A competition-proven deep learning-based dual-domain framework for low-dose CBCT reconstruction Authors: Xuzhi Zhao, Muliang Li, Yi Du Published: 2026 ABSTRACT BackgroundLow-dose cone-beam computed tomography (CBCT) imaging reduces radiation exposure but inevitably increases noise levels, leading to degradation of reconstructed image quality. This study proposes and systematically evaluates a deep learning-based dual-domain framework, named DLDF, for improving CBCT image quality. The effectiveness of DLDF is demonstrated by its first-place ranking in the low-dose task of the ICASSP-2024 3D-CBCT Challenge.MethodsThe proposed DLDF integrates a projection domain model (PDNet), based on a three-dimensional (3D) Res-UNet architecture, with an image domain model (IDNet), based on a two-dimensional (2D) Res-UNet architecture. Low-dose projections were first denoised by PDNet, followed by Feldkamp-Davis-Kress (FDK) reconstruction, and then further refined by IDNet. The DLDF was trained on data from 800 patients, validated on 100 patients, and tested on 110 patients. On the validation set, the DLDF was compared with 10 state-of-the-art (SOTA) deep learning methods using five quantitative metrics: mean squared error (MSE), peak signal-to-noise ratio (PSNR), feature similarity (FSIM), structural similarity (SSIM), and visual information fidelity (VIF). Statistical analysis was conducted using the Friedman test, followed by the post-hoc Durbin-Conover test with Holm-Bonferroni correction. On the test set, the DLDF was compared against 8 competing methods submitted by other participants.ResultsThe proposed DLDF preserved anatomical structures and fine details with notable improvements in overall visual integrity, while also demonstrating improved performance across quantitative metrics. On the validation set, the DLDF achieved average MSE, PSNR, FSIM, SSIM, and VIF values of 0.00162, 35.54 dB, 0.9709, 0.9186, and 0.8492, respectively, outperforming all 10 SOTA methods with statistically significant improvements across all metrics (P<0.05). On the test set, the DLDF achieved an average MSE of 0.00145, ranking first among all competing methods submitted by the participants.ConclusionsThe proposed DLDF effectively improves the quality of low-dose CBCT images. By enabling high-quality reconstructions from lower radiation exposures, this framework may help reduce patient dose and associated health risks in clinical CBCT applications. The code is publicly available at: https://github.com/Xuzhi-bjtu/ICASSP-2024-3D-CBCT-challenge-Low-dose-CBCT-reconstruction. Introduction Cone-beam computed tomography (CBCT) is a widely used imaging modality in various clinical applications, such as dentistry (1), otolaryngology (2), small joint imaging (3), image-guided radiotherapy (4), and image-guided surgery (5). As a compact X-ray-based imaging technique, CBCT enables high-resolution, three-dimensional (3D) visualization of patient anatomy (6,7). However, patient exposure to ionizing radiation during CBCT examinations poses a potential risk of radiation-induced cancer (8). Therefore, the development of low-dose CBCT imaging techniques is of great importance in clinical practice (9) and has received increasing attention from the international radiology community (10). Reducing the X-ray tube current, expressed in milliampere-seconds (mAs), is a straightforward and commonly used approach to lowering the radiation dose in CBCT imaging (11), as the dose scales approximately linearly with the mAs per projection (12). Nevertheless, reduced tube current leads to increased noise levels in the acquired projections, thereby degrading the quality of the reconstructed images (13). To mitigate this issue, various prior-based iterative reconstruction methods have been developed (14-22). For instance, Liu et al. (21) developed an iterative reconstruction method that incorporates the Hessian Schatten penalties into the objective function, effectively suppressing noise and preserving edge details in low-dose CBCT images. Despite their effectiveness, iterative reconstruction methods often suffer from high computational complexity and long processing times (23), which limit their practical use in clinical settings. In recent years, deep learning (DL) techniques have emerged as powerful tools for low-dose CT/CBCT imaging (24-28), demonstrating superior performance over traditional iterative reconstruction methods (29). These DL-based approaches are often used in conjunction with conventional filtered back projection (FBP) or Feldkamp-Davis-Kress (FDK) reconstruction algorithms (30,31) and can be categorized into three main groups according to their application stage relative to image reconstruction: (I) projection domain pre-processing (32-36), (II) image domain post-processing (37-50), and (III) projection-image dual-domain processing (51-57). The first category, projection domain pre-processing methods, aims to denoise low-dose projection data before FBP/FDK reconstruction, with representative works including SS-UNet (32), AttRDN (33), and SIST (34). Moreover, Geng et al. (35) and Choi et al. (36) proposed unsupervised learning models that enable network training without requiring clean reference projections. The second category, image domain post-processing methods, operates on the degraded CT/CBCT images reconstructed from low-dose projection data using the FBP/FDK algorithm. Representative convolutional neural network (CNN) models include RED-CNN (37), FBPConvNet (38), DD-Net (39), CaGAN (40), FRCNN (41), CNCL-UNet (42), and ACGNet-NSPL (43). To alleviate the over-smoothing effects commonly observed in CNN-generated images, diffusion-based models such as CoCoDiff (44) and CoreDiff (45) have been developed. Furthermore, to address the limited ability of CNNs to model long-range dependencies and capture global semantic information, Transformer-based models such as TransCT (46), STEDNet (47), CBCTformer (48), HybridFormer (49), and CT-Denoimer (50) have been proposed. The third category, projection-image dual-domain processing methods, incorporates a projection domain model, the FBP/FDK reconstruction algorithm, and an image domain model to progressively denoise low-dose projection data, reconstruct images, and refine structural details. By jointly exploiting information from both the projection and image domains, dual-domain