Assembling the evidence…
Strengths
Jointly predicting intensity and segmentation through a single shared implicit decoder directly targets the boundary-blurring smoothness bias of implicit representations, and yields consistent improvements over GAN- and implicit-representation baselines in both PSNR/SSIM and downstream Dice.
↳ Methods III-C; Table I; Table II
Incrementally adding structural information and contrasting the shared decoder against two-branch and two-stage alternatives provides concrete evidence that both the structural priors and the shared-decoder choice drive the observed gains.
↳ Section V-C; Figure 4; Table III
Applying a pretrained nnUNet to reconstructed volumes provides an indirect, clinically relevant assessment of anatomical fidelity that goes beyond standard image-quality metrics, with substantial Dice improvements across bones, muscles, and organs.
↳ Section V-B; Table II
Limitations
All biplanar X-rays are synthetically generated from CT via parallel-beam projection at 128x128 resolution, so the synthetic-to-real domain gap central to clinical use is untested and the clinical-utility claims are not yet substantiated.
↳ Section IV Experiments
Only three baselines are compared, one a 1984 classical algorithm, while recent implicit-representation methods cited in the related work are excluded from quantitative comparison, weakening the state-of-the-art positioning behind the 'first framework' claim.
↳ Section IV-A; Introduction contributions list
Figure 2 caption metrics do not match Table I values without explanation, and the manuscript contains text corruption, a duplicated sentence, and method-name typos, alongside a missing ethics statement for clinically sourced data.
↳ Figure 2 vs. Table I; Introduction; Section IV LUCY dataset
The strongest signal is the shared-decoder design, which addresses a real failure mode of implicit reconstruction and is supported by ablations that isolate the contribution of both the structural priors and the decoder architecture, plus a downstream segmentation evaluation that adds clinical meaning. Scores were held below the 'competently executed, weaknesses do not undermine conclusions' band primarily because every biplanar X-ray is simulated via parallel-beam projection at 128x128, so the evidence base does not match the clinical framing. The novelty claim of being the first such framework is not fully supported given a thin related-work survey and the omission of cited recent implicit-representation baselines from quantitative comparison. Reporting is constrained by unresolved figure-table numerical discrepancies and multiple editing artifacts that indicate an unpolished draft. Impact potential is genuine but preliminary, gated by the absence of real-data and prospective clinical validation.
Moderate3.2/5.0
The joint prediction of intensity and segmentation through a shared implicit decoder is a clear, mechanism-specific advance over prior biplanar baselines that addresses the smoothness-bias boundary problem, with consistent quantitative gains. The 'first framework' novelty claim is not rigorously substantiated given a thin related-work survey and the idea's proximity to multi-task reconstruction work.
structural and image features are predicted jointly via a shared decoder
The experimental design is competent with two datasets, three baselines, multiple metrics, and informative ablations on joint training and decoder sharing, but all biplanar X-rays are simulated via parallel-beam projection at 128x128 with no real-acquisition validation, and reproducibility details are insufficient for full reconstruction. The baseline set is limited and omits cited recent INR methods from quantitative comparison.
we used the Core Imaging Library (CIL) under a parallel-beam projection geometry
The argument flows logically from ill-posed-problem formulation through method to experiments and ablation, with clear mathematical notation. Multiple editing artifacts (text corruption, a duplicated sentence, method-name typos, and a loss section announcing three components but defining two) impair the reading experience without obscuring the core claims.
the integration oftion oftion of of supervised learning
The paper situates itself adequately against CNN-, GAN-, and INR-based biplanar reconstruction work and motivates the structural-prior approach, but it does not engage the synthetic-to-real domain-gap literature, the multi-task learning literature relevant to its shared decoder, or clinical deployment constraints. The limitations section is brief and omits the simulation-only and resolution gaps.
SPIDER is sensitive to the weights of the joint training loss function
Watch-outs3 of 4 checks
The methodological narrative is consistent across abstract, methods, and results, but the write-up contains presentation-level discrepancies: figure-caption PSNR/SSIM values do not match Table I, the loss section announces three components but defines two, and there are duplicated sentences and method-name typos.
The LUCY dataset is derived from archived clinical records at a named hospital but no ethics/IRB approval statement is provided, and no conflict-of-interest statement appears. Data and code availability are not claimed. These are material watch-outs that do not change interpretation of the reported ML results.
Flags: 0 declared / 5 total
Reference integrity check could not run: no references section detected in extracted text.
No retraction notice found in Retraction Watch.
Sources: Retraction Watch ✓PubPeer (coming soon)
Medium3.0/5.0
The work targets a genuine, clinically deployed need (low-dose 3D reconstruction for orthopedic and craniofacial surgical planning) with identifiable stakeholders, but it does not engage specific clinical decision contexts or the quality threshold required for adoption.
personalized treatment planning and image-guided surgical navigation
The system is proof-of-concept on retrospective and synthetic data, with no prospective clinical validation, no workflow integration, and no regulatory pathway discussion. The downstream segmentation evaluation adds clinical relevance but does not establish deployment readiness.
synthetic X-ray images were generated at a resolution of 128 x 128
Two anatomically distinct datasets and a jaw-open/jaw-close transfer experiment give some evidence of generalizability, but all data is CT-derived synthetic with no multi-site or real-acquisition validation, and the method requires segmentation annotations and is sensitive to loss weights.
an out-of-domain data (jaw-open data) biplanar Xray of the same body part
The work connects to an active sparse-view and INR reconstruction line and contributes a reusable downstream-segmentation evaluation framework, positioning it within a cumulative imaging-to-analysis pipeline. The evidence remains pre-deployment and the path to clinical use requires substantial further work.
downstream segmentation tasks further verify the superior visibility and positional accuracy
Nabu’s assessment, alongside the field’s view.
Evaluation Metadata
AI-powered, human-governed. Something look off? Contact us to request a review.