greater than the second symmetry threshold, it is determined to be a geometric offset artifact. Therefore, this invention classifies artifacts into structure-encoded artifacts (beam-hardened dark band artifacts, truncated ringing artifacts) and degradation artifacts (metal artifacts, motion artifacts, geometric bias artifacts), and assigns a quantified intensity level to each artifact, providing a precise classification and grading basis for the differentiated processing in step S4.

[0033] In this specific embodiment, step S3, which determines the artifact type and intensity level based on quantization parameters, includes: The quantization parameters are input into a classifier based on a physical rule base, which contains a deterministic mapping relationship between the parameter value range and the artifact type. When the discontinuity angle of the metal trajectory is greater than the first angle threshold, it is determined to be a metal artifact, and the intensity level is determined according to the ratio of the angle to the maximum angle value. When the penetration energy index of the dark band is greater than the first energy index threshold and the truncation profile asymmetry coefficient is less than the first symmetry threshold, it is determined to be a beam-hardened dark band artifact, and the intensity level is determined according to the magnitude of the energy index. When the local variance of the ringing phase angle is less than the second variance threshold, it is determined to be a truncated ringing artifact, and the intensity level is determined based on the amplitude of the phase angle. When the notch depth detected by the fuzzy kernel notch frequency vector in the frequency domain exceeds the first depth threshold, it is determined to be a motion artifact, and the intensity level is determined according to the notch depth and frequency distribution range. When the asymmetry coefficient of the truncated profile is greater than the second symmetry threshold, it is determined to be a geometric offset artifact, and the intensity level is determined based on the value of the asymmetry coefficient.

[0034] It should be noted that this embodiment further defines the specific implementation method of determining the artifact type and intensity level according to the quantization parameters in step S3. The present invention uses a classifier based on a physical rule base, rather than a black-box deep learning classifier. The rule base contains a deterministic mapping relationship from the quantization parameter value range to the artifact type. The specific judgment rules are as follows: (1) When the discontinuity angle of the metal trajectory is greater than the first angle threshold (e.g., 30°), it is judged as a metal artifact, and the intensity level is determined according to the ratio of this angle to the maximum possible angle value (e.g., 180°) (e.g., ratio <0.3 is mild, 0.3-0.7 is moderate, >0.7 is severe). (2) When the penetration energy index of the dark band is greater than the first energy index threshold, and the truncation contour asymmetry coefficient is less than the first symmetry threshold (indicating that the dark band has good symmetry and is not an artifact caused by bias), it is judged as a beam-hardened dark band artifact, and the intensity level is determined according to the magnitude of the energy index. (3) When the local variance of the ringing phase angle is less than the second variance threshold (indicating that the phase is highly consistent in the local region, conforming to the ringing law rather than random noise), it is determined to be a truncated ringing artifact, and the intensity level is determined according to the amplitude of the phase angle. (4) When the notch depth detected by the fuzzy kernel notch frequency vector in the frequency domain exceeds the first depth threshold, it is determined to be a motion artifact, and the intensity level is determined according to the notch depth and frequency distribution range (e.g., the deeper the notch and the more concentrated the frequency range, the higher the intensity level). (5) When the asymmetry coefficient of the truncated profile is greater than the second symmetry threshold, it is determined to be a geometric offset artifact, and the intensity level is determined according to the value of the asymmetry coefficient. The above rules are designed based on the physical causes and geometric characteristics of artifacts, and have interpretability and determinism, and can finely distinguish the types and intensities of artifacts.

[0035] Step S4: In response to the artifact type being a structure-encoded artifact, the structured features of the artifact in the original 3D medical image are extracted according to the artifact type and intensity level. These structured features are then injected into the 3D segmentation network as enhancement signals to improve the network's ability to identify target boundaries, thereby annotating the original 3D medical image. In response to the artifact type being a degenerative artifact, the original 3D medical image is repaired according to the artifact type and intensity level before being annotated.

