Four-dimensional cone beam CT image reconstruction method and system, computer device and storage medium

Abstract

The application discloses a four-dimensional cone beam CT image reconstruction method and system, computer equipment and a storage medium, which are applied to the technical field of image processing and have the method comprising the following steps: acquiring sparse projection data and corresponding angle information of a to-be-reconstructed object at each respiratory phase; inputting the sparse projection data and the corresponding angle information into a projection domain generation module to generate space-time enhanced features that fuse space geometric information and respiratory motion information, decoding the space-time enhanced features, and outputting dense projection data at each respiratory phase; analyzing and reconstructing the dense projection data to obtain an intermediate volume image; inputting the intermediate volume image into an image domain correction module for correction, and outputting a four-dimensional cone beam CT image; and through the dual-domain collaborative strategy of combining projection domain completion and image domain correction, the application realizes high-fidelity and low-dose four-dimensional cone beam CT image reconstruction while effectively maintaining the spatial continuity of anatomical structures and the time continuity of respiratory motion.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, system, computer device, and storage medium for four-dimensional cone-beam CT image reconstruction based on spatiotemporal consistency coding. Background Technology

[0002] In image-guided radiotherapy (IGRT), cone-beam computed tomography (CBCT) is widely used for patient positioning verification. However, CBCT imaging of the chest and abdomen is susceptible to the effects of respiratory motion, producing motion artifacts and blurring, reducing the accuracy of target localization. Four-dimensional CBCT (4D CBCT) technology effectively compensates for respiratory motion by grouping and reconstructing acquired X-ray projections according to respiratory phases to obtain a time-resolved three-dimensional image sequence. However, the core challenge of this technology lies in the fact that after distributing the total number of projections across multiple respiratory phases (typically 10), each phase has only a very small number of sparse spectral projections, leading to severe streak artifacts, structural defects, and noise in the reconstructed images.

[0003] Reference patent application number CN118015116A discloses a four-dimensional cone-beam CT image reconstruction method based on projection-image domain co-enhancement. Its specific content includes: obtaining computed tomography cone-beam CT data and raw projection data from a public database, and dividing the dataset into training, validation, and test datasets; constructing a projection domain enhancement, image domain enhancement, and motion compensation-based method. A co-enhanced four-dimensional cone-beam CT image reconstruction network, PICE, is constructed, and its loss function L is defined. The training dataset is input into the reconstruction network PICE, and the network is trained by backpropagation using the loss function L. The best-performing training model is selected using the validation dataset. The reconstruction results are obtained by inputting the test dataset into the selected training model.

[0004] Although this method employs a dual-domain strategy, its projection domain enhancement network mainly relies on the sliding window self-attention module to capture global information, failing to explicitly model the physical geometric relationships of the projection acquisition (such as the angular intervals on the circular trajectory). At the same time, its image domain enhancement and motion compensation networks process each phase image relatively independently, failing to fully utilize the temporal continuity information between different phases within the respiratory cycle. As a result, under extremely sparse projection conditions, the motion details and structural integrity of the reconstructed image still need to be improved.

[0005] To overcome these shortcomings, this application proposes a four-dimensional cone-beam CT image reconstruction method, system, computer equipment, and storage medium. Summary of the Invention

[0006] The purpose of this application is to provide a four-dimensional cone-beam CT image reconstruction method, system, computer equipment, and storage medium, aiming to solve the above-mentioned problems.

[0007] To achieve the above objectives, this application provides the following technical solution: In a first aspect, this application provides a method for reconstructing four-dimensional cone-beam CT images, the steps of which include: Obtain sparse projection data and corresponding angle information of the object to be reconstructed in each respiratory phase; The sparse projection data and corresponding angle information are input into the projection domain generation module to generate spatiotemporal enhancement features that fuse spatial geometric information and respiratory motion information. The spatiotemporal enhancement features are then decoded to output dense projection data for each respiratory phase. The dense projection data is analyzed and reconstructed to obtain an intermediate volume image; The intermediate volume image is input into the image domain correction module for correction, and a four-dimensional cone-beam CT image is output.

[0008] Secondly, this application provides a four-dimensional cone-beam CT image reconstruction system, specifically including: The data acquisition module is used to acquire sparse projection data of the object to be reconstructed in each respiratory phase and the corresponding angle information; The projection domain generation module is used to generate spatiotemporal enhancement features that fuse spatial geometric information and respiratory motion information based on the sparse projection data and the corresponding angle information, and to decode the spatiotemporal enhancement features to output dense projection data for each respiratory phase. The analytical reconstruction module is used to perform analytical reconstruction on the dense projection data to obtain an intermediate volume image; The image domain correction module is used to correct the intermediate volume image and output a four-dimensional cone-beam CT image.

[0009] Thirdly, this application provides a computer device, the computer device including a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing a four-dimensional cone-beam CT image reconstruction method; the processor is used to execute the program instructions stored in the memory to implement a four-dimensional cone-beam CT image reconstruction.

[0010] Fourthly, this application provides a computer-readable storage medium storing processor-executable program instructions for performing a four-dimensional cone-beam CT image reconstruction method.

