Organizations: School of Computing and Data Science, The University of Hong Kong, Hong Kong, China · Engineering Research Center of Industrial Computed Tomography Nondestructive Testing, Ministry of Education, Chongqing University, Chongqing 400044, China · Jiangxi Intelligent Non-Destructive Testing Innovation Center, Nanchang 330096, China · Key Laboratory of Nondestructive Testing (Ministry of Education), Nanchang Hangkong University, Nanchang 330063, China · Department of Stomatology, Peking University Third Hospital, Beijing 100191, China · Institute of Medical Technology, Peking University Health Science Center, Beijing 100191, China · Shenzhen Bay Laboratory, Shenzhen 518132, China · Institute of Science and Technology for Brain-inspired Intelligence, Fudan University, Shanghai 200433, China
Sparse-view cone-beam computed tomography is widely used for fast, low-dose volumetric imaging and lesion localization in dentistry and image-guided interventions, but it introduces severe artifacts. Feed-forward Gaussian methods transfer the reconstruction cost to training and decode each coordinate from projection-aligned features and neighboring anchors. However, information from the whole scan is written only when the field is constructed and cannot be retrieved during coordinate decoding, leaving the volume inconsistent within the scan. This is a common limitation of such methods: once the field is fixed, decoding uses only local features and reconstruction quality drops substantially. We propose SparseTokenGS, a Gaussian-based method for sparse-view cone-beam computed tomography reconstruction via a reusable token pool. Anchor Token Fusion and Point-Token Attention query this shared pool at the anchors and at the query coordinates, respectively. The local Gaussian prior is retained, and the volume is reconstructed in one forward pass, without per-scan optimization or a projection-domain loss. On three public chest and dental datasets, SparseTokenGS achieves the highest reconstruction performance among the compared methods. On clinical dental scans, reconstruction from only six projections reaches 29.51 decibels in peak signal-to-noise ratio and a structural similarity of 0.842 against the full-view reference, and off-the-shelf tooth-and-pulp segmentation keeps these structures separable for root-canal localization. Experiments on volcanic rock and chicken show performance after same-object adaptation. A reusable token pool gives feed-forward reconstruction a mechanism that keeps scan-wide context available through decoding and can benefit future rapid clinical lesion localization and analysis. The code is available at: https://github.com/FrankDengAI/SparseTokenGS.
Figures & tables
Fig. 1: Motivation for dual-level token conditioning in sparse-view CBCT.
Fig. 2: Architecture of SparseTokenGS. Sparse-view projections enter a weight-shared encoder, which keeps a full-resolution map for view alignment and a downsampled map for the anchors and the token pool. The pool is then read twice. In (a), Fuse updates the anchors and Attend updates each query, and the decoded intensities form the volume. In (b), each anchor queries the token pool and a zero-initialized residual is added to its local feature. In (c), a coordinate network embeds the query, and the view feature, the neighboring Gaussian feature, and the point-token feature are concatenated and mapped to one intensity.
Fig. 3: Construction of the scan-specific token pool. (a) Multi-view feature sampling. (b) The token matrix. (c) Cross-attention and self-attention. The grids are schematic.
Method
Parameters (M)
MACs (T)
FBPConvNet
1.9
2.4
DGR
7.1
8.3
DiffNR
1305.5
103.6
R 2 -Gaussian
0.97
3.0
X-GRM
172.6
4.2
DIF-Net
31.1
1.4
TABLE I: Parameters and MAC estimates for six 2562 projections and a 2563 volume. T denotes 1012 MACs. DGR, DiffNR, and R 2 -Gaussian include the six-view iteration budget.
