{"id":"1a611e59-19d1-4532-9588-b41b69428f0a","arxiv_id":"2607.19099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"BART now includes an open-source sequence framework with a Siemens driver for online parameter adjustment and Pulseq export, validated by matching T1 and R2*/B0 maps from online vs offline acquisitions.","lead":"A team extended the open-source BART MRI toolbox with a pulse-sequence programming framework that runs scans on a clinical Siemens scanner and can export the same sequence to the vendor-neutral Pulseq format for offline playback. This closes a reproducibility gap by letting the same computational MRI method be executed online with live parameter adjustment and reproduced offline from the same source code.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 'exact offline reproduction' claim is qualified by the paper's own phase-correction step and by an indirect, single-subject validation.","rationale":"The reader's weakest_assumption identified the same underlying point: faithful preservation of all acquisition details is assumed but not directly verified. My concern sharpens this by pointing to a known violation (the phase inconsistency) and to the indirect nature of the validation. The reader's CONDITIONAL verdict is appropriate: the paper is valuable, code and data are provided, but the 'exact' claim should be toned down to 'equivalent after retrospective correction' or further validated. I find no reason to change the verdict, so I mark it UNCHANGED. The proposed test would settle whether the correction is exact and whether other hidden discrepancies exist.","tokens_in":11695,"tokens_out":8455,"duration_ms":77935,"concrete_test":"Perform a phantom experiment with interleaved online and offline acquisitions (alternating every few seconds) under temperature-stabilized conditions, using a fixed FOV shift. After applying the paper's retrospective correction, compute the complex residual between online and offline raw k-space data (after phase alignment using a reference signal, not an optimal global scaling). Test whether the residual is consistent with the measured noise floor across all spokes (e.g., chi-squared test). If a systematic residual remains, the correction is incomplete and the 'exact' claim fails; if it is noise-like, the central claim holds for this scenario.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper claims that the exact same sequence can be reproduced offline (Abstract), but Section 3.2 states that Pulseq acquisitions of non-Cartesian sequences are 'affected by phase inconsistencies' and must be corrected retrospectively using the known FOV shift. Thus the offline raw data are not identical to the online acquisition; equivalence is only achieved after a data-dependent correction. The paper does not quantify the residual error after correction. The validation rests on comparisons of reconstructed quantitative maps, which are an indirect probe of sequence fidelity: model-based reconstruction can tolerate or compensate for small systematic differences. The in vivo results are from a single volunteer (Section 3.1), and R2*/B0 agreement is reported only qualitatively (Section 4). If test-retest variability is substantial, the observation that online-vs-offline differences are 'similar to test-retest' does not establish small absolute differences. The assumption that all gradient moments, RF phases, timings, and trajectories are faithfully preserved (load-bearing for the central claim) is therefore not rigorously supported. The phase correction addresses only FOV shifts; other sequence parameters (e.g., RF spoiling phases, trigger delays, block splitting) could introduce unmodeled discrepancies that the current validation cannot detect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes an extension of the BART toolbox into a full end-to-end framework for computational MRI: a sequence-programming layer, a vendor-specific driver sequence (BOOST) for online execution on Siemens scanners with real-time parameter adjustment, and an offline Pulseq export path. The authors implement two quantitative MRI applications—single-shot inversion-recovery radial FLASH for T1 mapping and multi-echo radial FLASH for joint R2*/B0 mapping—and validate reproducibility by comparing online acquisitions with offline Pulseq acquisitions in phantom and in a single healthy volunteer. The central claim is that the same sequence can be reproduced exactly offline via Pulseq, and that quantitative maps agree after a retrospective correction for non-Cartesian FOV-shift phase inconsistencies.","tokens_in":11936,"tokens_out":5293,"duration_ms":52215,"significance":"If the central claim holds, this is a valuable contribution to reproducible computational MRI. The framework integrates sequence design, reconstruction, and data storage in one open-source ecosystem, which is a practical step toward end-to-end reproducibility. The authors provide open-source code, a