{"id":"00674c79-b980-427f-99a2-4a7906bfffa6","arxiv_id":"2607.23343","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Patient-agnostic synthetic pretraining plus limited patient-specific fine-tuning reaches within about 1 point of from-scratch sub-millimeter success rate on two public datasets while cutting per-patient training time from ~20 h to ~0.5 h.","lead":"This paper tests a 'pretrain on synthetic X-rays from many patients, then fine-tune on one new patient' recipe for 2D/3D registration. It reports a roughly 40x reduction in per-patient training time with almost unchanged sub-millimeter registration accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim is load-bearing on unreported pretraining cost; 44x speed-up is per-patient adaptation only.","rationale":"I read the paper as making two related claims: (1) patient-agnostic synthetic pretraining can preserve registration accuracy after patient-specific adaptation, and (2) this reduces training cost, quantified as a >40x speed-up. Claim (1) is supported by Table 1 and Table 2: pretrained full fine-tuning reaches 84.7%/85.0% SMSR versus 86.1%/85.0% from scratch, and the domain-randomization ablation in Table 4 shows monotone benefit. Claim (2), however, is only supported if the omitted pretraining time is accounted for. The paper is transparent that Table 2's 'training time' is patient-specific fine-tuning time (§4.1), but transparency does not make the efficiency claim complete. The reader's weakest assumption identified exactly this: the 44x speed-up is per-patient adaptation time, not end-to-end time, and pretraining cost is unreported. I agree with this assessment. The concern is a missing measurement rather than an internal inconsistency or a contradiction of the paper's equations. It does not invalidate the accuracy claim, and it is addressable by reporting T_pretrain and a break-even patient count. Therefore the appropriate verdict remains CONDITIONAL, unchanged from the reader's verdict. I also note the rapid patient-specific baseline [40] is cited but not compared, which would have strengthened the comparison, but this is secondary to the cost-accounting issue.","tokens_in":17172,"tokens_out":3556,"duration_ms":35387,"concrete_test":"Measure and report the wall-clock time of the patient-agnostic pretraining stage (500k regressor + 200k similarity samples, same A6000 GPU and batch settings as Table 2). Then compute total training time for a deployment cohort of N patients: T_total(N) = T_pretrain + N × T_adapt, compared with N × T_scratch, and report the break-even N. If T_pretrain cannot be provided, Table 2 should include a 'Total time (first patient)' column, which would show whether the claimed >40x speed-up survives end-to-end accounting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim—'reducing patient-specific training time by more than 40 times' (Conclusion) and the 44.0x/43.5x rows in Table 2—compares only patient-specific fine-tuning time (~0.47/0.48 h) against from-scratch training (~20.70/20.90 h). The one-time patient-agnostic pretraining stage, described in §4.2 as 500k pose-regressor + 200k similarity-network samples on an A6000, has no reported wall-clock cost. If pretraining costs roughly the same as one from-scratch run (plausible given the same sample scale and hardware), then total cost for the first patient is T_pre + 0.47 h ≈ 21.2 h, which is not faster than from-scratch; only after amortizing over multiple patients does pretraining pay off. The abstract's 'significantly reduce patient-specific training requirements' is defensible literally, but the practical efficiency claim in the title and conclusion requires total-cost accounting. Without T_pre, neither the break-even number of patients nor the claimed 'efficiency' can be verified. This is a missing measurement, not an internal contradiction: accuracy results in Table 1 are not affected. Secondary but related: no repeated-seed variance is reported, so the 84.7% vs 86.1% gap on DeepFluoro cannot be assessed as comparable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a two-stage strategy for patient-specific 2D/3D registration: pretrain on synthetic DRRs from multiple CT volumes with a segmentation-free domain randomization, then fine-tune on a small set of synthetic projections from the target CT. The architecture follows the authors' prior spherical similarity learning framework with CNN-Transformer encoders, an E-CNN volume module, and differentiable Levenberg–Marquardt refinement. On DeepFluoro and Ljubljana, full fine-tuning achieves 84.7%/85.0% SMSR versus 86.1%/85.0% trained from scratch, while