acceptodds
Withdrawn from ICLR 2027

SAGA: Structure-Aware Gaussian Allocation for Ultra-Sparse-View CT Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Ultra-sparse-view computed tomography (CT) reconstructs a 3D attenuation field from only a few X-ray projections, where severe angular undersampling makes projection consistency insufficient for reliable volumetric recovery. Explicit Gaussian representations provide adaptive volumetric modeling, but their topology evolution is commonly driven by optimization gradients or heuristic criteria that cannot reliably distinguish anatomical structures from sparse-view artifacts and depth-ambiguous responses. We present SAGA, a stage-stabilized structure-aware Gaussian allocation framework for sparse-view CT reconstruction without paired volumetric supervision. SAGA combines a geometry-frequency constrained objective for stabilizing attenuation geometry and preserving structural details with a G+P structure field that integrates volumetric-gradient saliency and residual-backprojection evidence. The resulting structural evidence is converted into per-Gaussian importance scores to guide cloning, splitting, and pruning. A two-stage optimization schedule keeps Gaussian evolution conservative during early geometry formation and progressively activates structure-guided allocation as the reconstruction becomes more reliable. Experiments on five anatomies under 6, 8, and 10-view settings show that SAGA achieves the best average volumetric PSNR and SSIM across all three sampling regimes. Under the challenging 6-view setting, SAGA improves over the explicit Gaussian baseline by 1.44 dB PSNR and 0.0827 SSIM. Ablation studies further validate the contributions of the reconstruction objective, structure-aware allocation, and stage stabilization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Ultra-Sparse-View CT Reconstruction, 3D Gaussian Representation, Structure-Aware Gaussian Allocation, Self-Supervised Reconstruction, Differentiable Tomography.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_562iOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAGA for ultra-sparse-view CT reconstruction. The proposed method combines projection consistency and volumetric information to guide the Gaussian evolution. Furthermore, a two-stage schedule separates geometric stabilization from structural refinement. Experimental results demonstrate that SAGA achieves impressive performance under ultra-sparse settings. But I think the paper needs better refinement.

Critical strengths

The proposed SAGA is well motivated by the observations that unreliable gradients and residuals under ultra-sparse views can misguide Gaussian allocation. The proposed framework makes conceptual sense by coupling reconstruction objectives with structure-aware allocation and separating the early geometric from later refinement.

Critical weaknesses
  1. The authors introduced the projection cue (i.e., Eq.9) to guide the Gaussian evolution. However, the experiment results in Tab.14 shows a subtle improvement from adding the P cue, which only improve 0.05 in PSNR and 0.001 in SSIM compared with the G cue only. Tab.2 shows that most of the performance improvement comes from the reconstruction objective, while G+P allocation and the two-stage scheduler provide limited performance gains. This weakens the empirical support for the paper’s contributions.
  2. The authors propose a transition factor s (i.e., Eq.12) to control the objective terms for stage-aware optimization. However, iteration count alone may not reliably indicate whether the construction is sufficiently stable for the stage transition. A criterion based on reconstruction stability would provide a more direct justification. Although Tab.10 demonstrates improved final performance with the two-stage scheduler, the results in Fig.11 exhibit drops for PSNR and SSIM, raising questions about its claimed stabilization effect. Furthermore, the values and selection for p_a and p_b are not reported in the paper, and no sensitivity analysis is provided for these key hyperparameters.
Questions for authors

In addition to the weaknesses discussed above, I also have the following question:

  1. Does the adaptive Gaussian (i.e., section Adaptive Gaussian Evolution) introduce any specific operation to the standard rules? If the rules are inherited from existing methods, the authors should cite the relevant work in the corresponding section.
  2. The G+P variants in Tab.13 and Tab.14 appear inconsistent, reporting 31.22 and 30.50 in terms of PSNR under the same chest 6-view setting.
Reviewer_Ug6wOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper essentially proposes a new gaussian splatting densification strategy and stage-aware scheduler for ultra-sparse-view CT reconstruction (≤10\leq 10 views). The authors present a new loss function that is formulated to specifically deal with the severely under-constrained nature of ultra-sparse-view CT reconstruction, as well as a way to mitigate the problem that is often encountered with gaussian splatting, in that Gaussian primitives are often cloned, split or removed based on the gradients, and in sparse-view CT large gradients may originate from true anatomical boundaries, streak artificers or depth-ambiguous strictures. Additionally, densifying gaussians too early on when the reconstruction is poor may densify gaussians in the wrong regions, so a scheduler aware of what stage in the reconstruction process the model is at is introduced to mitigate this.

They validate their strategy on 5 anatomical scenes/CT volumes. Compared to a vanilla CT gaussian splatting reconstruction method, they achieve a gain of 1.44 dB and 0.08 SSIM, with these gains reducing as more views are given as input.

Critical strengths
  1. The motivation behind the paper is clear, and the authors point out that identifying what should govern primitive allocation under a fixed budget constraint is often overlooked, and they seek to address this.
  2. The method and explanation of each of the terms in the loss function is also generally clear and well explained. However in some it could do with some clarification - please see the Questions section. Generally, however, the paper is interesting and proposes a new loss function and gaussian densification strategy that is clearly tailored specifically to ultra-sparse CT reconstruction, and it appears to help in these cases, particularly when there are only 6 uniformly-spaced views.
Critical weaknesses
  1. This is a Gaussian splatting CT reconstruction paper, so whilst it is definitely useful to compare to traditional algorithmic methods and a variety of CT NeRF methods, it would be expected that a comparison to more than one gaussian reconstruction method is made, as comparison against other gaussian splatting methods is effectively what readers are interested in. Such a detailed comparison against older methods is not necessary, and should not come at the expense of comparison to current SOTA.

  2. The paper only tests on 5 CT volumes. A greater number of volumes would be useful and add more weight to the paper.

  3. Figs. 11 - 13 in the Appendix should be produced properly rather than inserting screenshots.

  4. No AI use statement was included - please include one.

Questions for authors
  1. What is the time penalty, if any, of your proposed method compared to say the 3DGR baseline you chose to compare against?
  2. Lines 282-290: Could you clarify how backprojecting rtr_t (which is equal to y^t−y\hat{y}_t - y) into a volume and then taking it's L1 norm is different from SgS_g which also acts on a volume (albeit the grad of the volume)? And why for example do you not just take SpS_p to be the L1 norm of rtr_t?
  3. Lines 334-337: The Gaussian parameters are updated and optimized using the loss (LtL_t) as per standard practice, but then can you please detail how the gaussian cloning, splitting and pruning is different from current SOTA. You mention that they are "...performed according to the stage state and the local importance scores αi\alpha_i", but this mechanism is not then further explained in the paper. In other words, how do you actually use the per-gaussian importance score αi\alpha_i to actually update the density of the gaussians?
Reviewer_dRB1OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper present SAGA, a stage-stabilized structure-aware Gaussian allocation framework for sparse-view CT reconstruction without paired volumetric supervision. SAGA combines a geometry-frequency constrained objective for stabilizing attenuation geometry and preserving structural details with a G+P structure field that integrates volumetric-gradient saliency and residual-backprojection evidence.

Critical strengths

SAGA addresses the important coupling between reconstruction uncertainty and Gaussian allocation with a clear stage-aware design. It consistently outperforms strong baselines across five anatomies and three view settings.

Critical weaknesses

Evaluation uses only synthetic projections from five normalized volumes, so robustness to realistic noise and clinical data remains unclear. Moreover, most gains come from the reconstruction objective, while the proposed G+P allocation provides only marginal improvement in a single-scene ablation.

Questions for authors

See above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.