A metal artifact removal method based on multi-level broad spectrum X-ray data

Abstract

The present invention relates to a metal artifact removal method based on multi-energy-level, broad-spectrum X-ray data. The method comprises: obtaining dual-energy-level CT image data and multi-energy-level CT image data of multiple patients containing metal artifacts, as well as standard image data after metal artifact removal, to construct a training sample set; performing metal region segmentation based on CT image data of different energy levels corresponding to the same patient in the training sample set to obtain a metal region mask image; training a model using the training sample set and the metal region mask image to obtain a converged metal artifact removal model; and inputting CT image data of any energy level to be processed into the metal artifact removal model to obtain CT image data after metal artifact removal. The present invention solves the problem that metal artifact removal methods in the prior art rely on high-end CT scanners or use analog data with a single data energy level, resulting in poor generalization ability of the existing methods and poor processing effect on low-energy-level CT scan images.

Description

Technical Field

[0001] The present invention belongs to the technical field of CT imaging, and in particular relates to a method for removing metal artifacts based on multi-energy-level broad-spectrum X-ray data. Background Art

[0002] With advances in materials science, the use of medical metal implants in the human body is becoming increasingly widespread, particularly in the treatment of spinal degenerative diseases. For patients with lumbar metal internal fixation implants, orthopedic surgeons often choose computed tomography (CT) as the imaging modality of choice for postoperative evaluation. However, the presence of metal artifacts significantly impacts CT image quality, significantly reducing the diagnostic efficacy of CT imaging.

[0003] To reduce metal artifact interference in CT scans, researchers have proposed a variety of methods for removing metal artifacts, including increasing the scanning tube voltage, using high-energy virtual monoenergetic images, and employing metal artifact removal algorithms (MAR). However, all existing commercial MAR technologies rely on high-end CT scanners and can only perform metal artifact removal based on each manufacturer's own raw projection domain data, lacking universality across different manufacturers. Furthermore, the vast number of mid- and low-end CT scanners currently in use in grassroots hospitals in China are largely unable to be upgraded to utilize MAR technology. In current medical imaging diagnostics, the universal image data format standard is the DICOM standard. Therefore, it is imperative to develop a new, broadly applicable MAR technology that is independent of specific CT scanner types and applicable to the post-processing of all DICOM images.

[0004] Against the backdrop of the rapid development and application of big data and artificial intelligence, the rapid advancement of deep learning technology has provided a new option to meet this demand. In recent years, deep-learning-based MAR technology (deep-MAR) has demonstrated promising performance in reducing metal artifacts. However, due to the difficulty in obtaining CT data pairs with identical anatomical structures affected and unaffected by metal artifacts, all deep-MAR studies have used simulated datasets. However, differences between simulated and clinical data are inevitable, and deviations from real-world data distributions undermine the effectiveness and generalizability of the method. To avoid these drawbacks, real-world clinical datasets should be used throughout the model training process. Furthermore, in actual clinical applications, common tube voltages for CT scans range from 100 kVp, 120 kVp, or 140 kVp, depending on the scan site. For pediatric scans, a tube voltage of 80 kVp is generally recommended. For pulmonary artery or head and neck artery examinations, a tube voltage of 70 kVp is currently recommended to reduce radiation exposure and increase image contrast. As awareness of radiation-induced carcinogenesis and teratogenesis deepens, clinical requirements call for lower tube voltages to be used for pediatric CT scans. In summary, in actual clinical applications, patients affected by metal artifacts may face numerous CT scan options with different tube voltages depending on clinical needs. However, in existing deep-MAR technology research, many studies focus solely on using a single simulated discrete X-ray spectrum, such as studying only 120kvp or 100kvp. This range is too small to reflect the complexity of real-world scenarios. These issues significantly reduce the practicality of the technology and limit its promotion from laboratory to clinical applications. Summary of the Invention

[0005] In view of the above analysis, the present invention aims to disclose a metal artifact removal method based on multi-energy-level broad-spectrum X-ray data, which solves the problem that the metal artifact removal methods in the prior art rely on high-end CT scanners or use analog data with a single data energy level, resulting in poor generalization ability of the existing methods and poor processing effect on low-energy-level CT scan images.