[0011] This application provides a method, system, computer equipment, and storage medium for four-dimensional cone-beam CT image reconstruction, which has the following beneficial effects: This application first expands sparse projection data into dense projection data through a projection domain generation module, alleviating the image artifact problem caused by insufficient angle sampling from the data source. During the expansion process, angle information is encoded using an encoding method that preserves circumferential geometric characteristics, enabling the network to explicitly perceive the circumferential geometric relationships between projections, thereby more accurately recovering the projection content of missing angles. Simultaneously, by performing temporal modeling of spatial features to generate time-step encoding, the network can utilize motion continuity information across breathing phases to improve the reconstruction quality of motion regions. Based on this, the generated dense projection data is parsed and reconstructed to obtain an intermediate volume image, which is then refined through an image domain correction module to further eliminate residual artifacts and enhance structural details. Compared with existing technologies, this application achieves dual optimization of projection data and image data through the cascaded collaboration of projection domain generation and image domain correction, effectively suppressing stripe artifacts caused by sparse viewpoints, improving the reconstruction accuracy and temporal coherence of motion regions, and avoiding the high computational overhead and excessive smoothing problems of iterative reconstruction by adopting a feedforward network architecture, thereby improving reconstruction efficiency while ensuring image quality. Attached Figure Description

[0012] Figure 1 This is a schematic flowchart of a four-dimensional cone-beam CT image reconstruction method according to Embodiment 1 of this application; Figure 2 This is a flowchart of a four-dimensional cone-beam CT image reconstruction method according to Embodiment 1 of this application; Figure 3 This is a schematic diagram of the structure of a four-dimensional cone-beam CT image reconstruction system according to Embodiment 2 of this application; Figure 4 This is a schematic diagram of the computer device structure according to Embodiment 3 of this application; Figure 5 This is a schematic diagram of the storage medium structure of Embodiment 4 of this application. Detailed Implementation

[0013] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0014] The following analysis, based on relevant technologies, examines existing solutions.

[0015] Currently, 4D CBCT reconstruction methods are mainly divided into three categories. The first category is motion estimation and compensation-based methods, which fuse cross-phase information by extracting deformation vector fields (DVF) between different phases. However, the accuracy of DVF estimation decreases when the respiratory amplitude and cycle are highly irregular, and it may introduce pseudo-structures or motion blur. The second category is iterative reconstruction methods based on regularization constraints, which introduce prior knowledge through regularization terms such as total variation (TV). However, excessive regularization may cause the target area motion to exceed the preset range, and the computational cost is high. The third category is deep learning-based methods, which use convolutional neural networks to learn spatial priors to compensate for undersampled information. However, most methods only operate in a single domain (projection domain or image domain), lack explicit modeling of the projection physical geometry, and do not fully utilize the temporal continuity of the respiratory cycle. Under extremely sparse conditions, structural loss and artifact residue problems still exist. In recent years, position coding technology has shown the potential to enhance the high-frequency detail learning ability of networks in implicit representation methods such as neural radiation fields. However, existing work on applying position coding to CT reconstruction is mostly aimed at static 3D scenes, without considering the circular geometric characteristics of the projection angle and the time periodicity of respiratory motion in 4D CBCT.

[0016] The closest approaches to this application mainly include: deep learning-based projection domain completion or image domain post-processing methods, such as CycN-Net which uses ungrouped original projections as global prior input networks to enhance image fidelity, and AttUNet which uses an attention mechanism to suppress image domain artifacts. However, these methods either operate only in the projection domain and ignore structural optimization in the image domain, or operate only in the image domain and lack constraints on the physical processes of projection, and none of them explicitly model the geometric relationship of the projection angle and the temporal dynamics of the breathing phase. Reconstruction methods based on implicit neural representations, such as NAF and NeRP, achieve volume recovery through continuous field modeling, but require sample-by-sample optimization, have high computational costs, lack explicit prior constraints, and their performance is limited under extremely sparse conditions.

[0017] Traditional analytical reconstruction (FDK) suffers from severe artifacts under sparse perspectives, with high-density structures such as ribs and spine appearing significantly blurred and broken, and soft tissue boundaries almost indistinguishable. While iterative reconstruction (SART, SART-TV) can partially alleviate artifacts, SART introduces high-frequency noise during iteration, leading to salt-and-pepper artifacts in low-contrast areas. The excessive smoothing effect of TV regularization causes blurring or even loss of fine structures such as vascular branches and small pulmonary nodules. Existing deep learning methods have the following shortcomings: First, most methods operate only in a single domain. The projection after completion by projection domain methods still requires FDK reconstruction, and residual artifacts cannot be further eliminated in the image domain. Image domain methods directly post-process the FDK reconstruction results containing severe artifacts, resulting in low input quality that makes it difficult for the network to recover the true structure. Second, there is a lack of explicit modeling of the physical geometry of the projection. Existing networks treat projections at different angles as ordinary multi-channel inputs, failing to utilize the geometric relationship of the projection angles on the circular trajectory, making it difficult for the network to accurately infer the content of the projection at missing angles. Third, the temporal continuity of the respiratory cycle was not fully utilized. Each phase was processed independently, ignoring the continuous changes in anatomical structures between adjacent phases and wasting valuable cross-phase complementary information.