frameworks typically achieve superior performance compared with single-domain methods. For fan-beam CT, representative methods include DP-ResNet (51), iBP-Net (52), DRCNN (53), TADDM-Net (54), and DDoS (55). For CBCT, Hu et al. proposed HDNet (56), a 3D CNN-based framework, while Chao et al. introduced Dual-AGNet (57), which employs two 3D CNN models incorporating spatial attention mechanisms to adequately utilize the spatial context information from consecutive projections and reconstructs slices in CBCT. Despite these advancements in DL methods for low-dose CT/CBCT imaging, several notable gaps remain. First, existing studies have primarily focused on image domain post-processing, whereas projection-image dual-domain methods remain limited with their potential to be further explored. Second, the majority of existing methods has been developed and evaluated using the 2016 National Institutes of Health (NIH)-American Association of Physicists in Medicine (AAPM)-Mayo Clinic Low-Dose CT Grand Challenge dataset (58), which provides simulated CT data at 25% of the normal-dose level. However, the performance of these methods on CT/CBCT data acquired at lower dose levels has not yet been thoroughly validated. Third, although some studies (35,36,45,54,57,59) have attempted to simulate lower-dose conditions by introducing additional noise into the projection data, they typically consider only Poisson noise, which does not fully capture the complex noise characteristics observed in clinical CT/CBCT acquisitions. Specifically, in practical imaging systems, due to their hardware limitations and imaging geometry, additional noise sources including electronic noise, quantization noise, and scatter are inevitably present (60-64). Therefore, these methods need to be further validated on low-dose CT/CBCT datasets that more closely resemble real-world clinical settings. To advance DL methods for real-world clinical CBCT imaging, the Institute of Electrical and Electronics Engineers (IEEE) International Conference on Acoustics, Speech, and Signal Processing (ICASSP) launched a dedicated 3D-CBCT Grand Challenge in 2024 (65,66). A key task in this challenge was to reconstruct CBCT images from low-dose projections corresponding to 10% of the clinical dose level. In response, we proposed an improved DL-based dual-domain framework, named DLDF, which integrates a projection domain denoising model (PDNet) with an image domain refinement model (IDNet). As reported in the official final ranking (65,66), the proposed DLDF exhibited robust performance, ultimately achieving first-place among the 9 participants in the challenge. Furthermore, based on the low-dose challenge dataset and five quantitative evaluation metrics, we compared the DLDF with 10 state-of-the-art (SOTA) DL methods reported in recent literature. The experimental results demonstrated that the DLDF consistently achieved the best performance among all 11 evaluated methods, with higher reconstruction fidelity for fine anatomical structures. As an extended challenge paper featuring a comprehensive comparison with SOTA methods, the highlights of this work are summarized as follows: We developed DLDF, an improved dual-domain framework for low-dose CBCT reconstruction that integrates projection and image domain models (PDNet and IDNet), both based on the Res-UNet architecture. We developed DLDF, an improved dual-domain framework for low-dose CBCT reconstruction that integrates projection and image domain models (PDNet and IDNet), both based on the Res-UNet architecture. Extensive comparisons demonstrated that the DLDF outperformed 10 SOTA methods reported in the literature on the validation set, achieving statistically significant improvements across all metrics (P<0.05) and exhibiting higher fidelity in reconstructing fine anatomical structures. Furthermore, the DLDF surpassed 8 competing methods submitted by other participants on the test set. Extensive comparisons demonstrated that the DLDF outperformed 10 SOTA methods reported in the literature on the validation set, achieving statistically significant improvements across all metrics (P<0.05) and exhibiting higher fidelity in reconstructing fine anatomical structures. Furthermore, the DLDF surpassed 8 competing methods submitted by other participants on the test set. All DL methods were evaluated on the official grand challenge CBCT dataset, which is characterized by an ultra-low-dose level (10% of the clinical dose) and incorporates complex noise sources (Poisson noise, electronic noise, quantization noise, and scatter). This dataset provides a realistic and challenging benchmark for assessing method performance under near-clinical conditions. All DL methods were evaluated on the official grand challenge CBCT dataset, which is characterized by an ultra-low-dose level (10% of the clinical dose) and incorporates complex noise sources (Poisson noise, electronic noise, quantization noise, and scatter). This dataset provides a realistic and challenging benchmark for assessing method performance under near-clinical conditions. We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0723/rc). Methods Overview of DLDF-based CBCT reconstruction Figure 1 shows the workflow of the proposed DLDF for low-dose CBCT reconstruction. First, the low-dose projections were denoised using the PDNet model. The denoised projections were then reconstructed into tomographic CBCT images using the FDK algorithm (31). Finally, the reconstructed images were further refined using the IDNet model. Overall workflow of the proposed DLDF for low-dose CBCT reconstruction. CBCT, cone-beam computed tomography; DLDF, deep learning-based dual-domain framework; FDK, Feldkamp-Davis-Kress algorithm. DL model architecture The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Using the low-dose 3D CBCT dataset provided by the ICASSP Challenge (65,66), we explored and evaluated more than 20 network architectures, including CNN-, diffusion-, and Transformer-based models (67-69). Our experiments showed that, among these, the Res-UNet architecture within the CNN family achieved the best generalization performance on the dataset. Therefore, the Res-UNet was selected as the final architecture for both the PDNet and IDNet models within the proposed DLDF. PDNet for projection denoising Figure 2 shows the architecture of the PDNet