Skip to content
Frontiers
Search

Original Research

Frontiers in Nuclear Medicine ·

Volume 6 - 2026 | https://doi.org/10.3389/fnume.2026.1711533

Cross-Domain TransNet for sparse-view CT reconstruction

  • Junling Wang1
  • Chunhua Zou2
  • Hongjie Yang3
  • Peixi Liao4*corresponding author
  • Xiaojun Mo5
  • Yingcui Gao1
  • Yanan Ning1
  • Jiahao Lai6*corresponding author
  1. 1Department of Radiology, Chengdu Sixth People's Hospital, Chengdu, China
  2. 2Department of Radiology, The First People’s Hospital of Chengdu, Chengdu, China
  3. 3School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, China
  4. 4Department of Stomatology, Chengdu Sixth People's Hospital, Chengdu, China
  5. 5Department of Radiology, General Hospital of the Western Theater Command of the Chinese People's Liberation Army, Chengdu, China
  6. 6Department of Interventional Radiology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China
Article metrics0 viewsView details
PDFCiteMetrics

Abstract

Sparse-view computed tomography (CT) reconstruction is crucial for clinical diagnostics, as reducing radiation exposure is essential to minimize risks to patients. Existing dual-domain reconstruction methods leverage both image and projection domains but often process them sequentially, overlooking their implicit correlations. To address this limitation, we propose Cross-Domain TransNet, a Transformer-based dual-domain framework for sparse-view CT reconstruction. The proposed model captures long-range dependencies within each domain and integrates image and sinogram representations through a hybrid self-attention mechanism. In addition, a Convolution Fusion Layer (CFL) is introduced to enhance feature interactions and facilitate more effective utilization of dual-domain information. Extensive experiments on the NIH-AAPM dataset demonstrate the superior performance and generalization capability of the proposed method under various sparse-view settings. The results show that Cross-Domain TransNet consistently improves reconstruction quality, effectively suppresses noise, and reduces artifacts, outperforming both conventional reconstruction algorithms and state-of-the-art deep learning approaches. Cross-Domain TransNet provides an effective and robust solution for sparse-view CT reconstruction. By fully exploiting complementary information from both image and projection domains, the proposed framework enhances diagnostic image quality while supporting radiation dose reduction.

Keywords

1 Introduction

Computed Tomography (CT) has become an indispensable tool in clinical diagnostics, playing a critical role during the COVID-19 pandemic by enabling rapid screening and assessment of disease severity (1, 2). However, increasing reliance on CT has raised concerns regarding radiation-induced health risks. Numerous studies have demonstrated that ionizing radiation from CT scans is associated with elevated cancer risk in both adults and children (3–5). For instance, exposure has been linked to higher incidences of leukemia and brain tumors, correlating with radiation doses to bone marrow and brain tissue (6). A long-term study of 31,462 patients reported that CT-related exposure increased lifetime cancer risk by 0.7% across the population and overall cancer mortality by 1%, with risks rising to 2.7%–12% in patients undergoing multiple scans (7). Moreover, children are particularly vulnerable, being three to four times more sensitive to radiation than adults, which may lead to more severe long term health effects (8).

In recent years, Sparse-View CT (SVCT) has gained wide-spread attention in clinical application, as its potential to reduce patient radiation exposure by lowering x-ray doses. However, lowering x-ray doses often leads to increased noise and artifacts in the reconstructed CT images, severely compromising image quality and affecting clinical diagnosis. Mitigating the adverse effects of reduced radiation doses in CT image reconstruction has therefore become a key focus of current research. Deep learning (DL), as an emerging artificial intelligence technology, has been demonstrated remarkable potential in the field of image processing and analysis (9). Numerous studies have confirmed the feasibility and effectiveness of using DL methods for sparse-view CT image reconstruction, yielding significant improvements in image quality over traditional approaches. As a result, employing DL methods for sparse-view CT image reconstruction is of significant research importance and offers promising practical applications. By reducing patient radiation exposure, enhancing the accuracy of medical image diagnostics, and advancing medical imaging technology, DL methods can provide safer, more precise, and higher-quality reconstructed medical images for clinical practice. This advancement ultimately contributes to reducing radiation exposure and enhancing the healthcare quality of patients.

By now, the advanced methods for the reconstruction of sparse-view CT images can be summarized into two main categories. Traditional methods typically involve minimizing a unified objective function iteratively to reconstruct images from noisy sinogram data. This objective function integrates prior information of the image domain with the statistical properties of the sinogram data. Total variation (TV) and its variants (10), non-local means (11), and dictionary learning (12) are image priors commonly used in recent researches. While these methods have significantly improved reconstruction quality, they depend on iterative forward and backward projection operations, requiring detailed knowledge of the specific scanner geometry. This results in high computational complexity and longer reconstruction times. In recent years, DL methods have achieved significant breakthroughs in sparse-view CT reconstruction tasks. These researchers utilize DL network structures such as Convolutional Neural Networks (CNNs) to acquire the mapping relationship between sparse-view projection data and high-quality CT images. DL techniques can effectively reduce noise, recover details during the reconstruction process and lead to substantial improvements in image quality through automatically learning complex image features from large data.

Current deep learning (DL) methods for sparse-view CT reconstruction can be broadly categorized into three groups, with CNN and U-Net based architectures being the most dominant. Representative works include RED-CNN (13), which combines deconvolution, autoencoders, and residual connections; DenseNet with deconvolution (14) to enhance image details; FCPRN (15), which integrates FCN with pyramidal residual blocks for multi-scale feature extraction; SR-CNN (16), designed to preserve edges and fine structures; and RNN-based models (17) for stripe artifact removal. Yan et al. further proposed TLD-CDL (18), combining convolutional dictionary learning, transfer learning, and dense connections for robust feature representation. Most of these DL methods operate in the image domain, as direct access to projection data is often impractical. Typically, sparse-view data are first reconstructed using filtered back projection (FBP), followed by artifact reduction via neural networks. However, this sequential strategy cannot recover information lost in projection data, limiting the potential of image-only approaches (19, 20).

