{"id":"75b1e886-d3d4-479b-b198-a9128eceae2f","arxiv_id":"2508.00781","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using Monte-Carlo Arithmetic, the authors rank SPM, FSL, and ANTs by numerical stability in linear MRI registration, and report similar variability in healthy and Parkinson's Disease cohorts.","lead":"This study uses Monte-Carlo Arithmetic simulations to measure how much small numerical errors in MRI registration tools (SPM, FSL, ANTs) change their outputs. It finds SPM the most stable, with FSL and ANTs more variable, and suggests these stability findings hold for Parkinson's patients as well as healthy controls.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MCA perturbation fidelity is the load-bearing assumption; the stability ranking and QC signal inherit its validity.","rationale":"The reader identified the fidelity of the MCA perturbation model as the weakest assumption, and the full-text abstract indeed offers no calibration or validation of that model. My stress-test agrees: the entire ranking and the proposed QC signal depend on MCA-induced dispersion being representative of true numerical uncertainty in real registrations. Because the full text is unavailable, I cannot check whether the authors calibrated perturbation levels, compared against hardware-induced variability, or reported a sensitivity analysis. The concern is neither circular nor ad hominem; it is a concrete empirical modeling risk. No other load-bearing issue is visible from the abstract: the HC/PD generalization is a secondary claim that also lacks power reporting, but the core ranking is more fundamental. The reader's verdict of UNVERDICTED with low confidence is appropriate, and my concern does not move it; it reinforces the need for full methods and reproducibility artifacts. The proposed concrete test would settle whether MCA is a faithful proxy: if the ranking survives a sweep of perturbation levels and aligns with real cross-platform variability, the central claim is supported; if not, the study's main conclusion collapses.","tokens_in":1036,"tokens_out":1898,"duration_ms":27099,"concrete_test":"On a subsample of ~10 HC and ~10 PD subjects, run each tool under three conditions: (1) MCA at multiple perturbation levels (e.g., 1, 2, and 4 mantissa bits), (2) the identical binary on different CPU architectures or with different BLAS implementations, and (3) compiler optimization flags varied. Compare the resulting distribution of transformation parameters and failure rates. If the MCA-driven spread is much larger than the hardware/software spread, or if the relative ranking of SPM/FSL/ANTs changes with perturbation level, then the MCA model is not a faithful proxy and the reported stability ranking is not evidence about real numerical uncertainty.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests entirely on Monte-Carlo Arithmetic (MCA) faithfully reproducing true numerical uncertainty of affine registration in SPM, FSL, and ANTs. Real IEEE-754 double-precision computation rounds deterministically; true run-to-run numerical variation arises from non-associativity of floating-point operations, platform-specific BLAS/FFT implementations, and compiler optimizations. MCA instead injects random rounding errors at every operation, which may inflate variability far beyond what real hardware produces, especially for well-conditioned linear registration. The abstract does not report how the perturbation magnitude (e.g., number of mantissa bits perturbed) was chosen or calibrated, making the SPM-highest-stability ranking, the FSL/ANTs overlap, and the ANTs occasional-failure rate potentially artifacts of the MCA hyperparameter. If a lower perturbation level produces a different ordering, or if real cross-platform variability is orders of magnitude smaller than MCA-induced dispersion, the headline result does not transfer to practice. This is not an internal inconsistency but a modeling assumption that cannot be validated from the abstract, and it is the single most load-bearing point of the study.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports an experimental study of numerical uncertainty in linear registration tools (SPM, FSL, ANTs) using Monte-Carlo Arithmetic. Based on the abstract, the authors claim that, with default similarity measures, SPM is the most stable, FSL and ANTs show greater and comparable variability with ANTs occasionally failing, that healthy controls (HC) and Parkinson's disease (PD) cohorts show no significant differences in numerical stability, and that numerical uncertainty measures can support automated quality control (QC) of linear registration. The study uses two brain templates, multiple similarity measures, and n=50 per cohort.","tokens_in":1022,"tokens_out":2581,"duration_ms":28256,"significance":"If the results are reliable, the paper would provide a useful empirical ranking of numerical robustness among widely used registration tools and a practical QC indicator. The use of a uniform external perturbation (MCA) across tools is a reasonable design choice, and the inclusion of a clinical cohort is a strength. The presented abstract alone, however, does not disclose the MCA calibration, the statistical power of the HC/PD comparison, or the validation of the proposed QC signal, so the significance currently depends on assumptions about the full text.","major_comments":[{"comment":"The central claim of tool-specific stability is conditioned on the fidelity of Monte-Carlo Arithmetic. The abstract does not report the perturbation magnitude (e.g., fraction of mantissa bits perturbed), the number of