{"id":"7bcf55b3-18eb-4701-ab07-4d8427fc8eda","arxiv_id":"2607.09828","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Differentiable SV-FBP is robust to discontinuous and multi-isocenter CBCT trajectories, competitive at moderate sparse views, and limited under severe undersampling.","lead":"A learned cone-beam CT reconstructor stays stable on messy, discontinuous, and multi-isocenter robot trajectories, with quality driven more by where samples sit than by path continuity. It is fast and competitive at moderate sparse views but falls behind iterative methods when sampling gets very thin.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection beyond the reader's already-identified synthetic/baseline caveats; the continuity-vs-distribution claim is internally well-supported.","rationale":"The paper is a systematic evaluation extension of prior differentiable SV-FBP work. Its central empirical claim is supported by the controlled reordering experiments (identical sampling locations, different continuity) and the weight visualizations that isolate spatial distribution from path order. The sparse-view transition and multi-isocenter applicability are presented with appropriate caveats. The reader's identification of synthetic data and the deliberately weak unregularized AIR baseline as the weakest assumption is accurate and already sufficient for a CONDITIONAL verdict; no stronger internal flaw (e.g., misapplication of Defrise–Clack operators, unacknowledged trajectory-derivative dependence, or contradictory metrics) appears. Therefore no verdict adjustment is warranted.","tokens_in":14435,"tokens_out":508,"duration_ms":5672,"concrete_test":"Re-run the RFNR Seed-0 sparse-view series (Table 2 / Fig. 5) after adding Poisson noise + scatter to the projections and replacing AIR-50 with a regularized MBIR (e.g., TV or learned prior, 100+ iterations). If the 300–400-view ranking or the claimed transition regime reverses, the operating-range claim weakens outside pure simulation; otherwise the reader's CONDITIONAL verdict stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No new load-bearing concern that would overturn the central claim. The paper's strongest claim (performance largely insensitive to ordering/continuity; spatial sampling distribution dominates; competitive at 300–400 views vs unregularized AIR; works on multi-isocenter without architecture change) is an empirical robustness map for an existing model, not a new method. Within the controlled simulation design of Sections 3.3–5.1, the evidence is coherent: RT/RNNR/RFNR yield nearly identical MSE/PSNR/SSIM (Table 1), learned weights align after re-indexing (Fig. 4), and the moderate quality drop relative to continuous sinusoidal is acknowledged. The multi-isocenter and sparse-view results are likewise consistent with the stated operating range. The reader's weakest assumption (synthetic primitives + weak AIR baseline) already captures the main external-validity limit; no internal inconsistency or hidden mathematical assumption in the SV-FBP pipeline is required for the reported conclusions to hold inside that setup.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"This paper systematically evaluates the previously introduced differentiable shift-variant FBP (SV-FBP) model under challenging CBCT acquisition conditions. Building on known-operator learning of redundancy weights, the authors test fixed-isocenter random, nearest-neighbor-reordered, and farthest-neighbor-reordered trajectories (Table 1, Fig. 4), sparse-view sampling from 400 down to 100 projections against unregularized AIR (Table 2, Fig. 5), and a multi-isocenter Lissajous-saddle geometry (Table 3, Fig. 6). The central empirical claims are that reconstruction quality is largely insensitive to trajectory ordering/continuity and is instead governed by the spatial distribution of sampling points; that the method remains competitive with iterative reconstruction at moderate sparsity (roughly 300–400 views) with substantially lower compute; and that it applies without architectural change to non-planar multi-isocenter paths.","tokens_in":14731,"tokens_out":1035,"duration_ms":9841,"significance":"If the reported operating range holds under more realistic conditions, the work supplies a useful robustness map for an efficient, interpretable alternative to both analytical SV-FBP (inapplicable to discontinuous orbits) and full iterative reconstruction for robotic/non-standard CBCT. Strengths include a controlled multi-seed design, explicit N/A marking of analytical SV-FBP for discrete trajectories, weight-map alignment after re-indexing (Fig. 4), and clear identification of a sparse-view transition regime rather than an unqualified superiority claim. The contribution is primarily empirical characterization of an existing architecture rather than a new reconstruction method, but that characterization is relevant to emerging robotic C-arm trajectories.","major_comments":[{"comment":"Sections 4.2 and 5.1–5.2: Training and primary quantitative claims rest almost entirely on synthetic geometric-primitive volumes (Gaussian-smoothed) with only five forward-projected patient scans used for inference-only evaluation. The continuity-vs-distribution claim and the sparse-view operating range (competitive at 300–400 views) are therefore established under idealized data; real noise, scatter, metal, or anatomical complexity could alter rankings. At