List of Papers Browse by Subject Areas Author List
Abstract
Reconstructing patient-specific 3D pelvic bone structures from biplanar X-ray images is attractive for routine clinical workflows, as it enables 3D understanding from widely available low-dose radiographs. Existing methods often lack effective cross-view interaction to fully exploit the complementarity between anterior–posterior (AP) and lateral (LAT) views, and provide limited global structural dependency modeling, which leads to structural ambiguity and missing details under limited projections. To address these issues, we propose X2Bone, an efficient end-to-end framework for 3D bone reconstruction from standard AP and LAT radiographs. X2Bone introduces a Cross-RWKV module that enables long-range, bidirectional cross-view interaction with linear complexity, achieving global semantic alignment and feature fusion across views. To bridge the 2D-to-3D representation gap, we further design a 3D Expand module, which lifts the fused 2D features into volumetric space via view-dependent depth probability modulation. Extensive experiments on the CTPelvic1K dataset demonstrate that X2Bone achieves the best segmentation performance while maintaining high efficiency. The results validate the effectiveness of long-range cross-view modeling and view-aware 2D-to-3D lifting for biplanar 3D reconstruction.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0798_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/pzhhhhh2263/X2Bone
Link to the Dataset(s)
CTPelvic1K dataset: https://github.com/MIRACLE-Center/CTPelvic1K
BibTex
@InProceedings{PanZha_X2Bone_MICCAI2026,
author = { Pan, Zhaohong AND Zhu, Jiahua AND Zhou, Haowei AND Yang, Guang AND Dai, Jingjing AND Xie, Yaoqin AND Liang, Xiaokun},
title = { { X2Bone: Reconstructing 3D Bone Structures from 2D Biplanar X-Rays via Cross-RWKV } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes X2Bone, an end-to-end deep learning framework designed to reconstruct patient-specific 3D pelvic bone structures from standard 2D biplanar (AP and LAT) X-ray images. The core technical contributions include: 1.Cross-RWKV Module: A novel interaction mechanism that utilizes a bidirectional Receptance Weighted Key-Value (Bi-WKV) operator to enable long-range, cross-view feature fusion with linear complexity, effectively capturing the complementarity between AP and LAT views. 2.3D Expand Module: A depth-attention-driven lifting strategy that bridges the 2D-to-3D representation gap by modulating volumetric features with learned depth probability weights, reducing information loss during dimensional expansion. 3.Efficiency and Performance: The authors demonstrate state-of-the-art (SOTA) performance on the CTPelvic1K dataset, achieving a Dice Similarity Coefficient (DSC) of 87.96% while maintaining a significantly lower parameter count (11.02M) compared to existing models.
- 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.Technical Novelty in Efficiency The integration of the RWKV (Receptance Weighted Key-Value) architecture into biplanar reconstruction is a timely and original contribution. By achieving global dependency modeling with linear complexity, it overcomes the quadratic scaling limitations of standard Transformers, making high-resolution 3D reconstruction more computationally feasible. 2.Superior Parameter Efficiency The model is remarkably lightweight. It outperforms the recent SdAoF method while using only 2.6% of its parameters (11.02M vs. 416.43M), which is highly relevant for resource-constrained clinical settings. 3.Effective Cross-View Interaction The proposed Cross-Shift operation and gated bidirectional aggregation explicitly address the inherent depth ambiguity in X-rays by exchanging and aligning local spatial contexts between orthogonal views. 4.Solid Benchmarking The paper provides a comprehensive evaluation against several SOTA architectures (UNet, SwinUNETR, BX2S-Net, etc. ) using standard metrics like DSC and HD95, consistently showing improvements in both global morphology and local structural details.
