{"id":"4655d32e-c6e4-4854-a15e-ef64a7f22695","arxiv_id":"2506.13545","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-domain diffusion model with analytic cone-beam projectors reconstructs CBCT volumes from 90-degree limited-angle scans, achieving MAE 35.5 HU and SSIM 0.84 on simulated pelvic data.","lead":"This paper shows a two-step AI pipeline that fills in missing X-ray views, then cleans up the resulting image, so a CT scan can be built from only a 90-degree sweep. The authors report strong scores on simulated X-ray data, but no test on real scans, so clinical value is not yet demonstrated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 2.2.2's split is not patient-independent: keeping one volume per patient for testing leaves 60 training volumes from the same 18 patients, so Table 1 cannot support the 'without retraining across patients' claim.","rationale":"The reader's weakest assumption was simulator fidelity: because training, inference, and ground truth all come from LEAP-CT, real-world transfer is unproven. That is a valid external-validity concern. The concern identified here is more fundamental because it affects internal validity: the reported test set appears to share patients with the training set, so even within the simulated setting the headline metrics do not demonstrate cross-patient generalization. The paper's own split description makes this leakage unavoidable given 78 volumes from 18 patients and one held-out volume per patient. This is not an accusation of misconduct; it may be a reporting error, but as written the central clinical claim is not supported by a valid patient-independent evaluation. A corrected patient-disjoint split and recomputation of Table 1 are necessary before the claim can be assessed. Therefore the appropriate verdict is unverified pending re-evaluation.","tokens_in":13900,"tokens_out":6493,"duration_ms":67695,"concrete_test":"Inspect the actual train/test patient assignment from DICOM metadata or the data manifest; if any patient ID appears in both sets, the split is not patient-independent. Then retrain both DDPMs from scratch with a strict patient-disjoint split (e.g., 14 patients for training and 4 entire patients for testing, preserving approximately 60/18 volume balance) and recompute Table 1. If MAE, SSIM, or PSNR on the held-out patients is materially worse than 35.5 HU, 0.84, and 29.8 dB, the reported numbers are inflated by within-patient leakage and the cross-patient claim is unsupported.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The paper's evaluation split is internally inconsistent. Section 2.2.2 states that 'a volume-level split was employed rather than a patient-level separation' and that 'one representative volume per patient (18 volumes) was held out for testing' while 60 volumes were used for training. Since the dataset contains 78 volumes from 18 patients, the 60 training volumes necessarily include additional scans from the same 18 patients whose one held-out volume forms the test set. The adjacent statement that 'no patient overlap exists between the training and testing subsets' is therefore not compatible with the described split. This is not a peripheral detail: the conclusion that LA-GICD works 'without retraining across patients' depends on patient-disjoint evaluation. Within-patient similarity across treatment fractions—same body habitus, similar applicator geometry, and correlated soft-tissue layout—can be memorized by the model and inflate the reported MAE, SSIM, and PSNR on held-out fractions. The simulator-fidelity concern raised by the reader is real, but before that can be assessed, the current quantitative comparison does not establish generalization to unseen patients.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LA-GICD, a dual-domain diffusion framework for limited-angle cone-beam CT (CBCT) reconstruction, in which a Projection-DDPM completes missing 90-degree projections, an analytic cone-beam forward/back-projector (GTM) maps them to image space, and an Image-DDPM refines the volume. The framework is trained on simulated projections from 78 planning CT volumes of 18 gynecologic HDR brachytherapy patients and evaluated against full-angle reconstructions using MAE, SSIM, and PSNR, plus visual review by four medical physicists. The authors report MAE 35.5 HU, SSIM 0.84, PSNR 29.8 dB and conclude that the method provides artifact-free, clinically applicable reconstructions from a single 90-degree scan, reducing acquisition time and dose four-fold without hardware changes.","tokens_in":14076,"tokens_out":2780,"duration_ms":27458,"significance":"If the results are validated, the proposed dual-domain, geometry-integrated diffusion approach