CBCT metal artifact removal method and system based on projection domain metal identification

Abstract

The application discloses a CBCT metal artifact removal method and system based on a projection domain metal recognition, which skips a traditional image domain segmentation and orthographic projection process, directly extracts key geometric features of a metal region in original projection images, and realizes high-precision recognition and spatial posture estimation of metal objects by matching with a pre-established metal model library, so that the processing flow is significantly simplified, and the recognition stability and artifact repair effect are improved. In view of the shielding and overlapping problems that may occur in a multi-metal structure, a step-by-step stripping recognition mechanism is proposed, and through round-by-round recognition, fitting and image updating, multi-target interference is effectively avoided, and the accuracy and integrity of model matching are ensured.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for removing metal artifacts in CBCT based on projection domain metal recognition. Background Technology

[0002] Currently, artifacts caused by high-density metal implants in cone-beam computed tomography (CBCT) images have become a key challenge in clinical image quality control. Traditional metal artifact removal methods generally rely on image-domain metal segmentation and mask reprojection. For example, CN103186889B proposes a metal artifact correction method based on image-domain metal detection and projection interpolation, and US9202296B2 further optimizes the projection interpolation strategy to reduce edge blurring. These methods typically involve multiple steps, including image-domain metal segmentation, projection-domain mask interpolation, and image reconstruction. While they possess a certain degree of artifact suppression capability, they suffer from problems such as error accumulation, computational complexity, and strong process dependency due to their reliance on reconstructed images for metal localization and post-processing.

[0003] In recent years, some studies have attempted to perform artifact restoration directly in the projection domain, bypassing image domain processing steps to reduce overall complexity. For example, CN103617598A proposes a metal artifact suppression method based on projection trajectory interpolation, which identifies high-density anomalous regions at the projection level and then performs region reestimation, thereby mitigating image artifact diffusion. US patent US11786193B2 designs a multi-stage optimization mechanism at the system architecture level, combining adaptive projection restoration with scanning parameters and local compensation strategies to handle metal artifacts in complex scenes. While these methods simplify some processes, they still rely on image domain metal initial localization or template matching modules, limiting their application capabilities in fully automated processing scenarios.

[0004] In summary, while existing technologies have alleviated the problem of metal artifacts in CBCT to some extent, they are limited by relying on reconstructed images for processing.

[0005] Most existing metal artifact removal methods rely on identifying metal regions in the image domain and then generating a mask for orthographic projection. This approach is highly dependent on the initial image quality and accurate segmentation of the metal region. However, artifacts themselves often interfere with image content, leading to blurred or even misidentified metal region boundaries and causing deviations in the initial mask. Since the mask is the basis for subsequent interpolation and reconstruction operations, any error in the initial identification will be amplified throughout the process, ultimately affecting image quality. This image domain-based processing also suffers from high subjectivity and low automation, making it difficult to adapt to the complex clinical scenarios where there are many types of metals with uncertain locations.

[0006] Secondly, most technical approaches employ a phased processing mechanism of "reconstruction—mask—orthographic projection—interpolation—reconstruction," with significant coupling between each stage. Projection domain interpolation operations are often based on previous reconstruction results, while the new reconstructed image, in turn, affects the next round of mask updates and artifact removal. This iterative processing chain makes the overall system bloated and complex, with data frequently transferred between multiple rounds of projection and reconstruction, increasing the risk of error accumulation and artifact persistence. Furthermore, the need to configure parameters and execution processes separately for each stage leads to high engineering implementation difficulty, low system efficiency, and high demands on computational resources and manual intervention in actual deployment. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a scheme that directly identifies and interpolates metal regions in the original projection domain, avoiding image domain dependence and reducing data conversion and reconstruction errors. It aims to achieve a metal artifact correction method with a simpler process, lower errors, and better image quality, and is particularly suitable for intraoperative CBCT and automated radiotherapy pathway planning scenarios.

[0008] To achieve the above objectives, this invention provides a CBCT metal artifact removal method based on projection domain metal recognition, comprising the following steps: S1: Acquire the original projection image and the angle information corresponding to each projection image, and simultaneously load model files from a metal model library; S2: Analyze the original projection image to identify whether there is metal stacking, and simultaneously mark the two orthogonal images with the most significant features. If metal stacking exists, proceed to step S3; otherwise, use a feature matching algorithm to directly identify the metal model most similar to the current projection image in the metal model library, and proceed to step S4; S3: Identify and match the metal models one by one. First, select the model with the highest matching degree, perform forward projection simulation on the corresponding fitting result, and peel off the corresponding projection area from the original projection image. The model matching and recognition process continues in the updated residual map until all metal models are correctly identified, and then proceeds to step S4; S4: Based on the identified metal models, further spatial pose estimation and position matching processes are performed to determine the three-dimensional spatial position and rotation direction of each metal model in the scanning scene; S5: The simulated projection images generated by each metal model are compared with the original projection images to detect and locate artifact occlusion areas. Then, for the grayscale distortion areas in the original projection image caused by metal occlusion, image content interpolation restoration processing is performed; S6: The interpolated and repaired projection image is input into the FDK reconstruction algorithm to complete the three-dimensional image reconstruction. Then, all identified metal models are located and superimposed on the reconstructed image to generate the final CBCT image.