[0036] It should be noted that in step S4, differentiated processing and annotation operations are performed based on the artifact type determined in step S3. When the artifact type is a structure-encoded artifact (beam-hardening dark band artifact or truncated ringing artifact), the structured features of the artifact are extracted as enhancement signals and injected into the 3D segmentation network. For beam-hardening dark band artifacts, the signed distance field of the dark band region and the gradient direction field pointing to the high-density region are extracted; for truncated ringing artifacts, the local phase angle, amplitude, and oscillation frequency are extracted. These structured features are concatenated with the original image along the channel dimension and then fed into a 3D segmentation network with a learnable gating module. The gating module automatically adjusts the contribution weight of the structured features to the segmentation result, thereby enhancing the network's ability to recognize target boundaries and ultimately outputting high-quality annotations. When the artifact type is degenerative (metallic artifact, motion artifact, or geometric bias artifact), the corresponding artifact repair strategy is selected according to the artifact type and intensity level: for metallic artifacts, linear interpolation, spline interpolation, or bi-energy decomposition correction is applied sequentially according to the intensity level; for motion artifacts, the number of iterations of non-blind deconvolution or multi-frame registration is adjusted according to the intensity level; for geometric bias artifacts, symmetrical extension reconstruction is performed within a corresponding range based on the magnitude of the asymmetry coefficient. The repaired image is then labeled using the corresponding segmentation annotation model. Through the above differentiated processing, this invention can both proactively utilize boundary information in beneficial artifacts to improve segmentation accuracy and specifically repair harmful artifacts, overcoming the shortcomings of the single strategy in existing technologies.

[0037] In this specific embodiment, step S4 extracts the structured features of artifacts in the original three-dimensional medical image based on the artifact type and intensity level, including: Based on the intensity level, the sampling density or feature dimension of feature extraction is adaptively adjusted; and, When the structure encoding artifact is a bundle hardening dark band artifact, the dark band region is detected, the dark band region is refined in three dimensions to obtain the skeleton, the signed distance from each voxel in the dark band to the skeleton is calculated to generate the distance field feature channel, the normalized gradient direction of each voxel in the dark band is calculated and forced to point to the adjacent high-density region to generate the gradient direction field feature channel. When the structure encoding artifact is a truncated ringing artifact, strong edges in the image are detected, an intensity profile is extracted in the direction perpendicular to the edge, local extreme points are detected, the local phase angle, amplitude and local oscillation frequency of the ringing are calculated, and phase angle feature channel, amplitude feature channel and frequency feature channel are generated. When the intensity level exceeds a preset threshold, the number of sampling points on the skeleton is increased or local extreme points are extracted at multiple scales.

[0038] It should be noted that this embodiment further defines the specific implementation method of extracting structured features based on artifact type and intensity level in step S4. First, the sampling density or feature dimension of feature extraction is adaptively adjusted according to the intensity level: when the artifact intensity level is low, a sparser sampling density or a lower feature dimension can be used to reduce computational overhead; when the artifact intensity level exceeds a preset threshold, the number of sampling points of the skeleton is increased or local extrema are extracted at multiple scales to obtain more refined structured features. For beam hardening dark band artifacts, the specific extraction process includes: detecting the dark band region (e.g., through morphological top-hat transformation or adaptive threshold segmentation), performing three-dimensional refinement on the dark band region to obtain its centerline skeleton, calculating the signed distance from each voxel in the dark band to the skeleton, and generating a distance field feature channel; simultaneously calculating the normalized gradient direction of each voxel and forcing the gradient direction to point to the adjacent high-density region (such as bone or metal), generating a gradient direction field feature channel. For truncated ringing artifacts, the specific extraction process includes: detecting strong edges in the image (e.g., using the Canny operator or phase consistency method); extracting an intensity profile in the direction perpendicular to the edge; detecting local extrema in the profile and calculating the local phase angle, amplitude, and local oscillation frequency of the ringing; and generating three feature channels for phase angle, amplitude, and frequency, respectively. This feature extraction method directly utilizes the spatial geometric or frequency domain phase characteristics of beneficial artifacts, providing effective boundary enhancement information for subsequent segmentation networks.

[0039] In this specific embodiment, step S4 involves injecting structured features as enhancement signals into the 3D segmentation network, including: The structured features are encoded into at least one enhanced feature channel and concatenated with the original 3D medical image in the channel dimension to form an enhanced input; The enhanced input is fed into a 3D segmentation network, which has a learnable gating module after at least one scale layer of its encoder. The gating module is used to concatenate the feature map of the encoder at the current scale with the downsampled enhanced feature channel, generate gating weights through convolution and activation functions, and multiply the gating weights element-wise with the encoder feature map to adaptively adjust the contribution of structured features to the segmentation result.