- 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.Reliance on Simulated Data A primary limitation is that experiments were conducted solely on Digitally Reconstructed Radiographs (DRRs) simulated from the CTPelvic1K CT dataset. While common in research, DRRs do not fully capture the noise, scattering, and artifacts present in real-world clinical X-rays, leaving the model’s true clinical utility unproven. 2.Limited Spatial Resolution The reconstruction resolution is limited to 128×128×128 voxels. While this is a standard benchmark, it may be insufficient for highly detailed preoperative planning or identifying fine fractures where higher resolution is critical. 3.Scope of Anatomical Validation The study focuses exclusively on the pelvis. Although the authors mention extending this to other skeletal structures as future work, the current framework’s generalizability to more complex or smaller bone structures (like the spine or carpal bones) remains unexplored. 4.Evaluation of the Cross-Shift Mechanism While the Cross-Shift operation is intended to align orthogonal views, the geometric justification for using a 3 × 3 re-parameterizable convolution to bridge views with 90° orientation differences could be more rigorously detailed or compared against more geometrically-aware projection methods.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The X2Bone framework proposed in this paper introduces the RWKV mechanism to the cross-view interactive dual-modeling of X-ray images; this approach is novel, and the overall methodological design is comprehensive. Cross-RWKV achieves global cross-view feature extraction while maintaining linear complexity, proving more efficient than Transformer-based alternatives. Notably, the model achieves state-of-the-art reconstruction accuracy with a mere 11.02M parameters—a highly desirable attribute for clinical deployment scenarios. The ablation studies and multi-view experiments also effectively substantiate the rationale behind the proposed methodology. However, the current experiments were conducted exclusively on the CTPelvic1K dataset using Digitally Reconstructed Radiographs (DRRs) to simulate X-rays; the lack of validation using—or even a discussion regarding the applicability to—actual X-ray data leaves the model’s generalization capability uncertain. Furthermore, the performance improvement yielded by the proposed optimization method is less than one percentage point; the absence of a statistical significance analysis suggests that the practical impact of this improvement may be marginal. The work would be significantly strengthened by the inclusion of comparisons against other linear-complexity models (such as Mamba), efficiency metrics (such as inference speed), and a comprehensive discussion regarding the method’s applicability to real-world data. Overall, while the work demonstrates promise, it exhibits a certain degree of bias and remains in a somewhat preliminary stage.
- 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.
The authors have provided a comprehensive, well-structured, and compliant rebuttal that addresses all my major concerns. All technical ambiguities have been resolved, and the authors have demonstrated a clear understanding of the limitations of their work. The tone is professional and constructive throughout.
Review #2
- Please describe the contribution of the paper
This paper presents a RWKV-based model for bi-planar X-ray to 3D bone reconstruction. The model employs a Cross-RWKV module to address inter-view feature fusion and intra-view long-range feature dependency modeling. Additionally, a 2D-lifting module is proposed to model output 3D features in a canonical 3D space for better 3D reconstruction. Quantitative improvements validate the effectiveness of the proposed architecture.
- 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.Applying RWKV modules to 3D bone reconstruction is an interesting topic. 2.Presentation logic is clear. 3.Quantitative performance is good.
- 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.Elaboration on method details needs to be polished. – 1.1.The details of Cross Spatial Mix are difficult to understand for readers not familiar with RWKV. – 1.2.The motivation of Eq. (6) and Eq. (7) is unclear. It is not well-justified why the depth probability is a D-dimensional vector. In my view, different 2D locations should have varying depth probabilities. Furthermore, the explanation of view-dependent alignment is too brief and confusing. – 1.3.Which 2D feature map the lifting module uses is somewhat unclear. From Sec. 2.2, I infer that the intermediate 2D features have been fused together in Eq. (2). However, Eq. (6) and Eq. (7) operate on view-specific features, which is inconsistent. 2.Missing experiments to validate the linear complexity benefit of the proposed backbone: A major benefit of RWKV is that it scales linearly with the token count. However, the authors do not present any experiments regarding this property. I suggest the authors provide experiments with increased input and output resolutions to demonstrate its advantages over prior art.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The main contribution is solid and clear, but some presentation details require further polish. Overall, I give a weak accept for the current manuscript.
- Reviewer confidence
Very confident (4)
- [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 authors address my major concerns.
Review #3
- Please describe the contribution of the paper
The paper adapts receptance-weighted KV, a recurrent version of KQV-style transformer, for Biplanar X-ray to 3d reconstruction. This architecture is meant to provide attention over long-range without the quadratic complexity of the KQV-style attention mechanism. The proposed architecture attends efficiently across the biplanar views, cross-attending same regions locally between the two views, and using recurrence for global attention. The dimension expansion from two 2D views to 3D using a learnt weighing is propsed. This is shown to perform competitively against state-of-the-art SwinUNetr using lot fewer trainable parameters and compares favorably to other popular off-the-shelf and domain-specific architectures on the CTPelvic1k dataset.