[0018] Therefore, this application proposes a four-dimensional cone-beam CT image reconstruction method, system, computer device, and storage medium. It explicitly captures the circumferential geometric relationship of the projection angle through position encoding based on physical process modeling to ensure spatial consistency, captures the dynamic relationship of the respiratory cycle across phases through time step encoding based on temporal modeling to ensure temporal consistency, and achieves end-to-end reconstruction from sparse projection to high-quality 4D CBCT volumetric images through a cascaded dual-domain strategy of projection domain generation and image domain correction.

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] Example 1 Please see Figure 1 This is a flowchart illustrating a four-dimensional cone-beam CT image reconstruction method according to Embodiment 1 of this application; the steps include: S1: Obtain the sparse projection data of the object to be reconstructed in each breathing phase and the corresponding angle information.

[0021] In this embodiment, sparse projection data of the object to be reconstructed under each respiratory phase and acquisition angle information corresponding to each projection are obtained. These data are used as input to the subsequent projection domain generation module to generate dense projection data and finally complete high-quality four-dimensional cone-beam CT image reconstruction.

[0022] Depending on the data source, two typical implementation methods are adopted: one is a simulated acquisition method based on public datasets, used for algorithm development and verification; the other is a real acquisition method based on clinical equipment, used for actual clinical applications. The following sections provide detailed explanations of each method.

[0023] First, high-quality 4D CT image data are obtained from publicly available databases. Taking the TCIA 4D lung cancer dataset as an example, this dataset contains 4D CT images of multiple patients. Each set of images consists of 3D volumetric images of 10 respiratory phases. The image of each phase clearly shows the anatomical structures at that moment, including bones, soft tissues, lungs, and tumors. These images can serve as a reference standard for reconstruction and are used for subsequent quantitative assessment.

[0024] Secondly, forward projection simulation is performed on the three-dimensional volumetric image of each respiratory phase to generate corresponding full-sample projection data. The forward projection simulation uses cone-beam CT scanning geometry parameters, specifically including: the distance from the X-ray source to the center of rotation, the distance from the source to the detector, the detector size, and pixel size. During the simulation, 256 projection angles are acquired at equal intervals within a 360° scan range, meaning one projection image is acquired approximately every 1.4° of rotation. Using a ray-driven forward projection algorithm, the volumetric image of each respiratory phase is projected, resulting in 256 projection images, forming a full-sample projection set. The size of each projection image is determined by the detector pixel array, for example, 1024×768 pixels.

[0025] A subset is extracted from the full sampled projection set at fixed angular intervals to simulate a sparse sampling clinical scenario. In this embodiment, the number of sparse projections per phase is set to 16, that is, 16 are uniformly extracted from 256 equally spaced projection angles, with an extraction interval of 22.5°. Specifically, projection images corresponding to angle values ​​of 0°, 22.5°, 45°, ..., 337.5° are selected to form the sparse projection set for each respiratory phase. This uniform sampling method ensures a uniform distribution of projection angles on the circumference, providing a foundation for subsequent geometric relationship modeling.

[0026] Simultaneously, the acquisition angle information corresponding to each selected projection is recorded. For each projection angle, in addition to recording its angle value itself, it is also converted into a normalized coordinate form on the rotation plane, i.e., a two-dimensional vector composed of sine and cosine values. This representation can explicitly preserve the circular geometry of the acquisition trajectory and accurately reflect the spatial positional relationship between different projection angles. The angle information of all selected projections is organized into a two-dimensional matrix, where each row corresponds to the normalized coordinates of a projection angle. Finally, the sparse projection image set of each respiratory phase is associated with and stored with the corresponding angle information matrix as input data for subsequent processing. The sparse projection data of all 10 respiratory phases together constitute the complete input dataset.

[0027] In clinical practice, data acquisition is performed simultaneously with the imaging scan. First, the patient lies on the treatment bed with free breathing, and the cone-beam CT imaging system rotates 360° around the patient for scanning. During the scan, projected images are continuously acquired at a fixed frame rate, for example, 10 frames per second, resulting in approximately 600 projected images over a 60-second scan. For each projected image acquired, the imaging system automatically records the current gantry rotation angle, i.e., the acquisition angle information corresponding to that projection. Simultaneously, respiratory monitoring equipment records the patient's respiratory signals. Commonly used respiratory monitoring methods include abdominal pressure sensors, infrared tracking systems, or optical surface monitoring systems. These devices record respiratory waveforms at the same sampling frequency as the projection acquisition, obtaining a respiratory signal sequence reflecting the rise and fall of the patient's chest and abdomen. The respiratory signal typically exhibits a periodic waveform, with a peak at the end of inspiration and a trough at the end of expiration.

[0028] Phase registration and grouping are performed on the acquired projection images and respiratory signals. First, the respiratory signals are processed to extract the respiratory phase value corresponding to each sampling moment. Common phase extraction methods include the Hilbert transform method and amplitude binning. Taking the phase binning method as an example, the respiratory waveform is converted into an instantaneous phase sequence through the Hilbert transform, and the phase value at each projection moment is normalized to... Within the range, and Corresponding to the end of inhalation, This corresponds to the end of expiration. Then, the entire respiratory cycle is evenly divided into multiple phase intervals. In this embodiment, it is divided into 10 phase intervals, each corresponding to a respiratory phase. Based on the respiratory phase value corresponding to the acquisition time of each projected image, it is assigned to the corresponding phase group. Due to the irregularity of the respiratory cycle, the number of projected images in each phase group may not be exactly equal, but with a sufficiently long scanning time, each phase group can usually obtain a sufficient number of projections.

