REVIEW 3 major objections 3 minor 20 references
Enhancing Diffusion-Weighted Images (DWI) for Diffusion MRI: Is it Enough without Non-Diffusion-Weighted B=0 Reference?
T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read This paper claims that super-resolving diffusion-weighted images with a pixel-wise MSE loss alone can leave the DWI/b=0 intensity ratio—the quantity diffusion metrics are computed from—uncontrolled, and that adding a logarithmic ratio…
desk verdict The proposed ratio-log loss in Eq. (3) is just a log-domain MSE on the DWI; the b=0 reference cancels algebraically, so the central claim that the method uses b=0 is contradicted by its own equations. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the DWI-to-b=0 intensity ratio S/S0, the quantity that carries diffusion information through the Stejskal-Tanner equation S/S0 = exp(-b g^T D g). The paper's proposed mechanism is the ratio-log loss, L_ratio_log = MSE(log(S_pred/(S0+eps)), log(S_GT/(S0+eps))), added to pixel MSE and FFT losses with weight 0.01. Taking the logarithm changes what the loss sees: instead of penalizing error in the exponential attenuation itself, it penalizes error in g^T D g, the inner product that tensor fitting actually solves for. The same mechanism lets the high-resolution b=0 image act as an anatomical anchor during training.
What would settle it
Re-run the same super-resolution experiment but give the ratio-log loss a degraded b=0 reference, for example a single un-averaged 4 mm anisotropic slice instead of the 5-repeat 2 mm isotropic average; if dratio no longer converges lower than the baseline or PSNR no longer improves, the benefit depends on the clean-reference assumption. Alternatively, compute FA and MD from the baseline and ratio-log outputs: if the maps are indistinguishable, the reduced ratio error does not propagate into the diffusion metrics the paper cares about.
Extended reading notes
Core claim
The central discovery is that PSNR and ratio fidelity decouple in DWI enhancement. In the baseline U-Net trained with 15 times the MSE loss plus the FFT loss, validation PSNR rises monotonically and converges at 34.80 dB, but the ratio error, called dratio, first falls and then climbs back to stabilize at 19860 after step 1842. The paper argues this divergence is not incidental: diffusion tensor fitting uses the attenuation S/S0 = exp(-b g^T D g), so error in the ratio, not in absolute DWI intensity, propagates into ADC and FA. The proposed remedy is the ratio-log loss, defined as the MSE between the logarithms of the predicted and ground-truth DWI/b=0 ratios, which by taking the logarithm targets the exponent -b g^T D g directly. With this term added at weight 0.01, dratio decreases steadily and converges at 16330, and PSNR improves slightly to 34.92 dB, which the authors attribute to the high-resolution b=0 image acting as a structural reference.
Load-bearing premise
The method assumes a clean, high-resolution b=0 image is available to form the ratio and to supervise training; if the b=0 reference is itself noisy or low-resolution, the log-ratio loss would fold that degradation into training and the reported gains could vanish.
Editorial extensions
If this is right
- Validation of DWI enhancement should report ratio error alongside PSNR, since a model with higher PSNR can still have worse DWI/b=0 fidelity and therefore unreliable diffusion metrics.
- The ratio-log loss can be added to any existing DWI super-resolution or denoising pipeline with negligible extra cost and no architectural change, because it is just one extra term in the total loss.
- The same loss principle transfers to other quantitative imaging modalities that normalize a measurement by a reference image, a direction the paper itself notes.
- Tensor-derived maps such as FA and MD, and tractography, are the natural place to test whether the reduced ratio error translates into clinical benefit; the paper explicitly leaves that as future validation.
Reading between the lines
- If the decoupling the paper reports is general, then PSNR-based model selection in dMRI may actively select outputs that degrade diffusion metrics, so a reference-aware term like the ratio-log loss should be part of the objective whenever a b=0 image is available.