Zenodo data release, and a continuous-integration test pipeline—these are concrete strengths that support the framework's maintainability. The quantitative applications are nontrivial and the empirical comparisons, while limited, are designed to address the online/offline equivalence question directly rather than by assumption. The main risk is that the strength of the 'exact reproduction' claim exceeds what the evidence supports, given the phase-correction step and the single-subject, partially qualitative validation.","major_comments":[{"comment":"The central claim of 'exact offline reproduction' (Abstract) is not supported as stated. Section 3.2 states that Pulseq acquisitions of non-Cartesian sequences are 'affected by phase inconsistencies' and require a retrospective correction using the known FOV shift to match online data. Thus the offline acquisition is not equivalent to the online one; equivalence is restored only by a data-dependent post-processing step. The residual error after this correction is not quantified. Because the quantitative-map agreement is an indirect probe of sequence fidelity (model-based reconstruction can tolerate small systematic errors), the paper should either (i) soften the 'exact' claim to 'equivalent after the prescribed correction' or (ii) provide a quantitative upper bound on the residual k-space/image/parameter differences after correction. This is load-bearing for the validation logic.","section":"Abstract; §3.2"},{"comment":"The in vivo validation rests on a single healthy volunteer (Section 3.1), and the R2*/B0 comparison in Figure 7 is presented only visually ('Good agreement ... demonstrated by low differences'), without numerical summary. The claim in the Abstract that 'quantitative parameter maps ... agree' is therefore not quantitatively supported for the R2*/B0 application. Please provide ROI-based or voxel-wise statistics (e.g., mean/SD of differences, Bland-Altman limits) for these maps, and ideally at least one additional subject or an explicit statement that the proof-of-concept is limited to single-subject demonstration.","section":"§3.1, §4, Figure 7"},{"comment":"For the T1 comparisons, the paper reports that online-vs-Pulseq differences are 'similar' to test-retest, but it does not report the actual test-retest variability numbers. In the cardiac T1 case, the online-vs-Pulseq mean difference is 10 ms with SD 20 ms; without the corresponding test-retest values, 'similar' is ambiguous. If test-retest SD is also ~20 ms, this does not establish small absolute differences between acquisition modes. Please report test-retest statistics alongside the online-vs-Pulseq statistics for all quantitative comparisons.","section":"§4"},{"comment":"The equivalence of online and offline acquisitions relies on the assumption that the vendor driver and the Pulseq exporter preserve all RF phases, gradient moments, timings, and block splitting. The validation checks raw data (Fig. 3) only after removing a global complex scaling factor, which can hide systematic phase errors, and only for isocenter/continuous readouts. This does not cover RF spoiling phases, trigger delays, or the block-splitting introduced in Pulseq export (§2.3). The manuscript should state this limitation explicitly and, if possible, validate a sequence with FOV shift and triggers directly (e.g., compare trajectories or measure a local phase map) to support the generalization of the equivalence claim.","section":"§2.2, §2.3"}],"minor_comments":[{"comment":"'BOOST' is used in the title but defined only in the Data Availability Statement; please define it at first use in the abstract or introduction.","section":"Abstract/Data Availability"},{"comment":"'Completely disentangle' is an overstatement given the required digital signature and mutual version check; suggest 'largely' or 'to the extent possible'.","section":"§2.2"},{"comment":"The symbol STEM is used before being fully defined; clarify that it denotes the signal equation for the m-th echo.","section":"§3.4, Eq. (5)"},{"comment":"In the code block, the comment says 'shift of 100mm in y' while the parameter is 0.1; make the units explicit (0.1 m = 100 mm).","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"This is a solid engineering contribution. The main risk is that the 'exact reproducibility' claim is too strong for the evidence presented. I recommend major revision rather than rejection because the limitations are addressable with added analysis and careful language."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a solid engineering paper that does something nobody else has done in one package—open-source sequence development, online scanner control, and offline Pulseq reproduction, all integrated with BART. The claimed 'exact' offline reproduction is over-stated, but the caveat is disclosed in the paper and is fixable.