[0029] To meet the requirements of sparse reconstruction, a fixed number of projection images are further selected from each breathing phase group. In this embodiment, 16 projections are selected for each phase. The selection principle is to prioritize projections that are evenly distributed along the circumferential angle to ensure geometric consistency in subsequent reconstruction. For example, the projection frames with the closest angular interval of 22.5° can be found in each phase group to form a sparse projection set for that phase. For each selected projection frame, in addition to retaining its projection image data, its acquisition angle information also needs to be recorded. Similarly, the angle values ​​are converted into normalized coordinates of the rotating plane, i.e., sine and cosine values, to preserve the circumferential geometric characteristics. All projection images also need to undergo necessary preprocessing before being input into the network, including dark field correction and gain correction to eliminate the influence of detector noise, negative logarithmic transformation to convert transmission intensity into line integral values, and size adjustment to adapt to the network input requirements.

[0030] After the above processing, a sparse projection image set for each respiratory phase and its corresponding angle information matrix are finally obtained. These data will be organized into a unified input format and input into the subsequent projection domain generation module for further processing. Depending on clinical needs and imaging conditions, the number of respiratory phases can be selected from 6 to 12, the number of projections per phase can be selected from 8 to 32, and the angle sampling method can also adopt different strategies such as equal interval sampling, golden angle sampling, or adaptive sampling, depending on the actual situation.

[0031] S2: Input the sparse projection data and the corresponding angle information into the projection domain generation module to generate spatiotemporal enhancement features that fuse spatial geometric information and respiratory motion information, and decode the spatiotemporal enhancement features to output dense projection data for each respiratory phase.

[0032] In this embodiment, the projection domain generation module adopts an encoder-decoder network architecture. By introducing a spatiotemporal consistency coding mechanism, it transforms sparse projection data and its angle information into spatiotemporal enhanced features containing spatial geometric information and respiratory motion information. After decoding, it outputs dense projection data for each respiratory phase. Specifically, it includes steps S21 to S24, which are described in detail below.

[0033] S21: Extract features from the sparse projection data to obtain spatial features.

[0034] The sparse projection data and corresponding angle information of each respiratory phase obtained in step S1 are input into the projection domain generation module. The projection domain generation module uses an encoder network as its front end to extract features from the input sparse projection data. The encoder network adopts a multi-scale hierarchical structure, extracting feature representations from the sparse projection image in progressively deeper layers through layer-by-layer convolution and downsampling operations, thus obtaining the spatial features corresponding to each respiratory phase. These spatial features contain both anatomical structural information from the projection image and preserve the intensity distribution characteristics of the projection data itself. During feature extraction, the encoder network can retain detailed information at different scales through skip connections and other methods, providing a rich feature foundation for subsequent feature fusion and decoding.

[0035] S22: Map the angle information to normalized coordinates of the rotation plane, where the normalized coordinates are represented by sine and cosine values. Perform multi-frequency Fourier feature mapping on the normalized coordinates to generate a position code that characterizes the circumferential geometric relationship between projections.

[0036] The projection domain generation module enhances features from two dimensions: spatial geometry and respiratory motion. In the spatial geometry dimension, the projection domain generation module generates position codes based on the input projection angle information. Specifically, for each projection, its acquisition angle is converted into normalized coordinates on the rotation plane. This coordinate representation can explicitly preserve the geometric positional relationship of the projection on the circular scanning trajectory.

[0037] Then, Fourier feature mapping is used to increase the dimension of the normalized coordinates and the angles. Convert to normalized coordinates Normalized coordinates are input into the Fourier feature mapping function for position encoding. Defined as: ; In this system, the frequency order L of the spatial coordinates is set to 10, and the frequency order L of the angular coordinates is set to 4; these are mapped to a high-dimensional feature space, thereby improving the network's ability to express high-frequency spatial changes.

[0038] The location encoding has the same dimension as the spatial features extracted by the encoder, and can inject geometric information into the feature representation through subsequent fusion operations.

[0039] S23: Arrange the spatial features in chronological order to form a temporal feature sequence, input the temporal feature sequence into a long short-term memory network, and obtain a time step code for characterizing respiratory motion dynamics through the long short-term memory network.

[0040] Spatial features corresponding to different respiratory phases are arranged chronologically to form a temporal feature sequence, which is then input into a Long Short-Term Memory (LSTM) network. The LTM network learns the dynamic patterns of respiratory movements by capturing the continuous changes in anatomical structures between adjacent phases, generating time-step codes that characterize the dynamics of respiratory movements. This time-step encoding reflects the anatomical state and its changing relationships at different moments in the respiratory cycle. LSTM operates in a compact feature embedding space, avoiding the gradient instability and computational inefficiency caused by directly applying cyclic operations to high-resolution projections, while effectively capturing the temporal continuity of respiratory motion.

[0041] S24: The spatial features, the location encoding, and the time step encoding are added element by element to generate the spatiotemporal enhancement features.

[0042] By employing an element-wise additive approach, positional and temporal encodings are superimposed on spatial features at the same locations, simultaneously incorporating geometric relationships between projections and dynamic motion information across phases into the spatial features. This fusion method is simple and efficient, enabling the network to explicitly perceive the geometric constraints of projection acquisition and the temporal continuity of breathing motion, thus allowing for more accurate inference of the projection content at missing angles during subsequent decoding. The resulting spatiotemporally enhanced features contain not only the content information of the original projection data but also embedded prior knowledge of spatial geometric consistency and temporal dynamic consistency.