It often results in severe streak artifacts and diminished image quality by applying traditional image reconstruction algorithms like FBP directly to sparse-view projection data. To address this issue, neural networks are employed to fill in the missing data in the projection domain, after which the completed projection data can be used in the next image reconstruction process. Liu et al. proposed a generative adversarial network (GAN) known as Pix2Pix to perform pixel-level reconstruction for sinogram data (21). Compared to traditional linear interpolation methods, Pix2Pix performed better while reconstructing high quality images. Kim et al. introduced an unsupervised training method based on deep neural networks in the projection domain (22). In which, it recovers sparse-view projection data and achieves performance comparable to supervised methods without full dose CT images. Besides, to address metal artifacts in CT, Ghani et al. proposed a framework named Deep-MAR (23). This method is based on GANs and treats projection data corresponding to dense metal regions as missing data.

Repairing data in the projection domain can mitigate signal loss due to sparser radiation doses, which enables the reconstruction process to commence under less noise conditions. Consequently, projection domain processing methods are more robust in handling noise. However, these methods require detailed knowledge of the CT device's hardware parameters typically. The wide variety of equipment manufacturers and models makes it hard to obtain and match these hardware parameters accurately, thereby limiting the practical applicability of these approaches. Moreover, in the three projection domain processing methods mentioned earlier, FBP is still used for direct reconstruction after repairing the completed projection data. To some extent, when DL methods are only employed for projection domain data repair, the limitations of traditional methods may still affect the reconstruction outcomes during image post-processing. Compared with single-domain deep learning methods, most current sparse-view CT reconstruction studies focus on dual domain processing, which simultaneously recovers projection data and reconstructs image data. Representative works include DuDoNet (24), which employs a mask pyramid U-Net for projection preprocessing and image post-processing; DuDoTrans (25), which integrates Swin-Transformer and CNN to repair sinograms; CAGAN (26), which combines adversarial autoencoders with coordinate attention; DuDoUFNet (27), which suppresses metal artifacts in the projection domain and reduces noise in the image domain; and DRONE (28), which embeds, refines, and enhances features to improve image quality. These approaches typically adopt a sequential workflow: first repairing projection data, then reconstructing CT images using algorithms such as FBP, followed by image domain refinement. While this improves quality over single domain methods, it essentially couples two independent processes without fully exploiting correlations between domains. Moreover, repeated use of FBP may introduce additional noise and artifacts, limiting reconstruction performance.

Recent progress in implicit neural representations has provided new solutions for sparse-view CT reconstruction by representing images or volumes as continuous functions optimized via coordinate-based multilayer perceptrons. For instance, NeRP (29) integrates patient-specific prior images into the network and optimizes this implicit representation under projection constraints, enabling robust reconstruction in highly sparse scenarios when longitudinal priors are accessible. TomoGRAF (30) extends NeRF-style volumetric rendering by embedding x-ray imaging physics into a generative radiance field. It achieves better geometric consistency in very sparse settings, including cases with only one or two views, through differentiable volume rendering that combines both projection and volumetric supervision. These methods based on implicit neural representations offer notable data efficiency, seamless integration of imaging physics, and the ability to produce structurally coherent reconstructions under severely limited data conditions. However, they also come with certain limitations, such as reliance on patient-specific priors in approaches like NeRP, high computational costs due to test-time optimization or volumetric rendering, and reported challenges in accurately recovering CT numbers and fine anatomical details under extreme sparsity.

In contrast, a transformer-based formulation adopts a different modeling strategy compared to the above methods. It learns global dependencies across projection views rather than explicitly parameterizing volumes or depending on patient-specific priors. Once trained, the model enables fast feed-forward reconstruction without requiring test-time optimization. Its global attention mechanism facilitates effective integration of information from sparse views and supports robust generalization. As a result, the transformer-based approach serves as a complementary alternative to the above techniques, which continue to excel in leveraging geometric priors and patient-specific information, especially under extremely sparse acquisition conditions.

2 Materials and methods

2.1 Datasets

The training and testing of the method used publicly available data from the 2016 NIH-AAPM-Mayo Low Dose CT Grand Challenge dataset (31), denoted as Mayo dataset. Besides, a scanning geometry with a fan-beam x-ray source comprising 800 detector elements was simulated following the settings described in references (32–34). Specifically, for this study, a total of 2,378 3 mm slices reconstructed using the B30 reconstruction kernel were chosen from 10 patients, with a resolution of 256 × 256. Out of these, 237 slices were randomly selected for testing, while the remaining 2,141 slices were used for training and validation for the models. The details were introduced in Supplementary Materials S1.

2.2 Cross-Domain TransNet architecture

We proposed a novel sparse-view CT reconstruction method based on the Vision Transformer architecture that incorporates a dual-domain cross-attention mechanism (Figure 1). The main idea is to progressively extract potential relationships between image domain and projection domain data, thereby complementing and enhancing the recovery of both domains. The method is primarily composed of three parts: initial Patch Embedding operation, stacked Cross-Domain Transformer modules (Figures 2, 3), and Patch Merging operation (Figure 4). Our model builds upon the previous models and incorporates new fusion ideas, enabling cross-domain information to be deeply integrated, thereby achieving better reconstruction results. The framework was introduced in Supplementary Materials S2 in details.