reported patient-specific fine-tuning time drops from ~20.7/20.9 h to ~0.47/0.48 h (~44×). The paper also ablates fine-tuned modules and domain randomization pools.","tokens_in":17445,"tokens_out":5460,"duration_ms":52290,"significance":"The contribution is potentially valuable: it offers an annotation-free way to convert expensive per-patient training into lightweight adaptation, and the experiments use public datasets with a leave-one-out protocol. Strengths include the controlled comparison of architectures, the domain-randomization ablation, and the direct wall-clock measurement of fine-tuning time. However, the headline efficiency claim is only partially supported because the one-time pretraining cost is not reported, and all accuracy comparisons are point estimates. The central idea is reasonable and empirically defensible pending these measurements.","major_comments":[{"comment":"The 44× speed-up compares only fine-tuning time (0.47/0.48 h) to from-scratch training (20.70/20.90 h). The pretraining stage, described in §4.2 as 500k pose-regressor + 200k similarity-network samples, is never given a wall-clock cost. The efficiency claim in the title and conclusion is therefore load-bearing on amortization: if pretraining costs about as much as one from-scratch run, the first patient is not faster. Please report pretraining wall-clock time and provide a break-even analysis (number of patients for which pretraining pays off), or qualify the claim as reducing per-patient adaptation time only.","section":"§4.2, Table 2"},{"comment":"All SMSR and mTRE numbers are point estimates from a single run. The claimed 'comparable accuracy' rests on 84.7% vs 86.1% on DeepFluoro and equality (85.0% vs 85.0%) on Ljubljana; without repeated-seed means, standard deviations, or a paired comparison, this difference cannot be assessed. Please report mean ± std over at least three seeds, or equivalent uncertainty quantification, for the main comparisons.","section":"Table 2 (and Tables 3–4)"},{"comment":"The core spherical similarity learning and the bi-invariant SO(4) pose-gradient supervision are described by reference to the authors' conference paper [28]. Since these components are central to the method and the manuscript is intended as a journal version, the paper should be self-contained: define the spherical exponential map in Eq. (7), the SO(4) embedding, and the geodesic discrepancy used in Eq. (9), or include them in a supplement. As written, a reader cannot reproduce the training objective without obtaining [28].","section":"§3.3, Eq. (9)"}],"minor_comments":[{"comment":"Define what '10% FT', '25% FT', etc. mean (e.g., fractions of the 20k-sample adaptation set). The 'Fine-tuning amount' column is ambiguous.","section":"Table 2"},{"comment":"'The learned deep similarity is visualized in Fig. 7' is a dangling reference; the figure should be described and interpreted in the text. Also add the missing period after 'Fig. 7'.","section":"§4.2, Fig. 7"},{"comment":"Define I0 and the valid intensity range after normalization; this affects reproducibility of the preprocessing.","section":"Eq. (18)"},{"comment":"'The baselines follow the same setting as we introduced in [28]' should be expanded so the reader does not need to consult [28] to understand the baseline evaluation protocol.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The missing pretraining cost is the main barrier; if the authors can supply it together with uncertainty quantification, the paper would be substantially stronger. The heavy reliance on [28] for core definitions is also worth addressing for a journal readership."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chen and Kong report that a patient-agnostic pretraining stage followed by limited fine-tuning on the target CT recovers essentially the same SMSR as training from scratch, at 44x less per-patient time. The core numbers hold up: Table 2 shows 84.7/85.0% SMSR with pretrained full fine-tuning versus 86.1/85.0% from scratch, and Table 4 gives a clean monotone benefit from adding domain randomization layers. The adaptation-cost analysis (Table 3) is genuinely useful — it says fine-tuning regressor+similarity network gets you almost everything, and adding the E-CNN is not worth it. I believe the main claim: pretraining transfers enough pose-sensitive representation that patient-specific training becomes a short fine-tuning problem.