[0040] It should be noted that this embodiment further defines the specific implementation method of injecting structured features as enhancement signals into the 3D segmentation network in step S4. First, the structured features extracted in step S2 (such as distance field, gradient direction field, ringing phase, etc.) are encoded into one or more enhancement feature channels, and concatenated with the original 3D medical image in the channel dimension to form an enhancement input. This enhancement input contains both the original image information and the structured prior information of artifacts. Then, this enhancement input is fed into the 3D segmentation network (such as 3D U-Net, nnFormer, etc.). In particular, a learnable gating module is set after at least one scale layer of the encoder of the 3D segmentation network. The gating module works as follows: the feature map of the current scale layer of the encoder is concatenated with the enhancement feature channels whose resolution is matched after downsampling; the concatenated feature map is passed through a 1×1×1 convolutional layer and a sigmoid activation function to generate a gating weight map with a value range of [0,1]; the gating weight is multiplied with the feature map of the current scale of the encoder channel by channel and spatial position to obtain the gated feature map, which is then passed on. In this way, the gating module can adaptively adjust the contribution weight of structured features to the segmentation result at each spatial location and each feature channel—enabling high weights in regions where structured features are reliable (such as where dark band boundaries are clear), and disabling weights in unreliable or artifact-free regions, thereby avoiding the introduction of noise interference.

[0041] In this specific embodiment, when the artifact type in step S4 is a degenerate artifact, artifact repair is performed according to the artifact type and intensity level, including: When the artifact type is metal artifact, different intensity projection domain interpolation repair is selected according to the intensity level corresponding to the discontinuity angle of the metal trajectory: linear interpolation is used when the intensity level is lower than the first metal threshold, spline interpolation is used when the intensity level is between the first metal threshold and the second metal threshold, and dual-energy material decomposition correction is used when the intensity level is higher than the second metal threshold. When the artifact type is motion artifact, non-blind deconvolution repair with different iteration numbers is selected according to the intensity level corresponding to the blur kernel notch frequency vector: the lower the intensity level, the fewer the deconvolution iterations; the higher the intensity level, the more the deconvolution iterations, supplemented by multi-frame registration. When the artifact type is geometric offset artifact, symmetrical continuation reconstruction with different continuation ranges is selected according to the intensity level corresponding to the truncation profile asymmetry coefficient: the higher the intensity level, the larger the continuation range.

[0042] It should be noted that this embodiment further defines the specific implementation method of artifact repair based on artifact type and intensity level in step S4 when the artifact type is a degenerative artifact. This invention employs differentiated repair strategies for different degenerative artifacts and their intensity levels. For metal artifacts: the intensity of projection domain interpolation repair is selected based on the intensity level corresponding to the discontinuity angle of the metal trajectory. When the intensity level is below the first metal threshold (mild artifact), linear interpolation is used to quickly fill the metal shadow in the projection domain; when the intensity level is between the first and second metal thresholds (moderate artifact), spline interpolation is used to maintain the smoothness of the projection curve; when the intensity level is above the second metal threshold (severe artifact), dual-energy material decomposition correction is used (if dual-energy CT data is available), and the metal matrix image is separated and projected forward to replace the original projection. For motion artifacts: the number of iterations for non-blind deconvolution repair is selected based on the intensity level corresponding to the fuzzy kernel notch frequency vector. Lower intensity levels (lighter motion blur) require fewer deconvolution iterations to conserve computational resources; higher intensity levels (severe motion blur) require more deconvolution iterations, supplemented by multi-frame registration, which further suppresses ghosting through registration fusion. For geometric offset artifacts: the extension range for symmetric extension reconstruction is selected based on the intensity level corresponding to the asymmetry coefficient of the truncated contour. Higher intensity levels (more severe truncation, larger asymmetry coefficient) result in a larger extension range to more fully restore the structure of the truncated region. After restoration, the restored image is output for subsequent segmentation and annotation models. This differentiated restoration strategy achieves a balance between restoration effectiveness and computational efficiency, avoiding the use of the same high-complexity or low-efficiency method for all artifacts.

[0043] In one specific application, the 3D medical image data annotation method is deployed on an agent-based collaborative annotation platform. This platform comprises an interaction layer and a data layer, a core layer of agent skill libraries, and a central coordination and control layer.

[0044] The interaction layer and data layer use the MONAI Label framework as the entry point for human-computer interaction and data management. It is responsible for receiving raw 3D medical images (such as CT / MRI sequences containing artifacts), displaying annotation results, and collecting feedback from manual reviewers.