minimum the manuscript should quantify performance on the real-patient set (or state that those results are deferred) and discuss how the synthetic design limits clinical transfer of the operating-range claim.","section":null},{"comment":"Section 5.2 / Table 2: The sparse-view competitiveness claim is made against unregularized AIR with a fixed 50 iterations. The paper itself notes that more iterations or lower learning-rate training can close the gap, so the reported transition regime is baseline-dependent. A stronger or regularized iterative reference (or an explicit statement that the comparison is only to this practical unregularized baseline) is needed before the “order-of-magnitude reduction with competitive quality” claim can be taken as a general operating-range result.","section":null}],"minor_comments":[{"comment":"Table 1: Analytical SV-FBP is correctly marked N/A; a brief footnote restating why continuous derivatives are required would help readers who skip Section 3.1.","section":null},{"comment":"Section 4.3: The Gaussian filter after the redundancy-weight layer (kernel 121, σ=20) and the SSIM weight γ=5e-3 are free hyperparameters; a short sensitivity note or justification would improve reproducibility.","section":null},{"comment":"Figure 3 caption mentions a sinusoidal trajectory for visual comparison, but the corresponding quantitative numbers appear only in the text (prior work); adding those numbers to a table or caption would make the moderate quality drop easier to assess.","section":null},{"comment":"Section 3.3.2: Lissajous-saddle parameters (Ax, Ay, Az, a, b, c, δ, f) are given; stating whether they were chosen to match a particular robotic system or purely for geometric stress-testing would clarify external relevance.","section":null},{"comment":"Minor typographical issues: “Yipen Sun” in one reference, inconsistent spacing around ±, and occasional missing spaces after periods in the related-work section.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is an incremental but solid robustness study of the authors’ own prior architecture (Ye et al. 2024/2025). Fit for MELBA/BVM special issue is reasonable; the main risk is over-claiming clinical readiness from synthetic data. I would not block on novelty if the authors tighten the external-validity discussion and baseline caveats."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean evaluation extension of the authors' own differentiable SV-FBP, not a new architecture. What is new is the stress map: RFNR reordering, the sparse-view transition versus AIR, and a Lissajous-saddle multi-isocenter case. That is useful for anyone working on robotic CBCT orbits.\n\nThey do the evaluation carefully inside their setup. Table 1 shows RT/RNNR/RFNR essentially identical on MSE/PSNR/SSIM across seeds. After re-indexing, the learned weights look the same (Fig. 4), which is the right check for the continuity-versus-distribution claim. The moderate drop relative to their earlier continuous sinusoidal numbers is acknowledged rather than papered over. Sparse-view results (Table 2, Fig. 5) mark a practical boundary around 300–400 views where they stay competitive with unregularized 50-iteration AIR at roughly 10× lower cost, then fall behind under severe undersampling. The multi-isocenter numbers (Table 3) are comparable to fixed-isocenter without any architecture change. Math is the standard Defrise–Clack pipeline with only the redundancy weights learned; citations are appropriate and self-citations correctly define the model under test.\n\nSoft spots are real but already mostly stated. Training is almost entirely synthetic primitives; the five real patient volumes are inference-only. AIR is deliberately weak (no regularization, 50 iterations). Free parameters (Gaussian smoother, SSIM weight, LR schedule, Lissajous amplitudes) are empirical. No code or data release. Those limit how far you can push the clinical-robustness language, but they do not break the internal claims.\n\nThis is for people already in non-standard CBCT reconstruction who need a fast FBP-type option with known-operator structure. It is not a method paper and not a clinical validation. I would send it to peer review as an evaluation study; the evidence is coherent and the limitations are proportionate. Worth reading if you care about robotic trajectories; cite if you need the operating-range numbers.","headline":"Solid empirical robustness map for their existing differentiable SV-FBP: continuity is secondary to sampling distribution, with a clear sparse-view operating range and multi-isocenter applicability.","tokens_in":15360,"tokens_out":528,"would_cite":true,"duration_ms":5589,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Differentiable SV-FBP reconstruction quality is driven by where views sample space, not by how continuous the path is, and stays competitive with iterative methods at moderate sparse-view counts.","keywords":["CBCT reconstruction","shift-variant FBP","known operator learning","non-standard trajectories","sparse-view CT","redundancy weights","robotic C-arm"],"falsifier":"Retrain and re-evaluate on real C-arm projections that include scatter, noise, and metal implants, and compare against a regularized iterative method with more iterations; if the ranking or the 300–400-view operating range reverses, the central robustness claim