\n\nThe weak spot is exactly where the concern lands. The '44x speed-up' compares 0.47h fine-tuning to 20.7h from-scratch. The patient-agnostic pretraining itself — 500k regressor samples plus 200k similarity-network samples on an A6000 — is never timed. Given that from-scratch training uses the same 500k/200k sample budget, pretraining plausibly costs about one from-scratch run. For the first patient, then, total time is ~21h either way; the benefit is amortized only after several patients. That is a fine story for a hospital doing many cases, but the paper should say it, and should report pretraining time or at least a break-even patient count. The omission matters because the title and conclusion sell the method as 'efficient' without that caveat.\n\nOther soft spots: all metrics are point estimates — no repeated seeds or error bars, so a 1.4-point gap (86.1 vs 84.7) might or might not be meaningful. The 'rapid patient-specific' baseline from [40] is cited but never compared, which is odd since it is the closest thing to the proposed pipeline. And the relationship between the 20k adaptation samples and the 500k/200k pretraining budget is described loosely, not specified exactly. None of this contradicts the central empirical result; it is all fixable reporting.\n\nThe math and methods look sound — the spherical similarity and LM refinement come from prior work, and this paper is honest about that. The paper reads as a careful empirical study, not an overclaim. It deserves a serious referee; the fixes are measurement and reporting, not rethinking. I'd send it out.","headline":"Solid, useful pretraining/adaptation study for patient-specific 2D/3D registration; the headline 44x speed-up only counts fine-tuning, not the one-time pretraining, which is never timed.","tokens_in":17988,"tokens_out":2192,"would_cite":true,"duration_ms":21500,"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":"Patient-agnostic synthetic pretraining can cut the per-patient training cost of intraoperative 2D/3D registration by more than 40 times while preserving sub-millimeter accuracy, this paper argues.","keywords":["2D/3D registration","patient-specific adaptation","synthetic pretraining","domain randomization","spherical similarity learning","differentiable pose refinement","intraoperative imaging","digitally reconstructed radiographs"],"falsifier":"Compute the total wall-clock time of pretraining plus patient-specific fine-tuning on the same GPU, then plot total time per patient against the number of patients treated. If the pretraining stage takes tens of hours, the break-even point may exceed realistic caseloads, which would overturn the paper's central efficiency claim. The from-scratch baseline time is given (20.7–20.9 h), so the comparison is directly testable.","tokens_in":16983,"feed_emoji":"🩻","tokens_out":4597,"duration_ms":41586,"temperature":0.7,"pith_summary":"The paper tries to show that the expensive step in patient-specific 2D/3D registration—training a separate model from scratch for each new patient—can be replaced by a two-stage recipe: pretrain once on synthetic X-ray-like images (digitally reconstructed radiographs, DRRs) generated from many CT volumes, then fine-tune briefly on a small set of synthetic projections from the target patient's own CT. The central question is whether this shortcut preserves the accuracy of full per-patient training. The authors report that with full fine-tuning the pretrained model reaches 84.7% sub-millimeter success on pelvic fluoroscopy and 85.0% on cerebral angiography, statistically close to the 86.1% and 85.0% achieved from scratch, while cutting per-patient training time from about 20.7–20.9 hours to 0.47–0.48 hours—roughly a 44-fold reduction. A segmentation-free domain randomization scheme (intensity, projection physics, occlusion, field-of-view perturbations) is what lets the pretraining generalize to real X-rays without anatomical labels. If the claim holds, per-patient registration becomes an adaptation problem instead of a training problem, which matters for clinical deployment where speed matters.","feed_headline":"Synthetic pretraining makes per-patient 2D/3D registration 40x faster","feed_subtitle":"A shared pretrained model reaches near from-scratch accuracy after under half an hour of per-patient fine-tuning.","key_machinery":"The framework combines three pieces: (1) patient-agnostic synthetic pretraining, in which DRRs generated from multiple CT volumes under randomized appearance, physics, occlusion, and field-of-view perturbations teach the network pose-sensitive features; (2) spherical similarity learning, in which image features are mapped to a hypersphere and the geodesic discrepancy between fixed and moving projections is minimized, with a bi-invariant SO(4) pose-gradient supervision aligning the learned metric's gradient to