Pith. sign in

REVIEW 4 major objections 7 minor 16 references

Weighted Mean Frequencies: a handcraft Fourier feature for 4D Flow MRI segmentation

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2506.20614 v1 pith:DJNHWEAH submitted 2025-06-25 eess.IV cs.AIcs.CV

Simon Perrin , Sébastien Levilly , Huajun Sun , Harold Mouchère , Jean-Michel Serfaty This is my paper · ORCID
classification eess.IVcs.AIcs.CV
keywords 4DFlowMRIWeightedMeanFrequenciesaortasegmentationhandcraftedfeatureFouriertransformPC-MRAthresholddeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

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.

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 (4)
  1. [Section 2, Eq. (2)] 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.
  2. [Section 3.1, Table 1] 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.
  3. [Section 4.2, Table 2] 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).
  4. [Section 2, spectral model] 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.
minor comments (7)
  1. [Abstract] 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.'
  2. [Section 2, Eq. (1)] 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.
  3. [Section 2, Figure 1] 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.'
  4. [Section 3.2, Table 1] 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.
  5. [Section 4.1] 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.
  6. [Section 3.1] 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.
  7. [Section 5] 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.

Circularity Check

1 steps flagged · score 4.0 of 10

Threshold-segmentation results are oracle-optimized on ground truth; WMF definition and deep-learning evaluation remain independent.

  1. fitted input called prediction [Section 3.1 (Method), Section 3.2 (Results), Table 1]
    "For each MRI in the dataset and each combination, the threshold is optimised by seeking the best intersection over union (IoU) between the ground truth and the resulting segmentation."

    The threshold is a fitted parameter chosen on the same ground-truth labels used to compute the reported IoU/Dice. Table 1's WMF IoU (0.262) is therefore max_t IoU(GT, 1-WMF > t), an in-sample optimum of the evaluation metric rather than an out-of-sample prediction. The reported 'factor of 2' improvement over PC-MRA(t) is partly a property of the threshold search itself, not solely of WMF. This does not affect the WMF definition (Eqs. 1-2 are parameter-free) and does not apply to the held-out deep-learning test (Table 2), so the circularity is confined to the threshold evaluation protocol.

full rationale

The WMF feature is defined parameter-free from the velocity Fourier spectra (Eq. 1) with a fixed min operator (Eq. 2); no parameter of WMF is fitted to segmentation labels. The central deep-learning comparison (Table 2) is evaluated on a separate test set and is independent of any fitted threshold, giving the central claim genuine predictive content. The only circular step is the threshold segmentation experiment: the per-image threshold is optimized to maximize IoU on the same ground truth used for evaluation, so Table 1 reports an oracle fit rather than a prediction. This inflates absolute numbers but applies symmetrically to all compared features. Self-citations are not load-bearing; reference [7] is used only as background for SNR effects and does not justify WMF's discriminative claim. Overall, the feature derivation is self-contained, with partial circularity in one secondary evaluation.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

WMF itself is a derived image feature, not a new physical entity, so the invented-entities ledger is empty. The central claim rests on the spectral-pulsatility signal model, the min-combination heuristic, the small manually selected test set, the label quality, and the temporal sampling of the MRI acquisition, all of which are assumptions the reader must accept from the paper's description.