model within the proposed DLDF, which is based on a 3D Res-UNet. The model takes a 3D patch extracted from the low-dose projections as input and generates the corresponding denoised 3D patch as final output. Overall architecture of the PDNet model within the proposed DLDF. DLDF, deep learning-based dual-domain framework; ReLU, Rectified Linear Unit. The model adopts an encoder-decoder structure. The encoder comprises four downsampling stages that progressively reduce an input of size D×H×W×1 to a latent representation of size (D/16)×(H/16)×(W/16)×512. The decoder, in turn, consists of four symmetric upsampling stages that upsample the encoded representation back to the original resolution D×H×W×1. Each encoder stage includes two consecutive 3×3×3 convolutional layers (zero-padded) with a stride of 1, each followed by a batch normalization (BN) layer and a Leaky Rectified Linear Unit (LReLU) activation function with a negative slope of 0.2. Residual learning is incorporated by adding the stage input to the output of the second LReLU activation function. Each stage concludes with a downsampling operation implemented via a 3×3×3 convolutional layer (zero-padded) with a stride of 2. In the decoder, each stage ends with an upsampling operation using a 3×3×3 transposed convolutional layer (zero-padded) with a stride of 2. Skip connections are employed to enable the decoder stages to fuse features from the corresponding encoder stages through channel-wise concatenation of feature maps. Furthermore, a global residual learning technique is adopted, where the model input is directly added to the output, and the resulting summation serves as the final output. IDNet for image refinement Figure 3 shows the architecture of the IDNet model within the proposed DLDF, which is based on a two-dimensional (2D) Res-UNet. The model takes a 2D axial image from the reconstructed 3D CBCT volume as input, aiming to refine it by suppressing the residual noise while preserving the delicate anatomical structures. Overall architecture of the IDNet model within the proposed DLDF. BN, batch normalization; DLDF, deep learning-based dual-domain framework; ReLU, Rectified Linear Unit. The encoder in the model progressively reduces an input of size H×W×1 to a latent representation of size (H/16)×(W/16)×1024, while the decoder symmetrically upsamples this representation back to the original resolution H×W×1. Each encoder and decoder stage contains two or three consecutive 3×3 convolutional layers with a stride of 1. Downsampling in the encoder is performed by a max-pooling layer with a stride of 2, whereas upsampling in the decoder is achieved by a 3×3 transposed convolutional layer with a stride of 2. All convolutional and transposed convolutional layers are followed by a BN layer and a Rectified Linear Unit (ReLU) activation function. Similar to PDNet, the model also employs skip connections and a global residual learning technique. Loss function Both the PDNet and IDNet models were trained using the mean squared error (MSE) loss function, which is defined as: Where F denotes the DL model with parameters Θ, and M denotes the mini-batch size. Specifically, for the PDNet model, X and Y denote the input projection patch and the reference patch, respectively, and N denotes the number of voxels within a patch. For the IDNet model, X and Y denote the input axial image and the ground truth (GT) axial image, respectively, and N denotes the number of pixels in an axial image. Data The ICASSP-2024 3D-CBCT Challenge provided a dataset comprising 1,010 lung nodule cases (65,66), derived from the publicly available LIDC-IDRI dataset (70). As described by the challenge organizers, clean CBCT projections were first generated by applying 3D cone-beam forward projections to thoracic CT images from the LIDC-IDRI dataset using the ASTRA Toolbox (71). Two types of noisy projections were then obtained from the clean projections by adding photon noise, electronic noise, quantization noise, and scatter at different levels, corresponding to 10% of the clinical dose (low-dose) and 100% of the clinical dose CBCT acquisitions. Only the noisy projections at these two dose levels were released to participants, while the clean projections were withheld. In addition, tomographic images reconstructed from the clean projections using the FDK algorithm (31) were provided as the GT. The geometric and reconstruction parameters of the simulated CBCT are summarized in Table 1 (65). In this study, the 100% clinical dose projections were used as supervision targets for training the PDNet model, whereas the GT images were used as supervision targets for training the IDNet model. Table 1Geometric and reconstruction parameters of the simulated CBCTParametersValuesScan range (degree)360Interval (degree)1Detector array dimension (row × column)256×256Detector pixel size (mm2)2.34375×2.34375Source-to-detector distance (mm)1,050Source-to-isocenter distance (mm)575Reconstructed volume dimension (depth × height × width)256×256×256Reconstructed voxel size (mm3)1.171875×1.171875×1.171875Reused with permission from Zhao et al. (72). CBCT, cone-beam computed tomography. Geometric and reconstruction parameters of the simulated CBCT Reused with permission from Zhao et al. (72). CBCT, cone-beam computed tomography. The 1,010 cases were divided into three subsets: 800 for training, 100 for validation, and 110 for testing. The training and validation sets were made available to participants for developing their DL methods, whereas the test set was withheld by the challenge organizers for independent evaluation of the submitted methods. Model training details In our experiments, the proposed DLDF weas implemented in Python (v3.8.18) using Pytorch (v2.0.1) with Cuda (v11.8.89). The computational platform comprised an Intel® Xeon® Platinum 8481C CPU @ 2 GHz and an NVIDIA RTX 4090 GPU with 24 GB of memory. The training pipeline of the PDNet and IDNet models within the DLDF was conducted as follows. First, PDNet is a 3D model; however, due to GPU memory limitations, the full 3D projection data could not be directly processed. Therefore, for PDNet training, input and reference patches of size 128×128×128 were extracted from the low-dose and 100% clinical dose full 3D projection data, respectively. Second, after PDNet inference, the predicted denoised patches were reassembled