fails.","tokens_in":15334,"feed_emoji":"🩻","tokens_out":954,"duration_ms":10869,"temperature":0.7,"pith_summary":"This paper asks how well a learned, geometry-aware cone-beam CT reconstructor holds up when the scan path is irregular, discontinuous, sparse, or multi-isocenter. The method keeps the classical shift-variant filtered-backprojection pipeline and learns only the redundancy weights that used to need an analytical trajectory derivative. Across random, nearest-neighbor, and farthest-neighbor reorderings of the same source positions, image quality barely changes, which the authors read as evidence that the spatial set of samples matters more than path continuity or ordering. At moderate undersampling (about 300–400 views) it matches an algebraic iterative baseline while running roughly an order of magnitude faster; under severe undersampling the lack of iterative data consistency shows, and quality falls behind. The same architecture also reconstructs a non-planar Lissajous-saddle trajectory whose rotation center moves continuously, without any redesign. Together these results map an operating range for non-standard and robotic CBCT acquisitions.","feed_headline":"CT path continuity barely matters; sampling geometry does","feed_subtitle":"Learned FBP stays competitive at 300–400 views and works on multi-isocenter orbits","key_machinery":"Differentiable shift-variant filtered backprojection (SV-FBP): the classical Defrise–Clack pipeline is kept as known operators, and only the trajectory-dependent redundancy weights are learned from data, removing the need for analytic derivatives of the source path.","core_discovery":"Differentiable shift-variant FBP remains stable under highly irregular and discontinuous fixed-isocenter trajectories; reconstruction performance is largely insensitive to ordering or continuity and is instead governed by the spatial distribution of sampling points. At moderate sparse-view densities it is competitive with unregularized iterative reconstruction at far lower cost, while a clear transition regime appears under severe undersampling. The same model applies without architectural change to multi-isocenter geometries such as a Lissajous-saddle path.","pith_inferences":["If spatial sampling density dominates continuity, trajectory-optimization objectives for robotic C-arms should weight angular coverage and view diversity more heavily than path smoothness.","The learned weight maps, once aligned to source position rather than acquisition order, could serve as a diagnostic of whether a candidate orbit is information-sufficient before any patient is scanned.","A natural next test is whether freezing the learned weights from synthetic data and only fine-tuning on a few real scans preserves the same operating range under metal and scatter.","The same known-operator skeleton may transfer to other incomplete-data CT settings (limited angle, truncated detector) where analytic redundancy weights are hard to write down."],"forward_implications":["Robotic or non-circular CBCT trajectories can be reconstructed without deriving new analytical weights for each path.","Path planners can prioritize good spatial coverage of source positions over continuous traversal order.","At roughly 300–400 projections the method offers a practical speed–quality trade-off versus iterative reconstruction.","Below that density, hybrid schemes that add data consistency become necessary.","Multi-isocenter and non-planar orbits (e.g., Lissajous-saddle) can use the same architecture without redesign."],"fun_headline_variants":["SV-FBP robust to discontinuous paths; sampling geometry decides quality","Trajectory continuity barely matters in differentiable cone-beam FBP","Spatial point distribution trumps path order for shift-variant FBP","Learned SV-FBP stable on irregular orbits without architecture change","Moderate sparse views keep SV-FBP competitive vs iterative recon"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That synthetic volumes of geometric primitives, plus a handful of forward-projected real patient scans used only at test time, and an unregularized 50-iteration iterative baseline, are enough to establish clinical robustness and the claimed sparse-view transition point.","fun_headline_variants_meta":{"raw":{"variants":["SV-FBP robust to discontinuous paths; sampling geometry decides quality","Trajectory continuity barely matters in differentiable cone-beam FBP","Spatial point distribution trumps path order for shift-variant FBP","Learned SV-FBP stable on irregular orbits without architecture change","Moderate sparse views keep SV-FBP competitive vs iterative recon"]},"model":"grok-4.5","effort":"low","cost_usd":0.00847,"raw_usage":{"total_tokens":2009,"prompt_tokens":800,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":84700000,"prompt_tokens_details":{"text_tokens":800,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1138,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":800,"tokens_out":71,"duration_ms":8986,"temperature":1.0,"reasoning_tokens":1138,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T15:13:59.369426+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain and re-evaluate on real C-arm projections that include scatter, noise, and metal implants, and compare against a regularized iterative method with more iterations; if the ranking or the 300–400-view operating range reverses, the central robustness claim fails.","supporting_citations":[],"review_version":1}