Dataset
View
FDK
SART
FBPConvNet
X-GRM
DIF-Net
DIF-Gaussian
Ours
LUNA16
3
14.40/0.2644/0.1905
18.81/0.5549/0.1157
19.44/0.5872/0.1076
19.24/0.6926/0.1091
22.41/0.7082/0.0758
26.06/0.8254/0.0498
27.52 / 0.8602 / 0.0421
6
15.17/0.2380/0.1744
23.04/0.6741/0.0711
22.26/0.6683/0.0777
24.07/0.7775/0.0626
23.98/0.7178/0.0633
28.16/0.8573/0.0391
30.11 / 0.8933 / 0.0319
10
16.30/0.2483/0.1531
25.10/0.7393/0.0561
23.86/0.6345/0.0648
24.86/0.7967/0.0571
26.42/0.7587/0.0477
29.77/0.8809/0.0325
31.70 / 0.9103 / 0.0264
14
16.80/0.2621/0.1445
26.42/0.7781/0.0483
24.45/0.7134/0.0612
23.28/0.7843/0.0686
27.08/0.7763/0.0443
30.33/0.8872/0.0305
32.32 / 0.9144 / 0.0242
18
17.14/0.2730/0.1389
27.33/0.8037/0.0434
25.06/0.7662/0.0569
24.55/0.7930/0.0592
27.91/0.7946/0.0402
30.91/0.8957/0.0285
32.99 / 0.9206 / 0.0224
ToothFairy
3
14.13/0.1829/0.1965
22.80/0.6680/0.0735
22.42/0.5892/0.0768
20.17/0.6453/0.0980
22.12/0.6506/0.0783
25.81/0.7883/0.0512
27.55 / 0.8538 / 0.0419
TABLE II: PSNR (dB) / SSIM / RMSE on the simulated chest and dental data. Bold marks the best in each row. Ours denotes SparseTokenGS.
Fig. 4: Coronal slices of one LUNA16 scan at 3 , 6 , 10 , 14 , and 18 views. Columns are FDK, X-GRM, DIF-Net, DIF-Gaussian, SparseTokenGS, and the reference. Display window: [−1000,1000] HU.
Fig. 5: Axial slices of the LUNA16 scan in Fig. 4 at fourteen and eighteen views. Columns follow that figure. Colored boxes mark enlarged lung regions. Display window: [−1000,1000] HU.
Fig. 6: Six-view ablation. (a) PSNR and SSIM on LUNA16 and ToothFairy for the settings in Table III . (b) LUNA16 validation PSNR at epoch 400 on the half-resolution grid, plotted against anchor-grid resolution.
Fig. 7: Analytical MAC estimates for a 2563 volume, using the same 1012 -MAC unit as Table I . (a) Six views as the projection side grows. (b) Cost as the number of views changes for the per-scan methods in that estimate.
Fig. 8: Six-view reconstructions. Rows are LUNA16, LUNA25, ToothFairy, and ToothFairy4. Columns are FDK, DGR, DiffNR, R 2 -Gaussian, DIF-Net, X-GRM, DIF-Gaussian, SparseTokenGS, and the reference. Numbers under the panels are PSNR (dB) and SSIM. Rows 1 and 2 use the chest window [−1000,1000] HU. Rows 3 and 4 use the dental window [−800,1200] HU.
Setting
LUNA16
ToothFairy
Main reference
30.1061 / 0.8933
30.6234 / 0.9043
No anchor fusion
29.3950 / 0.8877
30.2162 / 0.9028
20000 samples
29.2359 / 0.8855
30.1663 / 0.9022
No tokens, 20000
29.0458 / 0.8825
30.1534 / 0.9019
No similarity term
28.6847 / 0.8787
29.7013 / 0.8968
Anchor read only
23.7512 / 0.8016
24.7824 / 0.8096
TABLE III: Six-view ablation on LUNA16 and ToothFairy. Each cell is PSNR (dB) / SSIM.
Fig. 9: Six-view reconstruction of one held-out clinical dental scan. Left: axial, coronal, and sagittal sections of the scanner reconstruction, SparseTokenGS, and absolute error. Right: reference and predicted external views.
Fig. 10: TIPs on three SparseTokenGS dental reconstructions, without retraining [ 33 ] .
Fig. 11: Acquisition setup for the volcanic rock and chicken specimens. The eighteen-view reconstructions from this setup are reported in Table IV and Fig. 12 .
Specimen
FDK
X-GRM
DIF-Net
DIF-Gaussian
Ours
Volcanic rock
15.81/0.2724
19.79/0.5316
19.96/0.4874
20.24/0.3732
25.71 / 0.6802
Chicken
16.09/0.3109
21.21/0.5799
21.18/0.5630
21.50/0.4602
28.73 / 0.6861
TABLE IV: Eighteen-view PSNR (dB) / SSIM on volcanic rock and chicken. Ours denotes SparseTokenGS. Bold marks the highest value in each row.
Fig. 12: Eighteen-view slices after adaptation. Columns are the reference, SparseTokenGS, DIF-Gaussian, DIF-Net, X-GRM, and FDK. Rows 1 and 2 are volcanic rock, axial slice 167 and sagittal slice 182. Rows 3 and 4 are chicken, axial slice 81 and sagittal slice 124.