to recover the full projection data. The FDK algorithm was then applied to the denoised full projection data to reconstruct 3D CBCT images. Third, 2D axial images of size 256×256 from the reconstructed 3D CBCT images were paired with the corresponding GT axial images for IDNet training. The mini-batch sizes were set to two patches for PDNet and ten axial images for IDNet. The parameters of all convolutional and transposed convolutional layers were initialized using the Kaiming initialization method (73). The models were optimized with the Adam optimizer (74), with an initial learning rate of 10−3, β1=0.9, and β2=0.999. To enhance training stability and mitigate overfitting, both a learning rate scheduler and an early stopping criterion were employed, guided by the average loss value on the validation set. Specifically, the learning rate was reduced by a factor of ten if the validation loss did not decrease for 10 consecutive epochs, and the training process was terminated if the validation loss failed to decrease for 20 consecutive epochs. Each model was trained for a maximum of 100 epochs. The final model weights submitted to the challenge organizers correspond to those obtained at the training epochs where the validation loss reached its minimum. Evaluation metrics The performance of the proposed DLDF was quantitatively evaluated using five image quality metrics (27,75): MSE, peak signal-to-noise ratio (PSNR), feature similarity (FSIM) (76), structural similarity (SSIM) (77), and visual information fidelity (VIF) (78). The MSE and PSNR are defined as: where X^ and Xdenote the DLDF-generated axial image and the corresponding GT axial image, respectively. NX is the number of pixels in an axial image, and max(X) is the maximum pixel value in the GT axial image. The FSIM is defined as where Ω represents the whole image spatial domain, PCX^(x) and PCX(x) are the phase congruency values at pixel x for the DLDF-generated axial image and the GT axial image, respectively, and GX^(x) and GX(x) are the gradient magnitude values at pixel x for the two images, respectively. Constants T1, T2, α, and β are set to 0.85, 160, 1, and 1, respectively, following the settings in the original paper (76). The term max(PCX^(x),PCX(x)) ensures that the higher phase congruency values at pixel x between the two images are emphasized. The SSIM is defined as: where and represent the mean intensities of local windows in the DLDF-generated axial image and the GT axial image, respectively. σX^ and σX are the corresponding standard deviations, and σX^X denotes the covariance between the two windows. The constants C1 and C2 are set to (0.01L)2 and (0.03L)2, respectively, where L equals to the dynamic range of pixel values in the GT axial image (77). In this study, SSIM was computed using a sliding window of size 11×11. The VIF is an index based on natural scene statistics and image information theory (78). It is defined as: where I(C→N,j;F→N,j|sN,j) and I(C→N,j;E→N,j|sN,j) represent the information that could ideally be extracted by the human visual system from subband j of the DLDF-generated axial image and the GT axial image, respectively. Comparison study On the validation set, the proposed DLDF was compared with 10 SOTA DL methods, including 2 projection domain models (32,33), 6 image domain models (37,38,42,45,47,48), and 2 projection-image dual-domain frameworks (56,57). Detailed descriptions of these methods are provided in Table 2. Among them, Corediff is a diffusion-based model, STEDNet and CBCTformer are Transformer-based models, while the remaining methods are based on CNN architectures. As the source codes for SS-UNet, AttRDN, RED-CNN, FBPConvNet, STEDNet, CBCTformer, HDNet and Dual-AGNet were not publicly available, these methods were re-implemented and fine-tuned. Specifically, the network architectures and hyperparameters were configured according to the methodological descriptions reported in their original publications (32,33,37,38,47,48,56,57). In addition, to ensure a fair comparison, all SOTA methods were trained using the same strategy as DLDF described in Section 3.2, including a learning rate scheduler and an early stopping criterion. For performance evaluation, the metrics MSE, PSNR, FSIM, SSIM, and VIF were calculated. In addition, the complexity of all DL methods was assessed in terms of network parameters, floating point operations (FLOPs), and inference time (27). Table 2Descriptions of different comparison methodsMethods†DescriptionsGTFDK reconstruction with clean projectionsBaselineFDK reconstruction with low-dose projections(P) SS-UNet (32)SS-UNet projection denoising, then FDK reconstruction(P) AttRDN (33)AttRDN projection denoising, then FDK reconstruction(I) RED-CNN (37)FDK reconstruction with low-dose projections, then RED-CNN image denoising(I) FBPConvNet (38)FDK reconstruction with low-dose projections, then FBPConvNet image denoising(I) CNCL-UNet (42)FDK reconstruction with low-dose projections, then CNCL-UNet image denoising(I) CoreDiff (45)FDK reconstruction with low-dose projections, then CoreDiff image denoising(I) STEDNet (47)FDK reconstruction with low-dose projections, then STEDNet image denoising(I) CBCTformer (48)FDK reconstruction with low-dose projections, then CBCTformer image denoising(D) HDNet (56)PDU-Net projection denoising, FDK reconstruction, then IDU-Net image refinement(D) Dual-AGNet (57)Pre-AGNet projection denoising, FDK reconstruction, then Post-AGNet image refinement(D) DLDF (Proposed)PDNet projection denoising, FDK reconstruction, then IDNet image refinement†, the prefixes (P), (I), and (D) indicate deep learning methods in the projection domain, image domain, and projection-image dual-domain, respectively. DLDF, deep learning-based dual-domain framework; FDK, Feldkamp-Davis-Kress algorithm; GT, ground truth. Descriptions of different comparison methods †, the prefixes (P), (I), and (D) indicate deep learning methods in the projection domain, image domain, and projection-image dual-domain, respectively. DLDF, deep learning-based dual-domain framework; FDK, Feldkamp-Davis-Kress algorithm; GT, ground truth. To ensure statistically rigorous comparisons between the DLDF and the SOTA methods, non-parametric statistical tests were employed. First, the Friedman test (79) was used to determine whether significant differences existed among the