free parameters (3)
  • Per-MRI optimal threshold = Variable, step size 0.02, chosen per MRI to maximize IoU
    In Section 3.1 the threshold is optimized against ground truth for each feature and each image, so the threshold segmentation results are upper-bound discriminability scores rather than deployable decision rules. This parameter does not define WMF but directly affects the reported threshold numbers.
  • PC-MRA gamma = 0.2
    The gamma value for the PC-MRA implementation is taken from Bustamante et al. (2018) and is not fitted in this paper. It affects the PC-MRA baseline images.
  • U-Net training hyperparameters = Channels [64,128,256,512], patch 48x48x24, initial lr 0.05, momentum 0.9, weight decay 1e-4
    These values are chosen by hand for the deep learning experiment. They are not fitted to the test set, but the reported deep learning gains depend on this particular configuration and training setup.
assumptions (5)
  • domain assumption The energy-weighted mean of positive temporal frequencies is a valid pulsatility indicator: fluid voxels have a dominant low-frequency cardiac component and non-fluid voxels have noise-dominated flat spectra.
    Section 2 and Figure 1f support this signal model qualitatively through points #1-#4. If this spectral separation does not hold in some acquisition conditions, WMF would not mark the flow region.
  • ad hoc to paper The minimum across the three velocity components preserves the pulsatility signal: at every fluid voxel at least one velocity component has a detectable low-frequency pulsatile peak.
    Equation (2) defines WMFmin to handle components with no flow; the paper observes that WMFu is high at the top of the aortic arch and relies on the other components to keep the minimum low.
  • domain assumption The manually selected test set of six MRIs is representative of the heterogeneous acquisition conditions (field strength, VENC, plane, contrast).
    Section 4.1 states selection was manual with stratified sampling. With only six volumes, this representativeness assumption is load-bearing for the deep learning comparison.
  • domain assumption The retrospective labels produced by non-expert researchers are accurate enough to support relative feature comparisons.
    Section 3.1 describes labeling with ITK-SNAP by non-expert researchers. Label noise affects both features equally but sets a ceiling on the absolute reported scores.
  • domain assumption The cardiac-cycle time series is sampled finely and regularly enough for the temporal Fourier transform to resolve the cardiac frequency.
    Section 2 computes the temporal Fourier transform, but the paper does not report the number of cardiac phases, temporal resolution, or any interpolation used before the transform.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Weighted Mean Frequencies: a handcraft Fourier feature for 4D Flow MRI segmentation." pith.science (2026). https://pith.science/paper/DJNHWEAH

@misc{pith2026250620614,
  author       = {Pith},
  title        = {Pith review of: Weighted Mean Frequencies: a handcraft Fourier feature for 4D Flow MRI segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DJNHWEAH}},
  note         = {Machine review of arXiv:2506.20614}
}
read the original abstract

In recent decades, the use of 4D Flow MRI images has enabled the quantification of velocity fields within a volume of interest and along the cardiac cycle. However, the lack of resolution and the presence of noise in these biomarkers are significant issues. As indicated by recent studies, it appears that biomarkers such as wall shear stress are particularly impacted by the poor resolution of vessel segmentation. The Phase Contrast Magnetic Resonance Angiography (PC-MRA) is the state-of-the-art method to facilitate segmentation. The objective of this work is to introduce a new handcraft feature that provides a novel visualisation of 4D Flow MRI images, which is useful in the segmentation task. 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. Indeed, this feature is representative of the hull of all pulsatile velocity voxels. The value of the feature under discussion is illustrated by two experiments. The experiments involved segmenting 4D Flow MRI images using optimal thresholding and deep learning methods. The results obtained demonstrate a substantial enhancement in terms of IoU and Dice, with a respective increase of 0.12 and 0.13 in comparison with the PC-MRA feature, as evidenced by the deep learning task. This feature has the potential to yield valuable insights that could inform future segmentation processes in other vascular regions, such as the heart or the brain.

Figures

Figures reproduced from arXiv: 2506.20614 by the authors.

Figure 1
Figure 1. Images (a) shows the anatomy, (b), (c) and (d) represent the result of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 9 canonical work pages

  1. [1]

    https://doi.org/10.1002/mrm.28257

    Berhane, H., Scott, M., Elbaz, M., Jarvis, K., McCarthy, P., Carr, J., Malaisrie, C., Avery, R., Barker, A.J., Robinson, J.D., Rigsby, C.K., Markl, M.: Fully automated 3D aortic segmentation of 4D flow MRI for hemodynamic analysis using deep learning84(4), 2204–2218. https://doi.org/10.1002/mrm.28257

    work page doi:10.1002/mrm.28257

  2. [2]

    Magnetic Resonance in Medicine63(2), 330–338 (2010)

    Bock, J., Frydrychowicz, A., Stalder, A.F., Bley, T.A., Burkhardt, H., Hennig, J., Markl, M.: 4d phase contrast mri at 3 t: Effect of standard and blood-pool contrast agents on snr, pc-mra, and blood flow visualization. Magnetic Resonance in Medicine63(2), 330–338 (2010). https://doi.org/10.1002/mrm.22199

    work page doi:10.1002/mrm.22199 2010

  3. [3]

    Bustamante, M., Viola, F., Engvall, J., Carlhäll, C.J., Ebbers, T.: Automatic time- resolved cardiovascular segmentation of 4D flow MRI using deep learning. J. of Magn. Reson. Imaging57(1), 191–203 (2023). https://doi.org/10.1002/jmri.28221

    work page doi:10.1002/jmri.28221 2023

  4. [4]

    https://doi.org/10.1016/j.media.2018.08.003

    Bustamante, M., Gupta, V., Forsberg, D., Carlhäll, C.J., Engvall, J., Ebbers, T.: Automated multi-atlas segmentation of cardiac 4d flow mri49, 128–140 (2018). https://doi.org/10.1016/j.media.2018.08.003

    work page doi:10.1016/j.media.2018.08.003 2018

  5. [5]

