{"id":"c0c033b2-f3bf-4367-a59b-64d5c3edb421","arxiv_id":"2501.18567","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"CSI-bSSFP with phase cycling improves deuterium metabolic imaging SNR in the human brain at 9.4T by 18% for glucose and 27% for Glx, while ME-bSSFP improves resolution but not SNR.","lead":"This MRI methods paper tests balanced steady-state free precession (bSSFP) acquisitions for deuterium metabolic imaging of the human brain at 9.4 tesla. It reports that the CSI-bSSFP variant improves glucose and Glx signal-to-noise by 18% and 27% over the standard sequence, while a multi-echo variant offers higher resolution but lower SNR.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CSI-bSSFP SNR gain may be inflated by data-driven eigenvector combination in IDEAL-modes; a split-half or theoretical-weight reanalysis would settle whether the 18%/27% claim is robust.","rationale":"The reader correctly identifies the in vivo SNR comparison as the load-bearing part of the central claim, but the specific mechanism they name, the PSF voxel-volume scaling, is less critical for the CSI-bSSFP versus CSI-FISP comparison because the correction factor is 0.96, close to unity. The more serious unexamined risk is the IDEAL-modes eigenvector recombination: choosing the principal eigenvector from the same noisy data can introduce a positive SNR bias that is absent from the standard-IDEAL reference. The paper's own Supporting Figure S6 suggests the weights resemble theoretical mode decay, which is reassuring, but the resemblance is itself estimated from the data and does not rule out partial overfitting. A split-half or theory-fixed reanalysis is a concrete way to test this. Because the feasibility claim is well supported by phantom experiments, open code/data, and cross-method agreement, the verdict should remain conditional rather than being strengthened or reversed.","tokens_in":14630,"tokens_out":8571,"duration_ms":85728,"concrete_test":"Recompute the CSI-bSSFP glucose and Glx SNR maps with the IDEAL-modes combination weights fixed a priori: either use the theoretical SSFP mode decay shown in Supporting Figure S6, or estimate the weights on one half of the interleaved data and apply them to the other half. Then compare the resulting whole-brain SNR against CSI-FISP. If the gain drops below roughly 10% for glucose or below roughly 15% for Glx, the reported 18%/27% advantage is partly an artifact of data-driven eigenvalue selection; if the gain is unchanged, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the 18%/27% CSI-bSSFP SNR advantage over CSI-FISP. The PSF normalization is not the main risk for that specific comparison, since the CSI-bSSFP correction factor is only 0.96; the more consequential assumption is in the IDEAL-modes reconstruction. In the 'IDEAL-modes fit' section, metabolite amplitudes from the K=4 phase-cycled bSSFP data are estimated per SSFP configuration mode (C=K/2-1=1, i.e., three modes), and the complex weights for combining them are obtained by taking the principal eigenvector with the largest eigenvalue of the fitted mode amplitudes. Selecting a combination direction from the same noisy data can bias the resulting SNR upward, especially for low-SNR metabolites such as glucose and Glx. No such data-driven eigenvalue selection is applied to the CSI-FISP reference, which is processed with standard IDEAL, so the comparison is not symmetric. The reported gain is also an average over three subjects without confidence intervals. If the eigenvector weights are overfit to noise, the 18%/27% numbers could shrink or disappear; the feasibility conclusion would survive, but the quantitative sensitivity claim would not.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports a feasibility study of balanced SSFP (bSSFP) acquisitions for deuterium metabolic imaging (DMI) of the human brain at 9.4 T. Two bSSFP variants are implemented and compared with the vendor gradient-spoiled CSI-FISP sequence with IDEAL processing: an acquisition-weighted CSI-bSSFP and a multi-echo bSSFP (ME-bSSFP), both with RF phase cycling. The authors measure in vivo T1, T2, and T2* of deuterated water, glucose, Glx, and lactate; use these values in steady-state simulations to optimize the protocols; validate spectral separation on a phantom; and acquire interleaved 10-minute DMI scans in three healthy volunteers. They report that phase cycling improves off-resonance robustness and adds spectral encoding, that CSI-bSSFP yields an average SNR increase of 18% for glucose and 27% for Glx over CSI-FISP, and that ME-bSSFP provides higher nominal spatial resolution but lower or comparable SNR. The conclusion is that bSSFP DMI