[0009] Further, in step S2, the presence of stacked metal objects is identified through the following steps: S21: Obtain the original projection image sequence and construct a limited set of analysis images by setting an angle sampling strategy; S22: For each projection image at the sampling angle, perform metal region identification and feature extraction through grayscale threshold segmentation; S23: Perform statistical analysis on the feature sets extracted from all sampling angles, and determine the number of metal regions at certain angles. The number of high grayscale peaks in a single metal region. If so, it is determined that there are multiple metal projection areas or stacked structures.

[0010] Furthermore, in step S22, during the process of metal region identification and feature extraction, the two orthogonal images with the most significant features are marked based on the metal region contour, the angle difference between the two projected images, the change of the principal axis direction angle, and the redundancy of metal information.

[0011] Furthermore, step S3 specifically includes:

[0012] S31: Initialize the matching candidate set according to the following steps:

[0013] S311: Perform region detection on typical angle projection maps to extract a complete set of metal features.

[0014] ,

[0015] in, For the area of ​​the metallic region, The center of the metal region, The principal axis direction angle of the metal region. The number of high grayscale peaks in a single metal region;

[0016] S312: In the metal model library In the process, generate a simulated projection of each model at the current angle:

[0017]

[0018] S313: For each model Calculate its matching degree with the current image:

[0019]

[0020] S314: Match score Sort them from highest to lowest to form a candidate matching sequence. ;

[0021] S32: Perform initial matching and metal model stripping according to the following steps:

[0022] S321: Select the model with the highest score from the sorted sequence. The optimal spatial transformation parameters are obtained using attitude estimation methods. ;

[0023] S322: Perform forward projection to obtain the model in the corresponding pose. Simulated projection image:

[0024]

[0025] S323: Subtract the original projected image to remove the contribution of the metal model, obtaining the stripped image:

[0026]

[0027] in, Represents the original projected image;

[0028] S33: Verify the matching reliability. If the verification fails, remove the current model from the candidate set and re-match the next candidate model. If the verification passes, record the model number and pose. And proceed to the next round of processing;

[0029] S34: Extract the currently stripped image As new input, repeat steps S31-S33, updating the residual map after each round of stripping, until the number of residual metal regions is 0 or all models are matched.

[0030] Furthermore, step S3 also includes: S35: After all metal recognition is completed, the simulated projection of each model and the original image are checked for total error superposition; it is determined whether there is any physical unreasonable overlap in the spatial position between all models; if there are conflicts or gaps, the pose of each model is fine-tuned by introducing a global optimization algorithm to ensure consistency.

[0031] Further, in step S33, the reasonableness of the changes before and after stripping is evaluated by at least one of the following indicators: the concentration of the total energy change of the regional residual in the target region, the area of ​​abnormal gray values ​​in the residual image, and the degree of overlap between the metal region boundary extracted in the original projection image and the corresponding region boundary in the simulated projection image.

[0032] Furthermore, in step S4, for metal models that have not overlapped, the pose is calculated directly using typical angle projection images; for overlapping models that have been stripped, the pose information is obtained according to the following steps:

[0033] S41: Define the optimization objective function:

[0034]

[0035] in, For the first The true metal projection profile at each sampling angle To position the metal model in space ,attitude The projected contour is obtained by projecting the image onto the ground. Total number of projection angles A function representing the difference between projected images;

[0036] S42: Optimize the attitude parameter space search by minimizing the error function through group cooperative search. From multiple key angles, feature information of the metal region in the projected image is extracted and compared with the standard projection features of the corresponding model in the metal model library. The minimum error matching or attitude backpropagation method based on regression network is used to realize the spatial positioning and attitude estimation of the target metal model in the current scanning scene.

[0037] S43: Perform matching verification and result confirmation.

[0038] Further, step S43 specifically includes: S431: Projecting the calculated metal model in the forward orientation onto a set of auxiliary angles to obtain multiple simulated metal projection images; S432: Performing residual analysis and structural similarity comparison on corresponding positions in the simulated metal projection images and the original projection images; S433: Based on the degree of matching between the model projection contour and the boundary of the original image, the magnitude of grayscale error in the local area, and the randomness and structural distribution of the residual image, comprehensively evaluating the matching accuracy of the current model posture; S434: If the error exceeds a preset threshold or the local structure is inconsistent, triggering a matching failure backoff mechanism, replacing the angle image, and iterating optimization again; S435: After successful verification, recording the final posture parameters of the metal model.