Framework of CDTN. The overall framework of the method proposed is primarily composed of three parts: initial Patch Embedding operation, stacked Cross-Domain Transformer modules, and Patch Merging operation.
Figure 1. Framework of CDTN. The overall framework of the method proposed is primarily composed of three parts: initial Patch Embedding operation, stacked Cross-Domain Transformer modules, and Patch Merging operation.
Architecture of cross-domain transformer block. The Cross-Domain Transformer Block takes concatenated dual-domain tokens as input, first applying layer normalization to standardize the tokens and prevent data offset during training.
Figure 2. Architecture of cross-domain transformer block. The Cross-Domain Transformer Block takes concatenated dual-domain tokens as input, first applying layer normalization to standardize the tokens and prevent data offset during training.
Architecture of cross attention operation. The cross-attention module then splits the dual-domain tokens into single-domain tokens, where self-attention within each domain is synchronized with cross-attention between the domains.
Figure 3. Architecture of cross attention operation. The cross-attention module then splits the dual-domain tokens into single-domain tokens, where self-attention within each domain is synchronized with cross-attention between the domains.
Architecture of convolutional fusion layer (CFL). The CFL first maps input tokens into a spatial feature map through a linear projection and activation. A convolution layer is then applied to perform local feature fusion. The fused features are subsequently mapped back to the token representation via another activation and linear layer. This module enables seamless interaction between the token space and image space, allowing the model to exploit both global transformer representations and local convolutional priors.
Figure 4. Architecture of convolutional fusion layer (CFL). The CFL first maps input tokens into a spatial feature map through a linear projection and activation. A convolution layer is then applied to perform local feature fusion. The fused features are subsequently mapped back to the token representation via another activation and linear layer. This module enables seamless interaction between the token space and image space, allowing the model to exploit both global transformer representations and local convolutional priors.

2.3 Dual-domain loss function

This work applied a dual-domain loss function to achieve better reconstruction results. The dual-domain loss function comprises three components: image reconstruction loss, image perceptual loss, and reconstruction projection consistency loss. The details were shown in Supplementary Materials S2.

2.4 Experimental setup

In this section, we implemented the proposed method in Python using the PyTorch library and employed the Adam (35). The details of our experimental setup were shown in Supplementary Materials S3.

2.5 Evaluation

For metrics, we adopted PSNR, SSIM, and NRMSE as basic evaluation metrics. Besides, we further conducted histogram-based quantitative analyses to evaluate the intensity consistency and residual error distribution of different reconstruction methods. For each test case, the full-view reconstruction was used as the reference image, and the reconstructed image obtained by each sparse-view method was compared with it on a pixel-wise basis.

3 Results

A set of image data from the 3mmB30 dataset was displayed in Figure 5, including regular dose data and simulated noise of varying intensities. These data were reconstructed using the FBP algorithm to generate CT images. It is evident from the images that as the radiation dose in the projection data decreases, the noise and streak artifacts in the reconstructed images produced by the FBP method gradually increase.

A set of 5 simulated images from different views. It displays a set of image data from the 3mmB30 dataset. It is evident from the images that as the radiation dose in the projection data decreases, the noise and streak artifacts in the reconstructed images produced by the FBP method gradually increase.
Figure 5. A set of 5 simulated images from different views. It displays a set of image data from the 3mmB30 dataset. It is evident from the images that as the radiation dose in the projection data decreases, the noise and streak artifacts in the reconstructed images produced by the FBP method gradually increase.

3.1 Comparison

We compared the proposed method with three recent DL-based approaches: RED-CNN (13) and CNCL (36) in the image domain, and DuDoTrans (25) in the dual-domain. Specifically, RED-CNN is the first to introduce convolutional neural networks into low-dose CT image post-processing, making significant progress in denoising compared to traditional methods. CNCL is a method based on generative adversarial networks. DuDoTrans is a dual-domain processing method based on an improved Transformer framework. FBP is widely used in routine-dose CT reconstruction in clinical practice due to its low computational resource requirements and fast speed. FBP is widely used in routine-dose CT reconstruction in clinical practice due to its low computational resource requirements and fast speed. Besides, we used official or widely adopted open-source implementations and followed the original parameter settings, using the same dataset for training, validation, and testing.

3.2 Visualization results

To illustrate the method's performance, one slice per view was selected from the 96-view dataset. Figure 6 shows reconstruction results for four views, comparing different methods. Traditional FBP introduces severe stripe artifacts that degrade visual quality and diagnostic reliability, making it unsuitable for sparse-view CT. In contrast, DL-based methods produce clearer images with clinically acceptable quality. Figure 7 zooms into the yellow box in Figure 6 to highlight detailed differences. To further demonstrate the robustness and generalizability of the proposed method, we expanded the subjective visual comparison by including multiple representative slices from different anatomical levels, as shown in Figure 8. These slices were selected from the test set to cover different anatomical structures and image characteristics. Compared with FBP, RED-CNN, DuDoTrans, and CNCL, the proposed CDTransNet consistently suppresses sparse-view streak artifacts while preserving anatomical structures and edge details across different anatomical levels. In particular, CDTransNet produces clearer structural boundaries and fewer residual artifacts in the enlarged regions, indicating its robust reconstruction performance on different slices.

Qualitative comparison on mayo dataset, each row from top to bottom corresponds to 32, 64, 96, 128 views. As shown in the images, the traditional Filtered Back Projection (FBP) method introduces significant stripe artifacts, which are not only visually disruptive but also critically compromise diagnostic accuracy in clinical settings, rendering this approach unsuitable for low-dose CT reconstruction tasks.
Figure 6. Qualitative comparison on mayo dataset, each row from top to bottom corresponds to 32, 64, 96, 128 views. As shown in the images, the traditional Filtered Back Projection (FBP) method introduces significant stripe artifacts, which are not only visually disruptive but also critically compromise diagnostic accuracy in clinical settings, rendering this approach unsuitable for low-dose CT reconstruction tasks.
Zoomed ROI of the yellow rectangle in Figure 6. We can see our method not only effectively eliminates artifacts but also significantly enhances detail preservation.
Figure 7. Zoomed ROI of the yellow rectangle in Figure 6. We can see our method not only effectively eliminates artifacts but also significantly enhances detail preservation.
**(a)** visual comparison of different reconstruction methods on representative slices from different anatomical levels. **(b)** The enlarged ROI regions highlight local anatomical details. CDTransNet achieves better artifact suppression and structural preservation than the competing methods across different anatomical levels.
Figure 8. (a) visual comparison of different reconstruction methods on representative slices from different anatomical levels. (b) The enlarged ROI regions highlight local anatomical details. CDTransNet achieves better artifact suppression and structural preservation than the competing methods across different anatomical levels.