- A testable extension is to feed a degraded b=0 reference, such as a single un-averaged anisotropic slice, into the ratio-log loss; the paper only uses a 5-repeat 2 mm isotropic average, and the reported gains should shrink or vanish if the clean-reference assumption is doing the work.
- A second testable extension is to extend the ratio-log loss across multiple b-values and gradient directions simultaneously, enforcing consistency with the tensor model itself rather than with a single b=1000 ratio, which would turn the loss from a fidelity term into a physics regularizer.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This manuscript addresses the question of whether diffusion-weighted image (DWI) enhancement should account for the non-diffusion-weighted (b=0) reference image. The authors train a U-Net for DWI super-resolution on UK Biobank data and observe that training with pixel-wise MSE alone leads to a DWI/b=0 ratio error that initially decreases and then rises, whereas adding a proposed 'ratio-log loss' reduces this ratio error and slightly improves PSNR. The paper claims that this loss explicitly relates the generated DWI to the b=0 reference and that this is necessary for accurate downstream diffusion metrics.
Significance. The underlying question is practically important: if DWI enhancement ignores the b=0 reference, derived quantitative maps such as ADC, FA, and MD may be inaccurate. The empirical observation that a pixel-wise MSE loss does not automatically preserve the DWI/b=0 ratio is potentially valuable. However, the central technical claim is undermined by a mathematical error: the proposed loss in Eq. (3) is algebraically independent of the b=0 reference because the reference cancels in the logarithm. The evaluation is also thin: single training runs, no statistical tests, and no downstream diffusion metrics are computed despite the abstract's claim. As written, the manuscript does not establish its main conclusions.
major comments (3)
- [Section 2.3, Eq. (3)] The ratio-log loss is algebraically independent of the b=0 reference. Since the same S_b0^GT + epsilon appears in both the predicted and ground-truth ratio terms, log(P/(S0+epsilon)) - log(T/(S0+epsilon)) = log(P/T). Therefore L_ratio_log = MSE(log P, log T), and the derivative with respect to S0 is identically zero. The loss does not 'relate the generated DWI with the reference b=0' as claimed in Section 2.3 and the Introduction; it is simply a log-domain MSE on the DWI intensity itself. The paper never compares against an ordinary log-MSE loss, so the empirical improvement in Table 1, while possibly real, cannot be attributed to the proposed mechanism.
- [Section 3.2, Table 1] The abstract states that the method leads to 'improved dMRI super-resolution and better preservation of b=0 ratio-based features for the derivation of diffusion metrics,' but no downstream diffusion metrics (ADC, FA, MD) are computed anywhere in the paper. The authors themselves acknowledge in Section 3.2 that 'further validation will be done through diffusion tensor fitting and tractography,' which is an internal admission that the central claim is unverified. The reported PSNR gain of 0.12 dB (34.80 to 34.92) is small and, with a single training run per setting and no error bars or statistical tests, could easily be within run-to-run variation. The converged dratio values (19860 vs 16330) are also reported without uncertainty.
- [Section 3.1, Figure 1] The claim that 'MSE-only training causes the DWI/b=0 ratio error to diverge' is based on a single training curve. The described behavior—dratio initially decreasing and then rising as PSNR continues to improve—could reflect learning-rate dynamics, overfitting, or other training artifacts rather than a fundamental limitation of the MSE loss. Without multiple random seeds, a controlled comparison with different hyperparameters, or a statistical summary, the conclusion that higher PSNR does not imply better ratio preservation is not established.
minor comments (3)
- [Section 3 (headers)] There are two sections numbered 3: 'EXPERIMENT SETTING' and 'RESULT AND DISCUSSION'. Please renumber the sections sequentially.
- [Section 2.2, Eq. (2)] The superscripts 'RECON' and 'GT' in Eq. (2) are not defined in the text. Define these notations explicitly.
- [Section 3.1 Dataset] The UK Biobank participants are referred to as 'patients,' but the UK Biobank is a population health study of volunteers. Consider using 'participants' instead.
