List of Papers Browse by Subject Areas Author List
Abstract
Biplanar X-ray 3D reconstruction offers a low-cost, low-dose alternative to computed tomography (CT) by inferring volumetric anatomy from two orthogonal projections. However, this task is highly ill-posed due to severe information loss. Existing methods either hallucinate missing structures via generative priors or learn supervised end-to-end mappings from X-rays to CT volumes—both requiring paired data and retraining when imaging geometry or acquisition settings change.
We present ALFA, a two-stage framework that redefines the learning paradigm for biplanar reconstruction. First, ALFA learns a spatially continuous and anatomically consistent latent space derived from CT volumes using a shared implicit neural representation (INR), encoding each subject as a latent vector that decodes into a continuous 3D field. Then, given real X-rays, reconstruction is achieved through geometry-aware latent optimization, aligning synthesized and measured projections under true imaging geometry.
By decoupling anatomical representation from geometry adaptation, ALFA is able to generalize seamlessly across devices and acquisition setups without retraining.
ALFA uses a new decoupling paradigm instead of the previous black-box approach, advancing the dual-plane X-ray imaging model to “plug and play with any device” under ideal conditions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0037_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YuTia_ALFA_MICCAI2026,
author = { Yu, Tianqi AND Du, Chenhe AND Wen, Jie AND Wei, Hongjiang AND Zhang, Yuyao},
title = { { ALFA: Biplanar X-Ray Reconstruction via Anatomy-Latent Field Adaptation } },
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
The paper proposes ALFA, a decoupled two-stage framework for biplanar X-ray 3D reconstruction that separates geometry-agnostic anatomical prior learning (via an INR-based latent space) from test-time physics-driven adaptation (via differentiable forward projection). The key contributions are: (1) zero-shot generalization across unseen Source-to-Image Distances (SID 800/1200mm) without retraining, (2) faithful recovery of patient-specific pathologies (e. g. , mandibular underbite/overbite) where end-to-end methods revert to population averages, and (3) robustness to clinical non-idealities such as scattering, Poisson noise, and geometric misalignment.
- 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.Methodological novelty: Decoupling anatomical priors from imaging geometry is a principled departure from geometry-coupled black-box mappings, enabling plug-and-play generalization. 2.Clinical relevance: Preserving patient-specific deformities rather than producing “mean-shape” hallucinations addresses a critical limitation for orthodontic and surgical planning. 3.Strong zero-shot demonstration: The three-scenario experimental design (identical, cross-geometry, mixed-condition) convincingly shows geometry-agnostic performance. 4.Robustness validation: Stress tests under scattering, noise, and misalignment (Fig. 5) meaningfully extend beyond clean-data evaluations.
- 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.Resolution-agnostic claim not validated: While INR-based methods are theoretically resolution-agnostic, the paper provides no experiments at different resolutions (e. g. , 64^3, 128^3, 256^3) to empirically support this claim. 2.Missing efficiency metrics: No runtime comparison against baselines, no memory footprint analysis, and no parameter count for ALFA vs. competitors. This is essential for a 3D reconstruction method claiming clinical viability. 3.Lack of ablation studies: Critical design choices are not ablated including latent vector dimension, INR depth/width, number of positional encoding bands, and most critically the number of sampling points for ray rendering, which is not even stated, severely hampering reproducibility. 4.Limited baselines and simulated data only: Only two older baselines (X2CT-GAN, PerX2CT) are compared diffusion-based methods and Swin-X2S are absent. All experiments use simulated X-rays from CT, not real acquisitions.
- 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
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?
I was initially tending toward Strong Accept due to the paper’s elegant decoupled paradigm, strong zero-shot generalization results, and clinical relevance. However, after careful consideration of the above-mentioned issues, I have revised my recommendation to Weak Accept. Given these gaps, the paper merits acceptance but requires a thorough revision to address the above issues before it can be considered for a higher rating. With these improvements, the work would indeed become a Strong Accept.