[0006] The purpose of the present invention is mainly achieved through the following technical solutions:

[0007] The present invention discloses a method for removing metal artifacts based on multi-energy-level broad-spectrum X-ray data, comprising:

[0008] Obtain dual-level CT image data and multi-level CT image data of multiple patients containing metal artifacts, as well as standard image data after removing metal artifacts corresponding to each CT image data, perform normalization processing on each image data, and construct a training sample set;

[0009] Based on the CT image data of different energy levels corresponding to the same patient in the training sample set, metal area segmentation is performed to obtain a metal area mask image;

[0010] Using the standard image data and the metal area mask image as labels, the neural network model is iteratively trained in sequence using the normalized dual-level CT image data and multi-level CT image data in the training sample set to obtain a converged metal artifact removal model;

[0011] The CT image data of any energy level to be processed is normalized and then input into the metal artifact removal model to obtain the CT image data after the metal artifacts are removed.

[0012] Furthermore, the normalization processing of each image data includes:

[0013] Setting a CT value threshold, and setting the CT value of pixels in the CT image data that are smaller than the threshold to a value corresponding to the threshold;

[0014] Based on the CT value of each pixel point of the CT image data, the CT image data is normalized using the following formula:

[0015] U=CTvalue / 1000*Uwater(E)+Uwater(E);

[0016] Where U is the normalized energy function of each pixel, Uwater(E) is the radiation absorption energy function of water, E is the average energy of different energy spectra CT, and CTvalue is the CT value of each pixel.

[0017] Furthermore, the metal region is segmented by the following method to obtain a metal region mask image:

[0018] The difference between the characteristic values ​​of the corresponding pixel points in the normalized CT image data of the same patient at two energy levels is calculated to obtain a difference map;

[0019] Extract the pixels whose values ​​are greater than a preset threshold in the difference map and set them to 1, and set the other pixels to 0;

[0020] The area image of the CT image data containing metal artifacts that is greater than the metal CT threshold and is set to 0 in the difference map is the metal area mask image.

[0021] Furthermore, the metal artifact removal model is an encoder-decoder structure; the neural network model is iteratively trained in sequence using the dual-level CT image data and the multi-level CT image data, including:

[0022] Performing preliminary iterative training on the neural network model using the dual-level CT image data to obtain a primary metal artifact removal model;

[0023] Freezing the coding layer of the primary metal artifact removal model, and performing migration training on the metal artifact removal model using multi-level CT image data to obtain a final metal artifact removal model;

[0024] During the initial iterative training and transfer training, images containing metal artifacts and corresponding standard image data are iteratively trained using mean square error and structural similarity weighted loss; CT image data containing metal artifacts and metal area masks are iteratively trained using mask loss.

[0025] Furthermore, the encoder includes two convolution modules, a feature fusion extraction module, and two TokMLP modules arranged in sequence; the encoder is used to perform multiple size compression and feature extraction on the input CT image data in sequence to obtain multiple feature maps of different sizes;

[0026] The decoder includes two TokMLP modules, a feature fusion extraction module, and two convolution modules that are sequentially arranged to correspond to the modules in the encoder. The modules are used to sequentially upsample and extract features based on the feature maps output by the last layer of the encoder, and fuse the feature maps of corresponding sizes in the encoder through jump connections to decode and obtain the final image after removing metal artifacts.

[0027] Furthermore, the feature fusion extraction module includes an improved Transformer module, a parallel convolution module and a fusion module;

[0028] The improved Transformer module and the parallel convolution module are set in parallel to perform feature extraction on the received feature maps to obtain global feature maps and local feature maps corresponding to the CT image data;

[0029] The fusion module is used to perform an intra-channel addition operation on the global feature map and the local feature map; and perform feature extraction on the feature map after the addition operation through a convolution layer to obtain a feature map with local features and global features.

