{"id":"afbd90e0-80c8-49e7-b05b-7411a8557e09","arxiv_id":"2505.12978","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding a log-ratio loss between predicted and ground-truth DWI/b=0 images reduces ratio error and slightly improves PSNR in diffusion MRI super-resolution, though with weak statistical evidence.","lead":"This paper tests whether adding a loss term that preserves the log ratio between diffusion-weighted and non-diffusion-weighted (b=0) images improves super-resolution of diffusion MRI. It finds the extra ratio loss reduces the DWI/b=0 ratio error and slightly raises PSNR, but the evidence lacks error bars and downstream validation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3) ratio-log loss is algebraically independent of the b=0 reference: the denominator cancels, so the central claim that a b=0 image is needed is not supported by the proposed loss.","rationale":"The central claim is that MSE-only DWI enhancement leaves the DWI/b=0 ratio uncontrolled and that the new ratio-log loss fixes this by using the b=0 reference. The disclosed Eq. (3) undermines that mechanism before any experimental issues are considered: the b=0 denominator cancels in the logarithm, making the loss equivalent to log-intensity MSE on the DWI alone. This is an internal algebraic property, not a disagreement with the field's consensus. The reader's weakest assumption concerned the availability of a clean high-resolution b=0 image; the concern here is stronger, namely that even a perfect b=0 image has no influence on the loss. The empirical numbers might survive as a finding about log-domain training, but the paper's framing and novelty claim do not. A revised version that explicitly compares against log-MSE and reframes the contribution could be reconsidered; as written, the central claim is not supported.","tokens_in":6331,"tokens_out":6261,"duration_ms":70939,"concrete_test":"Re-run the with-ℒ_ratio_log experiment after replacing S_b0^GT in Eq. (3) with a fixed constant (e.g., 1) or with a randomly shuffled version of the true b=0 volume, keeping all other settings identical. Because the same denominator appears in both log terms, the numerical value of ℒ_ratio_log and its gradients are invariant to this replacement up to floating point; if the training and validation curves, including the Table 1 entries, are unchanged, then the loss does not use b=0. As a companion check, train a third model with an explicit log-MSE loss on DWI intensities alone (with no b=0 term anywhere); if its dratio and PSNR curves coincide with the ratio-log model, the paper's b=0 framing is refuted.","verdict_should_be":"REJECT","load_bearing_attack":"The proposed ratio-log loss in Eq. (3) divides both the predicted and ground-truth DWI by the same S_b0^GT + ε before taking logarithms. Therefore log(P/(S0+ε)) - log(T/(S0+ε)) = log(P/T), and ℒ_ratio_log = MSE(log P, log T). The derivative of this loss with respect to S0 is identically zero. This means the loss does not 'relate the generated DWI with the reference b=0' as claimed in Section 2.3; it is simply a log-domain MSE on the DWI intensity itself, with no dependence on any b=0 reference. The empirical improvement in Table 1 may still be real, but the paper attributes it to a mechanism that the mathematics shows cannot operate. The title and abstract claim that optimizing DWIs without considering b=0 is insufficient, yet the proposed fix does not consider b=0 either. The paper also never compares against an ordinary log-MSE loss, so the claimed mechanism is not empirically separated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6563,"tokens_out":4309,"duration_ms":41961,"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":[{"comment":"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":"Section 2.3, Eq. (3)"},{"comment":"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":"Section 3.2, Table 1"},{"comment":"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.","section":"Section 3.1, Figure 1"}],"minor_comments":[{"comment":"There are two sections numbered 3: 'EXPERIMENT SETTING' and 'RESULT AND DISCUSSION'. Please renumber the sections sequentially.","section":"Section 3 (headers)"},{"comment":"The superscripts 'RECON' and 'GT' in Eq. (2) are not defined in the text. Define these notations explicitly.","section":"Section 2.2, Eq. (2)"},{"comment":"The UK Biobank participants are referred to as 'patients,' but the UK Biobank is a population health study of volunteers. Consider using 'participants' instead.","section":"Section 3.1 Dataset"}],"recommendation":"reject","confidential_remarks":"The algebraic cancellation in Eq. (3) is a load-bearing error: the proposed loss does not use the b=0 reference at all, so the paper does not answer its own title question. Even if the empirical result is real, it would only demonstrate a benefit of log-domain MSE over linear MSE, which is a much weaker and less novel claim. The absence of any downstream diffusion metric evaluation further weakens the paper's fit for a medical imaging venue. I recommend rejection rather than major revision because the core contribution as formulated is mathematically incorrect."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read Eq. (3) carefully before believing the abstract. The ratio-log loss divides both predicted and ground-truth DWIs by the same S_b0^GT + ε before taking logs and MSE. That denominator cancels: the loss is MSE(log P, log T), and its derivative with respect to the b=0 image is zero. So the proposed mechanism—using the non-diffusion-weighted reference to improve the DWI/b=0 relationship—does not operate. The paper's own mathematics undermines its central claim.\n\nWhat is genuinely useful is the empirical divergence observation: training with only MSE on the DWI lets the DWI/b=0 ratio error rise while PSNR improves. That is worth noting, and a log-space loss is a reasonable thing to try. But the paper never compares against an ordinary log-MSE loss without the b=0 denominator, so we cannot separate the effect of the log transform from the claimed b=0 coupling. The reported gains are thin: single runs, no error bars, and the 0.12 dB PSNR improvement is within run-to-run variation. The abstract promises improved diffusion metrics, but no ADC, FA, or MD is computed; the paper itself admits that later.