3.3 Quantitative results

To illustrate the method's performance, one slice per view was selected from the 96-view dataset. To quantitatively compare the relevant methods, we performed a 5-fold cross-validation using the Mayo dataset, employing PSNR, SSIM and NRMSE as the key metrics for image quality assessment. R2 was also evaluated between all methods to provide a more comprehensive assessment of reconstruction performance. All the results presented in Table 1, indicate that all methods enhanced the quality of the reconstructed images according to these metrics. Besides, histogram-based quantitative analyses are presented in Figure 9. It shows the joint histograms between the full-view reference images and the reconstructed images under different sparse-view settings. For FBP, the joint histogram exhibits a relatively dispersed distribution, especially under the 32-view and 64-view settings, indicating severe inconsistency between the reconstructed image and the full-view reference. Although RED-CNN and DuDoTrans reduce the dispersion to some extent, their distributions still deviate from the diagonal line due to residual artifacts and intensity bias. For error histograms, the FBP produces a broad error distribution, indicating large reconstruction deviations caused by sparse-view artifacts. RED-CNN and DuDoTrans substantially reduce the error range, but their residual distributions remain wider than those of CNCL and CDTransNet. CDTransNet achieves the narrowest error distribution centered around zero, suggesting that the proposed method introduces less reconstruction bias and achieves more accurate pixel-wise recovery. Notably, our proposed method exhibited superior performance, surpassing the other methods in both metrics. Specifically, as illustrated in Figure 7, our method not only effectively eliminates artifacts but also significantly enhances detail preservation.

Table 1. Quantitative comparison of different reconstruction methods under various sparse-view settings.

Methods32 views64 views96 views128 views
PSNRSSIMNRMSER2PSNRSSIMNRMSER2PSNRSSIMNRMSER2PSNRSSIMNRMSER2
FBP12.180 ± 1.152*0.348 ± 0.032*0.246 ± 0.033*−3.601414.770 ± 2.430*0.456 ± 0.023*0.183 ± 0.051*−1.835317.201 ± 1.893*0.538 ± 0.018*0.138 ± 0.030*−0.650818.874 ± 2.013*0.598 ± 0.011*0.114 ± 0.026*−0.1326
RED-CNN28.829 ± 1.218*0.771 ± 0.012*0.036 ± 0.005*0.906532.332 ± 1.342*0.890 ± 0.018*0.024 ± 0.0040.954934.339 ± 1.2070.918 ± 0.0170.019 ± 0.003*0.970736.131 ± 1.047*0.941 ± 0.016*0.016 ± 0.002*0.9809
DuDoTrans33.624 ± 1.4260.886 ± 0.0210.021 ± 0.003*0.963737.503 ± 1.846*0.927 ± 0.030*0.013 ± 0.003*0.982639.498 ± 1.542*0.945 ± 0.032*0.011 ± 0.002*0.991140.802 ± 2.110*0.955 ± 0.019*0.009 ± 0.002*0.9946
CNCL40.063 ± 1.156*0.969 ± 0.015*0.010 ± 0.001*0.993141.781 ± 1.429*0.978 ± 0.042*0.008 ± 0.001*0.994342.422 ± 1.2620.981 ± 0.0260.008 ± 0.001*0.995543.452 ± 1.029*0.986 ± 0.012*0.007 ± 0.001*0.9964
Ours43.030 ± 1.015*0.980 ± 0.027*0.007 ± 0.001*0.996543.783 ± 1.452*0.982 ± 0.036*0.006 ± 0.001*0.996744.846 ± 1.340*0.985 ± 0.013*0.006 ± 0.001*0.997445.456 ± 2.011*0.987 ± 0.018*0.005 ± 0.001*0.9981

PSNR, SSIM and NRMSE are evaluated on the test dataset for 32, 64, 96, and 128 projection views. The proposed CDTransNet consistently outperforms existing image-domain and dual-domain baselines across all sampling conditions. Paired t-test between all baseline algorithms and our method was performed with metrics. *p＜0.05.

Histogram-based quantitative comparison of different reconstruction methods. The joint histograms illustrate the intensity consistency between the full-view reference images and the reconstructed images, while the error histograms show the pixel-wise residual distributions. CDTransNet produces joint histograms that are more concentrated along the diagonal and error histograms that are more sharply centered around zero, indicating better reconstruction fidelity.
Figure 9. Histogram-based quantitative comparison of different reconstruction methods. The joint histograms illustrate the intensity consistency between the full-view reference images and the reconstructed images, while the error histograms show the pixel-wise residual distributions. CDTransNet produces joint histograms that are more concentrated along the diagonal and error histograms that are more sharply centered around zero, indicating better reconstruction fidelity.

To investigate the effectiveness of the cross-domain attention mechanism, we replaced the cross-attention module in the model with a conventional multi-head attention operation. This alternative approach directly calculates multi-head self-attention across all input tokens instead of splitting and merging dual-domain data. All other settings remained unchanged, and this experiment is referred to as exp-1. The corresponding results are presented in Table 2. The quantitative evaluation results were presented in Table 3, and a selection of the reconstructed images is displayed in Figure 3. The comparison of GPU memory usage, training time, and model size among different learning-based methods is summarized in Table 4.

Table 2. Ablation study of the proposed architecture.