[0030] Furthermore, the improved Transformer module includes a one-dimensional sequence conversion layer, a window attention layer, an MLP layer, and a three-dimensional sequence restoration layer arranged in sequence;

[0031] The one-dimensional sequence conversion layer is used to convert the received feature map from three dimensions to a sequence to obtain one-dimensional sequence data;

[0032] The window attention layer is used to perform global feature data enhancement on the one-dimensional sequence data through a window attention operation;

[0033] Performing a first addition operation on the feature data after data enhancement and the one-dimensional sequence data output by the one-dimensional sequence conversion layer, and then inputting the result into the MLP layer for feature extraction to obtain one-dimensional sequence data containing global features;

[0034] The feature data obtained after the first addition operation is added to the feature data output by the MLP layer for a second addition operation, and the sequence is converted to three dimensions through a three-dimensional sequence restoration layer to obtain a three-dimensional feature map containing global features.

[0035] Furthermore, the mean square error and structural similarity weighted loss are expressed as;

[0036] Loss Total =α·MSE+β·(1-SSIM);

[0037]

[0038]

[0039] Among them, Loss Total is the mean square error and structural similarity weighted loss, α and β are weight coefficients, MSE is the mean square error, y i is the true value of the i-th pixel of the standard image data, is the predicted value, n is the number of pixels, SSIM is the structural similarity index, μ x and μ y are the means of x and y respectively, and are the variances of x and y, σ xy is the covariance of x and y, and C1 and C2 are constants.

[0040] Furthermore, the mask loss is expressed as:

[0041] L metal =L+γ·filter(metal_mask)·L;

[0042] Among them, L is the pixel difference loss; γ is the weight of the metal area, and filter is the filtering operation.

[0043] Furthermore, each dual-energy CT image data includes 70keV and 100keV image data of the same part of the same patient, and each multi-energy CT image data includes 70keV, 80keV, 100keV, 120keV and 140keV image data of the same part of the same patient.

[0044] The present invention can achieve at least the following beneficial effects:

[0045] 1. The present invention proposes a metal artifact removal method based on multi-energy-level broad-spectrum X-ray data. This method utilizes a real clinical dataset containing multiple X-ray spectra for model training, normalizes the data at each energy level based on the radiation absorption energy coefficient of water, and segmentes the metal region using a metal region extraction method based on multi-energy-level data matching. This method improves the metal region segmentation effect, achieves subjective evaluation of metal artifact reduction on virtual monochrome images at different energy levels, and better reduces metal artifacts in CT imaging data of patients with metal implants, thus resolving the problem of CT imaging diagnosis in patients' postoperative follow-up.

[0046] 2. The present invention realizes the connection between deep and shallow image information by extracting global features through an improved Transformer module, extracting local features through parallel convolution, and fusing features, thereby capturing the global information of the image, significantly enhancing the representation of complex spatial transformations and long-distance feature dependencies, and improving the effect of metal artifact removal. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.

[0048] Figure 1 Flowchart of a method for removing metal artifacts based on multi-energy-level broad-spectrum X-ray data in an embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of a CT image containing metal artifacts in an embodiment of the present invention;

[0050] Figure 3 Schematic diagram of metal area segmentation in an embodiment of the present invention;

[0051] Figure 4 Schematic diagram of the metal artifact removal effects of different models in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, which constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.

[0053] The embodiment of the present invention discloses a method for removing metal artifacts based on multi-level broad spectrum X-ray data, such as Figure 1 As shown, the following steps are included:

[0054] Step S1: Obtain dual-level CT image data and multi-level CT image data of multiple patients containing metal artifacts, as well as standard image data after metal artifact removal corresponding to each CT image data, perform normalization processing on each image data, and construct a training sample set;

[0055] Specifically, when constructing a training sample set, existing metal artifact removal algorithms all use a simulation approach, namely, adding a metal mask to the data without metal artifacts and generating simulated metal artifacts through the algorithm. However, this embodiment uses CT scans of different parts in a clinical multi-device environment as the original samples for the training dataset (to improve the robustness of the model, the original samples can include CT images with and without metal artifacts). Standard images obtained after processing using traditional metal artifact removal algorithms are used as labels. This has two advantages. First, existing traditional algorithms can directly access the raw data signals acquired by CT images and compare them with threshold segmentation, providing better judgment of the presence and shape of real metal artifacts. Second, existing methods for simulating metal artifacts often use fixed numerical values ​​for the CT values ​​of the metal itself, but the density and structure of actual metals are subject to numerous subtle variations and fluctuations. This embodiment uses clinical metal artifact data to provide a better perception of real metal artifacts.