[0039] Furthermore, in step S5, the three-dimensional geometric representation of the metal model and the corresponding spatial transformation parameters are used, combined with the scanning geometric parameters of the CBCT system and the scanning pose at each projection angle, to generate simulated orthogonal projection images of the metal model at each projection angle, for use in occlusion area localization and interpolation repair.

[0040] Further, in step S5, image content interpolation restoration processing is performed according to the following steps:

[0041] S51: Based on the identified set of metal regions and the metal mask generated by the model pose projection, determine the pixel regions in each projected image that are occluded by metal. ,in This represents the original projected image at the i-th sampling angle;

[0042] S52: For each occluded area Extracting the set of boundary pixels based on the boundary contour It extracts the image grayscale values ​​and gradient directions within a certain range around the boundary as boundary conditions for interpolation. Based on satisfying these boundary conditions, it then... Internal pixels are reconstructed to grayscale, and then image inpainting methods are used to achieve continuous gradient filling by solving the following equation:

[0043]

[0044] in, Represents the boundary gradient vector field. The gray level is known for the boundary.

[0045] S53: Joint grayscale estimation is performed on the corresponding regions in adjacent unobstructed projection images under multiple angles to achieve cross-angle consistent interpolation, and then edge transition processing is performed on the interpolated image regions.

[0046] The technical solution of this invention also provides a CBCT metal artifact removal system based on projection domain metal recognition, which includes the following modules: a data acquisition module: acquiring the original projection image and the angle information corresponding to each projection image, and simultaneously loading model files from a metal model library; a projection image analysis module: identifying whether there is metal stacking by analyzing the original projection image, and simultaneously marking the two orthogonal images with the most significant features; a metal model recognition module: in the absence of metal stacking, identifying the metal model most similar to the current projection image in the metal model library through a feature matching algorithm, and in the presence of metal stacking, recognizing and matching the metal models one by one, first selecting the model with the highest matching degree, performing forward projection simulation on the corresponding fitting result, and then peeling the corresponding model from the original projection image. The projection area is then used to continue the model matching and recognition process in the updated residual map until all metal models are correctly identified. The spatial pose estimation module performs spatial pose estimation and position matching based on the identified metal models to determine the 3D spatial position and rotation direction of each metal model in the scanning scene. The image interpolation module compares the simulated projection images generated by each metal model with the original projection image, detects and locates artifact occlusion areas, and performs image content interpolation restoration for areas of grayscale distortion caused by metal occlusion in the original projection image. The 3D reconstruction module inputs the interpolated and repaired projection image into the FDK reconstruction algorithm to complete 3D image reconstruction, and then overlays the locations of all identified metal models onto the reconstructed image to generate the final CBCT image. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the CBCT metal artifact removal method based on a model library for projection domain metal recognition according to the present invention;

[0049] Figure 2 This is a 3D schematic diagram of the implantation needle in the metal model library of this invention;

[0050] Figure 3 This is a schematic diagram illustrating the determination of the model's spatial position and orientation based on the original projected image in an embodiment of the present invention;

[0051] Figure 4 This is the projected image after repairing the original projected image based on the model's spatial position and orientation in this embodiment of the invention;

[0052] Figure 5A and Figure 5B These are comparison images of image quality between direct reconstruction and the method of this invention. Detailed Implementation

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] This invention addresses the common problem of metal artifacts in CBCT images by proposing a method for correcting metal artifacts that combines metal model identification with projection domain interpolation. This method is implemented through the following specific technical means:

[0055] During CBCT scanning, the system simultaneously acquires the scanning angle information corresponding to each projected image and extracts the original projected image. In some application scenarios, the shape of the metal model is relatively simple, and its type and quantity are known or controllable. By analyzing projection images from all or part of the angles, the system prioritizes selecting two projection images containing the metal model with the most obvious features and a large angle difference (e.g., images differing by approximately 90°) to improve recognition efficiency and pose calculation accuracy. Edge detection and local extremum analysis algorithms are used to identify metal regions in the projected images, and key feature parameters such as region boundaries, area, principal axis direction angle, and local grayscale peaks are further extracted to construct a geometric description of the metal projection.

[0056] While acquiring projection data, the system loads a pre-established metal model database. Then, a feature matching algorithm identifies the metal model most similar to the current projected image from the database. Based on the extracted multi-angle geometric features and projection position information, rigid body transformation and attitude fitting algorithms are used to achieve accurate position and attitude estimation of the target metal model in the 3D image space. If multiple metal models overlap or stack, the system introduces a stepwise identification and stripping mechanism: first, the model with the highest matching degree is selected, its fitting result is simulated by forward projection, and its corresponding projection region is stripped from the original projected image. Then, the model matching and identification process continues in the updated residual image until all metal structures are correctly identified and their attitudes estimated.