Circularity Check
Eq. (3) ratio-log loss cancels the b=0 denominator by construction, so the central claim that a b=0 reference is needed reduces to an ordinary log-MSE on DWI intensity.
-
renaming known result
[Section 2.3, Eq. (3)]
"By modifying the distance metric in Eq. (2), we calculate the MSE loss between the logarithms of the predicted and ground truth ratios instead of the ratios themselves as in Eq. (3). ℒratio_log = MSE(log(S_b1000^Pred/(S_b0^GT+ε)), log(S_b1000^GT/(S_b0^GT+ε)))"
Because log(P/(S0+ε)) - log(T/(S0+ε)) = log(P/T), the MSE over these logarithmic terms is identically MSE(log S_b1000^Pred, log S_b1000^GT). The b=0 reference S_b0^GT cancels exactly, so the proposed 'ratio-log loss' has no dependence on any b=0 image. The paper claims this loss 'relates the generated DWI with the reference b=0' and that improving DWI without b=0 is insufficient, but by construction the proposed loss is just a log-domain pixel-wise loss on the DWI itself, i.e. a renamed known loss. Since no experiment compares against ordinary log-MSE without b=0, the empirical improvement cannot be attributed to the claimed b=0 mechanism, and the central message that a b=0 reference is necessary is not supported by the paper's own equation.
full rationale
The strongest issue is the algebraic cancellation in Eq. (3): the b=0 denominator cancels when the log-ratio prediction and ground truth are subtracted, making ℒratio_log mathematically equivalent to MSE between log-transformed DWIs. This is a specific, quotable reduction of the paper's central mechanism to a known loss with no b=0 input. The empirical comparison against the MSE baseline is still meaningful and the reported dratio improvement may be real, but it is not evidence for the paper's stated need for a b=0 reference, because the same loss could be trained without b=0. No fitted parameters are hidden in the loss derivation, and the self-citations to the authors' prior work are contextual rather than load-bearing. The circularity is therefore partial: the method's claimed b=0 dependence reduces by construction, while the experimental observation itself retains independent content.
Assumptions & free parameters
free parameters (3)
- Ratio-log loss weight (lambda_ratio_log) =
0.01
- MSE loss weight (lambda_MSE) =
15
- FFT loss weight (lambda_FFT) =
0.0025
assumptions (4)
- domain assumption Stejskal-Tanner signal model S/S0 = exp(-b g^T D g)
- domain assumption The U-Net with multi-head self-attention from reference [14] is an appropriate backbone
- ad hoc to paper MSE-only training causes the DWI/b=0 ratio error to rise after an initial decrease
- domain assumption A clean, high-resolution b=0 image is available for the ratio loss and evaluation
Cite this review
Pith. "Pith review of Enhancing Diffusion-Weighted Images (DWI) for Diffusion MRI: Is it Enough without Non-Diffusion-Weighted B=0 Reference?." pith.science (2026). https://pith.science/paper/TON66DN6
@misc{pith2026250512978,
author = {Pith},
title = {Pith review of: Enhancing Diffusion-Weighted Images (DWI) for Diffusion MRI: Is it Enough without Non-Diffusion-Weighted B=0 Reference?},
year = {2026},
howpublished = {\url{https://pith.science/paper/TON66DN6}},
note = {Machine review of arXiv:2505.12978}
}
read the original abstract
Diffusion MRI (dMRI) is essential for studying brain microstructure, but high-resolution imaging remains challenging due to the inherent trade-offs between acquisition time and signal-to-noise ratio (SNR). Conventional methods often optimize only the diffusion-weighted images (DWIs) without considering their relationship with the non-diffusion-weighted (b=0) reference images. However, calculating diffusion metrics, such as the apparent diffusion coefficient (ADC) and diffusion tensor with its derived metrics like fractional anisotropy (FA) and mean diffusivity (MD), relies on the ratio between each DWI and the b=0 image, which is crucial for clinical observation and diagnostics. In this study, we demonstrate that solely enhancing DWIs using a conventional pixel-wise mean squared error (MSE) loss is insufficient, as the error in ratio between generated DWIs and b=0 diverges. We propose a novel ratio loss, defined as the MSE loss between the predicted and ground-truth log of DWI/b=0 ratios. Our results show that incorporating the ratio loss significantly improves the convergence of this ratio error, achieving lower ratio MSE and slightly enhancing the peak signal-to-noise ratio (PSNR) of generated DWIs. This leads to improved dMRI super-resolution and better preservation of b=0 ratio-based features for the derivation of diffusion metrics.