[0056] Specifically, in the training sample set of this embodiment, each dual-energy CT image data is 70keV and 100keV image data of the same part of the same patient, and each multi-energy CT image data is 70keV, 80keV, 100keV, 120keV, and 140keV image data of the same part of the same patient. Since multi-energy CT image data is difficult to obtain, the data volume is large, and storage is difficult, this embodiment collects dual-energy CT image data of the metal implant sites of a large number of metal implant patients during data acquisition for preliminary training of the model. Multi-energy CT image data of a small number of patients are also collected for transfer training; the metal implant site of each patient is not limited, and the patients corresponding to the multi-energy CT image data and the dual-energy CT image data are preferably different patients to increase the generalization ability of the model.

[0057] This embodiment utilizes real clinical CT image data at multiple energy levels to achieve high-quality metal artifact removal for metal artifact images of any energy level and any CT scanning device, thereby improving the scope of application and processing speed of the metal artifact removal method.

[0058] Furthermore, in order to make the model applicable to data at each energy level, it is necessary to pre-process the image data at each energy level. In this embodiment, each image data is normalized, including:

[0059] Setting a CT scan image threshold, and setting the CT values ​​of pixels in the CT image data that are smaller than the CT scan image threshold to values ​​corresponding to the CT scan image threshold;

[0060] Based on the CT value of each pixel point of the CT image data, the CT image data is normalized using the following formula:

[0061] U=CTvalue / 1000*Uwater(E)+Uwater(E);

[0062] Where U is the normalized energy function of each pixel, Uwater(E) is the radiation absorption energy function of water, E is the average energy of different energy spectra CT, and CTvalue is the CT value of each pixel.

[0063] For example, because the CT values ​​around metal artifacts vary significantly, before inputting the CT image data into the neural network, this embodiment sets the CT value threshold to -1000 and sets the CT values ​​of pixels with a value less than -1000 to -1000. The CT image (in Housefield Units) is then converted into a radiation absorption attenuation image for input. Specifically, the CT value of the background region (out of field of view) is replaced with an extremely small value. According to the definition of CT, the minimum CT value is -1000, so the CT value of the background region is set to -1000. Furthermore, in conventional CT image metal artifact removal, Uwater uses a single fixed energy value. This embodiment processes Uwater using characteristic values ​​from different energy spectra. Uwater has different values ​​for light of different energies. This embodiment uses the average energy E of the energy spectrum CT as a coefficient, i.e., Uwater = Uwater(E). The value of Uwater can be obtained through a table lookup. Due to data storage accuracy and the fact that the background region is outside the CT value calculation range, the pixel values ​​of some pixels may be less than the theoretical minimum CT value (i.e., CT). Different manufacturers have different solutions for processing this part of data. Through the above normalization process, the method can be applied to equipment of all manufacturers, thereby improving the scope of application of the metal artifact removal method of the present invention.

[0064] Step S2: performing metal region segmentation based on CT image data of different energy levels corresponding to the same patient in the training sample set to obtain a metal region mask image;

[0065] Specifically, the metal region is segmented by the following method to obtain a metal region mask image:

[0066] The difference between the characteristic values ​​of the corresponding pixel points in the normalized CT image data of the same patient at two energy levels is calculated to obtain a difference map;

[0067] Extract the pixels whose values ​​are greater than a preset threshold in the difference map and set them to 1, and set the other pixels to 0;

[0068] The area image of the CT image data containing metal artifacts that is greater than the metal CT threshold and is set to 0 in the difference map is the metal area mask image.