Sparse-view computed tomography (CT) reconstruction recovers 3D images from limited X-ray projections while suppressing artifacts and preserving anatomy.
Sparse-view Cone-Beam Computed Tomography (CBCT) reconstruction is essential for reducing radiation dose and scan effort, but remains underconstrained due to limited projection information. Recent 3D Gaussian Splatting (3DGS)-based CBCT reconstruction methods rely on gradient-driven densification, which becomes unstable under sparse acquisition and often leads to over-smoothed structures. We propose AD-GS, an Anatomical Density-guided 3D Gaussian Splatting framework that formulates Gaussian densification as an attenuation-aware Markov Chain Monte Carlo (MCMC) sampling process. By interpreting Gaussians as samples from an underlying anatomical attenuation distribution, AD-GS adaptively concentrates representation capacity in high-attenuation regions, enabling sharper structural boundaries under sparse projections. To further constrain reconstruction, we incorporate a semantic regularization based on zero-shot features from Segment Anything in Medical Images, enforcing multi-view consistency through semantic priors. Experiments on the ToothFairy3 dental CBCT dataset demonstrate that AD-GS achieves superior structural fidelity and boundary preservation compared with existing 3DGS-based approaches under sparse-view settings within the dental and maxillofacial setting. Code is available at https://github.com/yulinjiaxue/AD-GS-BMVC2026.
Zonglin Li, Kexian Xie, Feiran Li +5
Beijing University of Posts and Telecommunications · Department of Stomatology, Peking University Third Hospital · Independent Researcher +2
Feed-forward 3D Gaussian Splatting (3DGS) reconstructs a scene in a single forward pass, replacing per-scene optimization with a network trained across many scenes. Its quality, however, degrades sharply as the number of input images drops. The bottleneck is upstream of the reconstruction heads: from a few unposed views, the internal representation they read carries no evidence for unobserved regions, leaving holes, floaters, and blur. The common remedy supplies that evidence as pixels, synthesizing extra views with an image or video generator and re-encoding them, which is costly and not 3D-consistent by construction. We instead densify the evidence itself. We present DensiTok, a plug-in module for pretrained feed-forward 3DGS models that densifies their internal geometry tokens directly, making a frozen backbone behave as though it had observed many more views than it was given. DensiTok compresses those tokens into a compact latent space, completes the latents of the unobserved viewpoints in a single flow-matching step conditioned on camera geometry, and decodes them back into tokens that the original reconstruction heads. The same module design can be integrated into different pretrained predictors while keeping each backbone and its reconstruction heads frozen. Completion in a low-dimensional latent space requires no image synthesis or additional encoder passes. Across three pretrained backbones and two benchmarks, DensiTok consistently improves sparse-view reconstruction and recovers much of the gap to dense-view reconstruction.
Minhyeok Lee, Jungho Lee, Minseok Kang +3
Yonsei University · NAVER AI Lab · Korea Institute of Science and Technology (KIST)
While 3D Gaussian splatting (3DGS) offers explicit and efficient scene representations for cone-beam computed tomography reconstruction, conventional photometric optimization inherently suffers from spectral bias under ultra sparse-view conditions, leading to over-smoothing and a loss of high-frequency anatomical details. Since wavelet transforms provide rich high-frequency information and have been widely utilized to enhance sparse reconstruction, this work integrates wavelet multi-resolution analysis with 3DGS. To circumvent the mathematical mismatch between the strict non-negativity of physical X-ray attenuation and the bipolar nature of high-frequency wavelet coefficients, we propose Residual Gaussian Splatting (RGS). Methodologically, we introduce a spectrally-decoupled Gaussian representation that stratifies the volumetric field into a geometric base component and a residual detail component. This decomposition systematically transforms explicit high-frequency fitting into a physically consistent, implicit residual compensation task. Furthermore, we devise a spectral-spatial collaborative optimization strategy to coordinate the interplay between geometric anchoring and texture refinement, effectively preventing spectral crosstalk. Extensive experiments on clinical datasets demonstrate that RGS enables the reconstructed images to capture highly refined geometric textures. It successfully resolves the trade-off between artifact suppression and detail preservation, yielding superior visual fidelity in complex trabecular and vascular structures compared to existing neural rendering baselines.
Jian Lin, Jiancheng Fang, Shaoyu Wang +4
School of Information Engineering, Nanchang University, Nanchang 330031, China · Laboratory of Image Science and Technology, the School of Computer Science and Engineering, and the Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Ministry of Education, Southeast University, Nanjing 210096, China