is feasible at 9.4 T and has potential to improve sensitivity, although the measured gains are smaller than the simulated predictions.","tokens_in":14839,"tokens_out":8040,"duration_ms":81480,"significance":"If the 18%/27% SNR advantage is robust, this is a useful incremental advance for human DMI at 9.4 T: higher-resolution metabolic maps within the same 10-minute acquisition time would improve the practicality of DMI. The paper's strengths include open data and code (GitHub/Zenodo), phantom validation for all four relevant deuterium resonances, interleaved in vivo acquisitions to reduce metabolic-dynamics confounds, and candid reporting of results that do not match the simulations: the measured gains are explicitly lower than the 45%/69% predictions, the ME-bSSFP SNR shortfall is discussed, and subject-2 B0-related artifacts are shown. The main caveat is that the quantitative CSI-bSSFP advantage depends on a data-driven eigenvector combination step that is applied to the same data used for SNR evaluation and is not symmetrically applied to the CSI-FISP reference; the robustness of the headline numbers therefore needs to be demonstrated before the exact gain is accepted.","major_comments":[{"comment":"The IDEAL-modes method obtains the complex weights for combining SSFP configuration modes by taking the principal eigenvector of the fitted metabolite amplitudes from the same phase-cycled bSSFP data that are then used to compute the SNR maps. This is a data-driven matched filter: it selects the linear combination that maximizes the apparent signal in the very noise realization being analyzed, and the bias is expected to be largest for low-SNR metabolites such as glucose and Glx, which are exactly the metabolites for which the headline 18%/27% gains are reported. The CSI-FISP reference is processed with standard IDEAL and receives no equivalent data-driven selection, so the comparison is asymmetric. The resemblance of the eigenvectors to simulated SSFP mode decay (Supporting Information Figure S6) is suggestive but is not a quantitative guard against inflation. Please reanalyze the bSSFP data with weights estimated from one half of the interleaved averages and applied to the other half, or with weights fixed to the simulated SSFP mode amplitudes, and report the resulting glucose/Glx gains.","section":"Spectral fitting; IDEAL-modes fit"},{"comment":"The central quantitative claim — an average SNR increase of 18% for glucose and 27% for Glx for CSI-bSSFP over CSI-FISP — is based on three subjects and is reported without confidence intervals, a subject-level breakdown, or a precise definition of the whole-brain ROI used for the comparison. The text states a range of 9–29% for glucose, which is wide relative to the claimed 18% mean; this matters because the abstract presents 18%/27% as the main positive result. Please provide per-subject values, a pooled estimate with uncertainty (e.g., bootstrap or a t-interval), and a clear description of how the ROI was defined and whether the interleaved SNR maps were averaged before or after computing the percentage change.","section":"Results; In vivo DMI studies (Figure 5)"},{"comment":"The SNR maps are scaled to SNR units using the Euclidean norm of the complex linear combination coefficients of the individual metabolites during spectral separation. This norm is method-dependent: it is derived from the standard IDEAL separation for CSI-FISP and from the IDEAL-modes or linear-fit procedures for the bSSFP data. Since the intended comparison is acquisition-related SNR, the manuscript should state explicitly whether the same scaling convention is applied to all three methods and quantify how sensitive the 18%/27% numbers are to the choice of scaling, for example by recomputing with a fixed theoretical weighting rather than the eigenvector-derived norm. Without this, part of the reported gain could be a reconstruction-scaling effect rather than a true SNR improvement.","section":"SNR calculation"},{"comment":"The simulations predict a 45% SNR increase for glucose and 69% for Glx for bSSFP relative to FISP, whereas the measured values are 18% and 27%. The discussion offers plausible explanations (FISP already near-optimal, phase-cycling stopband loss, SAR-limited flip angles), but the gap is large and no quantitative decomposition is provided. Because the headline result is a measured gain, it would strengthen the paper to show a revised simulation that includes the actual achieved flip angles (stated as 85–90% of nominal), the phase-cycle averaging penalty, and the measured B0 distribution, and to compare that revised prediction with the observed gains.","section":"Simulation; Figure 