evaluated methods. When significance was detected, post-hoc pairwise comparisons between DLDF and each SOTA method were performed using the Durbin-Conover test (80). The resulting P values were adjusted using the Holm-Bonferroni correction (81) to control the family-wise error rate. A significance level of 0.05 was adopted for all statistical tests. On the test set, the DLDF was compared against 8 competing methods submitted by other participants. This comparison was conducted by the challenge organizers, using MSE as the sole evaluation metric (65,66). Ablation study on dual-domain strategy To validate the effectiveness of the projection-image dual-domain joint strategy adopted by the proposed DLDF, we conducted ablation experiments by removing either the projection domain processing component or the image domain processing component from the framework. Specifically, on the validation set, the performance of the complete DLDF was compared with two reduced variants: one incorporating only the PDNet model and another incorporating only the IDNet model. All five metrics were calculated. Results Performance comparison with SOTA methods Table 3 presents the quantitative comparison between the proposed DLDF and the 10 SOTA DL methods on the validation set. The DLDF consistently outperformed all SOTA methods across all five metrics. Compared to the best average metrics obtained by the SOTA methods, the DLDF showed improvements of 0.00012 in MSE, 0.34 dB in PSNR, 0.0008 in FSIM, 0.0051 in SSIM, and 0.0099 in VIF. Table 3Quantitative comparison between the proposed DLDF and state-of-the-art deep learning methods on the validation setMethods†MSE (×10-3)PSNR (dB)FSIM (×10-2)SSIM (×10-2)VIF (×10-2)Baseline79.60±21.5718.41±1.6969.24±4.4032.17±4.6644.35±4.44(P) SS-UNet28.38±10.1922.96±1.7589.79±1.9375.41±4.8067.65±7.45(P) AttRDN28.14±9.4322.97±1.7891.13±1.8877.97±4.8573.96±7.08(I) RED-CNN3.04±1.3532.71±1.9196.63±1.5686.68±4.2278.08±5.41(I) FBPConvNet2.39±0.8433.71±1.8497.06±1.3088.75±3.7680.63±5.22(I) CNCL-UNet2.25±0.9234.06±1.9596.29±1.8989.58±4.0681.07±5.39(I) CoreDiff2.05±0.8734.47±2.0296.73±1.7889.92±4.0082.94±5.22(I) STEDNet2.95±1.1732.88±1.8195.61±1.8088.05±4.1277.24±5.65(I) CBCTformer2.51±1.0533.61±1.9296.20±1.7088.96±4.0680.26±5.33(D) HDNet1.74±0.7635.20±2.0397.01±1.8591.35±3.7783.93±5.39(D) Dual-AGNet3.55±2.1332.18±2.0395.65±1.9486.34±4.9577.84±6.12(D) DLDF (Proposed)1.62±0.71*35.54±2.04*97.09±1.93*91.86±3.60*84.92±5.27*All values are reported as mean ± standard deviation. †, baseline refers to FDK-reconstructed images without any deep learning method. The prefixes (P), (I), and (D) indicate deep learning methods in the projection domain, image domain, and projection-image dual-domain, respectively. The best results are marked with asterisks (*). DLDF, deep learning-based dual-domain framework; FDK, Feldkamp-Davis-Kress algorithm; FSIM, feature similarity; MSE, mean squared error; PSNR, peak signal-to-noise ratio; SSIM, structural similarity; VIF, visual information fidelity. Quantitative comparison between the proposed DLDF and state-of-the-art deep learning methods on the validation set All values are reported as mean ± standard deviation. †, baseline refers to FDK-reconstructed images without any deep learning method. The prefixes (P), (I), and (D) indicate deep learning methods in the projection domain, image domain, and projection-image dual-domain, respectively. The best results are marked with asterisks (*). DLDF, deep learning-based dual-domain framework; FDK, Feldkamp-Davis-Kress algorithm; FSIM, feature similarity; MSE, mean squared error; PSNR, peak signal-to-noise ratio; SSIM, structural similarity; VIF, visual information fidelity. Table 4 presents the statistical analysis results comparing the DLDF with the SOTA methods. The performance gains delivered by the DLDF were statistically significant for all metrics. Specifically, the p-values from the post-hoc Durbin-Conover test for pairwise comparisons between the DLDF and each SOTA method were all below 0.05. Table 4Statistical analysis of DLDF versus state-of-the-art deep learning methodsStatistical analysis†MSEPSNRFSIMSSIMVIFFriedman testP<0.05P<0.05P<0.05P<0.05P<0.05Durbin-Conover testP<0.05P<0.05P<0.05P<0.05P<0.05†, Friedman test was used to determine whether statistically significant differences existed among the evaluated methods. When significance was detected, post-hoc pairwise comparisons between DLDF and each state-of-the-art method were performed using Durbin-Conover test, with Holm-Bonferroni correction applied to adjust for multiple comparisons. DLDF, deep learning-based dual-domain framework; FSIM, feature similarity; MSE, mean squared error; PSNR, peak signal-to-noise ratio; SSIM, structural similarity; VIF, visual information fidelity. Statistical analysis of DLDF versus state-of-the-art deep learning methods †, Friedman test was used to determine whether statistically significant differences existed among the evaluated methods. When significance was detected, post-hoc pairwise comparisons between DLDF and each state-of-the-art method were performed using Durbin-Conover test, with Holm-Bonferroni correction applied to adjust for multiple comparisons. DLDF, deep learning-based dual-domain framework; FSIM, feature similarity; MSE, mean squared error; PSNR, peak signal-to-noise ratio; SSIM, structural similarity; VIF, visual information fidelity. Figure 4 shows the representative axial images of two patients reconstructed under different settings: from clean projections (GT), from low-dose projections without any DL method (baseline), and from low-dose projections using the projection domain models, image domain models, and projection-image dual-domain frameworks (including the proposed DLDF). Figure 5 shows the corresponding absolute difference maps with respect to the GT for each method. Figure 6 reports the MSE, PSNR, FSIM, SSIM, and VIF metrics computed on the reconstructed images shown in Figure 4. As illustrated in Figure 4 (b1,b2) and Figure 5 (a1,a2), the baseline reconstructions without any DL method suffer from severe noise, which obscures the anatomical structures. In Figure 4 (c1-i1,c2-i2) and Figure 5 (b1-h1,b2-h2), the single-domain DL methods improve the overall image quality of low-dose reconstructions but still leave residual noise or introduce over-smoothing, leading to the loss of fine