CBCT artifact removal method based on cascaded feature-connected residual network

Abstract

The present invention discloses a CBCT artifact removal method based on a cascaded feature-connected residual network. The method involves registering each set of CBCT and CT images to the same spatial coordinate system and obtaining a projection image through forward projection. The projection image is then input into the cascaded feature-connected residual network. The network-generated projection data and the input CBCT-CT projection data are then rapidly extracted using a FasterNet network. The Cobi Loss and Perceptual Loss between the feature set generated by the network and the feature set of the input CBCT, and between the feature set generated by the network and the feature set of the input CT, are calculated. The loss values ​​are then weighted to optimize the network and obtain artifact-free projection data. The generated projection data are then back-projected to obtain a reconstructed image. The present invention addresses the problem in CBCT artifact removal methods that processing in the image domain results in the trained model being unable to adapt to different device protocols.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a CBCT artifact removal method based on a cascaded feature-connected residual network. Background Art

[0002] Cone-beam computed tomography (CBCT) is an important imaging technology widely used in medical fields such as radiotherapy, dentistry, and orthopedics. CBCT replaces the rotating surround in the imaging system with a rotating cone beam, providing high-resolution three-dimensional imaging. Compared to traditional computed tomography (CT), CBCT offers the advantages of lower radiation dose and safer and more feasible clinical applications. However, due to the inherent characteristics of the detector and the varying tissue structures of the patient, image quality deteriorates due to radiation scattering and photon starvation. Furthermore, the low imaging dose reduces soft tissue contrast, introducing noise into the reconstructed CBCT images. Visible artifacts can be observed in these images. Removing CBCT artifacts and improving image quality would greatly enhance the accuracy of medical imaging and the reliability of diagnosis, helping physicians more accurately detect diseases and plan treatments, thereby improving patient outcomes.

[0003] In the prior art, there are several commonly used methods for artifact removal. The first is to introduce a CBCT scatter correction method based on a rotating collimator and a scatter reduction algorithm using a modulator, which can reduce some artifacts, but increases hardware costs, and the image quality cannot be significantly improved. The second is a CBCT slice data correction method based on deep learning. By learning the nonlinear mapping between CBCT and CT, the Hounsfield unit (HU) of the CBCT image is converted into the HU value of the CT image, thereby improving the CBCT image quality. However, the imaging protocols of different manufacturers are different, and the slicing method for one protocol cannot guarantee the compatibility of multiple protocols. The third is to use the Monte Carlo calculation method to estimate the scattering and absorption effects by simulating photon transmission to remove artifacts. However, the Monte Carlo-based method is time-consuming and not very suitable for clinical application.

[0004] The information disclosed in this background technology section is only intended to enhance understanding of the overall background of the invention and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the Invention

[0005] In order to overcome the defects of the existing technology, a CBCT artifact removal method based on a cascaded feature connection residual network is provided to solve the problem that the existing CBCT artifact removal method is processed in the image domain, resulting in the inability of the training model to adapt to different device protocols.

[0006] To achieve the above object, a CBCT artifact removal method based on a cascaded feature-connected residual network is provided, comprising the following steps:

[0007] Performing deformable registration on the patient's CBCT image and CT image so that the CT image is sequentially mapped onto the CBCT image and calculating a CBCT and CT image pair using a forward projection algorithm;

[0008] Construct a feature connection residual subnetwork, the feature connection residual subnetwork includes eight residual dense connection blocks connected head to tail and two residual blocks, the two residual blocks are respectively connected to the head and tail ends of the eight residual dense connection blocks, the residual block includes a 1×1 convolution kernel and two 3×3 convolution kernels connected in sequence, the first half of the first residual dense connection block consists of a 1×1 convolution kernel and ReLu and a 3×3 convolution kernel and ReLu, and is feature fused with the output of the first 3×3 convolution kernel in the first residual block to extract the local part of the input data Features, the second half of the first residual dense connection block is composed of the data after feature fusion, which passes through a 3×3 convolution kernel and ReLu and is added with the output of the first residual block. The first half of the next residual dense connection block is feature fused with the data after feature fusion in the first half of the previous residual dense connection block. The second half of the next residual dense connection block is composed of the data after feature fusion, which passes through a 3×3 convolution kernel and ReLu and is added with the output of the previous residual dense connection block;

[0009] Connect multiple feature-connected residual sub-networks to each other in a cascade structure to form a projection domain cascaded feature-connected residual network;