[0043] Finally, the spatiotemporal augmentation features are input into the decoder network for decoding, generating dense projection data for each breathing phase. The decoder network employs a hierarchical structure symmetrical to the encoder, recovering the spatiotemporal augmentation features into a high-resolution projection image through progressive upsampling and feature recombination. During decoding, features from the corresponding layers of the encoder are introduced through skip connections to preserve more detailed information. The output layer of the decoder network uses an appropriate activation function to ensure that the output projection data has a reasonable numerical range. For each breathing phase, the projection domain generation module outputs a set of dense angle projection data, with a number of projections far exceeding the number of sparse projections input. For example, if the input is 16 sparse projections per phase, the output can be 256 dense projections per phase, covering continuous angles within a 360° range.

[0044] In this embodiment, the projection domain generation module optimizes its parameters through end-to-end joint training, employing a method that includes mean squared error loss during the training process. Structural similarity loss and gradient consistency loss A joint loss function is used to improve the fidelity, structural integrity, and spatial continuity of the generated projection. Specifically, the weighted combination of the three losses is: ,in , The weighting coefficients were determined through experimental optimization and are preferred in this embodiment. .

[0045] The above steps complete the conversion from sparse projection to dense projection. The output dense projection data will serve as the input for subsequent analytical reconstruction, laying the foundation for finally obtaining high-quality four-dimensional cone-beam CT images.

[0046] S3: Analyze and reconstruct the dense projection data to obtain an intermediate volume image.

[0047] In this embodiment, the analytical reconstruction adopts the filtered back projection algorithm, which has the advantages of high computational efficiency and simple implementation.

[0048] Specifically, for each respiratory phase, the projection domain generation module outputs a set of dense angular projection data. This projection data has undergone network enhancement processing, resulting in denser angular sampling and higher image quality compared to the original sparse projection, with preliminary suppression of fringe artifacts. Filtered backprojection reconstruction is then performed on the dense projection data for each respiratory phase. The reconstruction process is carried out on a pre-defined three-dimensional volumetric grid, the size and resolution of which are determined according to clinical needs. Taking chest and abdominal imaging as an example, the reconstruction volume can be set to 512×512×120 voxels with a voxel spacing of 1×1×2 mm, covering the patient's anatomical structures.

[0049] The specific implementation process of filtered backprojection reconstruction is as follows: For each projection angle, the projection image is first weighted to compensate for the influence of cone-beam scanning geometry; then, the weighted projection data is filtered using a ramp filter or its variant (such as a filter kernel with a window function) to enhance high-frequency information and suppress low-frequency artifacts; finally, the filtered projection data is backprojected along the ray path into a three-dimensional volume grid, and the contribution values ​​of all rays passing through that voxel are accumulated for each voxel. The above operation is repeated for all projection angles to finally obtain the three-dimensional volume image of the respiratory phase.

[0050] During implementation, the specific parameters of the filtered back-projection algorithm can be set according to the imaging geometry. These scanning geometric parameters include the distance from the X-ray source to the rotation center, the distance from the source to the detector, the detector size, and pixel size. These parameters have already been determined during the forward projection simulation or clinical acquisition in step S1. During reconstruction, the spatial position and direction of each X-ray need to be accurately calculated based on these geometric parameters to ensure the geometric accuracy of the reconstructed image.

[0051] S4: Input the intermediate volume image into the image domain correction module for correction, and output a four-dimensional cone-beam CT image.

[0052] In this embodiment, the intermediate volumetric image obtained in step S3 is input into the image domain correction module. Through further processing, residual artifacts are removed and anatomical details are enhanced, ultimately outputting a high-quality four-dimensional cone-beam CT image. The image domain correction module employs an attention-based encoder-decoder convolutional network architecture for layer-by-layer correction of the three-dimensional volumetric image.

[0053] Specifically, the image domain correction module employs an encoder-decoder network structure, taking a 3D volumetric image as input and outputting a corrected image after layer-by-layer processing. In the encoder section, the network progressively extracts deep features of the image through a series of convolution and downsampling operations, capturing structural information and artifact patterns at different scales. In the decoder section, the network restores the deep features to the original resolution image through upsampling and feature reconstruction operations, while simultaneously generating the corrected output.

[0054] Between the encoder and decoder, an image domain correction module introduces an attention gating mechanism. This mechanism enables the network to adaptively focus on key regions in the image, particularly the locations of residual artifacts and structural edges. The attention gating unit generates an attention weight map through learning, reflecting the urgency of correction at different locations in the image. During feature fusion, the attention weight map reweights the features extracted by the encoder, allowing the network to allocate more computational resources to regions with severe artifacts or blurred structures, thereby improving the targeting and efficiency of the correction.

[0055] Furthermore, the image domain correction module performs layer-by-layer correction processing on the intermediate volumetric images. For the three-dimensional volumetric images, they are treated as a stack of two-dimensional tomographic images. The network processes each tomographic image layer by layer, while utilizing the contextual information between adjacent tomographic images to maintain consistency in the three-dimensional space. During the layer-by-layer processing, the network simultaneously learns to remove artifacts and enhance structural details, so that the corrected image maintains the authenticity of the anatomical structure while having sharper edges and more natural textures.