[0069] A very important step in the metal artifact removal method is to extract the metal area. In the existing deep learning metal artifact removal framework, the metal area can only be determined by the metal threshold method. However, in practical applications, the metal threshold is affected by factors such as the scanning ray energy level and the reconstruction method, resulting in inaccurate metal area extraction. This embodiment proposes a metal area extraction method based on multi-energy level data matching. The metal area can be better separated by the difference in metal artifact morphology between different energy level data, that is, the normalized corresponding two energy level images u_1 and u_2 are subtracted to obtain the difference image du. When the maximum CT value of du is less than a certain threshold, it is considered that there is no metal artifact in the two images. When the maximum value of du is greater than a certain threshold, it is considered that there is a metal artifact in the image. At this time, by extracting a mask image greater than the threshold, it is possible to obtain an area where the radiation absorption rate changes greatly under different energy levels. The area in the original metal artifact image that is greater than the metal CT threshold and does not show a large absorption rate change in the mask image is the real metal area metal_mask; as shown in FIG. Figure 2 and Figure 3 As shown, Figure 2 is the CT image data containing metal artifacts, Figure 3 The image is the corresponding metal area segmentation. It can be seen from the figure that this method can better eliminate the influence of high-density bones and extract the metal area more accurately.

[0070] Step S3: using the standard image data and the metal area mask image as labels, and using the normalized dual-level CT image data and multi-level CT image data in the training sample set to iteratively train the neural network model in sequence to obtain a converged metal artifact removal model;

[0071] Specifically, the metal artifact removal model is an encoder-decoder structure; the neural network model is iteratively trained in sequence using the dual-level CT image data and the multi-level CT image data, including:

[0072] The neural network model is preliminarily iteratively trained using a large amount of dual-level CT image data in a training sample set to obtain a primary metal artifact removal model;

[0073] Freezing the coding layer of the primary metal artifact removal model, and performing migration training on the metal artifact removal model using a small amount of multi-level CT image data to obtain a final metal artifact removal model;

[0074] During the initial iterative training and transfer training, images containing metal artifacts and corresponding standard image data are iteratively trained using mean square error and structural similarity weighted loss; CT image data containing metal artifacts and metal area masks are iteratively trained using mask loss.

[0075] Furthermore, the encoder includes two convolution modules, a feature fusion extraction module, and two TokMLP modules arranged in sequence; the encoder is used to perform multiple size compression and feature extraction on the input CT image data in sequence to obtain multiple feature maps of different sizes;

[0076] The decoder includes two TokMLP modules, a feature fusion extraction module, and two convolution modules that are sequentially arranged to correspond to the modules in the encoder. The modules are used to sequentially upsample and extract features based on the feature maps output by the last layer of the encoder, and fuse the feature maps of corresponding sizes in the encoder through jump connections to decode and obtain the final image after removing metal artifacts.

[0077] More specifically, the feature fusion extraction module includes an improved Transformer module, a parallel convolution module and a fusion module;

[0078] The improved Transformer module and the parallel convolution module are set in parallel to perform feature extraction on the received feature maps to obtain global feature maps and local feature maps corresponding to the CT image data;

[0079] The fusion module is used to perform an intra-channel addition operation on the global feature map and the local feature map; and perform feature extraction on the feature map after the addition operation through a convolution layer to obtain a feature map with local features and global features.

[0080] The improved Transformer module includes a one-dimensional sequence conversion layer, a window attention layer, an MLP layer and a three-dimensional sequence restoration layer arranged in sequence;

[0081] The one-dimensional sequence conversion layer is used to convert the received feature map from three dimensions to sequences to obtain one-dimensional sequence data. That is, the three-dimensional feature map is sliced ​​along any dimension of depth, height or width, and each slice is regarded as an element in the sequence, and all slice data are reshaped into a one-dimensional sequence.

[0082] The window attention layer is used to perform global feature data enhancement on the one-dimensional sequence data through a window attention operation;

[0083] Performing a first addition operation on the feature data after data enhancement and the one-dimensional sequence data output by the one-dimensional sequence conversion layer, and then inputting the result into the MLP layer for feature extraction to obtain one-dimensional sequence data containing global features;

[0084] The feature data obtained after the first addition operation is added to the feature data output by the MLP layer for a second addition operation, and the elements in the one-dimensional sequence data obtained after the addition operation are spliced ​​along the slice dimension through the three-dimensional sequence restoration layer to obtain a three-dimensional feature map containing global features.