ModulesPSNRSSIMNRMSENote
BASE44.846 ± 1.3400.985 ± 0.0130.006 ± 0.001Base model
exp-144.695 ± 2.0920.984 ± 0.0750.006 ± 0.001Conventional multi-head attention
exp-242.036 ± 2.0510.981 ± 0.0160.008 ± 0.002Without cross attention
exp-343.701 ± 2.1030.981 ± 0.0090.007 ± 0.002Replace CFL with MLP
exp-443.780 ± 1.6240.980 ± 0.0200.006 ± 0.001Only image loss
exp-544.536 ± 1.6670.984 ± 0.0110.006 ± 0.002Both image and sinogram loss
exp-644.754 ± 1.8060.984 ± 0.0180.006 ± 0.001Both image and perceptual loss

Each variant removes or replaces a key component of the network, including multi-head attention design, cross-attention module, convolutional fusion layer (CFL), and different loss combinations. Results demonstrate the critical role of cross-attention and dual-domain losses in achieving optimal reconstruction performance.

Table 3. Evaluation of cross-domain robustness on datasets with different reconstruction kernels and slice thickness. .

Protocols32 views96 views
PSNRSSIMNRMSEPSNRSSIMNRMSE
1mmB3043.409 ± 1.1320.977 ± 0.0280.007 ± 0.00244.501 ± 1.4210.980 ± 0.0120.006 ± 0.001
1mmD4542.672 ± 1.2110.971 ± 0.0180.008 ± 0.00143.572 ± 1.0470.974 ± 0.0330.007 ± 0.001
3mmB3043.030 ± 1.2770.980 ± 0.0140.007 ± 0.00144.846 ± 1.2910.985 ± 0.0300.006 ± 0.001
3mmD4542.851 ± 1.2090.978 ± 0.0120.007 ± 0.00144.158 ± 1.1470.982 ± 0.0190.006 ± 0.001

The model maintains strong PSNR, SSIM and NRMSE performance across varying clinical imaging protocols (1 mm/3 mm slice thickness and B30/D45 kernels), indicating good robustness and generalizability.

Table 4. Comparison of GPU memory usage, training time, and model size with other learning based models.

MethodGPU memory usageTraining time per epoch/sModel Size (Parameters)
RED-CNN∼1.5–3 GB∼20–50∼1M–1.5M
DuDoTrans∼10 GB∼200–300∼20M–30M
CNCL∼4–8 GB∼50–100∼5M–10M
CDTransNet∼12 GB∼356∼40M

4 Discussion

CT technology has become the most widely used auxiliary diagnostic tool in modern clinical medicine. To minimize the potential harm caused by x-ray radiation, the application of sparse-view CT has garnered increasing attention. This study proposed a dual-domain reconstruction method based on a Transformer architecture specifically designed for sparse-view CT imaging in order to reconstruct high-quality CT images from sparse-view sinogram data.

We employed Transformer architecture to overcome the limitations of finite receptive fields inherent in conventional CNN structures. Our method used a parallel dual-domain reconstruction approach, incorporating mixed attention mechanisms to capture latent information between the projection and image domains. Instead of adopting typical Multi-Layer Perceptrons (MLPs) in Transformer models, we introduced a CFL to enable the model to effectively learn long-range dependencies across image patches. Furthermore, we utilized a joint loss function to simultaneously compute reconstruction loss in both domains and perceptual loss in the image domain, thereby preserving more image details. In the following four sections, we evaluated and validated the effectiveness of various components and parameter settings in the proposed model. This included an assessment of the hybrid attention module, evaluation of the convolution fusion layer, verification of the dual-domain loss function's efficiency, and determination of the optimal loss weight parameters. Additionally, we examined the model's robustness across different datasets and analyzed both the computational cost and parameter count required during training and inference. All experiments were conducted on a simulated dataset consisting of 96 views from the 3mmB30 dataset, with PSNR, SSIM and NRMS values calculated for all reconstructed CT images in the test set.

We compared our model with previous SOTA models in which the results are shown in Table 1. From the results, we can find CNCL is better than DuDoTrans. Dual-domain reconstruction frameworks tend to outperform pure image-domain networks because they incorporate physics-based constraints. However, the performance of a specific method is also affected by architecture design, training strategies, data domain, and hyperparameters. Dual-domain does not guarantee universal superiority; effectiveness depends on how well the sinogram branch is adapted to the specific acquisition protocol. DuDoTrans's sinogram encoder generalized poorly to our scans, whereas CNN-based methods are less sensitive to the mismatch.

Our method is better than DuDoTrans although both of them are based on Transformer and domains. Our Cross-Domain TransNet introduces a much stronger, explicitly designed dual-domain interaction mechanism. (1) Our model uses bidirectional cross-attention, not separate domain Transformers. DuDoTrans applies self-attention on each domain separately and relies on implicit mapping. In contrast, we introduce a Cross-Domain Attention Module that includes: synchronized self-attention in each domain and explicit cross-attention. This design forces image-domain features to guide sinogram restoration and vice versa, significantly enhancing both. (2) Progressive fusion in every Transformer block. Each Cross-Domain Transformer Block repeatedly, which splits tokens into two domains, performs domain-specific self-attention, performs cross-attention to exchange feature information and fuses the updated dual-domain tokens. This progressive multi-stage interaction is absent in DuDoTrans. (3) Convolutional Fusion Layer (CFL) replaces the MLP. These designs significantly improved reconstruction quality in Table 1. The histogram-based analysis further confirms that the proposed CDTransNet not only improves global image quality metrics such as PSNR and SSIM, but also reduces pixel-wise reconstruction deviations. The compact joint histogram distribution and the narrow error distribution demonstrate that CDTransNet can better preserve the intensity distribution of full-view CT images while suppressing sparse-view artifacts.