[0010] Inputting the CBCT and CT image pairs into the projection domain cascade feature connection residual network to obtain a corrected CBCT image;

[0011] Performing fast feature extraction on the corrected CBCT image and the CBCT and CT image pair by FasterNet to obtain a feature set;

[0012] Using the CBCT image as a network input and the CT image as a label, respectively calculating the Cobi Loss and Perceptual Loss of the network input and label, and the network generation and label, and weighted combination to obtain the total loss to train the projection domain cascade feature connection residual network;

[0013] Inputting the CBCT image for testing into the trained projection domain cascade feature connection residual network to obtain the artifact-removed CBCT image;

[0014] The artifact-removed CBCT image is back-projected to obtain an artifact-removed CBCT reconstructed image.

[0015] Furthermore, when implementing the step of deformably registering the patient's CBCT image with the CT image so that the CT image is sequentially mapped onto the CBCT image to obtain a CBCT and CT image pair, the Elastix toolbox in the Slicer software is used to perform deformable registration on the CBCT image and the CT image, and the CT image is sequentially mapped onto the CBCT image to obtain the CBCT and CT image pair.

[0016] Furthermore, before implementing the step of inputting the CBCT and CT image pairs into the projection domain cascaded feature connection residual network to obtain a corrected CBCT image, the CBCT and CT image pairs are forward projected into projection data using the Ax function in the TIGRE toolbox installed on MATLAB, and then the projection data is converted into a raw format suitable for network input.

[0017] Furthermore, the calculation formula of the Perceptual Loss is:

[0018]

[0019] Wherein, a is the projection data of CBCT image;

[0020] b is the projection data of the CT image;

[0021] F i (a) and F i (b) Feature representation of the i-th layer extracted by FasterNet;

[0022] N is the number of feature layers.

[0023] Furthermore, the calculation formula of the Cobi Loss is:

[0024]

[0025] in, Distance ij Feature point a extracted by FasterNet i and feature point b j The distance between them is standardized;

[0026] (h i ,w i ) and (h j ,w j ) represents the feature point a extracted by FasterNeti and feature point b j The spatial distance between them;

[0027] N is the number of feature point sets extracted by FasterNet.

[0028] The beneficial effects of the present invention are that the CBCT artifact removal method based on the cascaded feature-connected residual network of the present invention uses deformable registration to align each set of CBCT and CT images to the same spatial coordinate system; uses a forward projection algorithm to process the image; inputs the projection image into the cascaded feature-connected residual network; obtains network-generated projection data and input CBCT-CT projection data for rapid feature extraction through the FasterNet network; calculates the CobiLoss and Perceptual Loss between the feature set generated by the network and the feature set of the input CBCT, and between the feature set generated by the network and the feature set of the input CT, respectively, optimizes the network generation by weighted loss values ​​of the two, and obtains artifact-removed projection data; finally, the generated projection data is back-projected to obtain a reconstructed image. The CBCT artifact removal method based on the cascaded feature-connected residual network of the present invention proposes model training in the projection domain, solving the problem that the trained model cannot adapt to different device protocols due to processing in the image domain.

[0029] The CBCT artifact removal method based on cascaded feature-connected residual networks of the present invention introduces a cascade structure. By downsampling and upsampling to obtain images of different sizes and connecting multiple independent network structures in series in a feature fusion manner, the cascade structure network can gradually extract and refine the features of the input data, thereby improving the performance of the overall model.

[0030] The CBCT artifact removal method based on a cascaded feature-connected residual network of the present invention constructs independent networks connected in series in a cascade structure: a feature-connected residual subnetwork, in which the residual densely connected blocks use the connections between blocks to extract and propagate features from each layer of the original image. By continuously accumulating local features and stacking them with the global features at the beginning of the network, the global dense features of the entire image are more effectively obtained.

[0031] The CBCT artifact removal method based on the cascaded feature-connected residual network of the present invention is based on Cobi Loss and Perceptual Loss. Cobi Loss uses weighted spatial awareness to consider local context similarity on the basis of improving projection data misalignment. The combination with Perceptual Loss can not only improve the problem of data misalignment but also constrain the appearance of artifacts in the generated image. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0033] Figure 1 This is a flowchart of a CBCT artifact removal method based on a cascaded feature connection residual network according to an embodiment of the present invention.

[0034] Figure 2 Schematic diagram of the structure of the projection domain cascade feature connection residual network according to an embodiment of the present invention.

