{"id":"fa076a8f-8f3b-46b4-babe-49f6d91d2a5f","arxiv_id":"2506.20614","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"WMF, the energy-weighted mean frequency of the temporal Fourier transform of 4D Flow MRI velocity, separates pulsatile flow from background better than PC-MRA and improves aorta segmentation.","lead":"The authors propose Weighted Mean Frequencies (WMF), a handcrafted MRI feature built from the temporal Fourier transform of velocity, to highlight voxels touched by pulsatile blood flow. In their experiments WMF improved aorta segmentation over the standard PC-MRA feature, though the deep learning comparison rests on only six test patients.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The spectral-pulsatility separation is not quantitatively validated, and the minimum over velocity components is an order statistic biased toward low values, so non-fluid voxels with any low-frequency noise (drift, motion) may be misclassified as fluid.","rationale":"The paper's contribution is a handcrafted feature whose design rationale is the spectral-pulsatility model of Section 2. The entire interpretation, and any expectation of generalization, depends on that model. The reader's weakest assumption correctly identifies this. Our stress-test sharpens it: the minimum operator (Eq. 2) is not just a heuristic merge; it is an order statistic that is systematically biased toward low values. In non-fluid voxels, any component with low-frequency noise (e.g., respiratory motion, B0 drift, eddy currents) will be selected by the min and may yield a WMFmin in the fluid range. The paper's evidence for the separation is qualitative (two background points in Fig. 1f) and does not establish that the WMFmin distributions of fluid and non-fluid voxels are separated across the dataset's heterogeneity (VENC 1.5-6.0, isotropic/non-isotropic, 1.5T/3T). Thus the central claim's first part is not yet supported. The second part, the segmentation gains, is real in-sample, but the threshold protocol is oracle-based and the deep-learning test set has only six volumes; these issues are secondary to the mechanism. The proposed test, measuring distribution overlap on the existing dataset, can directly settle whether the separation holds. If the overlap is small, the premise is validated and the verdict could be upgraded; if large, the feature's empirical gains need a different explanation and the verdict should be tightened. Since the test is not yet performed, CONDITIONAL remains the correct verdict.","tokens_in":8212,"tokens_out":6236,"duration_ms":71590,"concrete_test":"On the 43 subjects, use the reference segmentations to label aorta-lumen voxels and a set of non-fluid voxels (e.g., lung, chest wall, air background, and static tissue with high anatomical signal). For each voxel compute WMFu, WMFv, WMFw and WMFmin. Report the histograms and the overlap: the fraction of non-fluid voxels with WMFmin below the 10th percentile of the fluid WMFmin distribution. If this overlap exceeds, say, 5%, the spectral separation premise fails. Additionally, repeat for the subset with VENC > 4.0 m/s and for the non-isotropic acquisitions to test the claimed robustness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the spectral model of Section 2: fluid voxels have a dominant low-frequency cardiac peak, non-fluid voxels have a flat noise spectrum, and the min over components (Eq. 2) preserves separation. This is supported only by four qualitative point examples (Fig. 1f). The minimum operator is a first-order statistic: in a non-fluid voxel, WMFmin is the smallest of three noisy WMF estimates. Even if each component's noise spectrum is flat individually, the minimum of three independent estimates is biased downward; the expected minimum falls below the median of any single component. More importantly, if any component in a static-tissue voxel contains low-frequency phase variation (respiratory motion, eddy currents, B0 drift), WMF for that component will be low, and the min will select it, producing a false positive 'pulsatile' label. The paper's claim that 'the non-fluid domain has been found to contain the highest WMF values' is based on two points (#3, #4) and does not quantify the overlap of the WMFmin distributions across the 43-subject heterogeneous dataset. If background voxels with low WMFmin are common, the threshold results (Table 1) reflect oracle-threshold fitting rather than a true separation, and the deep-learning gains may not transfer.