To further explore the impact of dual-domain data interaction on reconstruction quality, we removed the cross-linking in the attention mechanism, calculating self-attention independently for each domain. This experiment is labeled as exp-2, with the results also shown in Table 2. Lastly, to verify the influence of the CFL on the model's performance, the conventional MLP module was replaced with the proposed CFL module in the Transformer architecture. This experiment is referred to as exp-3. We proposed the Cross-Domain TransNet for sparse-view CT image reconstruction, utilizing a joint loss function. The choice of loss functions in DL significantly influenced the image reconstruction process. During our experiments, it became clear that the weights assigned to different loss function terms directly affect the quality of the reconstructed images. Initially, we aimed to establish the relationship between various parameters and the resulting image quality. To identify the optimal weights for each loss term within the joint loss function, experiments were conducted where one parameter was varied while the others were kept constant, thereby determining the best value for each parameter. More importantly, the ablation experiments show that: (1) Exp-1 without domain merging performs similarly to the BASE model because only self-attention is used; the two domains do not interact, so the model cannot leverage additional information. While Exp-2 without cross-attention severely degrades performance, confirming that simple concatenation or self-attention alone cannot effectively integrate projection of image relationships. The significant performance drop in Exp-2 demonstrates that cross-attention is the key operation for meaningful dual-domain interaction, enabling the sinogram features to guide image-space refinement and vice versa. This confirms the reliability and necessity of the cross-attention design in our method.

To further explore the roles of different loss function terms in image reconstruction and assess the effectiveness of the joint loss function, the following experiments were implemented: 1) Using only the image loss, denoted as exp-4.2) Using both image loss and sinogram loss, denoted as exp-5.3) Using both image loss and perceptual loss, denoted as exp-6. The experimental results list in Table 2. When without the dual-domain loss function in the training process, the quality of the image is lower comparing to the results using the dual-domain loss function. To validate the robustness of the proposed method across different datasets, the training and test on images reconstructed using four distinct reconstruction kernels provided by the Mayo dataset were conducted. The experiments were performed with 32 views and 96 views, following the same data preparation process as used for the 3mmB30 dataset. As shown in Table 3 and Figure 3, the proposed method achieved satisfactory evaluation metrics and delivered visually appealing results across all four datasets. Besides, as we can see the CDTN results appear to exhibit patch-edge artifacts in Figure 7. The mild patch-edge artifacts are caused by the fixed-size patch embedding used in the Transformer encoder. When anatomical structures span across patch boundaries, the global self-attention may capture long-range relationships, but local continuity between adjacent patches is not fully preserved, especially at higher sparsity levels. The proposed CFL module is designed to alleviate this issue by restoring spatial continuity and performing local feature fusion after attention. While the CFL significantly reduces patch-edge inconsistencies, some residual artifacts can remain under extremely sparse-view settings.

Nonetheless, this study has certain limitations. Firstly, the model heavily relies on high-quality paired training data with accurate ground-truth reconstructions. In practice, obtaining such paired data is challenging, especially for sparse-view CT where high-dose full-view reconstructions may not be available or feasible due to radiation concerns. Secondly, our current experiments use simulated projections because raw clinical projection data are not publicly available for the benchmark dataset used in this study. We acknowledge that simulation may not fully capture detector noise characteristics, system blur, scatter, or calibration imperfections. Therefore, a domain gap between simulated and real measurements may exist. We will include a discussion on this limitation. Thirdly, the dual-domain design, combined with attention-based mechanisms and the proposed Convolutional Fusion Layer (CFL), leads to increased memory consumption and inference time. This may hinder the deployment of the model in time-sensitive or resource-constrained environments, such as portable imaging devices or point-of-care systems. Future work could explore lightweight model variants or optimization techniques, such as knowledge distillation or model pruning, to reduce computational costs while maintaining performance.

5 Conclusion

In conclusion, we conducted extensive qualitative and quantitative analyses on publicly available datasets, comparing the proposed method with traditional approaches, single-domain DL methods, and state-of-the-art dual-domain DL techniques. Our approach not only achieves high quantitative metrics but also delivers superior visual results. It effectively preserves the structure and texture of CT images while removing noise and artifacts, even when the projection data is under-sampled. Finally, through multiple ablation experiments, we validated the effectiveness of the model components and identified the optimal parameter settings.

Correction Note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnume.2026.1711533/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Chenyang Shen, University of Texas Southwestern Medical Center, United States

Reviewed by: Ruizhi Zuo, University of Texas Southwestern Medical Center, United States

Hamidreza Rashidykanan, University of Texas Southwestern Medical Center, United States

Statements

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This work was supported by The Chengdu Key Research and Development Support Project (Grant No. 2024YF0500910SN).

Data availability

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding authors.

Author contributions

JW: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Writing – original draft. CZ: Conceptualization, Formal analysis, Resources, Visualization, Writing – review & editing. HY: Funding acquisition, Methodology, Project administration, Resources, Writing – review & editing. PL: Methodology, Project administration, Resources, Supervision, Writing – review & editing. XM: Methodology, Software, Visualization, Writing – original draft. YG: Data curation, Formal analysis, Resources, Writing – original draft. YN: Data curation, Formal analysis, Software, Validation, Writing – original draft. JL: Funding acquisition, Supervision, Validation, Writing – review & editing.