[0035] Figure 3 Schematic diagram of the structure of the feature-connected residual sub-network according to an embodiment of the present invention.

[0036] Figure 4 Schematic diagram of the structure of FasterNet according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0038] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0039] Reference Figures 1 to 4 As shown, the present invention provides a CBCT artifact removal method based on a cascaded feature connection residual network, comprising the following steps:

[0040] S1. Perform deformable registration on the patient's CBCT image and CT image so that the CT image is sequentially mapped onto the CBCT image and a CBCT and CT image pair is obtained by calculation using a forward projection algorithm.

[0041] Specifically, deformable registration is performed on the CBCT and CT of the same patient, and forward projection is performed on the registered CBCT-CT pair to obtain a registered CBCT-CT projection image.

[0042] Among them, when implementing the step of performing deformable registration of the patient's CBCT image and CT image so that the CT image is sequentially mapped onto the CBCT image to obtain a CBCT and CT image pair, the Elastix toolbox in the Slicer software is used to perform deformable registration of the CBCT image and the CT image, and sequentially map the CT image onto the CBCT image to obtain a CBCT and CT image pair.

[0043] Before implementing the step of inputting the CBCT and CT image pairs into the projection domain cascade feature connection residual network to obtain the corrected CBCT image, the CBCT and CT image pairs were forward projected into projection data using the Ax function in the TIGRE toolbox installed on MATLAB, and then the projection data was converted into a raw format suitable for network input.

[0044] Step S1 of the present invention specifically includes:

[0045] Step S11: Using the Elastix toolbox in the Slicer software, deformable registration is performed on the CBCT image and the CT image of the same patient. The CT images are sequentially mapped onto the corresponding CBCT images, and the CBCT image and the CT image are resampled to the same resolution of 1×1×2.5 to obtain a registered CBCT and CT image pair.

[0046] Step S12: Use the Ax function in the TIGRE toolbox installed on MATLAB to forward project the registered images into projection data. During the projection process, each set of data is projected at intervals of 1°, ranging from 0° to 359°. A total of 360 projection images are obtained for each set of data as input to the network.

[0047] Step S13: convert the projection data in the DCM format into a raw format suitable for network input, and adjust the size of the data to 512.

[0048] S2. Construct a feature-connected residual subnetwork. The feature-connected residual subnetwork includes eight residual dense connection blocks connected end to end and two residual blocks. The two residual blocks are respectively connected to the head and tail of the eight residual dense connection blocks. The residual block includes a 1×1 convolution kernel and two 3×3 convolution kernels connected in sequence. The first half of the first residual dense connection block consists of a 1×1 convolution kernel and ReLu and a 3×3 convolution kernel and ReLu, and is fused with the output of the first 3×3 convolution kernel in the first residual block to extract local features of the input data. The second half of the first residual dense connection block is composed of the data after feature fusion through a 3×3 convolution kernel and ReLu and added to the output of the first residual block. The first half of the next residual dense connection block is fused with the data after feature fusion in the first half of the previous residual dense connection block. The second half of the next residual dense connection block is composed of the data after feature fusion through a 3×3 convolution kernel and ReLu and added to the output of the previous residual dense connection block.

