acceptodds
Withdrawn from ICLR 2027

Sparse2View: Rotatory Novel View Synthesis From Scarce X-Ray Inputs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The X-ray computed tomography (CT) requires a large number of X-ray projections for construction, exposing patients to radiation and a prolonged imaging time, To address these concerns, this work proposes Sparse2View, a novel rotational X-ray view synthesis framework, able to generate full-view projections from only a few sparse view inputs. Sparse2View features two key innovations: (1) intra-view feature refinement, which employs a tailored anisotropic spectral operator to capture the frequency information within the projection sequences; and (2) inter-view geometric fusion, which enhances geometric representation with cross-view information integration. Additionally, we curate a large-scale, deep learning–ready dataset comprising 1,614 object-level (mouse skeleton) scans, yielding a total of 419,640 medical expert–annotated 2D X-ray images. Extensive experiments demonstrate that Sparse2View can synthesize rotatory novel views on-demand under extremely sparse settings. For instance, taking only 4 views as inputs, it can generate the full 260-view of X-ray views over a 208-degree rotating angle (dramatically reducing the radiation dose by 65 folds), with an average PSNR of 36.72 and an SSIM of 0.931. Moreover, results under the zero-shot settings validate that the Sparse2View can be well generalized to other anatomical regions, unseen fractured bones, and human organs, underscoring its potential medical applications.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
X-ray Projection Generation, Rotatory Novel View Synthesis, Medical Imaging

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mpkvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Sparse2View, a novel-view synthesis method for sparse X-ray projections using a Transformer architecture. It integrates per-view anisotropic spectral filtering with cross-view attention, leveraging Plücker ray embeddings and PRoPE transformations. The authors present MiceBoneX and evaluate projection synthesis, downstream reconstruction, and transferability to unseen skeletal anatomies.

Critical strengths
  • Useful data contribution. MiceBoneX contains 1,614 real micro-CT scans with rotational projections and camera metadata. Collecting these measurements provides a useful resource for studying sparse-view X-ray synthesis
  • Consistent quantitative improvements are observed. Table 1 shows PSNR gains of 2.15 dB and 2.26 dB over LVSM in the 4-view and 8-view settings, respectively, along with increases in SSIM and LPIPS.
Critical weaknesses
  • The method lacks significant novelty, mainly combining existing components. The paper does not clarify HRoPE's added value beyond merging PRoPE with Plücker embeddings, and a controlled comparison with LVSM using standard PRoPE is needed to support its claims.
  • Ablation results show PSNR drops when removing certain components, but lack comparisons to simpler or alternative filters, so the necessity of the proposed design is not established.
  • Improved projection metrics do not guarantee better anatomical reconstruction. The reconstruction and bone morphometry experiments compare the proposed method to ground truth but exclude results from learned baselines. Similarly, the fracture experiments lack direct evaluation of fracture gap or fragment boundary accuracy.
  • The framework description is incomplete. The empty ray and query boxes in Figure 1 appear to be schematic placeholders, but the figure omits the pose-conditioning mechanism introduced in Equation (3). More importantly, the formulation does not clearly explain how target-dependent filters are applied when multiple target views are generated jointly.
Questions for authors
  • How does HRoPE-MSA differ from standard PRoPE with Plücker embeddings? Please provide a controlled comparison with LVSM using PRoPE, ensuring matched model capacity, training schedules, and loss functions.
  • Is the proposed anisotropic filtering more effective than simpler spatial filters or non-factorized spectral filters? Please separate the effects of filter factorization and pose conditioning, rather than eliminating the entire module.
  • Do improvements in projection metrics lead to better anatomical reconstruction compared to LVSM or LVSM + PRoPE? Please provide comparative bone morphometry errors and quantitative fracture gap measurements on the same test cases, rather than relying solely on agreement with ground truth.
Reviewer_KVoYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors first introduce MiceBoneX, the first large-scale X-ray projection dataset of mice designed for deep learning. They then propose Sparse2View, a framework for synthesizing rotational X-ray views that leverages frequency and multi-view information to generate full-range projections from a limited number of sparse-view inputs. This framework achieves high-fidelity synthesis of novel rotational X-ray views on the proposed dataset and enables 3D CT reconstruction using a lower radiation dose.