[0057] After the spatial position and orientation of the model are determined, the simulated projection images generated by each metal model are compared with the original projection images to detect and locate artifact occlusion areas. For areas occluded by metal, methods such as directional weighted interpolation and edge-preserving interpolation are used to repair the projection data, ensuring natural grayscale transitions and structural continuity.

[0058] Finally, the interpolated and repaired complete set of projection images is input into the FDK reconstruction algorithm to complete the 3D image reconstruction. The system then locates and superimposes all identified metal models onto the reconstructed image, generating a complete, structurally accurate CBCT image with significantly suppressed metal artifacts. Through the above structured process, this invention not only significantly improves the image quality around the metal region but also effectively reduces reliance on manual identification and reconstruction errors, possessing good versatility and clinical adaptability, and is particularly suitable for patient image optimization and planning in afterloading radiotherapy.

[0059] Specifically, see Figure 1 The present invention mainly includes the following steps:

[0060] 1. Obtain the original projected image and the angles corresponding to each projected image, and simultaneously load the model files from the model library;

[0061] 2. By analyzing the original projected images, identify whether the metal objects are stacked, and simultaneously mark the two orthogonal images with the most significant features. The specific implementation process is as follows:

[0062] 1) Obtain raw projection data and set angle sampling strategy

[0063] The input is a sequence of projected images (e.g., 380 frames). The angle information for each frame is obtained through an acquisition system or geometric calibration. To avoid computational redundancy caused by analyzing all images one by one, the system sets an angle sampling step size Δθ (e.g., 6°~10°), performing "interval sampling + half-cycle traversal" within the range of 0°~180° to construct a finite set of images for analysis. This data is then stored as... tuples of the form For angle, This is the corresponding two-dimensional projection image. Simple noise filtering (such as Gaussian blur) preprocessing can be performed to enhance the metal edges.

[0064] 2) Metal region identification and feature extraction

[0065] For each projection image at the above sampling angle First, grayscale thresholding is needed to identify potential metal regions. Then, the number of metal regions after thresholding is extracted. ,area Boundary shape. Based on this, the number of high grayscale peaks in each region is detected. And calculate the position of the center of gravity. principal axis direction angle Geometric features. The above results are combined to form the geometric features of the metal at this angle:

[0066] 3) Determine if there is metal stacking.

[0067] The system is for all Statistical analysis is performed on the extracted feature set: if there are features at certain angles... or a single area If so, it can be determined that the projection has multiple metal projection areas or a stacked structure;

[0068] At this point, the "iterative peeling and recognition mode" is triggered to identify and fit the metal model one by one; otherwise, it is assumed that the metal structure in the image is clear and non-overlapping, and the simplified attitude calculation process can be directly entered.

[0069] 4) Simultaneously label the two orthogonal images with the most significant features.

[0070] During the process of metal region identification and feature extraction, the system will also select the two most representative images. , If the system satisfies the following conditions: the metal region has a complete outline and a clear shape; the angle difference between the two images is as close to 90° as possible (or other set angle difference); the principal axis direction angle changes significantly; and the redundancy of metal information is low, then the system automatically labels these two orthogonal images while determining whether stacking exists, for subsequent use in metal pose estimation.

[0071] 3. When multiple metal models are stacked in the projected image, the metal models are identified and matched one by one for stripping. The specific implementation process is as follows:

[0072] 1) Initialize the matching candidate set

[0073] 1.1) Projection diagrams of typical angles (such as two orthogonal diagrams) , Perform region detection and extract a complete set of metal features:

[0074]

[0075] 1.2) In the metal model library In the process, generate a simulated projection of each model at the current angle (or store its standard features):

[0076]

[0077] 1.3) For each model Calculate its matching degree with the current image:

[0078]

[0079] 1.4) Match score Sort them from highest to lowest to form a candidate matching sequence.

[0080] 2) Initial matching and metal model stripping

[0081] 2.1) Select the model with the highest score from the ranked sequence. The optimal spatial transformation parameters are obtained using attitude estimation methods (such as ICP + simulated projection comparison). .

[0082] 2.2) Perform forward projection to obtain the simulated projected image of the model in this pose:

[0083]

[0084] 2.3) Subtract the original projection from the original projection to remove the contribution of the metal model:

[0085]

[0086] 3) Verify matching reliability

[0087] 3.1. The degree of concentration of total residual energy change in the target region, the area of ​​abnormal grayscale values ​​in the residual image, and the overlap matching degree between the metal region boundary extracted from the original projected image and the corresponding region boundary in the model's forward projected image are used to determine the following:

[0088] 3.2. If validation fails (e.g., model edges remain after stripping), remove the current model from the candidate set and re-match the next candidate model. .

[0089] 3.3. If the verification passes, record the model number and pose. Then proceed to the next round of processing.

[0090] 4) Iterative processing of the remaining model

[0091] The currently stripped image As new input, repeat the process of "initializing the candidate matching set → initial matching and metal model stripping → verifying matching reliability". Update the residual map after each stripping round until the number of residual metal regions is 0 or all models are matched.