is a potentially useful contribution to limited-angle CBCT reconstruction, addressing a real clinical need in brachytherapy and other time- or geometry-constrained workflows. The paper's strengths include a fully differentiable pipeline, an explicit geometry module, a clear description of the diffusion formulation, and the use of a clinically motivated dataset. However, the current evidence is insufficient to support the paper's central claims of patient-independent generalization, artifact-free reconstruction, and four-fold dose reduction, because the evaluation is compromised by the data-split design and the absence of state-of-the-art baseline comparisons and ablations. The clinical significance claimed in the abstract and conclusion substantially exceeds what the experiments demonstrate.","major_comments":[{"comment":"The data split is not patient-independent, and the statement 'No patient overlap exists between the training and testing subsets' is incompatible with the described procedure. The paper states that 60 volumes were randomly chosen for training and one representative volume per patient (18 volumes) was held out for testing, with 78 volumes total from 18 patients. Because each patient contributes 4 or 5 volumes, the 60 training volumes necessarily include additional fractions from the same 18 patients whose one held-out volume forms the test set. Within-patient similarity across treatment fractions (same body habitus, similar applicator geometry, correlated soft-tissue layout) can be memorized, inflating the reported MAE, SSIM, and PSNR. Consequently, Table 1 and the conclusion that LA-GICD works 'without retraining across patients' do not establish generalization to unseen patients. A proper patient-level split (or leave-one-patient-out evaluation) is required.","section":"Section 2.2.2"},{"comment":"The quantitative comparison is limited to FDK, a classical analytical baseline. No comparison is made against learned iterative reconstruction methods, post-processing CNNs, transformer-based approaches, or other diffusion-based limited-angle methods discussed in the Introduction. Without such baselines, the claim that the dual-domain, geometry-integrated design yields superior performance is not supported. Furthermore, there is no ablation of the proposed components (Projection-DDPM alone, Image-DDPM alone, GICD cycle losses, or the GTM as a fixed operator). The reader cannot determine which part of the framework drives the reported improvement, and the abstract's attribution of the gains to the GICD strategy remains unsubstantiated.","section":"Table 1 and Section 3"},{"comment":"All training, validation, and test projections are generated by LEAP-CT simulation from planning CT volumes of 18 patients; no real CBCT acquisitions are used. The central clinical claim—that LA-GICD produces 'artifact-free, high-contrast reconstructions' and reduces dose four-fold in clinical practice—therefore rests on an untested assumption that the simulator faithfully reproduces real gantry behavior including scatter, noise, detector response, and patient motion. The manuscript should either add a real-data validation study (or a domain-shift experiment with real measured projections) or substantially temper the claims about clinical applicability and dose reduction. The four-fold dose reduction is also not measured; it is inferred from the 90-degree arc length, and no dosimetric analysis is provided.","section":"Section 2.2.1 and Abstract"},{"comment":"The claim that LA-GICD effectively suppresses hallucination is not quantitatively supported. Figure 5 is described in the text as showing variability 'typically less than 0.01 normalized units,' but the caption reports a displayed range of 0–0.06 normalized units; this inconsistency is not explained. No statistical summary, threshold, or comparison against a baseline is given for the uncertainty maps. The term 'artifact-free' in the abstract and conclusion is also inconsistent with the acknowledgment of minor inconsistencies near applicators and bone–soft-tissue interfaces in the Discussion. The wording throughout the paper should be aligned with the actual evidence.","section":"Figure 5 and Discussion, 'Hallucination' paragraph"}],"minor_comments":[{"comment":"The running title reads 'Limited-Angel CBCT Reconstruction' instead of 'Limited-Angle'; this typo should be corrected.","section":"Running title"},{"comment":"In the first sentence, 'is de designed to generate' contains a typo ('de designed' should be 'designed').","section":"Section 2.1.2"},{"comment":"The caption uses 'LA-GACD' in the first sentence, which is inconsistent with the framework name LA-GICD.","section":"Figure 5 caption"},{"comment":"The