\n\nWhat's actually new: gammaSTAR does online adjustment but isn't open-source; Pulseq alone doesn't do online parameter changes. This framework adds both to BART, and the ability to log the full command-line and export Pulseq is a real reproducibility feature. The two qMRI demonstrations (T1, R2*/B0) are sensible proof-of-concepts. Code and data are available, which matters.\n\nSoft spots: the word 'exact' in the abstract and conclusion is not supported by the paper's own results. Section 3.2 shows that Pulseq acquisitions of non-Cartesian sequences need a retrospective phase correction for FOV shifts; so the offline data are not identical until after a data-dependent correction. That's a known Pulseq interpreter issue, and the correction is principled, but 'exact' should be replaced with something like 'closely reproduced,' and the residual after correction should be quantified. In vivo data are a single healthy volunteer; R2*/B0 comparisons are qualitative (visual difference maps). That's acceptable for a software methods paper, but it limits the strength of the claim that online and offline agree within test-retest—especially since test-retest variability includes physiology, and absolute differences aren't reported for the brain maps. The BOOST driver is distributed via Siemens' C2P, so not fully open, though the BART side is.\n\nThe stress-test worry about unmodeled discrepancies in RF phases or trigger delays is plausible but speculative—nothing in the data suggests a problem, and the k-space difference experiment for the FOV shift is more direct than the reconstructed maps. Still, a direct comparison of raw k-space after phase correction for one of the qMRI sequences would close the loop.\n\nThis is a paper for MRI method developers, especially those in the BART/Pulseq ecosystem. It deserves a serious referee. My recommendation: send out, ask for a revision that softens 'exact,' quantifies the phase-correction residual, and reports numbers for the R2*/B0 comparison. Also note the single-subject limitation in the discussion.","headline":"A genuinely useful open-source sequence framework for computational MRI, with a slightly over-stated 'exact' reproduction claim; worth refereeing.","tokens_in":12429,"tokens_out":2727,"would_cite":true,"duration_ms":26351,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["87.61.-c"],"model":"deepseek-v4-flash","headline":"This paper extends the open-source BART toolbox so that the same MRI pulse sequence runs both live on a clinical scanner and as an exported Pulseq file, and it shows that the resulting quantitative maps match within test–retest variability.","keywords":["open-source","pulse sequence design","BART","Pulseq","quantitative MRI","model-based reconstruction","radial FLASH","reproducibility"],"falsifier":"Measure the same phantom with a deliberately introduced one-raster-step (e.g., 10 µs) delay on the readout gradient in the sequence definition, once online and once via the exported Pulseq file. If the raw k-space data from the two paths differ in a way that cannot be removed by a global phase/time shift (beyond B0 drift), the equivalence claim fails; if they remain identical, the two interpreters demonstrably share exact timing semantics.","tokens_in":11581,"feed_emoji":"🧲","tokens_out":7895,"duration_ms":75050,"temperature":0.7,"pith_summary":"Modern computational MRI couples acquisition and reconstruction so tightly that published methods are often irreproducible: sequence details live in vendor-specific code. This paper extends the open-source BART toolbox with a sequence-development framework that describes a scan as generic events (gradients, RF, ADC, triggers), then runs those events through two interpreters: a vendor-specific driver for live scanning with online parameter adjustment, and an offline tool that recomputes the same events and exports them in the Pulseq format. The central claim is that the offline Pulseq replay is physically equivalent to the online scan: for two quantitative methods—single-shot inversion-recovery radial FLASH for T1 and multi-echo radial FLASH for joint water/fat R2* and B0 mapping—model-based reconstructions from online and offline acquisitions agree, with differences comparable to test–retest variability. If true, this means one open-source implementation can serve both clinical practice (where parameters must adapt per patient) and long-term, scanner-independent reproducibility.","feed_headline":"Open-source MRI toolbox runs the same scan live and offline","feed_subtitle":"Quantitative maps from the scanner and exported Pulseq files match within test-retest variability.","key_machinery":"The load-bearing mechanism is the event abstraction plus its dual interpreters. A sequence is a list of parametrized events (gradient triangles, RF pulses, ADC periods, wait/trigger) grouped into blocks; gradients are integrated over the raster to preserve zeroth moments, and the midtime is the phase reference. The online path compiles BART as a dynamic library and links it to a vendor-specific