[0056] The image domain correction module requires training before use. The training process uses paired data: the input is the intermediate volumetric image obtained through training, and the label is the corresponding real high-quality image. A joint loss function is used during training, including a loss term to constrain pixel-level deviations between the corrected image and the real image, and a loss term to constrain structural similarity between the two. The pixel-level mean squared error loss ensures the corrected image is accurate in grayscale values, while the structural similarity loss ensures that the perceived quality of the image, such as edges and textures, is similar to that of the real image. The two are weighted and combined... The weighting coefficients and Based on experimental optimization, the preferred embodiment is determined to be... Through training on a large amount of data, the image domain correction module learns the mapping relationship from intermediate volumetric images with artifacts to high-quality images.

[0057] After processing by the image domain correction module, the intermediate volume images of each respiratory phase are converted into high-quality corrected images. Organizing all corrected images of the respiratory phases in chronological order yields the final four-dimensional cone-beam CT image sequence, which comprehensively records the dynamic changes in the anatomical structures of the object to be reconstructed within a complete respiratory cycle. The final output four-dimensional cone-beam CT images can be used for clinical applications such as target localization and respiratory motion assessment in image-guided radiotherapy. Because this application significantly improves reconstruction quality through spatiotemporal consistency coding and a dual-domain collaborative strategy, even under extreme conditions with only 16 sparse projections per phase, the output four-dimensional cone-beam CT images still maintain the spatial continuity of anatomical structures and the temporal coherence of respiratory motion, meeting clinical requirements for image quality and motion accuracy.

[0058] Please see Figure 2 This is a flowchart of a four-dimensional cone-beam CT image reconstruction method according to Embodiment 1 of this application. First, raw projection data of N=16 respiratory cycles and P=10 respiratory phases are obtained from 4D CBCT scan data. This data is then grouped by respiratory phase and... Interval sampling yields a sparse projection set for each respiratory phase. Each phase contains 16 sparse projections and their corresponding acquisition angle information.

[0059] The sparse projection set and angle information are input into the projection domain generation module. This module first extracts features from the sparse projection and then generates a position code based on the angle information. Each acquired angle is then... Normalized coordinates mapped to the plane of rotation The features are then extended to a high-dimensional space via Fourier feature mapping, where the frequency order of spatial coordinates is L=10 and the frequency order of angular coordinates is L=4. On the other hand, the spatial features of each respiratory phase are sequentially input into an LSTM network for temporal modeling, capturing the temporal relationship of respiratory motion across 10 phases and generating time-step codes. The spatial features, position codes, and time-step codes are fused and input into a Swin-UNet decoder. Through progressive upsampling and skip connections, a dense projection set for each respiratory phase is output, with the number of projections expanded to K=256. This dense projection set is input into an FDK reconstruction module, where a filtered backprojection algorithm reconstructs the intermediate volumetric image for each respiratory phase. Finally, the intermediate volumetric image is input into an image domain correction module, where a layer-by-layer correction is performed using an attention-based encoder-decoder convolutional network to suppress residual artifacts and enhance anatomical details, outputting a high-quality four-dimensional cone-beam CT volumetric image.

[0060] Furthermore, this application was validated on publicly available datasets and physical phantom data. The experiments used the TCIA 4D lung cancer dataset (containing 20 sets of high-quality 4D spiral CT reconstructed images, each set containing 10 phase volumetric images) and the CIRS chest physical phantom dataset. Forward projection was performed on each phase of each dataset to obtain 256 projections as references, and 16 projections were extracted at fixed intervals of 22.5° as sparse inputs.

[0061] Quantitative evaluation results show that this application achieves the best performance among all comparative methods. On the TCIA-LUNG dataset, the PSNR of this application reaches 33.311 dB, 32.695 dB, and 32.658 dB in three representative respiratory phases (Phase 2 / 5 / 8), respectively; SSIM reaches 0.845, 0.823, and 0.825, respectively; and MAE reaches 0.035, 0.032, and 0.039, respectively. All of these are significantly better than FDK (PSNR approximately 21.6 dB), SART (PSNR approximately 22.4 dB), SART-TV (PSNR approximately 23.4 dB), N-Net (PSNR approximately 22.1 dB), CycN-Net (PSNR approximately 22.5 dB), and AttUNet (PSNR approximately 23.8 dB). On the CIRS physical phantom dataset, the proposed method achieved PSNRs of 36.497 dB, 36.392 dB, and 36.261 dB, and SSIMs of 0.871, 0.874, and 0.877, further validating its generalization ability. Ablation experiments confirmed that the Projection Domain Generation Module (PDGM) and Image Domain Correction Module (IDCM) have a significant synergistic gain effect: using PDGM alone can improve the PSNR from the baseline of 23.916 dB to 30.352 dB, and using both modules together further improves it to 32.536 dB. Qualitative evaluation shows that the proposed method can effectively eliminate the stripe artifacts of FDK, achieving a natural and smooth transition between bone and soft tissue regions, and significantly outperforms all comparative methods in terms of anatomical fidelity and geometric consistency.