notation for the Image-DDPM losses is inconsistent: Eq. (7) defines ℒ_Recon-DDPM, but Eq. (10) refers to ℒ_θ_I^μ and ℒ_θ_I^Σ without defining them or relating them to Eq. (7). Please clarify the relationship among these loss terms.","section":"Equations (5)–(10)"},{"comment":"The cycle-domain loss in Eq. (10) uses weights γ1=0.05, γ2=0.5, and γ3=0.5 described as 'empirically set,' but no sensitivity analysis or justification is provided. A brief rationale or a supplementary experiment varying these weights would strengthen the presentation.","section":"Section 2.1.5"},{"comment":"The paired t-test is reported only as p < 0.01 for each metric; the number of paired samples (18 test volumes) should be stated in the table or text, and the test should be described precisely (e.g., two-sided paired t-test) to avoid ambiguity.","section":"Section 3, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's clinical claims substantially exceed the evidence base: the evaluation is entirely simulated, the data split is not patient-independent, and no SOTA baselines or ablations are provided. The core methodology may be sound, and the dual-domain diffusion idea is worth pursuing, but the current form would not be publishable in a serious journal without a major reworking of the evaluation protocol and a substantial rewriting of the claims. I recommend major_revision rather than reject because the methodological issues are fixable within scope, and the proposed framework has potential value if properly validated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Top line: the LA-GICD architecture is a real combination—two DDPMs in projection and image space linked by a fixed analytic cone-beam projector/back-projector, with cycle consistency. It builds on your earlier CG-DDPM/PC-DDPM work and is implemented cleanly with the LEAP-CT differentiable projector. On simulated pelvic projections it beats FDK by a wide margin, and the qualitative figures show the expected artifact reduction.\n\nThe soft spots are substantial. First, the patient split is not patient-independent. The paper says 78 volumes from 18 patients, and one volume per patient is held out for testing, leaving 60 for training. Those 60 come from the same 18 patients. The text claims 'no patient overlap exists'—that is inconsistent with the described split. So the reported MAE/SSIM/PSNR likely overstate the model's ability to generalize to a new patient. The 'without retraining across patients' conclusion is not supported.\n\nSecond, the validation is entirely simulated. The projections are generated with LEAP-CT from planning CTs, and the same simulator is used for training, for the analytic GTM, and for the ground truth. There is no real CBCT scan, no scatter, no detector noise, no motion. The sim-to-real transfer is untested, and the abstract's 'clinically feasible' and 'artifact-free' language goes beyond the evidence.\n\nThird, the only comparator is FDK. There is no state-of-the-art learned or iterative baseline, and no ablation of the Projection-DDPM, Image-DDPM, or the cycle loss. That makes it hard to know which component carries the weight.\n\nFourth, the hallucination analysis is weak. Running the same input three times and measuring voxel-wise standard deviation gives a stability map, not a hallucination detector; and the reported text says <0.01 while the figure caption shows 0-0.06 range.\n\nWho is this for? People working on limited-angle CT reconstruction will find the architecture interesting and may want to build on the cycle-domain idea. But as it stands, the quantitative claims about clinical deployment are not yet supported.\n\nFor peer review: I'd send it out. The method is coherent and the clinical question is worth pursuing. The reviewer should demand a patient-disjoint split, real or at least realistically corrupted data, modern baselines, and a chance for the authors to tone down the claims.","headline":"A sensible dual-domain diffusion pipeline for 90-degree CBCT, but the evaluation split does not support the claim of patient-independent generalization, and the lack of real data and strong baselines makes the clinical claims premature.","tokens_in":14687,"tokens_out":3249,"would_cite":false,"duration_ms":30972,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two linked diffusion models reconstruct cone-beam CT volumes from a single 90-degree scan.","keywords":["cone-beam CT","limited-angle reconstruction","denoising diffusion probabilistic model","projection completion","cycle-domain consistency","image-guided radiotherapy","short-arc acquisition","HDR brachytherapy"],"falsifier":"Acquire a real 90-degree CBCT scan on a gantry-mounted system for a phantom or