[0049] Figure 3To construct a feature-connected residual subnetwork, it consists of two residual blocks at the beginning and end and eight residual dense connection blocks. The residual dense connection blocks use the connections between blocks to extract and propagate features from each layer of the original image, effectively utilizing information from all convolutional layers (both high and low). By continuously accumulating local features and stacking them with the global features at the beginning of the network, a global dense feature of the entire image can be obtained. The first half of the first residual dense connection block in the feature connection residual subnetwork consists of a 1×1 convolution kernel and ReLu and a 3×3 convolution kernel and ReLu. After that, feature fusion is performed with the output of the first 3×3 convolution kernel in the first residual block to extract the local features of the input data. The second half of the first residual dense connection block is composed of the data after feature fusion through a 3×3 convolution kernel and ReLu, and then added to the output of the first residual block. The first half of the second residual dense connection block is composed of a 1×1 convolution kernel and ReLu and a 3×3 convolution kernel and ReLu. After that, feature fusion is performed with the data after feature fusion in the first half of the first residual dense connection block. The second half of the second residual dense connection block is composed of the data after feature fusion through a 3×3 convolution kernel and ReLu, and then added to the output of the first residual dense connection block. The above operation is repeated for subsequent residual dense connection blocks. Figure 3The specific network implementation process is as follows: a 3×4×4 size data is input into the feature connection residual sub-network, after the first residual block, the data output by the second 3×3 convolution kernel and the output of the first 1×1 convolution kernel are added to obtain a 128×4×4 size data, and then input into the residual dense connection block 1, after a 1×1 convolution kernel, a 3×3 convolution kernel and two ReLUs, the output 32×4×4 size data is combined with the 128×4×4 size data output by the first 3×3 convolution kernel in the first residual block for the third dimension feature fusion to obtain 160(32+128)×4×4 size data, after the second 3×3 convolution kernel and ReLU, the 128×4×4 size data output by the first residual block is added to the 128×4×4 size data output by the first residual block and input into the residual dense connection block 2, and after The 32×4×4 data output after passing through a 1×1 convolution kernel, a 3×3 convolution kernel, and two ReLUs is combined with the 160×4×4 data output from the feature fusion in the residual dense connection block 1 for feature fusion of the third dimension to obtain 192(32+160)×4×4 data. After that, the 128×4×4 data output after passing through the second 3×3 convolution kernel and ReLU is added to the 128×4×4 data output from the residual dense connection block 1 and input into the residual dense connection block 3. The above operation is repeated until the residual dense connection block 8, and the output data of 128×4×4 enters the last residual block. The data output by the second 3×3 convolution kernel and the output of the first 1×1 convolution kernel are added to obtain data of 128×4×4, which is the final output of the first feature densely connected residual network.

[0050] S3. Connect multiple feature-connected residual sub-networks to each other in a cascade structure to form a projection domain cascade feature-connected residual network.

[0051] Figure 2 In order to construct a projection domain cascade feature connection residual network, the network is a cascade structure that connects multiple feature connection residual sub-networks constructed in step S2. Specifically, the image size of 512 is gradually downsampled to 256, 128, 64, 32, 16, 8 and 4 by downsampling. The small-scale images are input into the feature connection residual sub-network in turn, fused with the higher-scale images after upsampling, and input into the feature connection residual sub-network again. Figure 2The specific network operation process: The input image size is 3×512×512 and is gradually downsampled to 256, 128, 64, 32, 16, 8, and 4. The feature densely connected residual network first operates on the smallest size 3×4×4 image to a size of 128×8×8. Then, through upsampling, the image size is changed to 128×8×8 to facilitate feature fusion with the 3×8×8 input image. This process is then repeated, resulting in an image of size 128×512×512. Finally, a 3×3 convolution kernel is used to resize the image back to its original size of 3×512×512. Both downsampling and upsampling are performed using bilinear interpolation.

[0052] S4. Input the CBCT and CT image pairs into the projection domain cascade feature connection residual network to obtain the corrected CBCT image.

[0053] Figure 4 In order to build a FasterNet fast feature extraction network suitable for the present invention, ordinary feature extraction networks will perform convolution on all input channels, which will lead to unnecessary calculations and memory access, greatly increasing the running time.

[0054] The present invention uses a modified FasterNet for fast feature extraction. FasterNet applies convolution operations to only some channels in the input image through PConv. Medical image data is single-channel and has to be replicated into three channels in order to adapt to the feature extraction network. In essence, the data of each layer of channels are the same, so there is no need to perform convolution for each layer of channels. This operation can effectively reduce computational redundancy and improve computational efficiency. Since only features are extracted and no classification tasks are performed, only the subsequent pooling and fully connected layer parts of FasterNet are cancelled, and the other parts remain unchanged.

[0055] Figure 4 Specific network operation process:

[0056] The 3×512×512 size data passes through an Embedding layer (regular 4×4 convolution with a stride of 4) to obtain 96×128×128 size data, and then passes through FasterNetBlock1. Each FasterNetBlock consists of a PConv layer, followed by two 1×1Conv layers to form a residual block, and the output is 96×128×128 size. After passing through a Merging layer (regular 2×2 convolution with a stride of 2), the output is 192×64×64 size. After that, three FasterNetBlocks and two Merging layers are repeated to output a feature map of size 768×16×16.

[0057] The registered CBCT projection image is input as input data into the constructed projection domain cascade feature connection residual network to obtain the corrected CBCT projection image initially generated by the network.

[0058] S5. FasterNet is used to perform fast feature extraction on the corrected CBCT image and the CBCT and CT image pairs to obtain feature sets.