structural details. By contrast, in Figure 4 (j1,l1,j2,l2) and Figure 5 (i1,k1,i2,k2), the dual-domain DL methods better preserve the anatomical structures and fine details, with lower absolute pixel errors relative to the GT images. Among all methods, as indicated by the red arrows in the upper-left regions of interest (ROIs) of Figure 4 (a2-l2), the proposed DLDF achieves the highest reconstruction fidelity, preserving tiny pulmonary vessel structures that most closely resemble those in the GT images. These qualitative observations are further supported by the quantitative results in Figure 6, where DLDF achieves the best performance across all five metrics (MSE, PSNR, FSIM, SSIM, and VIF). Representative axial images from two patients: (a1,a2) GT images reconstructed from clean projections; (b1,b2) baseline images reconstructed from low-dose projections without any deep learning method; images reconstructed from low-dose projections using (c1,c2) SS-UNet, (d1,d2) AttRDN, (e1,e2) FBPConvNet, (f1,f2) CNCL-UNet, (g1,g2) CoreDiff, (h1,h2) STEDNet, (i1,i2) CBCTformer, (j1,j2) HDNet, (k1,k2) Dual-AGNet, and (l1,l2) the proposed DLDF. A magnified ROI is displayed in the upper-left corner of each subfigure. The colorbar on the far right of each group applies to all subfigures within that group. DLDF, deep learning-based dual-domain framework; GT, ground truth; ROI, region of interest. Absolute different maps between the GT axial images and the reconstructed axial images obtained by different methods for two patients: (a1,a2) baseline (no deep learning method), (b1,b2) SS-UNet, (c1,c2) AttRDN, (d1,d2) FBPConvNet, (e1,e2) CNCL-UNet, (f1,f2) CoreDiff, (g1,g2) STEDNet, (h1,h2) CBCTformer, (i1,i2) HDNet, (j1,j2) Dual-AGNet, and (k1,k2) the proposed DLDF. A magnified ROI is displayed in the upper-left corner of each subfigure. The colorbar on the far right of each group applies to all subfigures within that group. DLDF, deep learning-based dual-domain framework; GT, ground truth; ROI, region of interest. Quantitative metrics for reconstructed axial images obtained by different methods for two patients, including (A,F) MSE, (B,G) PSNR, (C,H) FSIM, (D,I) SSIM, and (E,J) VIF. For better visualization, MSE values in (A,F) are presented on a logarithmic scale. DLDF, deep learning-based dual-domain framework; FSIM, feature similarity; MSE, mean squared error; PSNR, peak signal-to-noise ratio; SSIM, structural similarity; VIF, visual information fidelity. Figures 7,8 provide further comparisons of selected profiles from the GT axial images with those from the DL method-generated axial images shown in Figures 4,5, respectively. Specifically, Figure 7 shows profiles passing through high-contrast regions that include lung parenchyma and bronchial structures, while Figure 8 shows profiles passing through low-contrast regions that include cardiac tissue. Relative to the GT profiles, the profiles generated by the DL methods appear smoother, with reduced sharp variations between peak points. This indicates that although the DL methods suppress noise, they also smooth out some useful high-frequency components in the images. Nevertheless, among all methods, the proposed DLDF generates profiles that most closely match the GT profiles, yielding the lowest MSE values. Comparison of high-contrast profiles between the GT in Figure 4 (a1) and deep learning methods in Figure 4 (c1-l1). The profile location is marked by the horizontal red line in Figure 4 (a1), corresponding to pixel positions x=72-101 at y=150. MSE values were calculated to quantify the differences between the GT profile and the profiles generated by each method. DLDF, deep learning-based dual-domain framework; GT, ground truth; MSE, mean squared error. Comparison of low-contrast profiles between the GT in Figure 4 (a2) and deep learning methods in Figure 4 (c2-l2). The profile location is marked by the vertical red line in Figure 4 (a2), corresponding to pixel positions y=83-112 at x=124. MSE values were calculated to quantify the differences between the GT profile and the profiles generated by each method. DLDF, deep learning-based dual-domain framework; GT, ground truth; MSE, mean squared error. Complexity of DL methods Table 5 lists the complexity of the proposed DLDF and the 10 SOTA DL methods. Among these methods, the DLDF ranked 11th in network parameters, 10th in FLOPs, and 8th in inference time, with all rankings reported in ascending order. Table 5Complexity of the proposed DLDF and state-of-the-art deep learning methods, evaluated in terms of network parameters, FLOPs, and inference timeMethods†Parameters (M)FLOPs (T)‡Inference time (ms)§(P) SS-UNet40.0922.19683±28(P) AttRDN0.4510.63654±43(I) RED-CNN1.8531.06192±22(I) FBPConvNet34.5617.37472±21(I) CNCL-UNet43.8218.67755±22(I) CoreDiff4.83104.048,155±58(I) STEDNet74.261.832,781±38(I) CBCTformer99.587.897,187±36(D) HDNet42.3711.321,250±47(D) Dual-AGNet0.7215.11744±79(D) DLDF (Proposed)119.8144.771,486±69†, the prefixes (P), (I), and (D) indicate deep learning methods in the projection domain, image domain, and projection-image dual-domain, respectively. ‡, the FLOPs of each deep learning method are calculated based on obtaining a 3D CBCT reconstructed volume measuring 2563 voxels. §, the inference time of each deep learning method is measured based on obtaining 3D CBCT reconstructed volumes for each patient in the validation set. All values are reported as mean ± standard deviation. The inference time is measured with a mini-batch size of 1 on a computational platform equipped with an NVIDIA RTX 4090 GPU, excluding the execution time of the FDK algorithm. 3D, three-dimensional; CBCT, cone-beam computed tomography; DLDF, deep learning-based dual-domain framework; FDK, Feldkamp-Davis-Kress; FLOPs, floating point operations. Complexity of the proposed DLDF and state-of-the-art deep learning methods, evaluated in terms of network parameters, FLOPs, and inference time †, the prefixes (P), (I), and (D) indicate deep learning methods in the projection domain, image domain, and projection-image dual-domain, respectively. ‡, the FLOPs of each deep learning method are calculated based on obtaining a 3D CBCT reconstructed volume measuring 2563 voxels. §, the inference time of each deep learning method is measured based on obtaining 3D CBCT reconstructed volumes for each patient in the validation set. All values are reported as mean ± standard deviation. The inference time is measured with a mini-batch size of 1 on a computational platform equipped with an NVIDIA RTX 4090 GPU, excluding the execution time of the FDK algorithm. 