References

  1. 1. Li X, Zeng W, Li X, Chen H, Shi L, Li X, Xiang H, Cao Y, Chen H, Liu C, Wang J, CT Imaging changes of corona virus disease 2019 (COVID-19): a multi-center study in Southwest China. J Transl Med. (2020) 18 (1): 1 – 8. doi:10.1186/s12967-020-02324-w
  2. 2. García Santos JM, Plasencia Martínez JM, Fabuel Ortega P, Lozano Ros M, Sánchez Ayala MC, Pérez Hernández G, Menchón Martínez P, The role of CT imaging for management of COVID-19 in epidemic area: early experience from a University Hospital. Insights Imaging. (2021) 12 (1): 1 – 5. doi:10.1186/s13244-020-00954-8
  3. 3. Meulepas JM, Ronckers CM, Smets AMJB, Nievelstein RAJ, Gradowska P, Lee C, Jahnen A, Van Straten M, De Wit M-CY, Zonnenberg B, Klein WM, Merks JH, Visser O, Van Leeuwen FE, Hauptmann M, Radiation exposure from pediatric CT scans and subsequent cancer risk in The Netherlands. J Nat Cancer Inst. (2019) 111 (3): 256 – 63. doi:10.1093/jnci/djy104
  4. 4. Lee KH, Lee S, Park JH, Lee SS, Kim HY, Lee WJ, Cha ES, Kim KP, Lee W, Lee JY, Lee KH, Risk of hematologic malignant neoplasms from abdominopelvic computed tomographic radiation in patients who underwent appendectomy. JAMA Surg. (2021) 156 (4): 343 – 51. doi:10.1001/jamasurg.2020.6357
  5. 5. Feng S-T, Law MW-M, Huang B, Ng S, Li Z-P, Meng Q-F, Khong P-L, Radiation dose and cancer risk from pediatric CT examinations on 64-slice CT: a phantom study. Eur J Radiol. (2010) 76 (2): e19 – 23. doi:10.1016/j.ejrad.2010.03.005
  6. 6. Pearce MS, Salotti JA, Little MP, McHugh K, Lee C, Kim KP, Howe NL, Ronckers CM, Rajaraman P, Craft AW, Parker L, Berrington de González A, Radiation exposure from CT scans in childhood and subsequent risk of leukaemia and brain tumours: a retrospective cohort study. Lancet. (2012) 380 (9840): 499 – 505. doi:10.1016/S0140-6736(12)60815-0
  7. 7. Sodickson A, Baeyens PF, Andriole KP, Prevedello LM, Nawfel RD, Hanson R, Khorasani R, Recurrent CT, cumulative radiation exposure, and associated radiation-induced cancer risks from CT of adults. Radiology. (2009) 251 (1): 175 – 84. doi:10.1148/radiol.2511081296
  8. 8. Verdun FR, Bochud F, Gundinchet F, Aroua A, Schnyder P, Meuli R, Quality initiatives radiation risk: what you should know to tell your patient. Radiographics. (2008) 28 (7): 1807 – 16. doi:10.1148/rg.287085042
  9. 9. Wang G, Ye JC, De Man B. Deep learning for tomographic image reconstruction. Nat Mach Intell. (2020) 2 (12): 737 – 48. doi:10.1038/s42256-020-00273-z
  10. 10. Fang R, Zhang S, Chen T, Sanelli PC, Robust low-dose CT perfusion deconvolution via tensor total-variation regularization. IEEE Trans Med Imaging. (2015) 34 (7): 1533 – 48. doi:10.1109/TMI.2015.2405015
  11. 11. Green M, Marom EM, Kiryati N, Konen E, Mayer A. Efficient low-dose CT denoising by locally-consistent non-local means (LC-NLM). In: Ourselin S, Joskowicz L, Sabuncu MR, Unal G, Wells W, editors. Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part III. Cham: Springer International Publishing (2016). p. 423 – 31.
  12. 12. Chen Y, Shi L, Feng Q, Yang J, Shu H, Luo L, Coatrieux J-L, Chen W, Artifact suppressed dictionary learning for low-dose CT image processing. IEEE Trans Med Imaging. (2014) 33 (12): 2271 – 92. doi:10.1109/TMI.2014.2336860
  13. 13. Chen H, Zhang Y, Kalra MK, Lin F, Chen Y, Liao P, Zhou J, Wang G, Low-dose CT with a residual encoder-decoder convolutional neural network. IEEE Trans Med Imaging. (2017) 36 (12): 2524 – 35. doi:10.1109/TMI.2017.2715284
  14. 14. Zhang Z, Liang X, Dong X, Xie Y, Cao G, A sparse-view CT reconstruction method based on combination of DenseNet and deconvolution. IEEE Trans Med Imaging. (2018) 37 (6): 1407 – 17. doi:10.1109/TMI.2018.2823338
  15. 15. Lv X, Ren X, He P, Zhou M, Long Z, Guo X, Fan C, Wei B, Feng P, Image denoising and ring artifacts removal for spectral CT via deep neural network. IEEE Access. (2020) 8: 225594 – 601. doi:10.1109/ACCESS.2020.3044708
  16. 16. Jiang Z, Chen Y, Zhang Y, Ge Y, Yin F-F, Ren L, Augmentation of CBCT reconstructed from under-sampled projections using deep learning. IEEE Trans Med Imaging. (2019) 38 (11): 2705 – 15. doi:10.1109/TMI.2019.2912791
  17. 17. Bazrafkan S, Van Nieuwenhove V, Soons J, De Beenhouwer J, Sijbers J, To recurse or not to recurse: a low-dose CT study. Prog Artif Intell. (2021) 10: 65 – 81. doi:10.1007/s13748-020-00224-0
  18. 18. Yan R, Liu Y, Liu Y, Wang L, Zhao R, Bai Y, Gui Z, Image denoising for low-dose CT via convolutional dictionary learning and neural network. IEEE Trans Comput Imaging. (2023) 9: 83 – 93. doi:10.1109/TCI.2023.3241546
  19. 19. Tan X, Liu X, Xiang K, Wang J, Tan S. Deep filtered back projection for CT reconstruction. in IEEE Access. (2024) 12: 20962 – 72. doi:10.1109/ACCESS.2024.3357355
  20. 20. Ikuta M, Zhang J. A deep convolutional gated recurrent unit for CT image reconstruction. IEEE Trans Neural Netw Learn Syst. (2023) 34 (12): 10612 – 25. doi:10.1109/TNNLS.2022.3169569