Figures
Reference graph
Works this paper leans on
-
[1]
Looking into the functional architecture of the brain with diffusion MRI,
D. Le Bihan, “Looking into the functional architecture of the brain with diffusion MRI,” Nat Rev Neurosci, vol. 4, no. 6, pp. 469–480, Jun. 2003, doi: 10.1038/nrn1119
-
[2]
METHOD 2.1 Problem Formulation A typical dMRI acquisition includes repeated non-diffusion- weighted (b=0) reference image and several diffusion encoding images acquired with diffusion weightings (b- values) and several diffusion-encoding gradient directions. For each voxel, the signal attenuation 𝑆𝑆/𝑆𝑆0 for each diffusion-weighted image (DWI) is calculate...
-
[3]
EXPERIMENT SETTING 3.1 Dataset and Pre-processing Pre-processed dMRI data of 13 particip ants from the UK Biobank [13] were used for this study. The acq uisition methods and parameters detail can be found at (http://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri. pdf). The training dataset comprised their dMRI data, with a voxel size of 2 mm isotrop...
-
[4]
RESULT AND DISCUSSION 3.1 Does a Higher PSNR Guarantee Better Preservation of the DWI/b=0 Ratio? Initially, we trained the U -Net model without incorporating the ℒratio_log loss term, setting the total loss function as ℒtotal = 15 × ℒMSE + 0.0025 × ℒFFT . As shown in Figure 1 , PSNR increased steadily during training; however, the dratio metric initially ...
-
[5]
CONCLUS ION This study presents a novel approach to dMRI enhancement by introducing a ratio- log loss term, ℒratio_log , which explicitly optimizes the critical DWI/b=0 ratio. Our results demonstrate that incorporating ℒratio_log not only improves the fidelity of the DWI/b=0 relationship but also contributes to a slight increase in PSNR, which is traditio...
-
[6]
ACKNOW LEDGMENT This research has been conducted using the UK Biobank Resource under Application Number 100203. This work is funded in part by Imperial College London President’s PhD Scholarship, in part by Imperial College London I -X Moonshot Seed F und, and in part by UKRI Future Leaders Fellows Development Network. G. Yang’s work is funded by the UKRI...
