List of Papers Browse by Subject Areas Author List
Abstract
Three-dimensional (3D) reconstruction from biplanar X-rays has high clinical values and remains technically challenging. Existing methods emphasize larger datasets and stronger networks, with less focus on spatial modeling. This usually works well for X-ray inputs with dense structural cues (e.g., chest). However, for inputs with sparse foreground (e.g., long-bone), the impact of geometry unawareness might be significant and introducing such constraints may be helpful. Here, we introduce ExpGS (Explicit Gaussian Splatting), a spatial modeling framework that uses a fixed virtual biplanar geometry to eliminate per-case camera metadata at inference. By preserving non-degenerate anteroposterior/lateral ray constraints within a consistent frustum-intersection volume, it reduces background interference and stabilizes reconstruction. ExpGS reduces to canonical and implicit modeling as geometric constraints relax. Experiments were performed on two CT datasets with distinct content-sparsities, Femur (long bones) and LIDC (chest), whereby synthetic X-rays and matched virtual geometry were obtained from CT projections for training and testing. Results show that spatial modeling confers only minor benefits to information-dense LIDC but substantially improves reconstruction stability (SSIM 86.13%) on the sparse Femur dataset, comparing both with other methods and with the two ablated forms of ExpGS. Our work establishes the importance of geometric constraints for content-sparse planar inputs such as long bones. Implementation is available at: https://github.com/HKUSZH/ExpGS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3258_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HKUSZH/ExpGS
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LonNuo_From_MICCAI2026,
author = { Long, Nuo’er AND Liu, Songkang AND Chen, Cheng AND Tan, Tao AND To, Michael K. T. AND Chen, Peikai},
title = { { From Sparsity to Geometry: Spatial Modeling for 3D Reconstruction from Biplanar Bone X-rays } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposed to incorporate spatial modeling in 3D reconstruction from biplanar bone X-rays which had sparse contents.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
The idea is novel and motivated by an empirical observation of x-rays with sparse contents
-
The spatial modeling approach is intuitive and achieved by sampling multiple points along the 3D ray
-
The idea was validated using fermur bone x-ray dataset with significant performance gain achieved
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
Effectiveness of the spatial modeling techniques was only verified in x-rays with sparse contents, not in more general setting such as lung x-rays.
-
Not much details were provided for implementation
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
More implementations are expected so that the reproducibility can be facilitated. Open-source codes or demons will be helpful.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Very interested and effective idea which is validated with sufficient experimental evidence.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes ExpGS, a spatial modeling framework for 3D reconstruction from biplanar X-rays based on explicit geometric constraints. The key idea is to impose a fixed virtual AP/LAT camera geometry and perform reconstruction via Gaussian splatting, enforcing non-degenerate ray intersections in a predefined frustum. The work compares three regimes: explicit, canonical, and implicit spatial modeling, under a controlled setup and shows that explicit geometry improves reconstruction stability, particularly in sparse anatomical settings such as long bones. Training is performed using multi-view 2D projection consistency loss, without direct 3D supervision.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Clear motivation and well-defined hypothesis The paper isolates a specific factor: spatial/geometric modeling and studies its role in reconstruction quality under varying input sparsity. This is a focused and well-motivated research question. 2.Controlled experimental design (architectural side) The comparison between explicit, canonical, and implicit modeling under the same backbone and training protocol is methodologically sound and allows attribution of performance differences to spatial assumptions rather than confounding factors. 3.Insight on sparsity vs geometry The observation that explicit geometry matters more in low-information regimes (e.g., femur) than in dense settings (e.g., chest) is interesting and potentially valuable for the community.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Unclear and potentially flawed evaluation protocol: The most serious concern is the evaluation setup, which is both unclear and potentially misleading: The model is trained and evaluated entirely on projections simulated from CT volumes. The loss is defined only in projection space (2D), with no direct supervision on 3D geometry Evaluation is performed on “held-out AP/LAT projections generated from the corresponding CT volumes” However, the paper does not clearly specify: how these “held-out” projections differ from the input AP/LAT views. Whether they correspond to different camera parameters, perturbations, or simply re-renderings under identical geometry. This creates a fundamental ambiguity: It is unclear whether the model is evaluated on truly unseen projections or simply on projections generated under the same synthetic setup and geometry as the inputs. This raises a concerns about evaluation circularity. If held-out AP/LAT projections are not meaningfully different from the conditioning inputs, the task risks becoming partially redundant. 2.No direct evaluation of 3D reconstruction quality All metrics (SSIM, PSNR, LPIPS) are computed in projection space (2D), however, the method is framed as a 3D reconstruction approach, but it is not validated in 3D. 3.Synthetic-only setup Both training and testing rely entirely on synthetic projections from CT. No validation on: real radiographs, mismatched geometry or noisy acquisition conditions. 4.Dataset limitations Femur dataset is very small (34 volumes) Results on LIDC show only marginal gains, suggesting limited impact in realistic dense scenarios