  21. 21. Liu J, Li J. Sparse-sampling CT sinogram completion using generative adversarial networks. In: 2020 13th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI). Piscataway, NJ: IEEE (2020). p. 640 – 4.
  22. 22. Ghani MU, Karl WC. Fast enhanced CT metal artifact reduction using data domain deep learning. IEEE Trans Comput Imaging. (2019) 6: 181 – 93. doi:10.1109/TCI.2019.2937221
  23. 23. Kim K, Soltanayev S, Chun SY. Unsupervised training of denoisers for low-dose CT reconstruction without full-dose ground truth. IEEE J Sel Top Signal Process. (2020) 14 (6): 1112 – 25. doi:10.1109/JSTSP.2020.3007326
  24. 24. Lin WA, Liao H, Peng C, Sun X, Zhang J, Luo J, Chellappa R, Zhou SK. Dudonet: dual domain network for CT metal artifact reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE (2019). p. 10512 – 21.
  25. 25. Wang C, Shang K, Zhang H, Li Q, Hui Y, Zhou SK. DuDoTrans: dual-domain transformer provides more attention for sinogram restoration in sparse-view CT reconstruction. arXiv [Preprint] arXiv:2111.10790 (2021). Available online at: (Accessed June 25, 2026).
  26. 26. Sun C, Deng K, Liu Y, Yang H. A lightweight dual-domain attention framework for sparse-view CT reconstruction. In: 2022 IEEE 8th International Conference on Computer and Communications (ICCC). Piscataway, NJ: IEEE (2022). p. 2102 – 6.
  27. 27. Zhou B, Chen X, Xie H, Zhou SK, Duncan JS, Liu C, DuDoUFNet: dual-domain under-to-fully-complete progressive restoration network for simultaneous metal artifact reduction and low-dose CT reconstruction. IEEE Trans Med Imaging. (2022) 41 (12): 3587 – 99. doi:10.1109/TMI.2022.3189759
  28. 28. Wu W, Hu D, Niu C, Yu H, Vardhanabhuti V, Wang G, DRONE: dual-domain residual-based optimization network for sparse-view CT reconstruction. IEEE Trans Med Imaging. (2021) 40 (11): 3002 – 14. doi:10.1109/TMI.2021.3078067
  29. 29. Shen L, Pauly J, Xing L. NeRP: implicit neural representation learning with prior embedding for sparsely sampled image reconstruction. IEEE Trans Neural Netw Learn Syst. (2022) 35 (1): 770 – 82. doi:10.1109/TNNLS.2022.3177134
  30. 30. Xu D, Yang Y, Liu H, Lyu Q, Descovich M, Ruan D, Sheng K, TomoGRAF: an x-ray physics-driven generative radiance field framework for extremely sparse view CT reconstruction. PLoS One. (2025) 20 (8): e0330463. doi:10.1371/journal.pone.0330463
  31. 31. McCollough CH, Bartley AC, Carter RE, Chen B, Drees TA, Edwards P, Holmes DR, Huang AE, Khan F, Leng S, McMillan KL, Michalak GJ, Nunez KM, Yu L, Fletcher JG, Low-dose CT for the detection and classification of metastatic liver lesions: results of the 2016 low dose CT grand challenge. Med Phys. (2017) 44 (10): e339 – 52. doi:10.1002/mp.12345
  32. 32. Zhang H, Liu B, Yu H, Dong B, MetaInv-Net: meta inversion network for sparse view CT image reconstruction. IEEE Trans Med Imaging. (2020) 40 (2): 621 – 34. doi:10.1109/TMI.2020.3033541
  33. 33. Wang C, Zhang H, Li Q, Shang K, Lyu Y, Dong B, Zhou SK. Improving generalizability in limited-angle CT reconstruction with sinogram extrapolation. In: de Bruijne M, Cattin PC, Cotin S, Padoy N, Speidel S, Zheng Y, Essert C, editors. Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VI 24. Cham: Springer International Publishing (2021). p. 86 – 96.
  34. 34. Wang C, Shang K, Zhang H, Li Q, Zhou SK. Dudotrans: dual-domain transformer for sparse-view CT reconstruction. In: Wang G, Syben C, Ye JC, editors. International Workshop on Machine Learning for Medical Image Reconstruction. Cham: Springer International Publishing (2022). p. 84 – 94.
  35. 35. Kingma DP, Ba J Adam: a method for stochastic optimization. arXiv [Preprint] arXiv:1412.6980 (2014). Available online at: (Accessed June 25, 2026).
  36. 36. Geng M, Meng X, Yu J, Zhu L, Jin L, Jiang Z, Qiu B, Li H, Kong H, Yuan J, Yang K, Shan H, Han H, Yang Z, Ren Q, Lu Y, Content-noise complementary learning for medical image denoising. IEEE Trans Med Imaging. (2021) 41 (2): 407 – 19. doi:10.1109/TMI.2021.3113365
  37. 37. Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Cham: Springer International Publishing (2015). p. 234 – 41.

Citation

Wang, J., et al. (2026). Cross-Domain TransNet for sparse-view CT reconstruction. Frontiers in Nuclear Medicine, Volume 6. https://doi.org/
Received
Accepted
Published
Volume
6 - 2026

Copyright

© 2026 Wang, Zou, Yang, Liao, Mo, Gao, Ning and Lai.

License: https://creativecommons.org/licenses/by/4.0/

*Correspondence: Peixi Liao; Jiahao Lai

Article metrics

Views
0
Downloads
0
Citations
0
Shares
0

Related articles

ReviewFrom task-specific AI to multimodal foundation models in PET/CT: a readiness framework for clinical translationAmgad Muneer, Maliazurina B. Saad, Hui Xu, Jia WuOriginal ResearchAbsolute stress myocardial blood flow is reduced in patients with prior cardiac disease despite normal perfusion and ejection fraction on rubidium-82 PET: an Arabian Gulf cohortAhmad Alenezi, Masoud Garashi, Layla Ghadenfer, Satish Panchadar, Gautam Biswas, Kawther AliMini ReviewQuantification in bone SPECT/CTJovana Zivanovic, Jules Zhang-Yin