-
[7]
Diffusion Tensor Cardiovascular Magnetic Resonance Imaging,
Z. Khalique, P. F. Ferreira, A. D. Scott, S. Nielles-Vallespin, D. N. Firmin, and D. J. Pennell, “Diffusion Tensor Cardiovascular Magnetic Resonance Imaging,” JACC Cardiovasc Imaging, vol. 13, no. 5, pp. 1235–1255, May 2020, doi: 10.1016/j.jcmg.2019.07.016
-
[8]
Deep learning-based diffusion tensor cardiac magnetic resonance reconstruction: a comparison study,
J. Huang et al., “Deep learning-based diffusion tensor cardiac magnetic resonance reconstruction: a comparison study,” Sci Rep, vol. 14, no. 1, p. 5658, Mar. 2024, doi: 10.1038/s41598-024- 55880-2
Show all 20 references
-
[9]
DDM$^2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models,
T. Xiang, M. Yurt, A. B. Syed, K. Setsompop, and A. Chaudhari, “DDM$^2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models,” Feb. 2023, [Online]. Available: http://arxiv.org/abs/2302.03018
2023 arXiv
-
[10]
Super-resolution for multislice diffusion tensor imaging,
D. H. J. Poot et al., “Super-resolution for multislice diffusion tensor imaging,” Magn Reson Med, vol. 69, no. 1, pp. 103–113, Jan. 2013, doi: 10.1002/mrm.24233
2013 doi
-
[11]
SDnDTI: Self-supervised deep learning-based denoising for diffusion tensor MRI,
Q. Tian et al., “SDnDTI: Self-supervised deep learning-based denoising for diffusion tensor MRI,” Neuroimage, vol. 253, p. 119033, Jun. 2022, doi: 10.1016/j.neuroimage.2022.119033
2022
-
[12]
High-resolution reference image assisted volumetric super-resolution of cardiac diffusion weighted imaging,
Y. Wu et al., “High-resolution reference image assisted volumetric super-resolution of cardiac diffusion weighted imaging,” in Medical Imaging 2024: Image Processing, O. Colliot and J. Mitra, Eds., SPIE, Apr. 2024, p. 68. doi: 10.1117/12.3006008
2024 doi
-
[13]
Data- and Physics-driven Deep Learning Based Reconstruction for Fast MRI: Fundamentals and Methodologies,
J. Huang et al., “Data- and Physics-driven Deep Learning Based Reconstruction for Fast MRI: Fundamentals and Methodologies,” IEEE Rev Biomed Eng, pp. 1–20, 2024, doi: 10.1109/RBME.2024.3485022
2024
-
[14]
Deep learning reconstruction of diffusion-weighted brain MRI for evaluation of patients with acute neurologic symptoms,
S. I. Park, Y. Yim, J. Bin Lee, and H. S. Ahn, “Deep learning reconstruction of diffusion-weighted brain MRI for evaluation of patients with acute neurologic symptoms,” Sci Rep, vol. 14, no. 1, p. 24761, Oct. 2024, doi: 10.1038/s41598-024-75011-1
2024 doi
-
[15]
Accelerating Prostate Diffusion-weighted MRI Using a Guided Denoising Convolutional Neural Network: Retrospective Feasibility Study,
E. A. Kaye et al., “Accelerating Prostate Diffusion-weighted MRI Using a Guided Denoising Convolutional Neural Network: Retrospective Feasibility Study,” Radiol Artif Intell, vol. 2, no. 5, p. e200007, Aug. 2020, doi: 10.1148/ryai.2020200007
2020 doi
-
[16]
Wasserstein GAN,
M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein GAN,” Jan. 2017
2017
-
[17]
DAGAN: Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction,
G. Yang et al., “DAGAN: Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction,” IEEE Trans Med Imaging, vol. 37, no. 6, pp. 1310–1321, Jun. 2018, doi: 10.1109/TMI.2017.2785879
2018
-
[18]
Multimodal population brain imaging in the UK Biobank prospective epidemiological study,
K. L. Miller et al., “Multimodal population brain imaging in the UK Biobank prospective epidemiological study,” Nat Neurosci, vol. 19, no. 11, pp. 1523–1536, Nov. 2016, doi: 10.1038/nn.4393
2016 doi
-
[19]
Improved Denoising Diffusion Pr obabilistic Models,
A. Nichol and P. Dhariwal, “Improved Denoising Diffusion Pr obabilistic Models,” Feb. 2021
2021
-
[20]
Physics-Guided Loss Functions Improve Deep Learning Performance in Inverse Scattering,
Z. Liu, M. Roy, D. K. Prasad, and K. Agarwal, “Physics-Guided Loss Functions Improve Deep Learning Performance in Inverse Scattering,” IEEE Trans Comput Imaging, vol. 8, pp. 236–245, 2022, doi: 10.1109/TCI.2022.3158865. For the purpose of open access, the authors have applied ...
2022
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.