patient with a known full-arc reference, run LA-GICD as described, and compare against the full-arc reconstruction; if MAE, SSIM, and PSNR depart substantially from the reported values or visible streaks and shading remain, the central claim fails.","tokens_in":13656,"feed_emoji":"🩻","tokens_out":5097,"duration_ms":48542,"temperature":0.7,"pith_summary":"LA-GICD claims that a pair of denoising diffusion models, one that completes missing projection views and one that refines the reconstructed volume, connected by a fixed analytic cone-beam projector and back-projector, can turn a single 90-degree cone-beam CT acquisition into a volume nearly matching a full 360-degree scan. On 18 held-out gynecologic HDR brachytherapy volumes simulated from planning CTs, the pipeline reports a mean absolute error of 35.5 HU, SSIM of 0.84, and PSNR of 29.8 dB, with visibly fewer streak and shading artifacts than the analytic FDK baseline. If this carries over to real scanners, short-arc CBCT would reduce acquisition time and imaging dose four-fold while preserving image quality and requiring no hardware changes. That would make volumetric image guidance practical in time-constrained, collision-prone, or motion-prone radiotherapy settings.","feed_headline":"Short-arc CBCT scans can now match full-arc quality","feed_subtitle":"Dual diffusion models fill in missing views and refine the volume, cutting scan time and dose four-fold.","key_machinery":"The load-bearing object is the Geometry Transformation Module (GTM), a fixed, differentiable analytic cone-beam forward- and back-projector using an FDK-style ramp filter, with parameters matched to a gantry-mounted CBCT system. The GTM connects two learned diffusion modules: the Projection-DDPM synthesizes full-view projections conditioned on the limited-angle data, the GTM converts them into an intermediate volume, and the Image-DDPM refines that volume conditioned on the GTM output. A geometry-integrated cycle-domain (GICD) loss enforces agreement between the reconstructed volume, the ground-truth full-angle volume, and the re-projected projections, so every denoising update is anchored to measured rays rather than to an unconstrained image prior.","core_discovery":"The paper's central claim is that limited-angle CBCT reconstruction can be solved by enforcing geometry-aware cycle consistency between projection and image domains. A Projection-DDPM inpaints the missing projections from a 90-degree input; a fixed geometry transformation module back-projects the completed sinogram using exact scanner geometry; an Image-DDPM denoises and sharpens the resulting volume; and a cycle loss re-projects the refined volume and compares it with measured data. The authors claim this closed loop suppresses hallucinated anatomy and preserves soft-tissue boundaries, producing artifact-free, high-contrast reconstructions from a single 90-degree arc with a four-fold reduction in acquisition time and dose.","pith_inferences":["If the simulator-to-real gap is small, the same framework could be trained once on mixed arc lengths and applied to previously unseen arcs such as 60 or 45 degrees, a test the paper flags as future work.","The reported uncertainty maps, with low but nonzero variability near bone, soft-tissue boundaries, and applicator tips, suggest clinical deployment should display confidence maps alongside reconstructions rather than treat every voxel as certain.","The cycle-domain, geometry-integrated design could transfer to other ill-posed inverse problems with a fixed analytic forward model, such as sparse-view CT or metal artifact reduction, though the paper does not test these settings."],"forward_implications":["Clinicians could acquire CBCT over a 90-degree arc and reconstruct volumes with reported accuracy of 35.5 HU MAE, 0.84 SSIM, and 29.8 dB PSNR relative to full-arc CBCT, cutting scan time and imaging dose about four-fold without hardware changes.","Short-arc acquisition would reduce motion artifacts from respiration, peristalsis, and patient movement, and would avoid gantry collisions in crowded settings such as gynecologic HDR brachytherapy with metallic applicators.","Because LA-GICD acts as a software-only post-acquisition reconstruction module on existing flat-panel CBCT systems, it could be deployed across diverse clinical centers without scanner modification.","The authors claim the trained model works across patients without retraining in the tested 90-degree configuration, and suggest the dual-domain design could generalize across different angular spans and orientations."],"supporting_citations":[{"why":"Establishes the clinical role of CBCT in image-guided radiotherapy and motivates reducing scan