{"id":"6d92e9a7-d81d-4ee3-817a-066d9f76f6b1","arxiv_id":"2506.15750","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-domain diffusion model combining spatial and frequency guidance with a shared critic and uncertainty mask loss improves multi-contrast MRI synthesis over existing baselines.","lead":"D2Diff is a new MRI synthesis method that uses two diffusion networks, one guided by spatial features and one by frequency features, to generate missing MRI contrasts from the ones already acquired. It reports better image quality than current baselines, and shows the synthetic images can support brain tumor segmentation without losing diagnostic value.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The superiority claim is not statistically substantiated: Table 1 asserts p<0.05 without reporting p-values, and margins over SynDiff are often within one standard deviation; Table 2 dice differences lack error bars.","rationale":"The reader's CONDITIONAL verdict is appropriate, and this stress-test does not move it. However, the single most load-bearing concern is not the DCT feature complementarity identified in the reader's weakest_assumption, but the unsupported statistical significance claim in Table 1 and the absence of error bars in Table 2. The numbers as published do not demonstrate that D2Diff's improvements over SynDiff are real rather than noise, and the downstream segmentation 'superiority' is a gap of about 0.2 Dice points without uncertainty estimates. This is directly testable and addressable with additional analysis, consistent with a conditional rather than accept verdict. The reader did note the p-value issue in the rationale, so agreement is partial rather than full.","tokens_in":8757,"tokens_out":3185,"duration_ms":33924,"concrete_test":"Run a paired significance analysis on the per-subject metrics from the released code or by requesting the authors' predictions. For each contrast and metric in Table 1, compute paired t-tests and Wilcoxon signed-rank tests between D2Diff and the strongest baseline (SynDiff) across the 30 test subjects, applying multiple-comparison correction (e.g., FDR). If the majority of differences are not significant at p<0.05, the claim of universal superiority is not supported. For Table 2, compute Dice scores with bootstrap confidence intervals across subjects and test whether the D2Diff vs SynDiff and D2Diff vs Complete differences are significant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that D2Diff outperforms all baselines across tasks rests entirely on Table 1's mean±std values, yet the table's note 'statistical significance p<0.05 based on paired mean t-test' is not backed by any reported p-values, confidence intervals, or test details. The margins over the strongest baseline SynDiff are frequently small compared to the reported standard deviations: for BraTS T1CE, PSNR is 28.58±2.69 vs 28.16±2.36 and MAE is 1.97±1.14 vs 2.08±1.14; for FLAIR, PSNR is 27.57±2.18 vs 27.13±2.11. With only 30 test subjects, such differences may or may not be significant depending on the correlation of paired errors, but the paper provides no evidence. The downstream segmentation claim in Table 2 is similarly fragile: D2Diff's Dice (81.05) is only 0.22 points above the 'Complete' setup (80.83) and 0.14–1.0 points above SynDiff, with no standard deviations or significance testing. If these gaps are within noise, the claims of 'superiority' and 'diagnostic value' are unsupported. This concern is more load-bearing than the DCT-complementarity question: even if the frequency features are not perfectly complementary, the method could still work; but if the reported improvements are not statistically reliable, the headline empirical claim fails outright.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D2Diff, a dual-domain diffusion framework for multi-contrast MRI synthesis. The method trains two U-Net denoising networks jointly via a shared critic: one network is conditioned on spatial-domain features extracted from source contrasts, and the other on frequency-domain features obtained by applying multi-scale discrete cosine transforms. An uncertainty-aware mask loss derived from critic attention maps is used to focus learning on high-uncertainty regions, and inference averages the predictions of the two networks. The method is evaluated on BraTS2019 and a healthy brain dataset for synthesis of T1, T1CE, T2, and FLAIR contrasts, compared with GAN- and diffusion-based baselines, and further assessed through a downstream tumour segmentation task and an ablation study. The authors report consistent improvements over baselines on PSNR, SSIM, and MAE, and claim statistical significance at p<0.05.","tokens_in":9087,"tokens_out":2990,"duration_ms":30491,"significance":"If validated, the paper makes a useful contribution to multi-contrast MRI synthesis by explicitly combining spatial and frequency-domain conditioning inside a diffusion framework, with a shared critic for mutual learning and an uncertainty mask for region-focused training. The inclusion of a downstream segmentation evaluation and an ablation study is commendable, as is the stated intent to release code and hyperparameters. However, the central empirical claim of superiority over state-of-the-art baselines currently rests on a statistical assertion that is not substantiated, and several implementation details needed for reproducibility are missing. The work is therefore significant in scope and potential impact, but the evidence as presented is not yet sufficient to support the headline claims.","major_comments":[{"comment":"The note 'statistical