driver sequence that checks parameter feasibility, prepares events just-in-time, and exposes a custom UI; the offline path runs bart seq, which reproduces the same timings and waveforms and writes a Pulseq .seq file, splitting blocks to respect Pulseq's one-RF/one-ADC-per-block limit. A run-time versi","core_discovery":"The paper demonstrates that a single event-based sequence representation can drive a clinical scanner and an offline simulator-equivalent without loss of fidelity. Sequences are built from parametrized blocks—gradients decomposed into triangle splines, RF pulses with phase/frequency/shape, ADC with dwell-time and loop indices, trigger/wait events—prepared just-in-time by the online driver and recomputed offline by the bart seq command. For the two proof-of-concept qMRI methods, the authors report that T1 maps and R2*/B0 maps from online acquisition and from the exported Pulseq file agree at the level of test–retest repeats; for off-center FOVs, a provided correction script restores the phase","pith_inferences":["If the equivalence holds across harsher gradient loads (shorter raster times, higher slew rates), the same pair of interpreters could serve as a cross-vendor quality-assurance phantom: scan once online, once via Pulseq, and compare maps.","The event abstraction decouples sequence logic from hardware, so porting to a second vendor should require only rewriting the thin driver layer; that would let the field distribute one canonical sequence implementation instead of per-vendor rewrites.","Because the offline path is deterministic and fast, it enables gradient-based or reinforcement-learning sequence optimization in simulation, with the winning sequence then run online without translation to another language.","A natural stress test is to verify the equivalence for moving subjects (e.g., cardiac) with retrospective gating, where timing jitter between triggers and spokes might expose subtle differences between the two interpreters that static phantoms mask."],"forward_implications":["Researchers can distribute a single BART sequence definition that is both usable interactively on a scanner and replayable offline, eliminating the usual double implementation of sequence and reconstruction.","The Pulseq export makes a recorded acquisition portable to other vendors' scanners and to Bloch simulators, so a study's acquisition can be re-run or simulated exactly as published.","For the demonstrated qMRI methods, quantitative maps are not measurably degraded by the offline path, meaning multi-center or longitudinal studies could archive the .seq file plus reconstruction code as the method record.","The automatic log file and versioned library pairing make each scan self-documenting, aiding audits and debugging of failed acquisitions.","The offline command also recovers k-space trajectory, sample times, and phases needed for model-based reconstruction, so the reconstruction can consume faithful metadata rather than approximate trajectories."],"fun_headline_variants":["Open-source MRI toolbox unifies live and offline sequences","MRI sequence toolbox: scanner and simulation match","One MRI toolbox reproduces scans live and offline","Same MRI scan live or simulated with open toolbox","BART open toolbox: identical MRI sequences on scanner and PC"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim stands on the assumption that both the live driver and the Pulseq exporter execute the exact same event stream, preserving every gradient moment, RF phase, timing, and trajectory detail; if either interpreter misinterprets an event or loses a phase or timing during conversion, online and offline scans become physically different and the demonstrated agreement would not generalize.","fun_headline_variants_meta":{"raw":{"variants":["Open-source MRI toolbox unifies live and offline sequences","MRI sequence toolbox: scanner and simulation match","One MRI toolbox reproduces scans live and offline","Same MRI scan live or simulated with open toolbox","BART open toolbox: identical MRI sequences on scanner and PC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000668,"raw_usage":{"total_tokens":2893,"prompt_tokens":766,"completion_tokens":2127,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":2053}},"tokens_in":510,"tokens_out":2127,"duration_ms":15546,"temperature":1.0,"reasoning_tokens":2053,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T13:25:42.214715+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same phantom with a deliberately introduced one-raster-step (e.g., 10 µs) delay on the readout gradient in the sequence definition, once online and once via the exported Pulseq file. If the raw k-space data from the two paths differ in a way that cannot be removed by a global phase/time shift (beyond B0 drift), the equivalence claim fails; if they remain identical, the two interpreters demonstrably share exact timing semantics.","supporting_citations":[],"review_version":1}