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Weighted Mean Frequencies (WMF), a handcrafted feature for 4D Flow MRI segmentation. For each velocity component, WMF is the energy-weighted mean of strictly positive temporal Fourier frequencies (Eq. 1); the three component maps are combined by a per-voxel minimum (Eq. 2). The authors argue that low WMF values mark voxels that experienced pulsatile flow, while high values mark background, producing a 'hull' of pulsatile velocity voxels. They evaluate WMF in two settings: an optimal-threshold segmentation task on 43 patients and a 3D U-Net segmentation task on a test set of 6 manually selected volumes. They report that WMF alone or as an input improves IoU and Dice relative to PC-MRA-based inputs, with a 0.12/0.13 gain claimed in the deep learning task.","tokens_in":8415,"tokens_out":2696,"duration_ms":32417,"significance":"If the claims hold, WMF is a simple, parameter-free, and interpretable feature for 4D Flow MRI that could help both manual and automatic aortic segmentation, especially in diastolic frames where PC-MRA degrades. The definition in Eq. (1) is clear and requires no training or label fitting, and the threshold experiment on all 43 patients provides some evidence of discriminative power beyond anecdotal examples. However, the reported gains are currently supported by limited quantitative validation: the threshold experiment uses an oracle threshold optimized on ground truth, the deep learning test set consists of six manually selected volumes with no statistical testing, and the central spectral-separation assumption is only illustrated qualitatively. With a larger evaluation or explicit distributional validation, the feature could be a useful contribution to the segmentation toolbox.","major_comments":[{"comment":"The central claim that 'the non-fluid domain has been found to contain the highest WMF values' is supported only by four qualitative point examples (Figure 1f). This claim is load-bearing for both experiments, but the manuscript does not report the distributions of WMF_min in fluid versus non-fluid voxels across the 43-subject dataset. Because the minimum over three components is an order statistic biased toward low values, even a flat noise spectrum can produce low WMF_min in background voxels, and low-frequency phase artifacts in any single component would be selected by the min. Please provide a quantitative analysis of the WMF_min distributions (e.g., histograms, separability measures, or voxel-wise ROC curves) on the full dataset to demonstrate that the separation is not an artifact of the min operator.","section":"Section 2, Eq. (2)"},{"comment":"The threshold segmentation experiment optimizes the threshold to maximize IoU against the ground truth for each MRI and each feature. This is an oracle threshold and not a practical segmentation method; the reported IoU values are therefore upper bounds rather than achievable performance. The 2-fold improvement over PC-MRA(t) (0.262 vs. 0.129) is still informative about feature separability, but the manuscript should state this limitation explicitly and, if possible, add a train/test split or a fixed-threshold experiment to show how much of the advantage survives without oracle selection.","section":"Section 3.1, Table 1"},{"comment":"The deep learning claim rests on a test set of only six manually selected volumes (Section 4.1), with no confidence intervals, paired significance tests, or analysis of sensitivity to the manual selection. The differences between rows (e.g., row 3 vs. row 5: IoU 0.671 vs. 0.665; row 6 having std 0.269) are within the observed variability, so the headline 'increase of 0.12 and 0.13' cannot be distinguished from noise. Please report per-subject results, effect sizes with confidence intervals, and at least a paired permutation test for the key comparisons (WMF vs. PC-MRA inputs).","section":"Section 4.2, Table 2"},{"comment":"The spectral-pulsatility separation - fluid voxels have a dominant low-frequency cardiac peak while non-fluid voxels have a flat spectrum - is assumed rather than validated. The dataset is heterogeneous (Section 3.1: 1.5T/3T, VENC 1.5-6.0 m/s, isotropic and non-isotropic resolutions), and acquisitions with low temporal resolution or high VENC may not resolve the cardiac peak; in such cases WMF_min would not mark the pulsatile hull. Please quantify the frequency resolution and the spectral shape across the dataset, or at least state the acquisition parameters (number of cardiac phases, temporal resolution) and show that the separation holds for representative low-temporal-resolution cases.","section":"Section 2, spectral model"}],"minor_comments":[{"comment":"The sentence 'This feature, termed Weighted Mean Frequencies (WMF), is capable of revealing the region in three dimensions where a voxel has been passed by pulsatile flow' is grammatically awkward; consider revising to 'which voxels have been traversed by pulsatile flow.'","section":"Abstract"},{"comment":"The notation E_i(u_j) is used for |FT(u_j)|^2 