[0059] The constructed FasterNet is used to perform rapid feature extraction on the CBCT projection images generated by the network and the registered CBCT and CT projection images to obtain the corresponding feature sets.

[0060] S6. Use CBCT images as network input and CT images as labels. Calculate the Cobi Loss and Perceptual Loss of network input and labels and network generation and labels respectively, and combine them to obtain the total loss to train the projection domain cascade feature connection residual network.

[0061] The calculation formula of Perceptual Loss is:

[0062]

[0063] Wherein, a is the projection data of CBCT image;

[0064] b is the projection data of the CT image;

[0065] F i (a) and F i (b) Feature representation of the i-th layer extracted by FasterNet;

[0066] N is the number of feature layers.

[0067] Specifically, Cobi Loss and Perceptual Loss are calculated. Since the present invention uses deformable registration in data preprocessing, this will cause slight stretching deformation of the image data and thus lead to spatial misalignment between the projection data. Cobi Loss uses weighted spatial awareness to consider local context similarity on the basis of improving the misalignment of projection data. Experiments have found that if feature point matching is performed without considering the spatial structure, many artifacts will appear; the goal of Perceptual Loss is to minimize the distance between the input image and the target image in the feature space. The combination of the two can not only improve the problem of data misalignment but also constrain the generation of image artifacts.

[0068] Assume that the input CBCT projection data is a and the CT projection data is b. Then F i (a) and F i(b) is the feature representation of the i-th layer extracted by FasterNet, N is the number of feature layers, and the calculation formula of Perceptual Loss is as follows:

[0069]

[0070] In this embodiment, the calculation formula of Cobi Loss is:

[0071]

[0072] in, Distance ij Feature point a extracted by FasterNet i and feature point b j The distance between them is standardized;

[0073] (h i ,w i ) and (h j ,w j ) represents the feature point a extracted by FasterNet i and feature point b j The spatial distance between them;

[0074] N is the number of feature point sets extracted by FasterNet.

[0075] Specifically, let the feature point sets extracted by FasterNet be a i and b j Then, the cosine distance d between the two sets is defined as ij as follows:

[0076]

[0077] Where, β b It can be derived from the following formula, where N is the number of feature point sets extracted by FasterNet:

[0078]

[0079] The cosine distance is normalized using the following formula: ik The value is the smallest, indicating a i with b j The distance between them is the shortest, and ε is to prevent the denominator of the fraction from being 0:

[0080]

[0081] The distance is exponentiated using the following formula, where h is set to 0.5:

[0082]

[0083] The distance between points is normalized using the following formula:

[0084]

[0085] In summary, the calculation formula of Cobi Loss can be obtained:

[0086]

[0087] in (h i ,w i ) and (h j ,w j ) represents feature a i and b j The spatial distance between them.

[0088] The loss function in the experiment is:

[0089]

[0090] Where G(a) represents the image generated by the cascaded feature densely connected residual network based on the input CBCT projection data, a represents the features of the original CBCT projection data obtained by the feature extraction network, b represents the features of the CT projection data obtained by the feature extraction network, and θ is a constant used to balance the loss functions of both parties. After experimental verification, θ = 5 is selected.

[0091] Using CBCT projection images as network input and CT projection images as labels, the Cobi Loss and Perceptual Loss of the input and labels, as well as the network output and labels, are calculated using the formulas. The resulting weighted total loss is then used to continuously optimize the network training process.

[0092] S7. Input the CBCT image for testing into the trained projection domain cascade feature connection residual network to obtain the artifact-removed CBCT image.

[0093] S8. Obtain an artifact-removed CBCT reconstructed image by back-projecting the artifact-removed CBCT image.

[0094] Specifically, the artifact-removed CBCT projection image is back-projected using the Ax function in the TIGRE toolbox installed on MATLAB to obtain the artifact-removed CBCT reconstructed image.

[0095] The CBCT artifact removal method based on the cascaded feature connection residual network of the present invention proposes model training in the projection domain, which solves the problem that the trained model cannot be adapted to different device protocols due to processing in the image domain.

[0096] The CBCT artifact removal method based on cascaded feature-connected residual networks of the present invention introduces a cascade structure. By downsampling and upsampling to obtain images of different sizes and connecting multiple independent network structures in series in a feature fusion manner, the cascade structure network can gradually extract and refine the features of the input data, thereby improving the performance of the overall model.