Monte-Carlo runs per registration, or how the perturbation level was calibrated or validated against real hardware variability. Without this information, the SPM/FSL/ANTs ranking and the ANTs failure rate could be artifacts of the MCA hyperparameter rather than intrinsic numerical sensitivity.","section":"Abstract, first sentence"},{"comment":"The generalization claim ('no significant differences were observed between healthy and PD cohorts') is a null result. The abstract reports neither effect sizes nor confidence intervals nor a power analysis for this comparison. With n=50 per cohort, the null could simply reflect insufficient statistical power, making the suggested generalization to clinical populations unsupported.","section":"Abstract, HC/PD comparison sentence"},{"comment":"The QC demonstration is asserted without any reported validation metrics. The abstract says numerical uncertainty measures 'may support' automated QC but gives no sensitivity, specificity, or comparison to existing QC methods, so the claim is not substantiated at the level of the presented evidence.","section":"Abstract, final sentence"}],"minor_comments":[{"comment":"The phrase 'greater and similar ranges of variability' is ambiguous; please report the actual dispersion values (e.g., interquartile ranges or variance) for FSL and ANTs to support the comparison.","section":"Abstract, SPM/FSL/ANTs ranking"},{"comment":"The abstract does not state software versions, computing platform, or compiler/BLAS configuration, which are critical for a numerical reproducibility study.","section":"Abstract, methodology"},{"comment":"The sentence 'no significant differences were observed' should be accompanied by a clear statement that this is a null result with its confidence interval, not evidence of equivalence.","section":"Abstract, HC/PD comparison"},{"comment":"The term 'default similarity measures' is vague; please specify which measures were used for each tool in the abstract or clearly point to the full-text listing.","section":"Abstract, similarity measures"}],"recommendation":"major_revision","confidential_remarks":"This assessment is based solely on the abstract because the full text was not provided in the review materials. The editor should ensure that the full manuscript includes explicit MCA calibration details, a statistical analysis plan for the HC/PD comparison including power or effect sizes, and QC validation metrics. If those are present and sound, the paper may be acceptable; if not, the central claims are not yet supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The abstract promises a genuinely useful thing: a quantitative, tool-by-tool ranking of numerical stability for linear registration in SPM, FSL, and ANTs, plus a check that the ranking holds in a clinical cohort and a proposed QC use. That is a real gap in the neuroimaging literature, and applying Monte-Carlo Arithmetic (MCA) to this problem is a sensible, new use of an established technique. I appreciate that the perturbation model is external and uniform across tools, so the ranking isn't forced by construction. The experimental shape is clear, and the QC demonstration is a nice bonus if it holds up.\n\nThe soft spot is the one the stress-test flags, and it is load-bearing: MCA's fidelity as a proxy for true numerical uncertainty is simply assumed. Real IEEE-754 double-precision rounding is deterministic; run-to-run variation in practice comes from non-associativity of floating-point operations, BLAS/FFT differences, compiler optimizations. MCA injects random rounding at every operation, which can easily inflate variability beyond what real hardware produces, especially for well-conditioned affine registration. The abstract never reports how many mantissa bits were perturbed, how that was chosen, or whether the perturbation level was calibrated against real cross-platform variability. If a smaller perturbation flips the ordering, or if true variability is an order of magnitude smaller, the headline result does not transfer to practice. That is not an internal inconsistency; it is a modeling assumption that a referee absolutely must probe.\n\nThe second issue is smaller but real: the HC/PD null result is used to claim generalization to clinical populations, but the abstract reports no power, effect sizes, or confidence intervals. A null is only informative if the test could have detected a meaningful difference. That is a standard request, not a fatal flaw.\n\nI am reviewing the abstract only, so I cannot verify the statistics, the artifact availability, or the related-work coverage. But the question the paper asks is practical and important enough that the full text deserves a serious referee. My recommendation: send it out, but the editor should ask reviewers to focus on MCA calibration and validation against real numerical variability. If the authors can show the perturbation level is sensible and the ranking is robust to it, this becomes a citable reference. If they cannot, it is a cautionary tale about perturbation-based rankings. Either way, it earns referee time.","headline":"A useful practical comparison of linear registration stability under Monte-Carlo Arithmetic, but the ranking's validity hinges on a perturbation model the abstract never calibrates.","tokens_in":1740,"tokens_out":1209,"would_cite":false,"duration_ms":17480,"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":"The paper claims that Monte-Carlo Arithmetic perturbation reveals stable and unstable linear registration tools, with SPM the most stable.","keywords":["linear registration","numerical uncertainty","Monte-Carlo