2"}],"minor_comments":[{"comment":"The FOV entry for CSI-bSSFP is listed as '20x20x200' and appears to be a typo for '20x20x20'; please correct it.","section":"Table 1"},{"comment":"The phrase 'to improve maximize ADC duty cycle' is ungrammatical; it should read 'to improve ADC duty cycle' or 'to maximize ADC duty cycle'.","section":"Methods; DMI protocols"},{"comment":"The citation to Nguyen and Bieri is misspelled as 'Nyugen' in the text, and Reference 16 should read 'Oppelt' rather than 'Opplet'.","section":"References and text"},{"comment":"The sentence 'The bi-exponential T1 relaxation of in vivo water was not apparent in the inversion recovery data' is confusing because the following sentence and Table 2 concern bi-exponential T2 relaxation; please rephrase to avoid ambiguity.","section":"Discussion"},{"comment":"The claim that ME-bSSFP Glx maps 'consistently predict higher concentrations' than the CSI methods is presented without a cross-method calibration phantom; if this is intended as a resolution/partial-volume effect, a brief quantitative justification would help.","section":"Figure 8"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the eigenvector-based SNR combination. If the authors can demonstrate that the 18%/27% gains survive a split-half or theoretical-weight reanalysis, I would support publication; if not, the abstract should be revised to present the gains as preliminary and to emphasize the feasibility and qualitative benefits of phase-cycled bSSFP. The paper is otherwise within MRM scope, and the open data/code sharing is commendable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper for two reasons: it is the first demonstration of bSSFP deuterium metabolic imaging in the human brain at 9.4 T, and it ships open code and data, which is still rare enough in this field to matter. The authors also measure in vivo T1/T2 of deuterated metabolites and propose two new fitting methods for phase-cycled bSSFP data. That is the real contribution, and it is done carefully.\n\nThe paper is honest about its limits. They report that the measured SNR gains (18% glucose, 27% Glx) are far below the simulated 45%/69%, they show ME-bSSFP underperforming, and they flag B0 artifacts in one subject. The phantom and in vivo comparisons are clearly described, with PSF normalization and SNR-units reconstruction. I buy the feasibility claim.\n\nWhere I push back is on the quantitative SNR comparison. The stress-test note is on target: the IDEAL-modes method estimates the complex weights for combining SSFP modes by taking the principal eigenvector from the fitted amplitudes of the same noisy data. That is a data-driven SNR optimization, and it is not applied to the CSI-FISP reference, which uses standard IDEAL. So part of the reported gain could be noise fitting, especially for low-SNR metabolites like glucose and Glx. The authors even show IDEAL-modes beating the linear fit in Figure 6, which is consistent with some overfitting. This does not kill the paper, but it means the 18%/27% numbers should be treated as an upper bound until confirmed with a split-half analysis or theoretical weights. The n=3 sample with no confidence intervals makes this more than a minor caveat.\n\nOther soft spots are minor: the PSF scaling factors are reasonable, the water reference assumption is acknowledged, and the comparison protocols are matched in acquisition time. The citation pattern is fine, with proper credit to the preclinical bSSFP DMI work.\n\nWho is this for? Anyone working on DMI pulse sequence development or ultra-high-field metabolic imaging will get value from the data, the fitting methods, and the honest reporting. It deserves a serious referee; I would send it to MRM or NMR in Biomed with a request to address the eigenvector overfitting concern and report subject-level variability.\n\nMy recommendation: engage with it, cite it if you work in this area, and treat the SNR gain as provisional.","headline":"A solid feasibility study showing bSSFP DMI works in the human brain at 9.4 T, but the headline 18%/27% SNR gain is likely optimistic because the IDEAL-modes combination step can overfit noise.","tokens_in":15412,"tokens_out":1812,"would_cite":true,"duration_ms":19371,"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":"Phase-cycled balanced SSFP acquisitions are feasible for high-resolution deuterium metabolic imaging of the human brain at 9.4 T, with CSI-bSSFP improving glucose and Glx SNR by 18% and 27% over standard gradient-spoiled CSI with IDEAL…","keywords":["deuterium metabolic imaging","balanced SSFP","phase cycling","ultra-high-field MRI","SNR","IDEAL","glucose metabolism","human