[0097] The CBCT artifact removal method based on a cascaded feature-connected residual network of the present invention constructs independent networks connected in series in a cascade structure: a feature-connected residual subnetwork, in which the residual densely connected blocks use the connections between blocks to extract and propagate features from each layer of the original image. By continuously accumulating local features and stacking them with the global features at the beginning of the network, the global dense features of the entire image are more effectively obtained.

[0098] The CBCT artifact removal method based on the cascaded feature-connected residual network of the present invention is based on Cobi Loss and Perceptual Loss. Cobi Loss uses weighted spatial awareness to consider local context similarity on the basis of improving projection data misalignment. The combination with Perceptual Loss can not only improve the problem of data misalignment but also constrain the appearance of artifacts in the generated image.

[0099] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention herein is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in this application.

Claims

1. A CBCT artifact removal method based on cascaded feature-connected residual network, characterized in that: The following steps are involved: Performing deformable registration on the patient's CBCT image and CT image so that the CT image is sequentially mapped onto the CBCT image and calculating a CBCT and CT image pair using a forward projection algorithm; Construct a feature connection residual subnetwork, the feature connection residual subnetwork includes eight residual dense connection blocks connected head to tail and two residual blocks, the two residual blocks are respectively connected to the head and tail ends of the eight residual dense connection blocks, the residual block includes a 1×1 convolution kernel and two 3×3 convolution kernels connected in sequence, the first half of the first residual dense connection block consists of a 1×1 convolution kernel and ReLu and a 3×3 convolution kernel and ReLu, and is feature fused with the output of the first 3×3 convolution kernel in the first residual block to extract the local part of the input data Features, the second half of the first residual dense connection block is composed of the data after feature fusion, which passes through a 3×3 convolution kernel and ReLu and is added with the output of the first residual block. The first half of the next residual dense connection block is feature fused with the data after feature fusion in the first half of the previous residual dense connection block. The second half of the next residual dense connection block is composed of the data after feature fusion, which passes through a 3×3 convolution kernel and ReLu and is added with the output of the previous residual dense connection block; Connect multiple feature-connected residual sub-networks to each other in a cascade structure to form a projection domain cascade feature-connected residual network; Inputting the CBCT and CT image pairs into the projection domain cascade feature connection residual network to obtain a corrected CBCT image; Performing fast feature extraction on the corrected CBCT image and the CBCT and CT image pair by FasterNet to obtain a feature set; Using the CBCT image as a network input and the CT image as a label, respectively calculating the Cobi Loss and Perceptual Loss of the network input and label, and the network generation and label, and weighted combination to obtain the total loss to train the projection domain cascade feature connection residual network; Inputting the CBCT image for testing into the trained projection domain cascade feature connection residual network to obtain the artifact-removed CBCT image; The artifact-removed CBCT image is back-projected to obtain an artifact-removed CBCT reconstructed image.

2. The CBCT artifact removal method based on cascaded feature-connected residual network according to claim 1, characterized in that: When implementing the step of deformably registering the patient's CBCT image with the CT image so that the CT image is sequentially mapped onto the CBCT image to obtain a CBCT and CT image pair, the Elastix toolbox in the Slicer software is used to perform deformable registration on the CBCT image and the CT image, and the CT image is sequentially mapped onto the CBCT image to obtain the CBCT and CT image pair.

3. The CBCT artifact removal method based on cascaded feature-connected residual network according to claim 1, characterized in that: Before implementing the step of inputting the CBCT and CT image pairs into the projection domain cascaded feature connection residual network to obtain a corrected CBCT image, the CBCT and CT image pairs are forward projected into projection data using the Ax function in the TIGRE toolbox installed on MATLAB, and then the projection data is converted into a raw format suitable for network input.

4. The CBCT artifact removal method based on cascaded feature-connected residual network according to claim 1, characterized in that: The calculation formula of the Perceptual Loss is: Wherein, a is the projection data of CBCT image; b is the projection data of the CT image; F i (a) and F i (b) Feature representation of the i-th layer extracted by FasterNet; N is the number of feature layers.

5. The CBCT artifact removal method based on cascaded feature-connected residual network according to claim 4 is characterized in that: The calculation formula of the Cobi Loss is: in, Distance ij Feature point a extracted by FasterNet i and feature point b j The distance between them is standardized; (h i ,w i ) and (h j ,w j ) represents the feature point a extracted by FasterNet i and feature point b j The spatial distance between them; N is the number of feature point sets extracted by FasterNet.

Images