[0085] The metal artifact removal model of this embodiment performs global feature extraction through an improved Transformer module, performs local feature extraction through parallel convolution, and realizes the connection of top-level and bottom-level information through feature fusion, captures the global information of the image to be processed, significantly enhances the representation of complex spatial transformations and long-distance feature dependencies, and improves the effect of metal artifact removal.

[0086] Furthermore, during the initial iterative training and transfer training, the images containing metal artifacts and the corresponding standard image data are iteratively trained using mean square error and structural similarity weighted loss, which is expressed as;

[0087] Loss Total =α·MSE+β·(1-SSIM);

[0088]

[0089]

[0090] Among them, Loss Total is the mean square error and structural similarity weighted loss, α and β are weight coefficients, MSE is the mean square error, y i is the true value of the i-th pixel of the standard image data, is the predicted value, n is the number of pixels, SSIM is the structural similarity index, μ x and μ y are the means of x and y, respectively. x and y are the normalized eigenvalues ​​of the original CT image data and the standard image data, respectively. and are the variances of x and y, σ xy is the covariance of x and y, C1 and C2 are constants. In this embodiment, α is set to 1.0 and β is set to 0.1-0.5 to improve the image structure recovery capability.

[0091] For CT image data containing metal artifacts and metal area masks, iterative training is performed through mask loss, which is expressed as:

[0092] L metal =L+γ·filter(metal_mask)·L;

[0093] Wherein, L is the pixel difference loss; γ is the weight of the metal area part; filter is the filtering operation. This embodiment adopts Gaussian filtering as the filter, so that the weight of the area closer to the metal area mask is greater, and the weight of the area farther away from the metal area mask is smaller.

[0094] During model training, the Adam optimizer was used, with a learning rate set between 0.0001 and 0.001, momentum parameters Beta1 = 0.9, and Beta2 = 0.999. A cosine annealing schedule (periods 5-10) was used to dynamically adjust the learning rate to balance convergence speed and stability. During training, the batch size ranged from 16 to 64, the number of training epochs ranged from 50 to 400, and an early stopping strategy (with a patience value of 10-20) was used to prevent overfitting.

[0095] During the model evaluation phase, quantitative metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) were used to measure the quality of de-artifacted images. Qualitative visual analysis was also used to verify the model's performance in detail restoration and artifact removal. The model's generalization capability was further ensured by appropriately setting the dropout rate (0.2-0.5). By optimizing these parameters within this range, the model achieved efficient training and excellent performance in CT metal artifact removal.

[0096] Furthermore, this embodiment employs a cross-validation method to assess the effectiveness of metal artifact removal in CT image data at different keV energy levels. To comprehensively evaluate the effectiveness of the proposed method, this embodiment employs a combination of objective metrics and subjective analysis. Cross-energy evaluation is used to assess the generalization capability of the cross-domain learning method. A model trained in one keV setting is evaluated on datasets obtained at different keV settings. For example, a model trained on 70 keV images is tested on 100 keV data.