brain"],"falsifier":"Reanalyze the same raw data with an independent point-spread-function estimation and with matched nominal resolutions, or run the three protocols on a phantom with known deuterium concentrations over a range of B0 offsets; if CSI-bSSFP no longer shows a resolution-normalized SNR gain over CSI-FISP in a larger subject cohort, the central claim is not supported.","tokens_in":14455,"feed_emoji":"🧠","tokens_out":8388,"duration_ms":73050,"temperature":0.7,"pith_summary":"Deuterium metabolic imaging (DMI) tracks glucose uptake and downstream metabolism after oral intake of deuterated glucose, but its low signal has kept spatial resolution coarse. This paper tries to break that limit with balanced steady-state free precession (bSSFP), a steady-state sequence that refocuses magnetization every repetition and can give more signal per unit time for metabolites with favorable T2/T1. The authors show that phase-cycled CSI-bSSFP and ME-bSSFP both yield whole-brain dynamic quantitative DMI maps at 9.4 T in 10 minutes, with nominal voxels of 0.58 mL for the CSI variants. The central quantitative result is an average SNR gain of 18% for glucose and 27% for Glx (glutamate+glutamine) for CSI-bSSFP relative to the vendor's gradient-spoiled CSI-FISP with IDEAL processing. ME-bSSFP did not improve SNR but delivered a true 2.18 mL voxel volume and finer spatial detail, including resolved skull uptake.","feed_headline":"Balanced SSFP lifts deuterium glucose SNR 18-27% in human brain","feed_subtitle":"9.4 T study shows acquisition-weighted CSI-bSSFP beats standard CSI for glucose and Glx in 10-minute scans.","key_machinery":"The working object is phase-cycled balanced steady-state free precession (bSSFP), a steady-state sequence in which transverse magnetization is fully refocused each TR and the RF phase is stepped across a cycle; the phase cycling removes off-resonance banding and adds a spectral-encoding dimension. The IDEAL-modes fit separates the bSSFP configuration modes by a discrete Fourier transform over phase cycles and combines mode amplitudes with weights from a principal-eigenvector analysis, while the linear fit matches the data to the full analytical bSSFP signal equation including T1, T2, off-resonance, flip angle, and RF phase increment. The comparison is normalized to resolution by SNR-units reconstruction and point-spread-function voxel-volume scaling factors of 0.96 and 1.31 for CSI-bSSFP and ME-bSSFP, respectively.","core_discovery":"On its own terms, the paper establishes that bSSFP acquisitions with phase cycling are a viable sensitivity route for deuterium metabolic imaging of the human brain at 9.4 T. After oral [6,6'-2H2]-glucose intake, both CSI-bSSFP (four phase cycles, Hamming-weighted) and ME-bSSFP (eighteen phase cycles, elliptical scanning) produced high-resolution whole-brain maps of water, glucose, Glx, and lactate/lipids with no visible banding artifacts. Compared with the standard gradient-spoiled CSI-FISP plus IDEAL, CSI-bSSFP raised average SNR by 18% for glucose and 27% for Glx, while ME-bSSFP showed lower SNR but a true resolution close to its nominal 2 mL. Two proposed fitting methods, IDEAL-modes and linear fit, handle the amplitude and phase modulations introduced by phase cycling, and the authors demonstrate that phase cycling itself contributes extra spectral encoding that can separate all four deuterium resonances even with fewer echoes.","pith_inferences":["My inference: if the 18%/27% SNR gains reproduce in a larger cohort with confidence intervals, CSI-bSSFP could become the default DMI readout at ultra-high field, with the gain spent on reducing scan duration.","My inference: because the IDEAL-modes method only assumes chemical shift and B0, it should transfer to other field strengths and tissues with unknown relaxation times; a direct test would be comparing it with the linear fit in a tumor model.","My inference: the SNR comparison hinges on the PSF scaling factors 0.96 and 1.31; recomputing the comparison with an independent PSF measurement or matched nominal resolution would isolate how much of the reported gain is sequence physics versus normalization.","My inference: the additional spectral encoding from phase cycling implies TR can be shortened below the value needed for conventional spectral separation; this can be tested by reducing echo spacing while keeping four phase cycles and checking conditioning."],"forward_implications":["DMI at 9.4 T can be done at 0.58 mL nominal voxel size in 10 minutes, about five times finer than earlier 3 mL studies at the same site.","CSI-bSSFP with phase cycling should give 18% better glucose and 27% better Glx SNR than standard