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The paper would significantly benefit from explicit clarification of: how held-out projections are generated how they differ from input views Inclusion of true 3D evaluation metrics Validation on real X-ray data Comparison under geometry mismatch scenarios
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the paper addresses an interesting question regarding the role of spatial modeling in biplanar reconstruction and presents a controlled architectural comparison, the evaluation methodology is insufficient to support the main claims. The most critical issue is that both training and evaluation are conducted entirely in projection space using synthetic data, the notion of “held-out” projections is not clearly defined and no 3D reconstruction accuracy is measured. As a result, it is unclear whether the method truly improves 3D reconstruction or simply optimizes 2D projection consistency under a closed simulation setup. Given these concerns, the paper does not meet the standards expected for a MICCAI contribution in its current form.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the rebuttal, I revise my assessment toward acceptance. The rebuttal successfully clarifies the most important concern regarding the evaluation protocol and the interpretation of the “held-out projections.” In particular, the authors explain that the task is not AP/LAT-to-AP/LAT reprojection, but rather AP/LAT-conditioned reconstruction of a 3D Gaussian representation that is subsequently rendered across a full 72-view orbit. The evaluation therefore measures multi-view consistency on unseen test CT volumes, with the reported metrics dominated by novel-view renderings rather than the conditioning projections themselves. This clarification substantially resolves the concern about potential evaluation circularity. I also appreciate the clarification that the work intentionally follows a projection-supervised 3DGS/NeRF-style reconstruction paradigm rather than direct voxel-supervised reconstruction. Within this formulation, the use of rendered-view metrics is consistent with related recent literature. While the lack of direct 3D evaluation remains a limitation, I no longer consider it a fatal methodological issue. The rebuttal also adequately contextualizes the synthetic-only setup. The authors correctly position the work as a controlled methodological study designed to isolate the role of spatial modeling under known geometry and sparse anatomical conditions, rather than as a clinically deployable reconstruction system. The discussion regarding why explicit geometry provides larger benefits in sparse settings than in dense chest anatomy is coherent and aligns with the reported experimental observations. The paper still has limitations. Validation on real radiographs, robustness to geometry mismatch and acquisition noise, and direct 3D geometric evaluation would strengthen the claims further. In addition, the femur dataset remains relatively small. However, these limitations are now clearly acknowledged and do not invalidate the core methodological contribution. I believe the rebuttal resolved the main ambiguity underlying my initial concerns. The paper presents a focused and well-controlled study on the role of explicit spatial modeling for sparse biplanar reconstruction, and I support acceptance.
Review #3
- Please describe the contribution of the paper
The paper proposes a deep learning-based approach to 2D-to-3D reconstruction from X-ray images. The proposed method combines a transformer/MLP-based model with explicit spatial constraints, to improve performance on sparse datasets. The method is evaluated on two X-ray datasets.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
The paper presents a systematic definition of the proposed method.
-
Based on the images and quantitative results, the proposed method indeed seems to improve specifically the sparser datasets.
-
There is a small additional experiment with different dataset ratios.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
The two evaluation datasets are relatively small.
-
The source code is not public and will not be published, or at least this is not mentioned in the submission.
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I am not an expert in this topic, but this seems a nicely presented method for the sparse reconstruction problem, with a method that improves on prior work.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The paper presents a fairly systematic approach to an interesting problem. I agree with the other reviewers that sharing the code would be helpful, and I am happy to see that the rebuttal includes a promise to publish the code on acceptance.
Author Feedback
We thank all reviewers for their helpful, thoughtful feedback. Our responses are as below:
Q1 Verification of spatial modeling effectiveness (R1) A1: We also validated on LIDC chest/lung data, representing dense anatomy. While LIDC lacks the same controlled ablation as femur, the compared methods cover explicit (ExpGS), canonical (DIF-Gaussian, PerX2CT), and implicit (DiT, X2CT-GAN) paradigms, providing complementary evidence in dense anatomy.
Q2 Implementation details and reproducibility (R1, R3) A2: Implementation details: ResNet-50 encoder; 8-layer Transformer (8 heads, d=512); N=64 depth samples; 256^2 resolution; Adam (lr=1e-4, batch=4). Supervision: 72 views/CT, 4 sampled views/iteration. Code will be released upon acceptance.