\n\nThe soft spots are central, not minor. The loss is mischaracterized, the evidence is underpowered, and the comparison misses the natural control. A careful reader should not take the headline result as established. That said, the divergence observation could be recovered in a revised paper that either drops the b=0 framing and directly compares log-MSE, or builds a loss that genuinely depends on a degraded b=0 reference.\n\nFor now: not something I would cite, and not something I would send to review in its present form. If the authors fix the framing and add proper controls, it could become a small honest note. Serious thinkers should read it to see how easily a plausible-sounding physics-informed loss can be hollowed out by algebra.","headline":"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.","tokens_in":742,"tokens_out":1475,"would_cite":false,"duration_ms":35430,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["diffusion MRI","DWI super-resolution","b=0 reference","ratio-log loss","diffusion tensor metrics","quantitative MRI","loss function design"],"falsifier":"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.","tokens_in":6135,"feed_emoji":"🧠","tokens_out":7573,"duration_ms":68656,"temperature":0.7,"pith_summary":"Diffusion MRI's clinical value comes from ratios: each diffusion-weighted image (DWI) is divided by the non-diffusion-weighted b=0 image to get the signal attenuation that feeds diffusion tensor fitting and ADC, FA, and MD maps. This paper sets out to establish that super-resolution models trained only on pixel-wise error in the DWI can drift in that ratio even while their PSNR improves, so a better-looking DWI may be worse for downstream diffusion metrics. The authors propose a ratio-log loss—the mean squared error between predicted and ground-truth logarithms of the DWI/b=0 ratio—and show that adding it at small weight improves convergence of the ratio error and slightly raises PSNR. On a brain dMRI super-resolution task from anisotropic 4x2x2 mm slices to 2 mm isotropic, the converged ratio error drops from 19860 to 16330 while PSNR rises from 34.80 to 34.92 dB.","feed_headline":"Log-ratio loss cuts DWI/b=0 error by 18 percent","feed_subtitle":"Pixel-perfect DWI enhancement can still wreck the ratio used to compute ADC, FA, and MD; this loss keeps it intact.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Introduces the idea that a high-resolution b=0 reference image can assist DWI super-resolution, motivating the use of b=0 in the loss.","marker":"[7]"},{"why":"Represents the conventional DWI-enhancement approach with pixel-wise losses that the paper argues is insufficient without ratio consistency.","marker":"[9]"},{"why":"Another conventional DWI denoising or enhancement baseline trained without explicit DWI/b=0 ratio consistency.","marker":"[10]"},{"why":"Cited as the basis for using logarithms to stabilize training involving divisions, a stated motivation for the ratio-log loss.","marker":"[11]"},{"why":"Supplies the FFT loss term that the total loss combines with the ratio-log loss.","marker":"[12]"},{"why":"Provides the population brain dMRI dataset used for training and validation in the super-resolution experiments.","marker":"[13]"},{"why":"Supplies the U-Net backbone architecture with multi-head self-attention used in all experiments.","marker":"[14]"},{"why":"Documents that physics-informed losses often lower PSNR, which the paper invokes to highlight that its ratio-log loss slightly raises PSNR.","marker":"[15]"}],"fun_headline_variants":["Ratio loss fixes what PSNR misses in DWI enhancement","DWI super-res needs b=0 ratio, not just pixel perfection","Log-ratio loss beats pixel loss for ADC and FA fidelity","For diffusion MRI, ratio fidelity trumps DWI PSNR","Enhancing DWI alone? Ratio error diverges without b=0 loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Ratio loss fixes what PSNR misses in DWI enhancement","DWI super-res needs b=0 ratio, not just pixel perfection","Log-ratio loss beats pixel loss for ADC and FA fidelity","For diffusion MRI, ratio fidelity trumps DWI PSNR","Enhancing DWI alone? Ratio error diverges without b=0 loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1739,"prompt_tokens":1030,"completion_tokens":709,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":617}},"tokens_in":646,"tokens_out":709,"duration_ms":7429,"temperature":1.0,"reasoning_tokens":617,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:22:01.667962+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Diffusion Tensor Cardiovascular Magnetic Resonance Imaging,","cited_arxiv_id":null,"evidence_quote":"Introduces the idea that a high-resolution b=0 reference image can assist DWI super-resolution, motivating the use of b=0 in the loss."},{"cited_title":"Super-resolution for multislice diffusion tensor imaging,","cited_arxiv_id":null,"evidence_quote":"Another conventional DWI denoising or enhancement baseline trained without explicit DWI/b=0 ratio consistency."},{"cited_title":"SDnDTI: Self-supervised deep learning-based denoising for diffusion tensor MRI,","cited_arxiv_id":null,"evidence_quote":"Cited as the basis for using logarithms to stabilize training involving divisions, a stated motivation for the ratio-log loss."},{"cited_title":"High-resolution reference image assisted volumetric super-resolution of cardiac diffusion weighted imaging,","cited_arxiv_id":null,"evidence_quote":"Supplies the FFT loss term that the total loss combines with the ratio-log loss."},{"cited_title":"Deep learning reconstruction of diffusion-weighted brain MRI for evaluation of patients with acute neurologic symptoms,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net backbone architecture with multi-head self-attention used in all experiments."}],"review_version":1}