CSI-FISP, gains that could be spent on smaller voxels, lower tracer dose, or shorter exams.","ME-bSSFP, despite no SNR benefit, provides actual resolution near its nominal 2 mL and can reveal spatially detailed features such as skull glucose uptake.","Phase cycling supplies enough extra spectral encoding to separate all four deuterium resonances with fewer than four echoes, which could shorten TR in future protocol designs."],"supporting_citations":[{"why":"Prior preclinical bSSFP DMI work at 15.2 T that this paper extends to the human brain at 9.4 T.","marker":"13,14"},{"why":"Provides the bSSFP signal model and the T2/T1 rationale predicting SNR gains for deuterated metabolites.","marker":"15"},{"why":"Supplies the analytical bSSFP signal equation with RF phase cycling used in the linear fit method.","marker":"19"},{"why":"Supplies the SSFP configuration-space equation that the IDEAL-modes fit extends to multiple metabolites.","marker":"25"},{"why":"The IDEAL algorithm used for spectral separation of CSI-FISP data and for estimating the B0 off-resonance map.","marker":"24"},{"why":"Defines SNR-units reconstruction, the normalization that makes the reported SNR comparison fair across protocols.","marker":"27"},{"why":"Prior 9.4 T human DMI study at 3 mL nominal resolution that provides the resolution baseline improved here.","marker":"9"},{"why":"Establishes DMI methodology and the assumed 10.12 mM water reference used for metabolite quantification.","marker":"2"}],"fun_headline_variants":["CSI-bSSFP lifts deuterium glucose SNR 18-27% in brain at 9.4 T","Phase-cycled bSSFP boosts deuterium brain imaging SNR","ME-bSSFP reaches 2 mL resolution in deuterium brain DMI","bSSFP CSI variant outdoes standard CSI for brain glucose and Glx"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fairness of the headline SNR comparison depends on the point-spread-function voxel-volume scaling (factors 0.96 and 1.31) and the SNR-units normalization; if those corrections are biased, the reported 18 and 27 percent gains for CSI-bSSFP could change, and the in vivo comparison used only three subjects with no confidence intervals.","fun_headline_variants_meta":{"raw":{"variants":["CSI-bSSFP lifts deuterium glucose SNR 18-27% in brain at 9.4 T","Phase-cycled bSSFP boosts deuterium brain imaging SNR","ME-bSSFP reaches 2 mL resolution in deuterium brain DMI","bSSFP CSI variant outdoes standard CSI for brain glucose and Glx"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000907,"raw_usage":{"total_tokens":3956,"prompt_tokens":1059,"completion_tokens":2897,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":2809}},"tokens_in":675,"tokens_out":2897,"duration_ms":22931,"temperature":1.0,"reasoning_tokens":2809,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T22:58:13.758974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reanalyze the same raw data with an independent point-spread-function estimation and with matched nominal resolutions, or run the three protocols on a phantom with known deuterium concentrations over a range of B0 offsets; if CSI-bSSFP no longer shows a resolution-normalized SNR gain over CSI-FISP in a larger subject cohort, the central claim is not supported.","supporting_citations":[{"cited_title":"Fundamentals of balanced steady state free precession MRI","cited_arxiv_id":null,"evidence_quote":"Provides the bSSFP signal model and the T2/T1 rationale predicting SNR gains for deuterated metabolites."},{"cited_title":"Static susceptibility effects in balanced SSFP sequences","cited_arxiv_id":null,"evidence_quote":"Supplies the analytical bSSFP signal equation with RF phase cycling used in the linear fit method."},{"cited_title":"Motion -insensitive rapid configuration relaxometry","cited_arxiv_id":null,"evidence_quote":"Supplies the SSFP configuration-space equation that the IDEAL-modes fit extends to multiple metabolites."},{"cited_title":"Multicoil Dixon chemical species separation with an iterative least- squares estimation method","cited_arxiv_id":null,"evidence_quote":"The IDEAL algorithm used for spectral separation of CSI-FISP data and for estimating the B0 off-resonance map."},{"cited_title":"Deuterium metabolic imaging in the human brain at 9.4 Tesla with high spatial and temporal resolution","cited_arxiv_id":null,"evidence_quote":"Prior 9.4 T human DMI study at 3 mL nominal resolution that provides the resolution baseline improved here."},{"cited_title":"Deuterium metabolic imaging (DMI) for MRI-based 3D mapping of metabolism in vivo","cited_arxiv_id":null,"evidence_quote":"Establishes DMI methodology and the assumed 10.12 mM water reference used for metabolite quantification."}],"review_version":1}