Q3 Evaluation protocol, held-out projections, and potential circularity (R2) A3: We apologize for the imprecise wording. We confirm that Table 1 reports metrics averaged over all 72 rendered views per test CT, not just AP/LAT. The actual protocol: 72 DRRs per CT (5° intervals); AP/LAT are inputs; 4 random views sampled as supervision during training; at evaluation, the predicted Gaussian field is rendered to all 72 views and compared to ground truth. Thus, the task is AP/LAT –> 3D representation –> multiview rendering, not 2D-to-2D copying. To clarify, the 72-view evaluation spans the full angular orbit; two re-rendered views at the AP/LAT angles are kept to ensure full orbit coverage, and the reported metrics are dominated by novel-view renderings (70/72).
Here, “held-out” refers to unseen test CT volumes (e.g. 8 testing CTs in Femur, and 100 LIDC test cases), where AP/LAT views are given as inputs and the predicted 3D field is rendered around the orbit. We will correct the misleading phrase “held-out AP/LAT projections” in the manuscript.
Q4 Projection-space loss and no direct 3D supervision (R2) A4: Although CT volumes are available for DRR generation in this controlled study, our training objective intentionally follows projection-supervised 3DGS/NeRF-style reconstruction rather than voxel regression. We agree direct 3D evaluation would further strengthen the claim.
Q5 3D quality evaluation (R2) A5: We acknowledge the lack of direct 3D metrics is a limitation. Our focus is a geometry-consistent multi-view Gaussian representation; rendered-view metrics (PSNR/SSIM/LPIPS) are commonly used for 3DGS-based medical reconstruction (3DGR-CAR by Fu et al. 2024; Gaussian Pancakes by Bonilla et al. 2024). Direct voxel-level evaluation requires converting 3DGS into voxels with a standardized protocol, which may introduce extra variables. We will clarify this limitation and evaluation scope.
Q6 Synthetic-only setup (R2) A6: We acknowledge that synthetic projections cannot prove generalization to real clinical X-rays. We use CT-derived projections as a controlled setting to compare spatial modeling under known geometry, consistent views, and clear supervision. This setting is admittedly idealized, but avoids ethical (dosage) and physical challenges, and is commonly used in recent related methodological works, such as VolumeNeRF (MICCAI 2024) and DCT-Net (MICCAI 2025). Thus, our experiments are controlled projection-supervised validation, not clinical-deployment validation; real X-ray generalization, geometry mismatch, noise, and domain adaptation remain future work.
Q7 Small femur dataset and marginal gains on LIDC (R2, R3) A7: We agree that the Femur set is relatively small (n=34), reflecting the scarcity of long-bone CT data. Our LIDC scale ablation (Fig. 4) shows that even at 5% training data (45 cases, comparable to Femur scale), all methods remain stable, confirming that the conclusion is driven by spatial modeling rather than data scale. Marginal LIDC gains are expected: explicit geometry benefits sparse inputs (femur) more than dense anatomy (chest), consistent with the main findings of our work.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper presents a novel and well-motivated approach to incorporating explicit geometric constraints for biplanar 3D reconstruction, with promising results, particularly in sparse settings. While reviewers appreciate the methodological clarity and controlled experimental design, there are notable concerns regarding the evaluation protocol (e.g., reliance on synthetic data and projection-space metrics), limited validation of true 3D reconstruction quality, and reproducibility details. Given the mixed assessments, I invite the authors to provide a rebuttal addressing these issues, particularly clarifying the evaluation setup and strengthening evidence for 3D reconstruction performance.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have addressed the reviewers’ comments. Discussion on the evaluation should be expanded, especially regarding why 3D evaluation was not conducted.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The decision is to accept the paper. The authors propose an effective method incorporating explicit geometric constraints for biplanar 3D reconstruction which performs particularly well in sparse anatomical settings. During the rebuttal phase they successfully clarified critical misunderstandings regarding the evaluation protocol by explaining that the performance metrics reflect novel view renderings across a full orbit.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I recommend acceptance. The paper presents a focused and well-motivated study of explicit spatial modeling for 3D reconstruction from biplanar X-rays, with particular benefit in sparse anatomical settings. The core idea is intuitive, technically clear, and supported by controlled comparisons between explicit, canonical, and implicit spatial modeling strategies.
The rebuttal substantially addressed the main concern about the evaluation protocol. In particular, the authors clarified that AP/LAT views are used as conditioning inputs, while the predicted 3D Gaussian representation is evaluated through rendering over a 72-view orbit on unseen test CT volumes. This resolves the earlier concern that the method may only be performing redundant AP/LAT reprojection. The authors also clarified implementation details and committed to code release.