- 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 adapts the lightweight RWKV attention popular from NLP - and adapted to Medical image segmentation by U-RWKV - by modifying 3D-to-3D end-to-end training pipeline to two 2D input to 3D output pipeline. This adaptation shows improved, if not comparative, performance at 6x lesser #parameters compared to the attention-based SOTA SwinUNetr, almost as efficient parameterwise as the popular Unet architecture.
- 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.
As is, this paper is in the right direction and could be a starting point for exploring efficient global attention mechanism for reconstructing large volumetric structure that benchmark #trainable parameters vs performance (and/or training/inference time), but currently the the thoroughness of the experimental evidence is lacking (see details in the comments section).
Their performance not only on image-based metrics (such as HD, DSC) but also on clinically relevant metrics (which may be evaluated automatically), confidence interval reporting, validating on multiple anatomies, to support the claims better is required.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
Limited experiment and reporting: Validation on only single anatomy seem limiting. Reconstructing other large anatomy such as Rib (RigSeg) or even whole spine (CTSpine1k) using readily available dataset could be added to strengthen the evidence. No mean+/-sd confidence interval means we do not know if the performance is statistically significant.
Limited Reporting metrics: Clinically-meaningful image-based metrics such as NSD (that does not penalize reconstruction error upto clinically acceptable limit, say 2mm) should also be reported. In addition, fidelity of Clinical parameters extracted from these reconstruction, such as Pelvic Tilt, slope etc., may also be explored.
Ablation of the 3D Expansion Module: Could you please provide a ablation study comparing the proposed learnt weighted expansion to other baselines such as unweighted expansion, repetition etc.
Writeup: No concrete evidence is provided to support this argument – “In contrast, the proposed method effectively alleviates topological discontinuities, producing reconstructions that exhibit higher consistency with the ground-truth annotations in both global morphology and local structural details.” Quantitative metrics for topology such as betti number may be used to support this argument. Certain words seem too verbose: depth probability modulation, depth-attention More measured wording: “We design a cross-RWKV…, effectively alleviating structural ambiguity under limited projection conditions.” effectively alleviating -> reduce.
- 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?
The adaptation of the RWKV mechanism itself may not provide enough value to the community especially with lacking experimental thoroughness – multiple validation, complete reporting of the metrics and meaningful performance or efficiency evidence. On the other hand, the paper has good potential in the direction of exploration of cross-view feature learning specific to biplanar x-ray outside of CNN- and transformer-based off-the-shelf architectures.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
With regards to other reviews and rebuttal from the authors, i will stick to reject. The suggestion of careful wording of some of the aspects of the conclusions has been incorporated. On the other hand, major aspects such as confidence interval reporting and clinically relevant metric such as NSD (which does not require new experiments), and which would have helped provide evidence of significance of the results, is not incorporated. Overall, i think, this work could benefit from additional work as suggested in the reviews, beyond current somewhat preliminary stage.