- 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
This paper proposes ALFA, a framework that leverages resolution-independent implicit neural representations (INR) combined with geometry-aware optimization. The method aims to separate anatomical prior learning from test-time physical adaptation, enabling zero-shot generalization across diverse source-to-detector distances (SIDs). While the proposed framework is conceptually interesting, the lack of clarity in experimental design and clinical motivation weakens the overall contribution.
- 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.The proposed framework introduces a clear separation between anatomical prior learning and physical consistency modeling, which enables zero-shot generalization across different acquisition geometries. 2.The use of INR as a frozen anatomical prior is well-motivated, allowing the model to focus on geometry-specific adaptation during inference, which improves robustness to geometric variations. 3.The method demonstrates strong robustness to noise and scattering, which is important for real-world deployment.
- 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.Limited discussion of clinical relevance of geometric variability: While the method emphasizes robustness to varying geometries (e.g., different SIDs), the clinical motivation for such variability is not sufficiently explained. It would be helpful to clarify in which real-world scenarios such variations occur and why zero-shot adaptation is necessary. 2.Potential issue in experimental setup (Section 3, S3 Mixed-Condition Robustness): The authors state that baselines are trained on mixed geometries including parallel-beam with different SIDs (800/1200 mm). However, under parallel-beam geometry, changing SID does not affect the imaging formation, making this setting questionable and potentially unfair or misleading. 3.Loss of high-frequency details compared to baseline: In Fig. 2, ALFA shows improved reconstruction in certain regions (e.g., sagittal view, underbite and deep overbite cases). However, globally the results appear smoother compared to PerX2CT, suggesting a loss of high-frequency details. It would be important for the authors to discuss: oWhy the method appears globally smoother, oWhy certain local regions still benefit despite the overall smoothing effect.
- 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 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents an interesting idea by decoupling anatomical prior learning from geometry-specific adaptation using INR and test-time optimization, which enables zero-shot generalization across varying acquisition geometries. This is a meaningful direction with potential practical impact.
However, several key issues limit the strength of the contribution. In particular, the motivation and realism of the experimental setup are not sufficiently justified (e.g., the use of parallel-beam simulations for bi-plane imaging and the treatment of SID variations), raising concerns about the practical relevance of the evaluation. In addition, the proposed method appears to trade off high-frequency details for smoother reconstructions, and this behavior is not adequately analyzed or discussed.
Overall, while the idea is promising, the current presentation and experimental design leave important questions unanswered, which leads to my score.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 #3
- Please describe the contribution of the paper
This paper proposes ALFA for 3D reconstruction from biplanar X-rays. Its main contribution is a two-stage decoupled framework: it first learns a continuous anatomical prior from CT volumes using a shared INR decoder and subject-specific latent codes, and then reconstructs new cases at test time by freezing the decoder and optimizing only the latent code under differentiable forward projection and known imaging geometry in the experimental setup.
- 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.The problem setting is meaningful, as the paper directly targets generalization across imaging geometries rather than only fixed-geometry 2D-to-3D regression.
2.The method is conceptually clean, separating anatomical prior learning from physics-guided test-time adaptation.
3.Under the authors’ experimental setup, ALFA consistently outperforms X2CT-GAN and PerX2CT in S1, S2, and S3, with particularly strong stability in the unseen-geometry S2 setting, suggesting good geometry robustness in simulation.
- 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.The paper makes strong claims about clinical feasibility, but the experiments are limited to simulated biplanar X-rays.
2.The evaluation is somewhat narrow, and it is not fully clear how much of the gain comes from the learned prior versus test-time geometry adaptation.
3.The pathology-preservation claim is supported mainly by qualitative results and global metrics, without stronger pathology-specific analysis.
- 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 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.