[0045] The core layer of the agent skill library comprises a skill library registry center and several specialized skill agents. The skill library registry center manages all available agents, each declaring its specialized skills (such as "bone edge enhancement segmentation," "metal artifact projection domain repair," and "ringing phase-guided super-resolution"). Specialized skill agents are independent, microservice-based functional units that encapsulate the core algorithm modules of this invention, including: an artifact quantization parameter extraction agent (for performing the alternative physical domain transformation and parameter extraction in steps S1 to S3), a structure-encoded artifact enhancement segmentation agent (for performing structured feature extraction and injection of beam-hardened dark bands / truncated ringing artifacts in step S4), and a degradation artifact repair agent (for performing differentiated repair of metal / motion / geometric bias artifacts in step S4).

[0046] The central coordination and control layer comprises a central task scheduler, a workflow engine, and a result fusion module. The central task scheduler receives complex annotation requests from users (such as "multi-organ segmentation of a whole-abdominal CT scan containing metal artifacts") and parses them into atomic task sequences. The workflow engine dynamically schedules specialized skill agents based on the task sequence and the actual detected artifact types and intensity levels in the images, forming an adaptive workflow: for example, it first calls the artifact quantization parameter extraction agent to identify the artifact type; if identified as a bundle-hardening dark band artifact, it calls the structure-encoding artifact enhancement segmentation agent; if identified as a metal artifact, it calls the corresponding intensity projection domain repair agent based on the intensity level before segmentation. The result fusion module integrates the outputs of each agent, handles overlaps and conflicts, and generates a final, consistent annotation result.

[0047] By embedding the artifact adaptive processing method of this invention into the above-mentioned intelligent agent cooperative architecture, the system can dynamically select the optimal processing path according to the physical type and intensity of artifacts in the image.

[0048] In this specific application, the data annotation map of three-dimensional medical images can be as follows: Figure 3 As shown.

[0049] This invention accurately identifies artifact types and their intensity levels by extracting quantization parameters related to the physical causes of artifacts. For structure-coded artifacts such as bundle-hardening dark band artifacts and truncated ringing artifacts, this invention extracts their structured features as enhancement signals injected into the segmentation network, actively utilizing the geometric or phase information contained in the artifacts to improve boundary recognition accuracy. For degenerate artifacts such as metal artifacts, motion artifacts, and geometric offset artifacts, corresponding repair strategies are selected based on their intensity levels. This differentiated processing avoids the loss of useful information or insufficient repair caused by a single processing strategy in existing technologies.

[0050] This invention transforms images into alternative physical domains such as projection domain sine waves, frequency domain Fourier spectra, multi-energy channel virtual monoenergetic spectra, and image geometric domains to extract parameters such as the discontinuity angle of the metal trajectory, the penetration energy index of the dark band, the ringing phase angle, the frequency vector of the blur kernel notch, and the asymmetry coefficient of the truncated contour. These parameters directly map the physical causes of artifacts, enabling precise quantification of artifact types and intensity levels, providing a reliable decision-making basis for differentiated processing, and overcoming the coarseness of existing technologies that rely on image domain statistical features or black-box classification.

[0051] For structured encoding artifacts, this embodiment of the invention directly injects extracted structured features such as distance field, gradient direction field, and ringing phase into the 3D segmentation network, and adaptively adjusts the contribution weights of these features through a learnable gating module, enabling the segmentation network to actively utilize boundary information in the artifacts. For degradation artifacts, the repaired image is directly input into the segmentation network, and the choice of repair strategy depends on the needs of the downstream segmentation task. This deep coupling mechanism significantly improves the overall annotation accuracy compared to the existing technology where artifact correction and segmentation are independent processes.

[0052] In summary, the embodiments of the present invention can adaptively switch processing strategies according to the type of artifact in three-dimensional medical images containing artifacts, effectively utilize the structured information of beneficial artifacts, and at the same time, specifically repair harmful artifacts, overcoming the shortcomings of traditional methods that uniformly correct all artifacts or ignore artifact information.