significance p<0.05 based on paired mean t-test' is not backed by any reported p-values, confidence intervals, or test details. Several D2Diff-vs-SynDiff margins are smaller than the reported standard deviations (e.g., BraTS T1CE PSNR 28.58±2.69 vs 28.16±2.36; BraTS FLAIR PSNR 27.57±2.18 vs 27.13±2.11), and with only 30 test subjects the significance of such small differences depends on the correlation of paired errors, which is not reported. Please provide the actual p-values with the test details for every comparison in Table 1, or temper the superiority claims to what the data support.","section":"Section 4, Table 1"},{"comment":"The description 'DCT with kernel sizes 3, 5, and 7' is ambiguous. A discrete cosine transform is a global or blockwise transform; the term 'kernel size' suggests block-wise DCT with patch sizes 3×3, 5×5, and 7×7, but the paper does not specify patch overlap, whether coefficients are truncated or fully retained, or how the resulting 'frequency features' are structured. This ambiguity is load-bearing because the claimed complementarity of frequency-domain features is a central design choice, and the current text does not allow the operation to be reproduced or compared with other frequency-domain approaches. Please define the exact DCT implementation and, ideally, provide an ablation or reference justifying the chosen kernel sizes.","section":"Section 2, Eq. (6) and Fig. 1B"},{"comment":"The experimental setup omits essential training and architectural details: the total number of diffusion timesteps T, the noise schedule beta_t, the loss weights lambda_s, lambda_m, lambda_c, learning rate, batch size, number of training iterations, optimizer, and the specific U-Net configurations for both denoising networks and the critic. The abstract states that 'Code and model hyperparameters are available,' but the paper itself does not report them. Without these details, the fair comparison with baselines cannot be assessed and the experiments are not reproducible from the manuscript alone.","section":"Section 3 and Section 4"},{"comment":"The downstream segmentation results are presented as Dice scores without standard deviations, confidence intervals, or significance tests. The differences are small (D2Diff 81.05 vs Complete 80.83; D2Diff 81.05 vs SynDiff 80.91), and the claim that 'superior downstream tumour segmentation highlights the diagnostic value of the synthetic images' is not supported without an estimate of variability. Please report error bars from multiple segmentation training runs or a statistical comparison, and adjust the interpretation accordingly.","section":"Section 4, Table 2"},{"comment":"The ablation study reports very small margins between D2Diff and the 'w/o freq. feat. adaptation' and 'Spatial guidance (H2)' variants (e.g., PSNR 28.58 vs 28.54 and 28.39; SSIM 91.84 vs 91.34 and 91.82; MAE 1.97 vs 2.07 and 2.02). Without significance testing or multiple runs, the claim that 'every component contributes to enhancing overall synthesis quality' is not established. Please provide statistical evidence or effect sizes for the ablation comparisons.","section":"Section 4, Table 3"}],"minor_comments":[{"comment":"The notation in Eq. (7) is difficult to parse: the expression 'f1 = [ Wi · sum ... ]' mixes a summation over k and a subscript i in a way that is not mathematically clean. Please rewrite this equation with clear indexing over the n source contrasts.","section":"Section 2, Eq. (7)"},{"comment":"The equations for the critic loss and adversarial loss contain unclear notation such as 'xpj t−1' and the use of eta; please clarify the dependence on j and define all variables before first use.","section":"Section 2, Eq. (10)"},{"comment":"The healthy dataset description says 'we extracted 100 middle slices from 85 healthy brain MRI scans' and then 'We allocated 50, 20 and 15 subjects for training, validation, and testing.' This is inconsistent: slices are counted but subjects are allocated, and 50+20+15=85. Please clarify whether the splits are per subject or per slice and reconcile the numbers.","section":"Section 3, Datasets"},{"comment":"The table reports MAE values for some baselines (e.g., MM-GAN on T2 and T1) that are several times larger than the PSNR values would suggest; please check whether these values are percentages or normalized and state the metric definitions explicitly in a footnote.","section":"Section 4, Table 1"},{"comment":"Reference [17] is formatted incorrectly as 'Al., M.'; the correct citation is to Menze et al., and the reference should be updated. Also, reference [25] lists 'Others Pytorch' and should be cleaned up.","section":"References"},{"comment":"The qualitative segmentation figure would benefit from clearly visible tumor boundaries and a longer caption explaining what is shown in each column, especially since the claimed segmentation improvement is central to the diagnostic-value argument.","section":"Section 4, Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially interesting and well-scoped MICCAI-style contribution, but the empirical claims currently outrun the evidence. The lack of reported p-values, confidence intervals, and error bars for the small margins over SynDiff and the 'Complete' segmentation setup is the main barrier. If the authors can supply the statistical detail and clarify the DCT implementation and training configurations, the paper could become acceptable. I would not reject outright, as the core idea is reasonable and the experimental design is otherwise standard."