(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 paper presents a well-motivated and conceptually clean framework for an important problem, and the reported improvements under the presented setup are substantial. However, the current evidence is still limited by simulation-only validation, a relatively narrow evaluation scope, and insufficient pathology-specific analysis. These concerns weaken the paper’s claims, but they do not appear to be fundamental flaws in the overall approach.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Author Feedback
We sincerely thank the reviewers for recognizing ALFA’s conceptually clean framework, zero-shot generalization, and robustness. Below we address your concerns:
1.Imaging Physics & Setup (R2) We agree that changing SID in parallel beams does not affect image formation and apologize for the ambiguous phrasing in Sec 3.2.To clarify: we did NOT vary SID for parallel beams. The S3 dataset simply merges three distinct subsets: (1) pure parallel-beam, (2) cone-beam (SID=800mm), and (3) cone-beam (SID=1200mm). Supervised baselines were trained on this mixed dataset solely to test their robustness to dataset-level geometric variance. No physical laws were violated, and we will explicitly clarify this.
2.Clinical Motivation & Simulated Data (R1, R2, R3)
Motivation (R2): In multi-center settings, SIDs vary widely (e.g., 1000-1800mm) due to vendor differences or patient positioning, causing end-to-end models to fail catastrophically. ALFA provides a zero-shot, plug-and-play solution across heterogeneous clinics.
Simulated Data (R1, R3): Acquiring paired biplanar X-rays and 3D CTs in identical poses is practically infeasible due to radiation hazards, making high-fidelity DRRs the standard benchmark. Crucially, baselines require paired data, whereas ALFA only needs unpaired CTs to learn the prior, ensuring high scalability to real clinical X-rays.
3.Efficiency & Multi-resolution (R1) ALFA is highly lightweight (only 15.3M parameters). Inference via Test-Time Optimization (TTO) takes ~15s/case, which is a clinically acceptable overhead for pre-operative planning where accuracy is paramount. While we omitted multi-resolution results due to space limits, this is a proven inherent advantage of INR-based methods, which we will detail in future work.
4.Pathology Metrics (R3) We agree that specialized pathological metrics provide deeper quantitative insights. Currently, our PSNR and SSIM metrics capture global similarity and reflect morphological distortions from pathologies. We will incorporate more specialized, in-depth quantitative pathological analyses in our subsequent work.
5.Smoothness vs. Topological Correctness (R2) ALFA appears smoother due to the “spectral bias” of INRs, which prioritizes fitting low-frequency structural topologies over high-frequency noise. End-to-end models (e.g., X2CT-GAN) hallucinate sharp but anatomically incorrect average teeth. ALFA trades high-frequency texture for topological faithfulness, ensuring that critical macro-level deformations (e.g., underbite/overbite) are strictly and anatomically valid.
6.Hyperparameters & Baselines (R1) Hyperparameters: During training, we randomly sample 51,200 points/subject per epoch. For rendering, we sample 1,024 rays/epoch and 128 points/ray to match real resolution. We will conduct detailed ablation experiments on the components and settings of the model in subsequent work. Baselines: Swin-X2S and Diffusion methods target meshes/segmentations or lack open-source paired training scripts for our volumetric setup. ALFA’s core contribution is the decoupled paradigm; future advanced priors (e.g., diffusion) can seamlessly replace INRs within our framework.
7.Source of Gain: Prior vs. TTO (R3) Both are indispensable. Without the learned prior, TTO is a severely ill-posed 2D-to-3D optimization, causing massive ray-aliasing artifacts. Without physics-guided TTO, performance drops drastically on unseen SIDs (matching S2 baseline failures). ALFA’s gain stems from restricting physical TTO strictly within the valid anatomical manifold defined by the prior.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 introduces ALFA, a decoupled framework for biplanar X-ray 3D reconstruction that separates anatomical prior learning from test-time physics-guided adaptation. The paper received overall positive reviews with scores of 4, 3, and 4.Reviewers appreciated the clean methodological design, the focus on geometry generalization, and the strong simulated results, especially under unseen geometry and robustness settings. The main concerns are about simulation-only validation, limited efficiency/ablation analysis, and the need to better justify some experimental choices. Overall, the contribution is technically sound, relevant to MICCAI, and addresses an important reconstruction problem. I thus recommend early acceptance, and encourage the authors to address the reviewers’ concerns in the final version.