[0053] Based on the above-mentioned agent-based automated 3D medical image data annotation method, this invention also provides an agent-based automated 3D medical image data annotation system, such as... Figure 2 As shown, the system includes: a physical domain conversion module 201, a parameter extraction module 202, an artifact type and level determination module 203, and a labeling module 204; The physical domain conversion module 201 is used to acquire the original three-dimensional medical image containing artifacts and perform a replacement physical domain conversion on the original three-dimensional medical image; wherein, the replacement physical domain includes one or more of the following: projection domain sine curve, frequency domain Fourier spectrum, multi-energy channel virtual monoenergy spectrum, and image geometry domain; The parameter extraction module 202 is used to extract at least one quantization parameter related to the physical cause of the artifact in the alternative physical domain; wherein the quantization parameter includes at least one of the following: a metal trajectory discontinuity angle for characterizing metal artifacts, a dark band penetration energy index for characterizing beam hardening artifacts, a ringing phase angle for characterizing truncated ringing artifacts, a fuzzy kernel notch frequency vector for characterizing motion artifacts, and a truncated profile asymmetry coefficient for characterizing geometric offset artifacts. The artifact type and intensity level determination module 203 is used to determine the artifact type and intensity level corresponding to the artifact in the original three-dimensional medical image based on the quantization parameters. The artifact type includes structural coding artifacts and degradation artifacts. Structural coding artifacts include at least one of bundle hardening dark band artifacts and truncated ringing artifacts. Degradation artifacts include at least one of metal artifacts, motion artifacts and geometric offset artifacts. The annotation module 204 is used to, in response to the artifact type being structure-encoded artifacts, extract the structured features of artifacts in the original 3D medical image according to the artifact type and intensity level, and inject the structured features as enhancement signals into the 3D segmentation network to enhance the segmentation network's ability to identify target boundaries, thereby annotating the original 3D medical image; in response to the artifact type being degenerative artifacts, perform artifact repair on the original 3D medical image according to the artifact type and intensity level, and then annotate it.

[0054] Optionally, the annotation module 204 extracts the structured features of artifacts in the original 3D medical image based on the artifact type and intensity level, specifically including: Based on the intensity level, the sampling density or feature dimension of feature extraction is adaptively adjusted; and, When the structure encoding artifact is a bundle hardening dark band artifact, the dark band region is detected, the dark band region is refined in three dimensions to obtain the skeleton, the signed distance from each voxel in the dark band to the skeleton is calculated to generate the distance field feature channel, the normalized gradient direction of each voxel in the dark band is calculated and forced to point to the adjacent high-density region to generate the gradient direction field feature channel. When the structure encoding artifact is a truncated ringing artifact, strong edges in the image are detected, an intensity profile is extracted in the direction perpendicular to the edge, local extreme points are detected, the local phase angle, amplitude and local oscillation frequency of the ringing are calculated, and phase angle feature channel, amplitude feature channel and frequency feature channel are generated. When the intensity level exceeds a preset threshold, the number of sampling points on the skeleton is increased or local extreme points are extracted at multiple scales.

[0055] Optionally, the annotation module 204 injects structured features as enhancement signals into the 3D segmentation network, specifically including: The structured features are encoded into at least one enhanced feature channel and concatenated with the original 3D medical image in the channel dimension to form an enhanced input; The enhanced input is fed into a 3D segmentation network, which has a learnable gating module after at least one scale layer of its encoder. The gating module is used to concatenate the feature map of the encoder at the current scale with the downsampled enhanced feature channel, generate gating weights through convolution and activation functions, and multiply the gating weights element-wise with the encoder feature map to adaptively adjust the contribution of structured features to the segmentation result.

[0056] Optional, the artifact type level determination module 203 is specifically used for: The quantization parameters are input into a classifier based on a physical rule base, which contains a deterministic mapping relationship between the parameter value range and the artifact type. When the discontinuity angle of the metal trajectory is greater than the first angle threshold, it is determined to be a metal artifact, and the intensity level is determined according to the ratio of the angle to the maximum angle value. When the penetration energy index of the dark band is greater than the first energy index threshold and the truncation profile asymmetry coefficient is less than the first symmetry threshold, it is determined to be a beam-hardened dark band artifact, and the intensity level is determined according to the magnitude of the energy index. When the local variance of the ringing phase angle is less than the second variance threshold, it is determined to be a truncated ringing artifact, and the intensity level is determined based on the amplitude of the phase angle. When the notch depth detected by the fuzzy kernel notch frequency vector in the frequency domain exceeds the first depth threshold, it is determined to be a motion artifact, and the intensity level is determined according to the notch depth and frequency distribution range. When the asymmetry coefficient of the truncated profile is greater than the second symmetry threshold, it is determined to be a geometric offset artifact, and the intensity level is determined based on the value of the asymmetry coefficient.