but Eq. (1) writes E_i(u_j); please define the index i (the i-th frequency bin) before its first use, and clarify whether the denominator is the sum over all strictly positive frequencies of the same component's spectrum.","section":"Section 2, Eq. (1)"},{"comment":"In Figure 1f, the text refers to 'the w-velocity component' and its WMF_w, but the curves are labeled u, v, w without a legend; please add a legend and label the subfigures more clearly. Also, the phrase 'an patient' in the figure caption should be 'a patient.'","section":"Section 2, Figure 1"},{"comment":"The table rows appear misnumbered: the row for 'Mag(t)*(1-WMF)' is labeled 8, and there is no row 8 elsewhere; please renumber the rows sequentially. Also, the statement that WMF 'exhibits the best performance' is true for IoU/Dice, but the recall of WMF (0.569) is actually among the lowest in the table; please reconcile or clarify why recall is not emphasized.","section":"Section 3.2, Table 1"},{"comment":"The manual selection of test and validation sets to 'represent the variety of data' is a potential source of selection bias; please describe the selection criteria more concretely and state whether the same six test volumes were used for all input combinations, and whether any hyperparameter tuning was performed on the test set.","section":"Section 4.1"},{"comment":"The dataset is labeled by 'non-expert researchers.' Since the ground truth is the reference for both experiments, please provide some assessment of label quality or inter-rater variability, even a small pilot study, to contextualize the reported IoU/Dice values.","section":"Section 3.1"},{"comment":"The discussion notes that 'the distribution of fluid domain velocity noise is dependent on anatomical SNR' and that WMF sensitivity to SNR 'requires confirmation using a larger dataset.' This is an important caveat; please move this point earlier or at least emphasize it in the conclusion, since it directly limits the generalizability claim.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a concise, useful proposal, and the WMF definition is refreshingly simple and reproducible. The main gap is quantitative validation: the threshold experiment is oracle-based and the deep learning experiment is underpowered. I would be supportive if the authors add distributional validation of WMF_min separation across the full dataset, a non-oracle threshold evaluation, and per-subject statistics for the deep learning comparison. I do not see a fundamental error in Eq. (1); the min operator in Eq. (2) is the main theoretical risk and should be examined directly. The fit to the journal's scope (medical image analysis / eess.IV) is good, though the manuscript is closer to a technical note than a full paper in its current depth."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The Weighted Mean Frequencies feature is exactly what it sounds like: the energy-weighted mean of positive temporal frequencies from the velocity time series, combined by taking the minimum across the three flow-encoding directions. That's a clean, parameter-free definition, and the min-combination is a sensible way to get a single pulsatility map. The threshold experiment on 43 patients is encouraging—WMF roughly doubles the IoU of PC-MRA(t)—and the paper honestly flags the small test set and the need for larger validation. So the core idea is worth taking seriously.\n\nThe soft spots are real, though. The threshold experiment uses an oracle threshold optimized on ground truth, which inflates the numbers and doesn't reflect how anyone would use this in practice. The deep-learning gain of 0.12 IoU comes from a single split with only six test volumes, and the standard deviations overlap, so that result is fragile. More substantively, the stress-test's concern about the min operator is legitimate: the minimum of three noisy WMF estimates is biased low, so any non-fluid voxel with low-frequency noise (respiratory motion, eddy currents, B0 drift) could be misclassified as pulsatile. The paper's evidence for the spectral separation is four point examples, not a distribution-overlap analysis across the 43-patient dataset. That gap matters because the whole method rests on that separation. No code or data is released, which makes independent verification harder.\n\nThe stress-test's worry is theoretically sound, but the empirical consistency in the threshold experiment gives me some confidence that the separation is real, at least for this dataset and acquisition range. The paper would be much stronger with a direct plot of WMF distributions in fluid vs. background voxels, a non-oracle threshold protocol, and a larger deep-learning test set.