    Magnetic Resonance in Medicine9(1), 139–149 (1989)

    Dumoulin, C.L., Souza, S.P., Walker, M.F., Wagle, W.: Three-dimensional phase contrast angiography. Magnetic Resonance in Medicine9(1), 139–149 (1989). https://doi.org/10.1002/mrm.1910090117

    work page doi:10.1002/mrm.1910090117 1989

  6. [6]

    In: 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1748–1758 (2022). https://doi.org/10.1109/WACV51458.2022.00181

    arXiv 2022

  7. [7]

    https://doi.org/10.1016/j.mri.2020.08.017

    Levilly, S., Castagna, M., Idier, J., Bonnefoy, F., Le Touzé, D., Moussaoui, S., Paul- Gilloteaux, P., Serfaty, J.: Towards quantitative evaluation of wall shear stress from 4D flow imaging74, 232–243 (2020). https://doi.org/10.1016/j.mri.2020.08.017

    work page doi:10.1016/j.mri.2020.08.017 2020

  8. [8]

    https://doi.org/10.48550/arXiv.2403.05246

    Liao, W., Zhu, Y., Wang, X., Pan, C., Wang, Y., Ma, L.: Lightm-unet: Mamba assists in lightweight unet for medical image segmentation (2024). https://doi.org/10.48550/arXiv.2403.05246

    arXiv doi:10.48550/arxiv.2403.05246 2024

Show all 16 references
  1. [9]

    arXiv preprint arXiv:2401.04722 (2024)

    Ma, J., Li, F., Wang, B.: U-mamba: Enhancing long-range dependency for biomed- ical image segmentation. arXiv preprint arXiv:2401.04722 (2024)

    2024 arXiv

  2. [10]

    Magnetic Resonance Imaging 99, 20–25 (2023)

    Marin-Castrillon, D.M., Lalande, A., Leclerc, S., Ambarki, K., Morgant, M.C., Cochet, A., Lin, S., Bouchot, O., Boucher, A., Presles, B.: 4D segmentation of the thoracic aorta from 4D flow mri using deep learning. Magnetic Resonance Imaging 99, 20–25 (2023). https://doi.org/10...

    2023 doi

  3. [11]

    Markl, M., Frydrychowicz, A., Kozerke, S., Hope, M., Wieben, O.: 4D Flow MRI. J. Magn. Reson. Imaging36(5), 1015–1036 (2012). https://doi.org/10.1002/jmri.23556, http://dx.doi.org/10.1002/jmri.23556 10 S. Perrin et al

    2012 doi

  4. [12]

    Pelc, N.J., Bernstein, M.A., Shimakawa, A., Glover, G.H.: Encoding strate- gies for three-direction phase-contrast MR imaging of flow. J. Magn. Re- son. Imaging1(4), 405–413 (1991). https://doi.org/10.1002/jmri.1880010404, http://dx.doi.org/10.1002/jmri.1880010404

    1991 doi

  5. [13]

    https://doi.org/10.1016/j.cmpb.2021.106236

    Pérez-García, F., Sparks, R., Ourselin, S.: TorchIO: A Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images in deep learning208, 106236. https://doi.org/10.1016/j.cmpb.2021.106236

  6. [14]

    In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. pp. 234–241. Springer Internation...

    2015 doi

  7. [15]

    IEEE Transactions on Medical Imaging43(9), 3377–3390 (2024)

    Shaker, A., Maaz, M., Rasheed, H., Khan, S., Yang, M.H., Shahbaz Khan, F.: Unetr++: Delving into efficient and accurate 3d medical image seg- mentation. IEEE Transactions on Medical Imaging43(9), 3377–3390 (2024). https://doi.org/10.1109/TMI.2024.3398728

    2024

  8. [16]

    Neuroimage31(3), 1116– 1128 (2006)

    Yushkevich, P.A., Piven, J., Cody Hazlett, H., Gimpel Smith, R., Ho, S., Gee, J.C., Gerig, G.: User-guided 3D active contour segmentation of anatomical struc- tures: Significantly improved efficiency and reliability. Neuroimage31(3), 1116– 1128 (2006). https://doi.org/10.1016/...

    2006 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.