Author Feedback
We thank the reviewers for their constructive comments and for recognizing the novelty/interesting idea (R1,R2), clear presentation (R1,R2,R3), and lightweight design (R1,R3). We address the main concerns below. Simulated data(R1, AC): We agree that real clinical X-ray validation is important for deployment. This work aims to validate the feasibility of the proposed framework and therefore uses CT-derived DRRs, a common and reproducible setting in X-ray-to-3D benchmarks. Paired real AP/LAT X-rays with patient-matched 3D CT annotations are difficult to obtain, and additional acquisition may introduce radiation exposure and ethical constraints. For clinical translation, domain adaptation or unsupervised style transfer could help reduce the DRR-to-X-ray domain gap. Additional experiments/statistics(R1, R3, AC): The rebuttal rules do not allow new experimental results, so we cannot add results on additional datasets or metrics. The submitted manuscript already supports the main claims: Table 1 shows that X2Bone achieves the best average DSC and HD95 with only 11.02M parameters, and Table 2 shows that Cross-RWKV and 3D Expand provide independent gains. X2Bone is not anatomically restricted to the pelvis; we focus on the pelvis because it is the core target of existing X-ray-to-3D benchmarks and SdAoF, and involves depth ambiguity, hip-bone occlusion, and complex sacral anatomy. Eqs. (6)–(7) (R2): The depth weights in Eq. (6) are not pixel-wise depth predictions, but global weights along the missing spatial axis after 2D features are lifted into 3D. We use a D-dimensional vector rather than location-wise depth distributions to keep the lifting stable and parameter-efficient under limited-view supervision, while spatial details are carried by the expanded 3D features and refined by the 3D decoder. Eq. (7) performs view-dependent alignment because AP and LAT views have different missing axes; the learned weights are therefore permuted to the corresponding 3D axis before modulation. 3D Expand ablation/clarification (R2,R3): In Table 2, “w/o 3D Expand” replaces our learned weighted expansion with simple feature replication. UNet-based baselines similarly rely on repeated copying of 2D input images for 2D-to-3D expansion, but the results suggest that such replication is less effective in bridging the 2D-to-3D dimensional gap. We thank R2 for noting this ambiguity. Cross-RWKV does not collapse AP/LAT into a single 2D feature map before 3D Expand. It maintains two view-specific output streams: the AP stream uses AP features as key/value and LAT features as the gate, while the LAT stream uses LAT features as key/value and AP features as the gate. 3D Expand then lifts the two streams along their respective missing axes. After Eq. (7) aligns them into the same 3D coordinate system, the two volumes are added before the decoder. Cross-Shift(R1): Cross-Shift is not intended for explicit geometric alignment of orthogonal views. It introduces lightweight local cross-view exchange before Bi-WKV scanning by combining the current-view token, neighboring tokens via a 3×3 convolution, and the corresponding token from the other view via a 1×1 convolution. This allows same-position, local-neighborhood, and cross-view contexts to be modeled jointly. Limited spatial resolution (R1): We acknowledge this limitation. We used $128^3$ to follow the benchmark and ensure fair comparison. Higher resolutions would make large baselines such as SwinUNETR difficult to run fairly on one RTX 3090.Unsupported description(R3): The statement refers to the qualitative comparison in Fig. 3 rather than a separate topological metric. The yellow-box regions show visually more complete sacral foramina structures. We will use more measured wording. Writing clarity (R2, R3): We will clarify Cross Spatial Mix and use more measured wording, e.g., replacing “effectively alleviating” with “reducing.” We hope these clarifications address the concerns.
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 proposes X2Bone, a Cross-RWKV-based framework for reconstructing 3D pelvic bone structures from biplanar X-ray views, and the method is promising because it achieves competitive results with a small parameter count. However, the evaluation is limited to simulated DRRs from one pelvic CT dataset, so its generalization to real X-rays and other anatomies remains uncertain. I therefore recommend Invite for Rebuttal, with the rebuttal focusing on clinical generalization. In addition, the rebuttal should address Reviewer #2’s concern by providing clearer ablations, including whether each component brings independent and necessary gains.
- 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.
After considering the reviews and rebuttal, I recommend acceptance. The paper proposes a promising and efficient framework for biplanar X-ray to 3D bone reconstruction, with a novel use of Cross-RWKV and favorable parameter efficiency. Two reviewers were satisfied by the rebuttal and moved/maintained an accept recommendation. Although concerns remain regarding validation on real X-ray data, statistical uncertainty, and clinically oriented metrics, the core methodological contribution and empirical results appear sufficient for acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
I agree with most of R2’s concerns, although R1 and R2 recommend acceptance. In particular, the missing metrics are important which do not require additional experiments, but the authors leave them mainly as future work. This weakens the reliability of the results.
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.
The rebuttal allowed the authors to address major concerns, including resolving technical ambiguities and demonstrating awareness of the method’s limitations. I acknowledge the remaining concern that the paper would be strengthened by confidence interval reporting and additional clinically relevant metrics such as NSD to better establish significance and clinical relevance. However, I find the core contribution sufficiently strong: the paper proposes an efficient Cross-RWKV framework for biplanar X-ray to 3D bone reconstruction, shows competitive reconstruction performance with a lightweight model, and addresses an important problem in medical image reconstruction. I encourage the authors to incorporate the remaining reviewer suggestions in the final version.