\n\nThis paper is for researchers in cardiovascular MRI who want a cheap, handcrafted feature to add to segmentation pipelines or a pulsatility map for clinical review. It deserves peer review: the idea is plausible, the experiments are honestly reported, and a serious referee can push for the missing analyses. I'd send it out, expecting major revision, not a desk reject.","headline":"WMF is a simple, parameter-free spectral feature that plausibly helps 4D Flow MRI aorta segmentation, but the current evidence is limited by oracle thresholds and a small deep-learning test set.","tokens_in":9009,"tokens_out":2417,"would_cite":false,"duration_ms":30070,"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":"This paper proposes Weighted Mean Frequencies (WMF), a Fourier feature that marks voxels traversed by pulsatile flow and improves 4D Flow MRI aorta segmentation over PC-MRA.","keywords":["4D Flow MRI","Weighted Mean Frequencies","aorta segmentation","handcrafted feature","Fourier transform","PC-MRA","threshold segmentation","deep learning segmentation"],"falsifier":"Acquire a 4D Flow MRI volume of a straight vessel with the flow nearly orthogonal to all three encoding axes, or with fewer than about eight cardiac phases and a high VENC, and compute the WMF distributions inside the lumen and in surrounding tissue; if the two distributions overlap so that no threshold separates them, the claimed separation is false.","tokens_in":7936,"feed_emoji":"🫀","tokens_out":7837,"duration_ms":70538,"temperature":0.7,"pith_summary":"Blood-flow MRI records velocity in every volume voxel over the cardiac cycle. This paper proposes Weighted Mean Frequencies (WMF), computed by taking the energy-weighted average of the temporal frequencies of each velocity component and then taking the minimum across the three components. The claim is that this single number separates voxels swept by pulsatile blood from background: fluid voxels concentrate energy at the low cardiac frequency, so they get low WMF values, while non-fluid voxels have flatter spectra and get high values. If that is true, WMF gives radiologists and algorithms a contrast that does not depend on anatomy or on a single systolic time frame. The paper reports that adding WMF as a U-Net input raises aortic segmentation IoU by 0.12 and Dice by 0.13 over the standard PC-MRA feature, and that a simple threshold on WMF roughly doubles the IoU of thresholded PC-MRA computed per time frame.","feed_headline":"One Fourier feature lifts 4D Flow MRI aorta segmentation by 0.12 IoU","feed_subtitle":"The Weighted Mean Frequencies image stays stable in diastole, where PC-MRA loses contrast and segmentation degrades.","key_machinery":"The load-bearing object is the energy-weighted mean frequency (WMF) defined in Eq. (1), together with the minimum-over-components fusion rule of Eq. (2). Per voxel and per velocity component, WMF takes the temporal Fourier transform of the velocity waveform, weights each strictly positive frequency by its energy, and averages; the min across the $u$, $v$, $w$ components keeps the component that is most flow-sensitive at that location and discards components where the vessel orientation or noise produces high frequencies. This makes WMF a one-pass, parameter-free, anatomy-independent image whose low values delineate the pulsatile lumen.","core_discovery":"On the paper's own terms, the central claim is that the Weighted Mean Frequencies image, defined by $$\\mathrm{WMF}(u_j)=\\frac{\\sum_i E_i(u_j) f_i}{\\sum_i E_i(u_j)}$$ with $E_i(u_j)=|\\mathrm{FT}(u_j)|^2$ for each velocity component and then combined as $\\mathrm{WMF}_{\\min}=\\min(\\mathrm{WMF}_u,\\mathrm{WMF}_v,\\mathrm{WMF}_w)$, reveals the region through which pulsatile blood has passed. Fluid voxels show a dominant low-frequency cardiac peak and therefore low WMF values; non-fluid voxels show flat noise spectra and high WMF values, whether or not they sit in a bright anatomical region. The paper argues this separation is what makes WMF useful: it beats the standard PC-MRA feature in both a threshold segmentation and a deep-learning segmentation of the aorta, with the deep-learning task reporting IoU and Dice gains of 0.12 and 0.13.","pith_inferences":["Inference: WMF could be tested directly as a pulsatility biomarker, independent of segmentation, since its per-voxel value is a continuous index of spectral content rather than a binary mask.","Inference: The minimum-over-components rule means a vessel whose flow is almost orthogonal to all three encoding axes could produce no low-WMF component; a rotation-invariant alternative would be to combine components by low-frequency energy fraction instead of the minimum.","Inference: The claimed SNR robustness should be re-checked on a multi-manufacturer, multi-site dataset with more than six test volumes; the paper itself