[0097] This embodiment uses 70KeV energy level data, 100KeV energy level data and the multi-energy level data of this embodiment to train three metal artifact removal models, namely model70, model100 and modelmix (i.e., the metal artifact removal model of this embodiment); Figure 4Axial images reconstructed using a soft tissue window of a patient who underwent lumbar metal internal fixation, as well as the effects of metal artifact removal using three metal artifact removal models and the commercial modelMAR technology. Among them, 70keV NoMAR is the original image of metal artifacts under 70keV conditions, 70keV modelMAR is the image with metal artifacts removed using the commercial modelMAR technology under 70keV conditions, 70keV model70 is the image with metal artifacts removed using the model70 technology under 70keV conditions, 70keV model100 is the image with metal artifacts removed using the model100 technology under 70keV conditions, 70keV modelmix is ​​the image with metal artifacts removed using the modelmix technology under 70keV conditions, 100keV NoMAR is the original image of metal artifacts under 100keV conditions, 100keVmodelMAR is the image with metal artifacts removed using the commercial modelMAR technology under 100keV conditions, 100keVmodel70 is the image with metal artifacts removed using the model70 technology under 100keV conditions, 100keV model100 is the image with metal artifacts removed using the model100 technology under 100keV conditions, 100keV ModelMix shows images obtained at 100 keV using the ModelMix technology for metal artifact removal. A comparison of these images shows that, regardless of whether the images were acquired at 70 keV or 100 keV, the CT images processed using the metal artifact removal model of this embodiment have fewer artifacts around metal fixtures than other images, and the spinal canal structure is displayed more clearly. This demonstrates the effectiveness of the metal artifact removal method of this embodiment.

[0098] Step S4: performing normalization processing on the CT image data of any energy level to be processed and inputting the data into the metal artifact removal model to obtain the CT image data after the metal artifacts are removed.

[0099] After training the metal artifact removal model using dual-energy and multi-energy CT image data, metal artifact removal operations can be performed on CT image data of any energy level, thereby improving the efficiency and auxiliary diagnosis effect of computed tomography technology in clinical applications.

[0100] In summary, the metal artifact removal method based on multi-energy-level wide-spectrum X-ray data of the present invention utilizes real clinical CT image data containing multiple energy levels for model training, normalizes the data of each energy level based on the radiation absorption energy coefficient of water, and segments the metal region through a metal region extraction method based on multi-energy-level data matching, thereby improving the segmentation effect of the metal region, achieving the reduction of subjective evaluation of metal artifacts on virtual monochrome images of different energy levels, better reducing metal artifacts in CT image data of patients with metal implants, and solving the problem of CT image diagnosis in postoperative follow-up of patients. In addition, the present invention realizes the connection of deep and shallow image information through global feature extraction using an improved Transformer module, local feature extraction through parallel convolution, and feature fusion, thereby capturing the global information of the image, significantly enhancing the representation of complex spatial transformations and long-range feature dependencies, and improving the effect of metal artifact removal.

[0101] Those skilled in the art will appreciate that all or part of the process steps of the methods in the above embodiments can be implemented by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0102] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for removing metal artifacts based on multi-level broad spectrum X-ray data, characterized in that: include: Obtain dual-level CT image data and multi-level CT image data of multiple patients containing metal artifacts, as well as standard image data after removing metal artifacts corresponding to each CT image data, perform normalization processing on each image data, and construct a training sample set; Based on the CT image data of different energy levels corresponding to the same patient in the training sample set, metal area segmentation is performed to obtain a metal area mask image; Using the standard image data and the metal area mask image as labels, the neural network model is iteratively trained in sequence using the normalized dual-level CT image data and multi-level CT image data in the training sample set to obtain a converged metal artifact removal model; The metal artifact removal model is an encoder-decoder structure; The iterative training includes: performing preliminary iterative training on the neural network model using the dual-level CT image data to obtain a primary metal artifact removal model; freezing the coding layer of the primary metal artifact removal model, and performing transfer training on the metal artifact removal model using multi-level CT image data to obtain a final metal artifact removal model; during the preliminary iterative training and transfer training, iteratively training the image containing metal artifacts and the corresponding standard image data using mean square error and structural similarity weighted loss; and iteratively training the CT image data containing metal artifacts and the metal area mask using mask loss; The CT image data of any energy level to be processed is normalized and then input into the metal artifact removal model to obtain the CT image data after the metal artifacts are removed.

2. The metal artifact removal method based on multi-level broad spectrum X-ray data according to claim 1, characterized in that: The normalization processing of each image data includes: Setting a CT value threshold, and setting the CT value of pixels in the CT image data that are smaller than the threshold to a value corresponding to the threshold; Based on the CT value of each pixel point of the CT image data, the CT image data is normalized using the following formula: U=CTvalue / 1000*Uwater(E)+Uwater(E); Where U is the normalized energy function of each pixel, Uwater(E) is the radiation absorption energy function of water, E is the average energy of different energy spectra CT, and CTvalue is the CT value of each pixel.