Arithmetic","MRI preprocessing","SPM","FSL","ANTs","quality control"],"falsifier":"Run the same SPM, FSL, and ANTs linear registrations on real hardware while varying compiler flags, rounding modes, or CPU architectures that change floating-point evaluation order, and check whether ANTs produces occasional failures and whether the same stability ranking appears. If the ranking and failure rate do not reproduce under genuine hardware-dependent rounding, the Monte-Carlo Arithmetic model is not faithful enough to transfer.","tokens_in":645,"feed_emoji":"🧠","tokens_out":3356,"duration_ms":33418,"temperature":0.7,"pith_summary":"The paper sets out to quantify numerical uncertainty in linear registration of brain images, a preprocessing step whose rounding errors are rarely studied. It claims that with default similarity measures SPM is the most stable tool, FSL and ANTs are more variable, and ANTs can fail under numerical perturbation. The authors argue that these uncertainty measures can support automated quality control, and that stability rankings observed in healthy subjects carry over to a Parkinson's disease cohort. A sympathetic reader would care because registration errors propagate through every downstream analysis in neuroimaging pipelines.","feed_headline":"SPM is the most numerically stable brain registration tool","feed_subtitle":"Monte-Carlo perturbation finds FSL and ANTs vary more than SPM, with ANTs occasionally failing","key_machinery":"Monte-Carlo Arithmetic (MCA) is the mechanism: instead of a single deterministic run, each floating-point operation's low-order bits are randomly perturbed, and the registration is repeated many times. The spread of the resulting transformation parameters estimates the numerical uncertainty of the tool and similarity measure. This perturbation-based dispersion is what the paper uses to compare tools and to demonstrate a quality-control signal.","core_discovery":"The paper reports that, across 50 healthy and 50 Parkinson's disease subjects, two templates, and several similarity measures, Monte-Carlo Arithmetic perturbations of floating-point operations produce measurable output variability in all three major linear registration tools. With default similarity measures, SPM shows the smallest dispersion, FSL and ANTs show larger and similar dispersion, and ANTs occasionally fails entirely under perturbation. The study finds no significant difference in numerical stability between healthy and PD cohorts, and it shows that dispersion measures can flag registrations whose results should not be trusted.","pith_inferences":["A practical extension would be to convert MCA dispersion into a per-subject QC score with a threshold tuned on real failing registrations.","The ANTs failures suggest a bifurcation in its optimization path under rounding; reproducing that bifurcation with reduced precision would identify the exact arithmetic step responsible.","If the healthy-to-clinical generalization holds beyond PD, numerical stability audits could be run on open healthy datasets and safely reused for many clinical studies."],"forward_implications":["Users of SPM's default settings can expect more reproducible linear registration results under numerical noise.","ANTs users should treat registration outputs as potentially unstable and should consider reruns or alternative tools when perturbation tests show large dispersion.","Automated QC pipelines could flag registrations with high numerical uncertainty without waiting for downstream artifacts.","Numerical stability findings from healthy cohorts can inform clinical studies of Parkinson's disease, reducing the need to repeat such analyses on every patient group.","Choice of similarity measure interacts with tool choice for numerical stability, so reproducible pipelines should pin both."],"supporting_citations":[],"fun_headline_variants":["SPM most stable brain registration tool under numeric noise","MCA tests: SPM stable, FSL/ANTs vary, ANTs fails","Numerical stability: SPM wins, ANTs fails in MRI registration","Perturbation study ranks SPM first in registration reliability","Brain registration numeric noise: SPM robust, ANTs fragile"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole ranking depends on the assumption that perturbing low-order arithmetic bits in a simulation faithfully reproduces the true numerical errors these tools make on real MRI scans.","fun_headline_variants_meta":{"raw":{"variants":["SPM most stable brain registration tool under numeric noise","MCA tests: SPM stable, FSL/ANTs vary, ANTs fails","Numerical stability: SPM wins, ANTs fails in MRI registration","Perturbation study ranks SPM first in registration reliability","Brain registration numeric noise: SPM robust, ANTs fragile"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1275,"prompt_tokens":842,"completion_tokens":433,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":342}},"tokens_in":458,"tokens_out":433,"duration_ms":5249,"temperature":1.0,"reasoning_tokens":342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:57:00.686904+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same SPM, FSL, and ANTs linear registrations on real hardware while varying compiler flags, rounding modes, or CPU architectures that change floating-point evaluation order, and check whether ANTs produces occasional failures and whether the same stability ranking appears. If the ranking and failure rate do not reproduce under genuine hardware-dependent rounding, the Monte-Carlo Arithmetic model is not faithful enough to transfer.","supporting_citations":[],"review_version":1}