[0092] 5) Global consistency check

[0093] After all metals are identified, the simulated projection of each model is superimposed with the original image to verify the total error; it is determined whether there is any physical overlap in the spatial position between all models; if there are conflicts or gaps, a global optimization algorithm can be introduced to fine-tune the pose of each model to ensure consistency.

[0094] 4. Based on the identified metal models, the system further performs spatial pose estimation and position matching to determine the 3D spatial position and rotation direction of each metal model in the scanned scene. This step is applicable to two types of models: one is non-overlapping metal models, whose pose can be directly calculated from typical angle projection maps; the other is overlapping models that have been stripped, whose feature information is extracted from the stripping residual map. The specific implementation process is as follows:

[0095] 1) Define the optimization objective function

[0096] set up For the first Real metal projection contours (binary images or boundary point sets) from various angles; : Place the model in space ,attitude The projected outline obtained by projecting the image onto the ground; : Total number of projection angles; redefine the error function as:

[0097]

[0098] in, The difference measurement function between projected images is specifically quantified by the intersection-over-union ratio (IoU) of the contour regions, the Hausdorff distance, the matching area error, the centroid deviation of the contour, and the weighted sum of the principal axis direction angle differences.

[0099] 2) Optimize attitude parameter space search

[0100] Since this problem is a non-convex multivariable optimization problem, each particle represents a combination of attitude parameters. Minimize the error function through group collaborative search Furthermore, based on the extracted information such as the centroid position, contour shape, and principal axis direction angle of the metal region in the projected images at multiple angles, the deflection angle of the metal model in space (including the rotation angle around the principal axis) and its three-dimensional position coordinates in the scanning coordinate system are calculated. In practical implementation, this can be achieved by analyzing the image features of the metal region in the projected images at multiple angles. By comparing the model with the standard projection features of each metal model in the database, the pose estimation and model localization are performed using minimum error matching or a deep learning-based regression network method, thereby accurately determining the spatial position and pose of the metal model in the scanning scene.

[0101] 3) Matching Verification and Result Confirmation: To improve matching reliability, the system needs to perform a matching verification step after completing the initial attitude estimation. The specific process is as follows:

[0102] 3.1) The metal model under the calculated attitude Multiple simulated metal projection images are obtained by projecting them forward onto a set of auxiliary angles;

[0103] 3.2) Perform residual analysis and structural similarity comparison on corresponding positions in the simulated projection image and the original projection image;

[0104] 3.3) The matching accuracy of the current model pose is comprehensively evaluated based on indicators such as the degree of fit between the model projection contour and the boundary of the original image, the magnitude of grayscale error in local regions, and the randomness and structure distribution of the residual map.

[0105] 3.4) If the error exceeds the preset threshold or the local structure is inconsistent, trigger the matching failure backoff mechanism, replace the angle image, and iterate and optimize again.

[0106] 3.5) After successful verification, record the model. final attitude parameters Entering the occlusion repair phase

[0107] 5. After completing the type identification and spatial attitude estimation of the metal model, the system utilizes the three-dimensional geometric representation of the model. Its spatial transformation parameters By combining the scanning geometry parameters of the CBCT system with the scanning pose at each angle, simulated orthographic projection images of the metal model at various projection angles are generated for occlusion region localization and interpolation repair. This process mainly includes the following steps:

[0108] 1) Importing system geometric parameters

[0109] The system reads the CBCT system configuration file and the angle information file attached to the raw scan data to obtain the projection angle sequence. Information such as source-image distance (SID), source-axis distance (SAD), detector physical size and pixel size (including pixel spacing in the X / Y directions), relative position of detector center point and origin of scanning coordinate system, and any existing geometric non-ideal factors (such as detector tilt angle and scanning trajectory eccentricity) are used to construct the projection transformation matrix of the spatial construction projection plane.

[0110] 2) Projection transformation calculation

[0111] For each angle The system will use metal models The coordinates in space are first transformed by attitude transformation (i.e., rotation matrix). With translation vector The coordinates are transformed to the global coordinate system, and then mapped onto the detector plane using perspective projection or parallel beam approximation.

[0112] When using a perspective projection model, the following transformation relationship is used:

[0113]

[0114] in:

[0115] Homogeneous coordinate representation of model points;

[0116] : Attitude transformation matrix (composed of rotation and translation);

[0117] Projection angle The corresponding projection matrix;

[0118] Pixel coordinates on the projection plane.

[0119] 3) Generation of multi-angle metal projection images

[0120] The system maps the projected image from each angle to a resolution and format consistent with the original projection data. Optional actions during this process include:

[0121] Z-buffer depth testing or Ray Casting methods can be used to accurately depict the projected edges of the model surface.

[0122] When multiple models (combinations) are overlaid, occlusion is determined based on spatial proximity.