3D, three-dimensional; CBCT, cone-beam computed tomography; DLDF, deep learning-based dual-domain framework; FDK, Feldkamp-Davis-Kress; FLOPs, floating point operations. Ablation study on dual-domain strategy Table 6 presents the quantitative results of ablation studies on projection-image dual-domain joint strategy adopted by the proposed DLDF, evaluated on the validation set. Compared to the individual PDNet and IDNet models, the complete DLDF showed superior performance across all metrics. Table 6Ablation studies on the projection-image dual-domain joint strategy.Methods†MSE (×10-3)PSNR (dB)FSIM (×10-2)SSIM (×10-2)VIF (×10-2)(P) PDNet27.59±10.0023.09±1.7891.41±1.8478.83±4.7375.24±6.92(I) IDNet2.09±0.8434.38±1.9696.58±1.7289.74±3.8382.92±5.16(D) DLDF1.62±0.71*35.54±2.04*97.09±1.93*91.86±3.60*84.92±5.27*All experiments were conducted on the validation set and all values are reported as mean ± standard deviation. The best results are marked with asterisks (*). †, the prefixes (P), (I), and (D) indicate deep learning methods in the projection domain, image domain, and projection-image dual-domain, respectively. DLDF, deep learning-based dual-domain framework; FSIM, feature similarity; MSE, mean squared error; PSNR, peak signal-to-noise ratio; SSIM, structural similarity; VIF, visual information fidelity. Ablation studies on the projection-image dual-domain joint strategy. All experiments were conducted on the validation set and all values are reported as mean ± standard deviation. The best results are marked with asterisks (*). †, the prefixes (P), (I), and (D) indicate deep learning methods in the projection domain, image domain, and projection-image dual-domain, respectively. DLDF, deep learning-based dual-domain framework; FSIM, feature similarity; MSE, mean squared error; PSNR, peak signal-to-noise ratio; SSIM, structural similarity; VIF, visual information fidelity. Figure 9 shows the representative axial images of another patient reconstructed from low-dose projections using the individual PDNet and IDNet models, as well as the complete DLDF. As illustrated in the ROIs of all subfigures, the complete DLDF produces anatomical structures and details more consistent with the GT image than those obtained using PDNet or IDNet alone, while also exhibiting lower absolute pixel errors. Representative axial images from another patient: (A) GT image reconstructed from clean projections; (B) images reconstructed from low-dose projections using PDNet, IDNet, and DLDF; (C) absolute difference maps between (B) and (A). A magnified ROI is displayed in the upper-left corner of each subfigure. The colorbar on the far right of each group applies to all subfigures within that group. DLDF, deep learning-based dual-domain framework; GT, ground truth; ROI, region of interest. Figure 10 further compares a selected high-contrast profile from the GT axial image with those from the PDNet-, IDNet-, and DLDF-generated axial images shown in Figure 9. Among the three methods, the profile produced by DLDF shows the closest agreement with the GT profile and achieves the lowest MSE value. Comparison of high-contrast profiles between the GT in Figure 9A and deep learning methods in Figure 9B. The profile location is marked by the vertical red line in Figure 9A, corresponding to pixel positions y=135-164 at x=164. MSE values were calculated to quantify the differences between the GT profile and the profiles generated by each method. GT, ground truth; MSE, mean squared error. Performance comparison with grand challenge competitors Table 7 presents the quantitative MSE-based comparison of the competing DL methods submitted by all participants on the test set (65,66). The proposed DLDF, developed by our team BJTU_PKUCH, achieved first-place in the low-dose task of the ICASSP 3D-CBCT Challenge. Table 7Quantitative MSE-based comparison of competing deep learning methods submitted by all participants on the test setParticipating teamsMSE (×10-3)Mikael Brix33.81Congcong Zhu20.48Thomas Blumensath16.10Sasidhar Alavala8.84Austin Yunker1.94Andreas Hauptmann21.71Andreas Hauptmann11.68Mikael Brudfors1.48BJTU_PKUCH (Proposed, Winner)1.45*All values are reported as mean. The best result is marked with asterisk (*). MSE, mean squared error. Quantitative MSE-based comparison of competing deep learning methods submitted by all participants on the test set All values are reported as mean. The best result is marked with asterisk (*). MSE, mean squared error. Discussion Although various DL methods have been developed for low-dose CBCT reconstruction, several notable gaps remain. First, projection-image dual-domain approaches remain underexplored, and their potential advantages have not been fully investigated. Second, most existing methods are evaluated on the Mayo-LDCT dataset at a simulated 25% dose level (58), with limited validation under more challenging lower-dose CBCT settings that better reflect clinical conditions. In this study, we propose a dual-domain DL framework, DLDF, which achieved first place in the low-dose task of the ICASSP 3D-CBCT Challenge (65,66), an international competition aimed at advancing DL methods for real-world clinical CBCT imaging. The comparison with 10 SOTA methods and 8 competing submissions highlighted the improved performance of the proposed DLDF in low-dose CBCT reconstruction. Based on the corresponding results presented in Tables 3,4,7 and Figures 4-8, two major findings can be summarized. First, the dual-domain design of DLDF showed considerable promise in improving the quality of low-dose CBCT images. The consistent improvements achieved by DLDF over single-domain methods across five quantitative metrics and in terms of visual fidelity suggested that effectively exploiting complementary information from both domains is beneficial for low-dose CBCT reconstruction, providing valuable insights for the development of future DL methodologies. Second, the DLDF demonstrated stronger adaptability and robustness when applied to ultra-low-dose CBCT datasets acquired under near-clinical conditions. Its favorable performance over all comparison methods indicated