time.","marker":"[1]"},{"why":"Supplies the analytic cone-beam reconstruction algorithm that the GTM uses for back-projection and that serves as the comparison baseline.","marker":"[7]"},{"why":"Documents the risk of hallucinated anatomy in tomographic reconstruction, the failure mode the GICD loop is designed to prevent.","marker":"[11]"},{"why":"Represents the learned iterative reconstruction approach whose sensitivity to step sizes and regularization the paper contrasts with its own design.","marker":"[15]"},{"why":"Earlier diffusion model for limited-angle CT with weak geometric conditioning, the gap this paper claims to close.","marker":"[19]"},{"why":"Previous cycle-domain geometry-integrated DDPM for single-view CBCT reconstruction that the limited-angle framework extends.","marker":"[21]"},{"why":"Patient-specific physics-integrated DDPM work that shares the cycle-domain geometry-integrated design lineage.","marker":"[22]"},{"why":"Supplies the denoising diffusion probabilistic model machinery used by both the Projection-DDPM and Image-DDPM.","marker":"[23]"},{"why":"Supplies the differentiable forward projector used to simulate projections from planning CTs and to build the GTM.","marker":"[24]"}],"fun_headline_variants":["90° CBCT scans rival full-arc quality with dual diffusion","Cycle-domain diffusion completes CBCT from a 90° arc","Geometry-aware diffusion bridges missing views for CBCT","Dual diffusion models cut CBCT dose 4x without quality loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on simulated projections produced by a CT simulator faithfully matching real cone-beam acquisitions, including scatter, noise, detector response, and patient motion; the same simulator generates the training data, test inputs, and ground truth, so a real-gantry discrepancy would change the reported numbers.","fun_headline_variants_meta":{"raw":{"variants":["90° CBCT scans rival full-arc quality with dual diffusion","Cycle-domain diffusion completes CBCT from a 90° arc","Geometry-aware diffusion bridges missing views for CBCT","Dual diffusion models cut CBCT dose 4x without quality loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00113,"raw_usage":{"total_tokens":4737,"prompt_tokens":1028,"completion_tokens":3709,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":3638}},"tokens_in":644,"tokens_out":3709,"duration_ms":23879,"temperature":1.0,"reasoning_tokens":3638,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:58:27.884761+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire a real 90-degree CBCT scan on a gantry-mounted system for a phantom or patient with a known full-arc reference, run LA-GICD as described, and compare against the full-arc reconstruction; if MAE, SSIM, and PSNR depart substantially from the reported values or visible streaks and shading remain, the central claim fails.","supporting_citations":[{"cited_title":"geometry -integrated","cited_arxiv_id":null,"evidence_quote":"Establishes the clinical role of CBCT in image-guided radiotherapy and motivates reducing scan time."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the analytic cone-beam reconstruction algorithm that the GTM uses for back-projection and that serves as the comparison baseline."},{"cited_title":"The method reconstructs high - quality CBCT volumes from a 90° short-arc projection data, without retraining across angles or patients","cited_arxiv_id":null,"evidence_quote":"Documents the risk of hallucinated anatomy in tomographic reconstruction, the failure mode the GICD loop is designed to prevent."},{"cited_title":"Cone beam computed tomography -based monitoring and management of target and organ motion during external beam radiotherapy in cervical cancer","cited_arxiv_id":null,"evidence_quote":"Represents the learned iterative reconstruction approach whose sensitivity to step sizes and regularization the paper contrasts with its own design."},{"cited_title":"Improved total variation -based CT image reconstruction applied to clinical data","cited_arxiv_id":null,"evidence_quote":"Previous cycle-domain geometry-integrated DDPM for single-view CBCT reconstruction that the limited-angle framework extends."},{"cited_title":"Deep learning image reconstruction for improving image quality of contrast -enhanced dual -energy CT in abdomen","cited_arxiv_id":null,"evidence_quote":"Patient-specific physics-integrated DDPM work that shares the cycle-domain geometry-integrated design lineage."},{"cited_title":"Data consistent artifact reduction for limited angle tomography with deep learning prior","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable forward projector used to simulate projections from planning CTs and to build the GTM."}],"review_version":2}