[0057] Optionally, when the artifact type is a degenerate artifact, the annotation module 204 performs artifact repair based on the artifact type and intensity level, specifically including: When the artifact type is metal artifact, different intensity projection domain interpolation repair is selected according to the intensity level corresponding to the discontinuity angle of the metal trajectory: linear interpolation is used when the intensity level is lower than the first metal threshold, spline interpolation is used when the intensity level is between the first metal threshold and the second metal threshold, and dual-energy material decomposition correction is used when the intensity level is higher than the second metal threshold. When the artifact type is motion artifact, non-blind deconvolution repair with different iteration numbers is selected according to the intensity level corresponding to the blur kernel notch frequency vector: the lower the intensity level, the fewer the deconvolution iterations; the higher the intensity level, the more the deconvolution iterations, supplemented by multi-frame registration. When the artifact type is geometric offset artifact, symmetrical continuation reconstruction with different continuation ranges is selected according to the intensity level corresponding to the truncation profile asymmetry coefficient: the higher the intensity level, the larger the continuation range.

[0058] In summary, the embodiments of the present invention can adaptively switch processing strategies according to the type of artifact in three-dimensional medical images containing artifacts, effectively utilize the structured information of beneficial artifacts, and at the same time, specifically repair harmful artifacts, overcoming the shortcomings of traditional methods that uniformly correct all artifacts or ignore artifact information.

[0059] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0060] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0061] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for automated three-dimensional medical image data annotation based on intelligent agents, characterized in that, The method includes: Step S1: Obtain the original three-dimensional medical image containing artifacts, and perform a substitution physical domain transformation on the original three-dimensional medical image; wherein, the substitution physical domain includes one or more of the following: projection domain sine curve, frequency domain Fourier spectrum, multi-energy channel virtual monoenergy spectrum, and image geometry domain; Step S2: Extract at least one quantization parameter associated with the physical origin of the artifact in the alternative physical domain; wherein the quantization parameter includes at least one of the following: metal trajectory discontinuity angle for characterizing metal artifacts, dark band penetration energy index for characterizing beam hardening artifacts, ringing phase angle for characterizing truncated ringing artifacts, fuzzy kernel notch frequency vector for characterizing motion artifacts, and truncated profile asymmetry coefficient for characterizing geometric offset artifacts. Step S3: Determine the artifact type and intensity level corresponding to the artifacts in the original three-dimensional medical image according to the quantization parameters; wherein, the artifact type includes structural coding artifacts and degradation artifacts, the structural coding artifacts include at least one of bundle hardening dark band artifacts and truncated ringing artifacts, and the degradation artifacts include at least one of metal artifacts, motion artifacts and geometric offset artifacts. Step S4: In response to the artifact type being the structure-encoded artifact, the structured features of the artifact in the original 3D medical image are extracted according to the artifact type and the intensity level, and the structured features are injected into the 3D segmentation network as an enhancement signal to enhance the segmentation network's ability to identify target boundaries, thereby annotating the original 3D medical image; In response to the artifact type being the degradation artifact, the original 3D medical image is repaired according to the artifact type and the intensity level, and then annotated.

2. The method for automated three-dimensional medical image data annotation based on intelligent agents according to claim 1, characterized in that, In step S4, based on the artifact type and intensity level, the structured features of the artifacts in the original three-dimensional medical image are extracted, including: Based on the intensity level, the sampling density or feature dimension of the feature extraction is adaptively adjusted; and, When the structure encoding artifact is a bundle hardening dark band artifact, the dark band region is detected, the dark band region is refined in three dimensions to obtain the skeleton, the signed distance from each voxel in the dark band to the skeleton is calculated to generate a distance field feature channel, the normalized gradient direction of each voxel in the dark band is calculated and forced to point to the adjacent high-density region to generate a gradient direction field feature channel. When the structure encoding artifact is a truncated ringing artifact, strong edges in the image are detected, an intensity profile is extracted in the direction perpendicular to the edge, local extreme points are detected, the local phase angle, amplitude and local oscillation frequency of the ringing are calculated, and phase angle feature channel, amplitude feature channel and frequency feature channel are generated. When the intensity level exceeds a preset threshold, the number of sampling points of the skeleton is increased or the local extreme points are extracted at multiple scales.