an enhanced capability to handle the complex degradation factors present in real-world clinical CBCT imaging, including various sources of noise and scatter contamination. The ablation study further confirmed the effectiveness of the projection-image dual-domain joint strategy. Since the PDNet and IDNet components within the proposed DLDF differed in architecture from the single-domain methods used in the comparison study, it is insufficient to attribute the performance gains of DLDF solely to the dual-domain design. Nor can it be asserted that the complete DLDF necessarily outperforms PDNet or IDNet when operated independently. To more fairly evaluate the contribution of each component, we directly compared the complete DLDF with PDNet-only and IDNet-only variants. As shown in Table 6 and Figures 9,10, DLDF achieved the best overall performance, providing strong evidence for the contribution of the dual-domain design. This observation is consistent with previous findings that multi-domain joint strategies are particularly beneficial for low-dose CBCT imaging, especially at dose levels of one-eighth of the normal dose or lower (27). As shown in Table 3, the existing dual-domain framework HDNet (56) achieves the second-best performance, following the proposed DLDF. Compared with HDNet, the advantages of DLDF can be attributed to several architectural differences. In the projection domain network, PDNet additionally introduces residual learning at both the input and output of two consecutive convolutional layers. Moreover, a global residual learning strategy is incorporated from the network input to the output. In contrast, the image domain network IDNet adopts a 2D architecture and also employs a global residual learning strategy from the network input to the output. These key differences in residual learning design significantly enhance the trainability and representational capacity (67) of the proposed DLDF, thereby improving optimization stability and overall performance. Consequently, the DLDF achieves higher reconstruction fidelity, particularly in preserving fine anatomical structures, highlighting the practical advantages of the proposed architectural improvements. The clinical implications of the proposed DLDF are twofold. First, as shown in Figure 4 (b1, b2), the images reconstructed without any DL method are severely degraded by noise, rendering anatomical structures and details indistinguishable. Such poor image quality may hinder physicians from accurately identifying lung nodules and performing reliable diagnostic assessments, thereby compromising clinical decision-making. In contrast, as shown in Figure 4 (l1, l2), the images reconstructed using the DLDF demonstrate substantially improved quality. The DLDF effectively suppresses noise while preserving anatomical structures and details, thereby potentially enhancing the diagnostic reliability and clinical utility of low-dose CBCT imaging. Second, because the GT images used to train the IDNet model were reconstructed with the FDK algorithm rather than iterative methods, the DLDF generates refined images that retain the intrinsic texture patterns and visual characteristics typical of FDK reconstructions. Since FDK-reconstructed CBCT images are widely used and well recognized by clinicians in routine practice, this visual consistency helps minimize the risk of introducing additional cognitive workload during image interpretation. This study has five main limitations. First, the proposed DLDF was trained and evaluated exclusively on a thoracic CBCT dataset at a simulated 10% dose level using a specific noise-injection pipeline provided by the challenge organizers. Future efforts will focus on evaluating the external generalizability of the method using low-dose CBCT data acquired from real clinical environments, covering different scanner systems and diverse anatomical regions. Second, the DLDF still exhibits a degree of smoothing in anatomical details (e.g., pulmonary vessels), which may lead to potential hallucination artifacts in the reconstructed images. Future studies will systematically compare GT images and DLDF-generated reconstructions across different scanner systems and anatomical regions to assess the reliability of the method under varying conditions. Additionally, structural-preserving attention mechanisms will be incorporated into the model architecture to mitigate the loss of fine anatomical details. Third, the DLDF was not compared with the first-ranked model in another clinical dose CBCT reconstruction task released by the challenge organizers (72). Since the clean projections required for supervised learning were not provided in the released dataset, consequently, projection domain models could not be trained, and the dual-domain strategy could not be applied to this task. Future studies will seek to obtain the clean projections from the challenge organizers and further validate the performance of the DLDF for clinical dose CBCT reconstruction. Fourth, the DLDF exhibits relatively high complexity in terms of both the number of network parameters and FLOPs compared to other DL methods included in the comparison study (see Table 5). Future research will aim to design a lightweight dual-domain framework that maintains strong representational capability. Fifth, due to the limited computational capacity of a single 24 GB GPU, the PDNet and IDNet models within the DLDF were trained independently in two stages, which may limit global optimization of the entire framework. Future work will investigate integrating a PyTorch-based differentiable FDK algorithm as a bridging module/layer (82) between the lightweight projection domain and image domain models, enabling an end-to-end trainable dual-domain framework with globally optimized parameters. In addition, future work will further refine the framework to correct motion artifacts commonly encountered in clinically relevant CBCT reconstruction (83,84). Conclusions The proposed DLDF, which integrates the projection domain model PDNet with the image domain model IDNet, exhibits prevailing performance over SOTA and competing DL methods in improving the quality of low-dose CBCT images. This framework may help reduce the clinical CBCT imaging dose, thereby mitigating the associated health risks for patients. Supplementary The article’s supplementary files as