flags this as needed.","Inference: In principle the same construction applies to any pulsatile vascular territory, including heart and brain, but the paper only evaluates the aorta; a direct test in the carotid or cerebral circulation would show whether the spectral separation generalizes."],"forward_implications":["Adding WMF as an input channel to a 3D U-Net raises aortic segmentation IoU by 0.12 and Dice by 0.13 on the paper's test set compared with the PC-MRA feature.","A single threshold on WMF more than doubles the IoU of per-time-frame PC-MRA thresholding (0.262 vs 0.129), so WMF alone can support fast interactive or automatic aorta outlining.","Because WMF values stay stable between systole and diastole while PC-MRA(t) Dice drops from 0.327 to 0.129, WMF can carry diastolic segmentation where PC-MRA loses contrast.","WMF can be combined with PC-MRA rather than replacing it; the best threshold combination and the deep-learning combination both include both features.","WMF's weaker dependence on anatomical signal makes it less sensitive to contrast variations, as shown by the smaller mean-median gap in the deep-learning results."],"supporting_citations":[{"why":"Defines 4D Flow MRI and the velocity-field data WMF is computed from.","marker":"[11]"},{"why":"Supplies the PC-MRA implementation used as the baseline in both segmentation experiments.","marker":"[4]"},{"why":"Introduces three-dimensional phase-contrast angiography, the basis of PC-MRA that WMF is compared against.","marker":"[5]"},{"why":"Describes three-direction phase-contrast encoding, grounding the per-component velocity analysis and noise behavior behind the minimum operator.","marker":"[12]"},{"why":"Supplies the U-Net architecture used in the deep learning segmentation experiment.","marker":"[14]"},{"why":"Supplies the labeling tool used to create the ground-truth aortic annotations.","marker":"[16]"},{"why":"Shows how contrast agents affect PC-MRA and flow visualization, motivating an anatomy-independent feature.","marker":"[2]"}],"fun_headline_variants":["WMF feature boosts 4D Flow MRI aorta segmentation by 0.12 IoU","New Fourier feature lifts MRI flow segmentation, +0.12 IoU","Pulsatile flow hull feature improves 4D Flow MRI segmentation","Weighted Mean Frequencies: better aorta segmentation in MRI","Aorta segmentation gains 0.12 IoU with WMF handcrafted feature"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that in every fluid voxel at least one of the three velocity components carries a strong low-frequency pulsatile peak, while non-fluid voxels have a flat spectrum; if an acquisition, vessel orientation, or noise level breaks that contrast, WMF no longer separates lumen from background.","fun_headline_variants_meta":{"raw":{"variants":["WMF feature boosts 4D Flow MRI aorta segmentation by 0.12 IoU","New Fourier feature lifts MRI flow segmentation, +0.12 IoU","Pulsatile flow hull feature improves 4D Flow MRI segmentation","Weighted Mean Frequencies: better aorta segmentation in MRI","Aorta segmentation gains 0.12 IoU with WMF handcrafted feature"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000679,"raw_usage":{"total_tokens":3129,"prompt_tokens":1032,"completion_tokens":2097,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":1997}},"tokens_in":648,"tokens_out":2097,"duration_ms":16846,"temperature":1.0,"reasoning_tokens":1997,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:44:52.097647+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire a 4D Flow MRI volume of a straight vessel with the flow nearly orthogonal to all three encoding axes, or with fewer than about eight cardiac phases and a high VENC, and compute the WMF distributions inside the lumen and in surrounding tissue; if the two distributions overlap so that no threshold separates them, the claimed separation is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines 4D Flow MRI and the velocity-field data WMF is computed from."},{"cited_title":"https://doi.org/10.1016/j.media.2018.08.003","cited_arxiv_id":null,"evidence_quote":"Supplies the PC-MRA implementation used as the baseline in both segmentation experiments."},{"cited_title":"Magnetic Resonance in Medicine9(1), 139–149 (1989)","cited_arxiv_id":null,"evidence_quote":"Introduces three-dimensional phase-contrast angiography, the basis of PC-MRA that WMF is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes three-direction phase-contrast encoding, grounding the per-component velocity analysis and noise behavior behind the minimum operator."},{"cited_title":"Magnetic Resonance in Medicine63(2), 330–338 (2010)","cited_arxiv_id":null,"evidence_quote":"Shows how contrast agents affect PC-MRA and flow visualization, motivating an anatomy-independent feature."}],"review_version":1}