[0123] The output can be a binary mask, a floating-point grayscale image, or a structured mask, supporting subsequent image reconstruction processes.

[0124] 4) Projection output and binding

[0125] The final system output is a simulated metal projection atlas. Comparable to the original projection diagram One-to-one correspondence is used for occluded area localization, interpolation repair reference, residual calculation, and re-identification. Simulated projection maps can also be bound to the original image as independent channels to form structured multi-channel input.

[0126] 6. After identifying and determining the spatial position and orientation of the target metal model, image content interpolation restoration is performed on the areas of grayscale distortion caused by metal occlusion in the original projected image. This process mainly includes the following steps:

[0127] 1) The set of metal regions extracted based on the aforementioned steps By combining the metal mask map generated by the model pose projection, the pixel regions obscured by metal in each projected image are accurately determined. ,in This represents the original projected image at the i-th angle.

[0128] 2) For each occluded area Extract the set of boundary pixels based on its boundary contour. The image grayscale values ​​and gradient directions within a certain range around the boundary are extracted as boundary conditions for interpolation. Based on satisfying the boundary conditions, the... Internal pixels are reconstructed to grayscale. Then, an image inpainting method based on the Poisson equation is used, which achieves continuous gradient filling by solving the following equation:

[0129]

[0130] in Represents the boundary gradient vector field. The gray level is known for the boundary.

[0131] 3) Joint grayscale estimation is performed on corresponding regions in adjacent unoccluded projection images from multiple angles to achieve cross-angle consistent interpolation. Edge transition processing is applied to the interpolated image region to ensure consistency in brightness, texture, and structure between the restored region and the original image edges, avoiding artifacts. Finally, an interpolated image is generated for subsequent reconstruction or processing. .

[0132] 7. After interpolating and repairing the occluded areas in all projected images, the FDK algorithm was used to perform 3D reconstruction of the repaired projected image sequence. Since the original metal occluded areas have been effectively restored through the aforementioned interpolation method, the generated reconstructed images are superior to traditional direct reconstruction results in terms of structural continuity, density uniformity, and edge sharpness. This significantly reduces the impact of metal artifacts, improves the overall image quality, and provides a more reliable image basis for subsequent diagnosis or treatment planning.

[0133] 8. Based on the spatial location and orientation information of the metal model determined in the preceding steps, the 3D metal model is embedded into the repaired reconstructed image. During the localization and overlay process, the metal model is precisely registered to the corresponding occluded area according to the spatial coordinate relationship, and coordinated processing is performed in terms of grayscale values, edge continuity, and local structural consistency to achieve organic fusion of the model and the reconstructed image. Finally, a target image containing the complete metal structure is obtained, providing more accurate image data support for subsequent clinical diagnosis or treatment planning.

[0134] Key aspects of this invention include:

[0135] 1. A method for extracting geometric features of metal regions based on multi-angle projection images

[0136] By performing angle separation processing on the metal region in the partial angle projection image with a fixed step size within a half-cycle, key geometric features, including contour boundaries, centroid position, area, principal axis direction angle, and local gray-level peaks, are extracted to form a feature set. This enables high-precision modeling of metallic regions, improving the robustness and accuracy of metal recognition and providing reliable data support for subsequent registration and recovery.

[0137] 2. Direct Recognition and Interpolation Repair Method Based on Metal Model Library

[0138] This invention innovatively skips the metal region identification step in the image domain, directly extracting the geometric features of the metal region from the original projected image and matching them with a pre-built 3D metal model in a model library. After calculating the spatial position and orientation of the model through an optimized algorithm, it is projected forward, and interpolation is performed to repair occluded areas in the projection domain. This technique avoids misidentification problems in the image domain caused by image artifacts, reduces error accumulation, and improves the automation level of the processing flow and the stability of the repair results.

[0139] 3. Supports multi-model stacking recognition and progressive peeling matching mechanism

[0140] To address the complex situations common in clinical settings where multiple metal objects coexist, and where there is partial occlusion and overlap, this invention introduces a progressive stripping metal recognition strategy: In each round of matching, the most prominent metal region is first identified, and the projection data corresponding to this metal model is simulated and removed from the original projection image. The remaining image is then used for the next round of recognition and matching until all metal regions are identified and spatially modeled. This strategy effectively solves the recognition interference problem caused by mutual occlusion of multiple metal objects, significantly improving matching accuracy and system robustness, and is particularly suitable for complex metal structure scenarios such as template implantation and multiple implantation needles.

[0141] 4. Optimal attitude estimation mechanism based on key projection angles

[0142] Compared to traditional methods that require extracting metallic features from all or most projection images to fit 3D attitude parameters, this invention proposes to directly deduce the position and attitude parameters of a metallic model in space using feature information extracted from only two or a few of the most representative projection images. This key angle-driven attitude estimation method significantly reduces the computational burden while maintaining matching accuracy, simplifies the overall processing flow, and improves system response speed, making it particularly suitable for intraoperative navigation and real-time image processing scenarios.