3. The method for automated three-dimensional medical image data annotation based on intelligent agents according to claim 1, characterized in that, Step S4, in which the structured features are injected as enhancement signals into the 3D segmentation network, includes: The structured features are encoded into at least one enhanced feature channel, which is then concatenated with the original three-dimensional medical image in the channel dimension to form an enhanced input. The enhanced input is fed into a 3D segmentation network, which has a learnable gating module after at least one scale layer of its encoder. The gating module is used to concatenate the feature map of the encoder at the current scale with the downsampled enhanced feature channel, generate gating weights through convolution and activation functions, and multiply the gating weights element-wise with the encoder feature map to adaptively adjust the contribution of the structured features to the segmentation result.

4. The method for automated three-dimensional medical image data annotation based on intelligent agents according to claim 1, characterized in that, Step S3, which determines the artifact type and intensity level based on the quantization parameters, includes: The quantization parameters are input into a classifier based on a physical rule base, which contains a deterministic mapping relationship between parameter value ranges and artifact types. When the discontinuity angle of the metal trajectory is greater than the first angle threshold, it is determined to be a metal artifact, and the intensity level is determined according to the ratio of the angle to the maximum angle value. When the penetration energy index of the dark band is greater than the first energy index threshold and the truncation profile asymmetry coefficient is less than the first symmetry threshold, it is determined to be a beam-hardened dark band artifact, and the intensity level is determined according to the magnitude of the energy index. When the local variance of the ringing phase angle is less than the second variance threshold, it is determined to be a truncated ringing artifact, and the intensity level is determined according to the amplitude of the phase angle. When the notch depth detected by the fuzzy kernel notch frequency vector in the frequency domain exceeds the first depth threshold, it is determined to be a motion artifact, and the intensity level is determined according to the notch depth and frequency distribution range. When the asymmetry coefficient of the truncated contour is greater than the second symmetry threshold, it is determined to be a geometric offset artifact, and the intensity level is determined according to the value of the asymmetry coefficient.

5. The method for automated three-dimensional medical image data annotation based on intelligent agents according to claim 1, characterized in that, In step S4, when the artifact type is a degenerate artifact, artifact repair is performed according to the artifact type and intensity level, including: When the artifact type is a metal artifact, different intensity projection domain interpolation repairs are selected according to the intensity level corresponding to the discontinuity angle of the metal trajectory: linear interpolation is used when the intensity level is lower than the first metal threshold, spline interpolation is used when the intensity level is between the first metal threshold and the second metal threshold, and dual-energy material decomposition correction is used when the intensity level is higher than the second metal threshold. When the artifact type is motion artifact, non-blind deconvolution repair with different iteration numbers is selected according to the intensity level corresponding to the blur kernel notch frequency vector: the lower the intensity level, the fewer the deconvolution iterations; the higher the intensity level, the more the deconvolution iterations, supplemented by multi-frame registration. When the artifact type is a geometric offset artifact, symmetrical continuation reconstruction with different continuation ranges is selected according to the intensity level corresponding to the truncation profile asymmetry coefficient: the higher the intensity level, the larger the continuation range.

6. An automated three-dimensional medical image data annotation system based on intelligent agents, characterized in that, The system includes: a physical domain conversion module, a parameter extraction module, an artifact type and level determination module, and a labeling module; The physical domain conversion module is used to acquire the original three-dimensional medical image containing artifacts and perform a substitution physical domain conversion on the original three-dimensional medical image; wherein, the substitution physical domain includes one or more of the following: projection domain sine curve, frequency domain Fourier spectrum, multi-energy channel virtual monoenergy spectrum, and image geometry domain; The parameter extraction module is used to extract at least one quantization parameter associated with the physical cause of artifacts in the alternative physical domain; wherein the quantization parameter includes at least one of the following: a metal trajectory discontinuity angle for characterizing metal artifacts, a dark band penetration energy index for characterizing beam hardening artifacts, a ringing phase angle for characterizing truncated ringing artifacts, a fuzzy kernel notch frequency vector for characterizing motion artifacts, and a truncated profile asymmetry coefficient for characterizing geometric offset artifacts. The artifact type and intensity level determination module is used to determine the artifact type and intensity level corresponding to the artifacts in the original three-dimensional medical image according to the quantization parameters; wherein, the artifact type includes structural coding artifacts and degradation artifacts, the structural coding artifacts include at least one of bundle hardening dark band artifacts and truncated ringing artifacts, and the degradation artifacts include at least one of metallic artifacts, motion artifacts and geometric offset artifacts. The annotation module is configured to, in response to the artifact type being the structure-encoded artifact, extract the structured features of the artifact in the original 3D medical image according to the artifact type and the intensity level, and inject the structured features as an enhancement signal into the 3D segmentation network to enhance the segmentation network's ability to recognize target boundaries, thereby annotating the original 3D medical image; and in response to the artifact type being the degradation artifact, perform artifact repair on the original 3D medical image according to the artifact type and the intensity level, and then annotate it.