Critical strengths

The Sparse2View method proposed by the authors achieves sota performance compared to recent work and enables efficient 3D modeling.

Critical weaknesses
  1. The ablation studies are insufficient to demonstrate the effectiveness of the novel modules proposed by the authors. For instance, regarding the geometric and anisotropic inductive biases introduced, there is a lack of comparative analysis against other positional encoding or spatial processing mechanisms to highlight their advantages.
  2. The study’s conclusions regarding the fidelity of clinically relevant structures remain subject to limited validation. Morphometric analysis was based on only 20 selected cases and utilized just a single cross-section of the tibia, while the fracture experiments relied primarily on global image quality metrics and qualitative comparisons.
Questions for authors
  1. When synthesizing multiple target views simultaneously, does the system recompute input features for each target view, or does it employ a shared representation conditioned on pose?
  2. The authors need to clarify whether the scan data used to evaluate the results after fine-tuning were entirely independent of the dataset used during the fine-tuning phase.
  3. Please specify whether the 65-fold reduction in radiation dose mentioned in the paper is supported by actual dose measurement data, or if it is based on simulation calculations derived from a specific acquisition protocol.
Reviewer_6VDyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies rotational X-ray novel view synthesis from extremely sparse input projections. It proposes Sparse2View, a Transformer-based framework that combines intra-view feature refinement and inter-view geometric fusion. Specifically, the method introduces an anisotropic spectral operator with pose-conditioned Fourier filtering to capture spatial dependencies, and an HRoPE-MSA module that integrates Plücker ray embeddings and camera geometry for cross-view feature fusion. The authors also introduce MiceBoneX, a real-world X-ray dataset containing 1,614 microCT scans and 419,640 projections. Experiments show that Sparse2View outperforms existing baselines under 4-view and 8-view settings.

Critical strengths
  1. Practical relevance and valuable data contributions. The paper addresses the problem of X-ray view synthesis from extremely sparse observations and introduces a substantial real-world dataset containing 1,614 microCT scans. This dataset could be a useful resource for further research.
Critical weaknesses
  1. Limited methodological novelty. The framework mainly combines an LVSM architecture, PRoPE-based geometric attention, and Fourier-domain filtering. The paper does not clearly establish how HRoPE differs substantively from existing hybrid camera encodings. The pose-conditioned separable filters also appear to be an incremental modification of established spectral processing. Overall, the methodological contribution beyond integrating and adapting existing techniques remains insufficiently established.
  2. Insufficient analysis of the proposed designs. Table 2 only removes entire components under the 8-view setting. It neither compares HRoPE with existing PRoPE-based encodings nor isolates the contributions of separable filtering and pose conditioning. Moreover, no corresponding zero-shot ablations support the claimed generalization benefit of anisotropic processing. These gaps make it difficult to determine whether the reported gains arise from the proposed innovations or from established components.
Questions for authors
  1. Novelty of HRoPE-MSA. How does the proposed HRoPE-MSA differ from the existing PRoPE-based hybrid geometric encoding? Could the authors provide a direct comparison with LVSM equipped with PRoPE and Plücker ray embeddings under the same experimental settings?
  2. Effectiveness of Anisotropic Processing. Could the authors provide more detailed ablations to isolate the contributions of separable spectral filtering and pose-conditioned modulation? In particular, how does the proposed design compare with standard Fourier filtering, and does it consistently improve zero-shot generalization?
  3. Clinical significance. The paper reports promising projection synthesis results, but how well do the synthesized views preserve fine anatomical details in reconstructed 3D volumes? Can the author provide a quantitative evaluation of downstream tasks such as organ segmentation?

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

Related papers

Discussion (0)

Sign in to comment.