[0143] Beneficial technical effects of the present invention:

[0144] This invention proposes a metal artifact removal method based on projection domain metal recognition and model matching. It skips the traditional image domain segmentation and orthographic projection process, directly extracting key geometric features of the metal region from the original projection image. By matching these features with a pre-established metal model library, it achieves high-precision recognition and spatial pose estimation of metal objects, significantly simplifying the processing flow and improving recognition stability and artifact repair effectiveness. To address the potential occlusion and overlap issues in multi-metal structures, a progressive peeling recognition mechanism is proposed. Through round-by-round recognition, fitting, and image updating, multi-target interference is effectively avoided, ensuring the accuracy and completeness of model matching. Furthermore, the two most representative projection images are used as input for pose inversion, replacing the traditional full-angle fitting method. This significantly reduces computational overhead while maintaining accuracy, greatly improving algorithm efficiency. Finally, the recognized model and the repaired image are fused to reconstruct a high-quality image with realistic structure, clear edges, and uniform grayscale, providing reliable image support for subsequent image-guided radiotherapy, intraoperative navigation, and clinical diagnosis. This method has a high degree of automation, a compact processing flow, and good versatility and engineering feasibility.

[0145] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for removing CBCT metal artifacts based on projection domain metal recognition, characterized in that, Includes the following steps: S1: Obtain the original projected image and the angle information corresponding to each projected image, and load the model file in the metal model library at the same time; S2: By analyzing the original projected image, identify whether there is a stack of metal objects, and simultaneously mark the two orthogonal images with the most significant features. If there is a stack of metal objects, proceed to step S3; otherwise, use the feature matching algorithm to directly identify the metal model most similar to the current projected image in the metal model library, and proceed to step S4. S3: Identify and match the metal models one by one. First, select the model with the highest matching degree, perform forward projection simulation on the corresponding fitting result, and peel off the corresponding projection area from the original projection image. Then, continue to perform the model matching and identification process in the updated residual map until all metal models are correctly identified, and then proceed to step S4. S4: Based on the identified metal models, further perform spatial pose estimation and position matching processes to determine the three-dimensional spatial position and rotation direction of each metal model in the scanned scene; S5: Compare the simulated projection images generated by each metal model with the original projection images to detect and locate the artifact occlusion areas. Then, perform image content interpolation restoration processing on the grayscale distortion areas in the original projection images caused by metal occlusion. S6: Input the interpolated and repaired projected image into the FDK reconstruction algorithm to complete the 3D image reconstruction. Then, locate and overlay all identified metal models onto the reconstructed image to generate the final CBCT image; where... Step S3 specifically includes: S31: Initialize the matching candidate set according to the following steps: S311: Perform region detection on typical angle projection maps to extract a complete set of metal features. ， in, For the area of ​​the metallic region, The center of the metal region, The principal axis direction angle of the metal region. The number of high grayscale peaks in a single metal region; S312: In the metal model library In the process, generate a simulated projection of each model at the current angle: S313: For each model Calculate its matching degree with the current image: S314: Match score Sort them from highest to lowest to form a candidate matching sequence. ; S32: Perform initial matching and metal model stripping according to the following steps: S321: Select the model with the highest score from the sorted sequence. The optimal spatial transformation parameters are obtained using attitude estimation methods. ; S322: Perform forward projection to obtain the model in the corresponding pose. Simulated projection image: S323: Subtract the original projected image to remove the contribution of the metal model, obtaining the stripped image: in, Represents the original projected image; S33: Verify the matching reliability. If the verification fails, remove the current model from the candidate set and re-match the next candidate model. If the verification passes, record the model number and pose. And proceed to the next round of processing; S34: Extract the currently stripped image As new input, repeat steps S31-S33, updating the residual map after each round of stripping, until the number of residual metal regions is 0 or all models are matched.

2. The method according to claim 1, characterized in that, In step S2, the presence of a stack of metal objects is identified through the following steps: S21: Obtain the original projected image sequence and construct a finite set of analysis images by setting the angle sampling strategy; S22: For each projected image at the sampling angle, metal region identification and feature extraction are performed by grayscale thresholding. S23: Perform statistical analysis on the feature sets extracted from all sampling angles, including the number of metal regions at certain angles. The number of high grayscale peaks in a single metal region. If so, it is determined that there are multiple metal projection areas or stacked structures.

3. The method according to claim 2, characterized in that, In step S22, during the process of metal region identification and feature extraction, the two orthogonal images with the most significant features are marked based on the metal region contour, the angle difference between the two projected images, the change of the principal axis direction angle, and the redundancy of metal information.