7. The agent-based automated three-dimensional medical image data annotation system according to claim 6, characterized in that, The annotation module extracts the structured features of artifacts in the original 3D medical image based on the artifact type and intensity level, specifically including: Based on the intensity level, the sampling density or feature dimension of the feature extraction is adaptively adjusted; and, When the structure encoding artifact is a bundle hardening dark band artifact, the dark band region is detected, the dark band region is refined in three dimensions to obtain the skeleton, the signed distance from each voxel in the dark band to the skeleton is calculated to generate a distance field feature channel, the normalized gradient direction of each voxel in the dark band is calculated and forced to point to the adjacent high-density region to generate a gradient direction field feature channel. When the structure encoding artifact is a truncated ringing artifact, strong edges in the image are detected, an intensity profile is extracted in the direction perpendicular to the edge, local extreme points are detected, the local phase angle, amplitude and local oscillation frequency of the ringing are calculated, and phase angle feature channel, amplitude feature channel and frequency feature channel are generated. When the intensity level exceeds a preset threshold, the number of sampling points of the skeleton is increased or the local extreme points are extracted at multiple scales.

8. The agent-based automated three-dimensional medical image data annotation system according to claim 6, characterized in that, The annotation module injects the structured features as enhancement signals into the 3D segmentation network, specifically including: The structured features are encoded into at least one enhanced feature channel, which is then concatenated with the original three-dimensional medical image in the channel dimension to form an enhanced input. The enhanced input is fed into a 3D segmentation network, which has a learnable gating module after at least one scale layer of its encoder. The gating module is used to concatenate the feature map of the encoder at the current scale with the downsampled enhanced feature channel, generate gating weights through convolution and activation functions, and multiply the gating weights element-wise with the encoder feature map to adaptively adjust the contribution of the structured features to the segmentation result.

9. The intelligent agent-based automated three-dimensional medical image data annotation system according to claim 6, characterized in that, The artifact type level determination module is specifically used for: The quantization parameters are input into a classifier based on a physical rule base, which contains a deterministic mapping relationship between parameter value ranges and artifact types. When the discontinuity angle of the metal trajectory is greater than the first angle threshold, it is determined to be a metal artifact, and the intensity level is determined according to the ratio of the angle to the maximum angle value. When the penetration energy index of the dark band is greater than the first energy index threshold and the truncation profile asymmetry coefficient is less than the first symmetry threshold, it is determined to be a beam-hardened dark band artifact, and the intensity level is determined according to the magnitude of the energy index. When the local variance of the ringing phase angle is less than the second variance threshold, it is determined to be a truncated ringing artifact, and the intensity level is determined according to the amplitude of the phase angle. When the notch depth detected by the fuzzy kernel notch frequency vector in the frequency domain exceeds the first depth threshold, it is determined to be a motion artifact, and the intensity level is determined according to the notch depth and frequency distribution range. When the asymmetry coefficient of the truncated contour is greater than the second symmetry threshold, it is determined to be a geometric offset artifact, and the intensity level is determined according to the value of the asymmetry coefficient.

10. The agent-based automated three-dimensional medical image data annotation system according to claim 6, characterized in that, When the artifact type is a degenerate artifact, the annotation module performs artifact repair based on the artifact type and intensity level, specifically including: When the artifact type is a metal artifact, different intensity projection domain interpolation repairs are selected according to the intensity level corresponding to the discontinuity angle of the metal trajectory: linear interpolation is used when the intensity level is lower than the first metal threshold, spline interpolation is used when the intensity level is between the first metal threshold and the second metal threshold, and dual-energy material decomposition correction is used when the intensity level is higher than the second metal threshold. When the artifact type is motion artifact, non-blind deconvolution repair with different iteration numbers is selected according to the intensity level corresponding to the blur kernel notch frequency vector: the lower the intensity level, the fewer the deconvolution iterations; the higher the intensity level, the more the deconvolution iterations, supplemented by multi-frame registration. When the artifact type is a geometric offset artifact, symmetrical continuation reconstruction with different continuation ranges is selected according to the intensity level corresponding to the truncation profile asymmetry coefficient: the higher the intensity level, the larger the continuation range.

Images