3. The metal artifact removal method based on multi-level broad spectrum X-ray data according to claim 2, characterized in that: The metal area is segmented by the following method to obtain the metal area mask image: The difference between the characteristic values ​​of the corresponding pixel points in the normalized CT image data of the same patient at two energy levels is calculated to obtain a difference map; Extract the pixels whose values ​​are greater than a preset threshold in the difference map and set them to 1, and set the other pixels to 0; The area image of the CT image data containing metal artifacts that is greater than the metal CT threshold and is set to 0 in the difference map is the metal area mask image.

4. The metal artifact removal method based on multi-level broad spectrum X-ray data according to claim 1, characterized in that: The encoder includes two convolution modules, a feature fusion extraction module, and two TokMLP modules arranged in sequence; the encoder is used to perform multiple size compression and feature extraction on the input CT image data in sequence to obtain multiple feature maps of different sizes; The decoder includes two TokMLP modules, a feature fusion extraction module, and two convolution modules that are sequentially arranged to correspond to the modules in the encoder. The modules are used to sequentially upsample and extract features based on the feature maps output by the last layer of the encoder, and fuse the feature maps of corresponding sizes in the encoder through jump connections to decode and obtain the final image after removing metal artifacts.

5. The metal artifact removal method based on multi-energy level broad spectrum X-ray data according to claim 4, characterized in that: The feature fusion extraction module includes an improved Transformer module, a parallel convolution module and a fusion module; The improved Transformer module and the parallel convolution module are set in parallel to perform feature extraction on the received feature maps to obtain global feature maps and local feature maps corresponding to the CT image data; The fusion module is used to perform an intra-channel addition operation on the global feature map and the local feature map; and perform feature extraction on the feature map after the addition operation through a convolution layer to obtain a feature map with local features and global features.

6. The metal artifact removal method based on multi-level broad spectrum X-ray data according to claim 5, characterized in that: The improved Transformer module includes a one-dimensional sequence conversion layer, a window attention layer, an MLP layer and a three-dimensional sequence restoration layer arranged in sequence; The one-dimensional sequence conversion layer is used to convert the received feature map from three dimensions to a sequence to obtain one-dimensional sequence data; The window attention layer is used to perform global feature data enhancement on the one-dimensional sequence data through a window attention operation; Performing a first addition operation on the feature data after data enhancement and the one-dimensional sequence data output by the one-dimensional sequence conversion layer, and then inputting the result into the MLP layer for feature extraction to obtain one-dimensional sequence data containing global features; The feature data obtained after the first addition operation is added to the feature data output by the MLP layer for a second addition operation, and the sequence is converted to three dimensions through a three-dimensional sequence restoration layer to obtain a three-dimensional feature map containing global features.

7. The metal artifact removal method based on multi-energy level broad spectrum X-ray data according to claim 1, characterized in that: The mean square error and structural similarity weighted loss are expressed as; Loss Total =α·MSE+β·(1-SSIM); Among them, Loss Total is the mean square error and structural similarity weighted loss, α and β are weight coefficients, MSE is the mean square error, y i is the true value of the i-th pixel of the standard image data, is the predicted value, n is the number of pixels, SSIM is the structural similarity index, μ x and μ y are the means of x and y respectively, and are the variances of x and y, σ xy is the covariance of x and y, and C1 and C2 are constants.

8. The metal artifact removal method based on multi-energy level broad spectrum X-ray data according to claim 1, characterized in that: The mask loss is expressed as: L metal =L+γ·filter(metal_mask)·L; Among them, L is the pixel difference loss; γ is the weight of the metal area, and filter is the filtering operation.

9. The metal artifact removal method based on multi-energy level broad spectrum X-ray data according to claim 1, characterized in that: Each dual-energy CT image data includes 70keV and 100keV image data of the same part of the same patient, and each multi-energy CT image data includes 70keV, 80keV, 100keV, 120keV and 140keV image data of the same part of the same patient.

Images