[0062] In summary, Embodiment 1 of this application achieves high-quality image reconstruction through the following steps: First, sparse projection data and corresponding angle information of the object to be reconstructed at each respiratory phase are acquired; second, the sparse projection data and corresponding angle information are input into the projection domain generation module, spatial features are obtained by feature extraction from the sparse projection data, and position codes are generated by encoding the angle information using an encoding method that preserves the circumferential geometric characteristics, and time step codes are generated by temporal modeling of the spatial features. After fusing the three, a spatiotemporal enhancement feature that integrates spatial geometric information and respiratory motion information is obtained. Then, the spatiotemporal enhancement feature is decoded to output dense projection data for each respiratory phase; next, the dense projection data is parsed and reconstructed to obtain an intermediate volume image; finally, the intermediate volume image is input into the image domain correction module for correction, and the final four-dimensional cone-beam CT image is output. This application achieves high-fidelity, high-temporal-coherence four-dimensional cone-beam CT image reconstruction under extremely sparse projection conditions by explicitly modeling the spatial geometric relationship of the projection and the cross-phase respiratory motion dynamics through a spatiotemporal consistency encoding mechanism, combined with a dual-domain collaborative strategy of projection domain generation and image domain correction.

[0063] Example 2 Please see Figure 3 This is a schematic diagram of the structure of a four-dimensional cone-beam CT image reconstruction system according to Embodiment 2 of this application; the specific content includes: The data acquisition module 100 is used to acquire sparse projection data of the object to be reconstructed in each respiratory phase and the corresponding angle information; The projection domain generation module 200 is used to generate a spatiotemporal enhancement feature that integrates spatial geometric information and respiratory motion information based on the sparse projection data and the corresponding angle information, and to decode the spatiotemporal enhancement feature to output dense projection data for each respiratory phase. The analytical reconstruction module 300 is used to perform analytical reconstruction on the dense projection data to obtain an intermediate volume image; The image domain correction module 400 is used to correct the intermediate volume image and output a four-dimensional cone-beam CT image.

[0064] In this embodiment, the data acquisition module 100 is used to acquire sparse projection data of the object to be reconstructed at each respiratory phase and the angle information corresponding to each projection. In practical applications, the data acquisition module 100 can receive the acquired projection images and synchronously recorded respiratory signals in real time from the cone-beam CT scanning system. After phase registration and grouping processing, it obtains the sparse projection set of each respiratory phase and its corresponding angle information; alternatively, it can read a pre-prepared dataset from the storage medium for algorithm verification or offline processing. After the acquired data is organized into a unified format, it is transmitted to the projection domain generation module 200 for subsequent processing.

[0065] The projection domain generation module 200 receives sparse projection data and corresponding angle information from the data acquisition module 100. The projection domain generation module 200 includes a feature extraction unit, a position encoding unit, a temporal encoding unit, a feature fusion unit, and a decoding unit. The feature extraction unit uses an encoder network to extract multi-scale features from the input sparse projection data to obtain spatial features. The position encoding unit maps the input angle information to normalized coordinates on the rotation plane based on the angle information, and performs multi-frequency Fourier feature mapping on the normalized coordinates to generate position codes representing the geometric relationships between projections. The temporal encoding unit arranges the spatial features of each respiratory phase in chronological order and inputs them into a long short-term memory network. By capturing the temporal dependencies across phases, it generates time-step codes representing the dynamics of respiratory motion. The feature fusion unit fuses the spatial features, position codes, and time-step codes element-wise to generate spatiotemporally enhanced features that simultaneously contain spatial geometric information and respiratory motion information. The decoding unit decodes the spatiotemporally enhanced features and outputs dense projection data for each respiratory phase through progressive upsampling and feature recombination.

[0066] The analytical reconstruction module 300 is used to convert the generated dense projection data into a three-dimensional volumetric image. A filtered backprojection algorithm is employed to analyze and reconstruct the dense projection data for each respiratory phase. During the reconstruction process, based on preset scanning geometry parameters, the projection images at each projection angle are weighted and filtered before being backprojected onto a three-dimensional volumetric grid, ultimately obtaining the intermediate volumetric image corresponding to each respiratory phase.

[0067] The image domain correction module 400 employs an attention-based encoder-decoder convolutional network architecture to perform layer-by-layer correction processing on the input intermediate volume image. During the correction process, the encoder extracts deep features of the image, the attention gating unit adaptively focuses on residual artifact regions and structural edges in the image, and the decoder restores the corrected image from the enhanced features. This processing effectively removes residual stripe artifacts in the intermediate volume image, enhances the clarity of fine anatomical structures such as blood vessels and nodules, while maintaining the natural texture and spatial continuity of the image. The high-quality image obtained after correction is the final four-dimensional cone-beam CT image, which is output after being organized according to the respiratory phase sequence for clinical diagnosis and image-guided treatment.

[0068] Through the coordinated operation of the above modules, the system can achieve high-quality four-dimensional cone-beam CT image reconstruction with only a very small amount of projection data in each respiratory phase, effectively maintaining the spatial continuity of anatomical structures and the temporal coherence of respiratory motion, and providing reliable imaging support for clinical applications such as image-guided radiotherapy.

[0069] For other details regarding the implementation techniques of each module in the four-dimensional cone-beam CT image reconstruction system of the above embodiments, please refer to the description in the four-dimensional cone-beam CT image reconstruction method of the above embodiments, which will not be repeated here.