4. The method according to claim 2, characterized in that, Step S3 also includes: S35: After all metal recognition is completed, the total error superposition verification is performed between the simulated projection of each model and the original image; it is determined whether there is any physical unreasonable overlap in the spatial position between all models; if there are conflicts or gaps, the pose of each model is fine-tuned by introducing a global optimization algorithm to ensure consistency.

5. The method according to claim 4, characterized in that, In step S33, the reasonableness of the changes before and after stripping is evaluated by at least one of the following indicators: the concentration of the total energy change of the regional residual in the target region, the area of ​​abnormal gray values ​​in the residual image, and the degree of overlap between the metal region boundary extracted in the original projection image and the corresponding region boundary in the simulated projection image.

6. The method according to claim 2, characterized in that, In step S4, for metal models that have not overlapped, the pose is calculated directly using typical angle projection images; for overlapping models that have been stripped, the pose information is obtained according to the following steps: S41: Define the optimization objective function: in, For the first The true metal projection profile at each sampling angle To position the metal model in space ,attitude The projected contour is obtained by projecting the image onto the ground. Total number of projection angles A function representing the difference between projected images; S42: Optimize the attitude parameter space search by minimizing the error function through group cooperative search. From multiple key angles, feature information of the metal region in the projected image is extracted and compared with the standard projection features of the corresponding model in the metal model library. The minimum error matching or attitude backpropagation method based on regression network is used to realize the spatial positioning and attitude estimation of the target metal model in the current scanning scene. S43: Perform matching verification and result confirmation.

7. The method according to claim 6, characterized in that, Step S43 specifically includes: S431: Project the calculated metal model in the forward orientation onto a set of auxiliary angles to obtain multiple simulated metal projection images; S432: Perform residual analysis and structural similarity comparison on corresponding positions in the simulated metal projection image and the original projection image; S433: Based on the degree of fit between the model projection contour and the boundary of the original image, the magnitude of grayscale error in local regions, and the randomness and structure distribution of the residual map, the matching accuracy of the current model pose is comprehensively evaluated. S434: If the error exceeds the preset threshold or the local structure is inconsistent, trigger the matching failure backoff mechanism, replace the angle image, and iterate and optimize again. S435: After successful verification, record the final attitude parameters of the metal model.

8. The method according to claim 6, characterized in that, In step S5, the three-dimensional geometric representation of the metal model and the corresponding spatial transformation parameters are used, combined with the scanning geometric parameters of the CBCT system and the scanning pose at each projection angle, to generate simulated orthogonal projection images of the metal model at each projection angle, which are used for occlusion area localization and interpolation repair.

9. The method according to claim 8, characterized in that, In step S5, the image content interpolation restoration process is performed according to the following steps: S51: Based on the identified set of metal regions and the metal mask generated by the model pose projection, determine the pixel regions in each projected image that are occluded by metal. ,in This represents the original projected image at the i-th sampling angle; S52: For each occluded area Extracting the set of boundary pixels based on the boundary contour It extracts the image grayscale values ​​and gradient directions within a certain range around the boundary as boundary conditions for interpolation. Based on satisfying these boundary conditions, it then... Internal pixels are reconstructed to grayscale, and then image inpainting methods are used to achieve continuous gradient filling by solving the following equation: in, Represents the boundary gradient vector field. The gray level is known for the boundary. S53: Joint grayscale estimation is performed on the corresponding regions in adjacent unobstructed projection images under multiple angles to achieve cross-angle consistent interpolation, and then edge transition processing is performed on the interpolated image regions.

10. A CBCT metal artifact removal system based on projection domain metal recognition, characterized in that, The system for performing the method as described in any one of claims 1-9 includes the following modules: Data acquisition module: acquires the original projected image and the angle information corresponding to each projected image, and simultaneously loads the model files in the metal model library; Projection Image Analysis Module: Identifies the presence of stacked metal objects by analyzing the original projection image, and simultaneously marks the two orthogonal images with the most significant features; Metal model recognition module: In the absence of metal stacks, the module directly identifies the metal model most similar to the current projected image from the metal model library using a feature matching algorithm. In the presence of metal stacks, the module identifies and matches the metal models one by one. First, it selects the model with the highest matching degree, performs forward projection simulation on the corresponding fitting result, and peels the corresponding projection area from the original projection image. Then, it continues to perform the model matching and recognition process in the updated residual map until all metal models are correctly identified. Spatial attitude estimation module: Based on the identified metal models, it performs spatial attitude estimation and position matching processes to determine the three-dimensional spatial position and rotation direction of each metal model in the scanned scene; Image interpolation processing module: Compare the simulated projection images generated by each metal model with the original projection images, detect and locate the artifact occlusion areas, and perform image content interpolation restoration processing for the grayscale distortion areas in the original projection images caused by metal occlusion. 3D Reconstruction Module: The interpolated and repaired projected image is input into the FDK reconstruction algorithm to complete the 3D image reconstruction. Then, all the identified metal models are located and superimposed on the reconstructed image to generate the final CBCT image.

Images