pose-space geodesic directions; and (3) differentiable Levenberg-Marquardt refinement, which iteratively updates the pose using Jacobians of the learned residual. The patient-specific a","core_discovery":"The paper's central claim is that a model pretrained on synthetic projections from multiple patients learns a pose-sensitive representation that transfers to a new patient, so that only a short fine-tuning step on the target CT's synthetic projections is needed. Direct transfer without adaptation fails (about 20% sub-millimeter success), but fine-tuning recovers nearly all of the from-scratch accuracy: on pelvic data, 10% of the usual fine-tuning budget reaches 42.4% SMSR, 50% reaches 74.3%, full fine-tuning reaches 84.7% versus 86.1% from scratch; on cerebral data, full fine-tuning matches 85.0% exactly. The authors attribute the transfer to the spherical similarity landscape and the SO(4)-","pith_inferences":["Editorial extension: the reported 44x speed-up counts only patient-specific fine-tuning; if the one-time pretraining cost is large, the total time savings depend on the number of patients treated. A deployment site seeing few patients might not break even.","Editorial extension: the same pretraining-plus-adaptation logic could apply to other per-patient tasks, such as instrument tracking or reconstruction from sparse views, wherever a shared geometry prior can be learned from synthetic projections.","Editorial extension: the domain randomization pools could be tuned per anatomy; adding surgical-tool occlusion specifically for interventional workflows might further close the synthetic-to-real gap, though the paper does not test this."],"forward_implications":["Per-patient training time for learning-based 2D/3D registration drops from about 20.7–20.9 hours to under half an hour, making same-day adaptation between cases realistic.","A single patient-agnostic pretrained model can serve as a reusable starting point across many patients, so the one-time pretraining cost is amortized over the caseload.","Segmentation-free domain randomization provides robustness to real fluoroscopic appearance without requiring anatomical labels or segmentation masks, removing a major annotation bottleneck.","Fine-tuning only the pose regressor and similarity network recovers most of the accuracy of full fine-tuning, so the most efficient adaptation setting is partial fine-tuning.","The learned spherical metric supplies a differentiable landscape that supports Levenberg-Marquardt refinement, so final accuracy depends jointly on the quality of the initialization and the learned similarity."],"fun_headline_variants":["Synthetic pretraining slashes per-patient fine-tuning cost for 2D/3D registration","One pretrained model cuts adaption time for patient-specific registration","Patient-agnostic pretraining preserves accuracy with a fraction of the training","Shared model adapts to new patients in minutes for intraoperative registration"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The efficiency claim rests on the assumption that the one-time patient-agnostic pretraining cost is small enough to be amortized across the patients a site actually treats—the paper reports only fine-tuning time (0.47–0.48 h) and never the pretraining wall-clock time, so if pretraining is expensive and caseload is small, the full pipeline could cost more than training from scratch.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic pretraining slashes per-patient fine-tuning cost for 2D/3D registration","One pretrained model cuts adaption time for patient-specific registration","Patient-agnostic pretraining preserves accuracy with a fraction of the training","Shared model adapts to new patients in minutes for intraoperative registration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000676,"raw_usage":{"total_tokens":2954,"prompt_tokens":829,"completion_tokens":2125,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2044}},"tokens_in":573,"tokens_out":2125,"duration_ms":17474,"temperature":1.0,"reasoning_tokens":2044,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:40:38.550228+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the total wall-clock time of pretraining plus patient-specific fine-tuning on the same GPU, then plot total time per patient against the number of patients treated. If the pretraining stage takes tens of hours, the break-even point may exceed realistic caseloads, which would overturn the paper's central efficiency claim. The from-scratch baseline time is given (20.7–20.9 h), so the comparison is directly testable.","supporting_citations":[],"review_version":1}