[0070] Example 3 Please see Figure 4 This is a schematic diagram of the computer device structure according to Embodiment 3 of this application. The computer device 50 includes a processor 51 and a memory 52 coupled to the processor 51.

[0071] The memory 52 stores program instructions for implementing the above-described four-dimensional cone-beam CT image reconstruction method.

[0072] The processor 51 is used to execute program instructions stored in the memory 52 to implement a four-dimensional cone-beam CT image reconstruction.

[0073] The processor 51 can also be referred to as a CPU (Central Processing Unit).

[0074] Processor 51 may be an integrated circuit chip with signal processing capabilities. Processor 51 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0075] Example 4 Please see Figure 5 This is a schematic diagram of the storage medium in Embodiment 4 of this application. The storage medium in this embodiment stores a program file 61 capable of implementing all the above methods. This program file 61 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or devices such as computers, servers, mobile phones, and tablets.

[0076] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0077] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0078] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

[0079] Of course, the present invention may have many other embodiments. Based on this embodiment, other embodiments obtained by those skilled in the art without any creative effort are all within the scope of protection of the present invention.

Claims

1. A method for reconstructing four-dimensional cone-beam CT images, characterized in that, include: Obtain sparse projection data and corresponding angle information of the object to be reconstructed in each respiratory phase; The sparse projection data and corresponding angle information are input into the projection domain generation module to generate spatiotemporal enhancement features that fuse spatial geometric information and respiratory motion information. The spatiotemporal enhancement features are then decoded to output dense projection data for each respiratory phase. The dense projection data is analyzed and reconstructed to obtain an intermediate volume image; The intermediate volume image is input into the image domain correction module for correction, and a four-dimensional cone-beam CT image is output.

2. The four-dimensional cone-beam CT image reconstruction method according to claim 1, characterized in that, The step of inputting the sparse projection data and corresponding angle information into the projection domain generation module to generate spatiotemporal enhanced features that fuse spatial geometric information and respiratory motion information specifically includes: Feature extraction is performed on the sparse projection data to obtain spatial features; The angle information is mapped to normalized coordinates of the rotation plane, where the normalized coordinates are represented as sine and cosine values. Multi-frequency Fourier feature mapping is performed on the normalized coordinates to generate a position code that characterizes the circumferential geometric relationship between projections. The spatial features are arranged in chronological order to form a temporal feature sequence. The temporal feature sequence is then input into a long short-term memory network to obtain a time step encoding for characterizing respiratory motion dynamics. The spatial features, the location encoding, and the time step encoding are added element by element to generate the spatiotemporal enhanced features.

3. The four-dimensional cone-beam CT image reconstruction method according to claim 2, characterized in that, The projection domain generation module adopts an encoder-decoder network architecture with multi-scale feature extraction capability, and the encoder-decoder network adopts a Transformer architecture based on a shift window self-attention mechanism.

4. The four-dimensional cone-beam CT image reconstruction method according to claim 3, characterized in that, The projection domain generation module is trained using a first joint loss function, which includes: Mean squared error loss used to constrain pixel-level deviations between generated dense projection data and real projection data; The structural similarity loss used to constrain the structural consistency between the generated dense projection data and the real projection data; And gradient consistency loss used to constrain the continuity between projected data.

5. The four-dimensional cone-beam CT image reconstruction method according to claim 1, characterized in that, The image domain correction module is an attention-based encoder-decoder convolutional network used to perform layer-by-layer correction on the intermediate volume image.

6. The four-dimensional cone-beam CT image reconstruction method according to claim 5, characterized in that, The image domain correction module is trained using a second joint loss function, which includes: Mean squared error loss used to constrain pixel-level deviations between the corrected image and the real image; And a structural similarity loss used to constrain the structural similarity between the corrected image and the real image.

7. The four-dimensional cone-beam CT image reconstruction method according to claim 1, characterized in that, The step of parsing and reconstructing the dense projection data to obtain an intermediate volume image specifically includes: The dense projection data is reconstructed using a filtered back projection algorithm to obtain an intermediate volume image.

8. A four-dimensional cone-beam CT image reconstruction system, characterized in that, For performing the four-dimensional cone-beam CT image reconstruction method according to any one of claims 1 to 7, the four-dimensional cone-beam CT image reconstruction system comprises: The data acquisition module is used to acquire sparse projection data of the object to be reconstructed in each respiratory phase and the corresponding angle information; The projection domain generation module is used to generate spatiotemporal enhancement features that fuse spatial geometric information and respiratory motion information based on the sparse projection data and the corresponding angle information, and to decode the spatiotemporal enhancement features to output dense projection data for each respiratory phase. The analytical reconstruction module is used to perform analytical reconstruction on the dense projection data to obtain an intermediate volume image; The image domain correction module is used to correct the intermediate volume image and output a four-dimensional cone-beam CT image.

9. A computer device, characterized in that, The computer device includes a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing the four-dimensional cone-beam CT image reconstruction method according to any one of claims 1-7; the processor is used to execute the program instructions stored in the memory to implement four-dimensional cone-beam CT image reconstruction.

10. A computer-readable storage medium, characterized in that, The device stores processor-executable program instructions for performing the four-dimensional cone-beam CT image reconstruction method according to any one of claims 1-7.

Images