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious referee. It builds a genuinely new combination: two jointly trained denoising diffusion networks, one conditioned on DCT-based frequency features and one on spatial features, linked by a shared critic and an uncertainty-weighted mask loss. The components are known, but the integrated architecture is something I have not seen in the prior work they cite. They also report experiments on two datasets, a downstream tumor segmentation check, an ablation, and they release code. That is a solid, honest package.\n\nThe main empirical weakness is exactly what the stress-test note says. Table 1 claims p<0.05 from a paired t-test, but no p-values, confidence intervals, or test details are given. With 30 test subjects, the margins over SynDiff are often around 0.3-0.5 dB PSNR while the standard deviations are 2-2.7 dB, so the significance cannot be checked. Table 2 Dice scores have no error bars at all, and D2Diff's advantage over the Complete setup is 0.22 points, which is likely noise. That is a serious gap: the paper's central \"outperforms all methods\" claim rests on statistics the paper does not actually provide.\n\nThere are two smaller issues worth flagging. The DCT \"kernel sizes\" in Eq. 6 are ambiguous: DCT is normally a transform over the whole image or an image block, not a convolution with a kernel, and the paper never clarifies whether this is block-DCT with different block sizes or something else. Also, the BraTS T2 MAE value for D2Diff in Table 1 is 1.65±0.09, which has a standard deviation about a tenth the size of every other entry in that column; that looks like a typo and should be fixed. Baseline hyperparameters and training details are not reported, which makes the comparison harder to reproduce.\n\nThe design itself is fine. The dual-domain idea is not obviously flawed; even if the frequency features are not perfectly complementary, the framework still synthesizes images somewhat better than the strongest baseline in most rows. The fix is not conceptual, it is evidentiary: run proper paired tests, report p-values or CIs, add error bars to the Dice results, and clean up the table. I would accept this for peer review and push for a conditional acceptance after those revisions. If the statistics do not hold, the claim shrinks from \"superiority\" to \"a modest but possibly real improvement,\" which would still be a reasonable contribution to the MRI synthesis subfield.","headline":"A plausible new dual-domain diffusion architecture for multi-contrast MRI synthesis, with real but modest gains over SynDiff; the main weakness is that the headline superiority claim is not backed by reported statistics.","tokens_in":9598,"tokens_out":1461,"would_cite":true,"duration_ms":16477,"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":"Conditioning diffusion on frequency and spatial features improves multi-contrast MRI synthesis.","keywords":["multi-contrast MRI synthesis","diffusion models","dual-domain learning","frequency domain features","discrete cosine transform","uncertainty-aware mask loss","brain tumour imaging","medical image synthesis"],"falsifier":"Run the same training pipeline on the T1CE task with the frequency branch replaced by a second spatial branch of matched parameter count. If synthesis metrics do not drop meaningfully, the claim that frequency-domain features carry complementary information is not supported.","tokens_in":8576,"feed_emoji":"🧠","tokens_out":7183,"duration_ms":69243,"temperature":0.7,"pith_summary":"Multi-contrast MRI synthesis asks a model to create one contrast, such as T1CE or FLAIR, from the other available contrasts. This paper claims that the missing contrast is better estimated when denoising is conditioned on both spatial features and frequency features at once, rather than on spatial features alone. It proposes D2Diff, a diffusion model with two jointly trained denoising networks—one fed multi-scale DCT frequency features, the other fed spatial residual features—plus a shared critic that aligns them and an uncertainty mask that focuses learning on difficult regions. On brain tumour and healthy datasets, D2Diff reports higher PSNR and SSIM and lower MAE than the compared methods, and tumour segmentation on synthetic images stays close to segmentation on real images. The point of the claim is that dual-domain conditioning yields clinically usable synthetic contrasts, reducing scan time and contrast-agent exposure.","feed_headline":"Dual-domain diffusion sharpens synthetic MRI contrasts","feed_subtitle":"DCT frequency plus spatial features reduce synthesis error and keep tumour segmentation on par with real scans.","key_machinery":"The load-bearing mechanism is dual-domain conditioning inside a denoising diffusion process. The frequency branch computes DCT features with kernel sizes 3, 5, and 7 for each source contrast, weights them with a softmax attention module, and fuses them into a conditioning vector; the spatial branch extracts per-contrast residual features. Two encoder-decoder denoisers then reverse the diffusion process, one guided by each branch, while a shared critic provides adversarial consistency and uncertainty maps. During sampling, the two networks' posterior means are averaged, so both domains contribute to every denoising step instead of being fused only at the input level.","core_discovery":"The central discovery is that a diffusion model for multi-contrast MRI can be made more accurate by feeding two separate denoising networks complementary views of the conditioning contrasts: frequency-domain features extracted by discrete cosine transforms at three kernel scales, and spatial-domain residual features. The two networks are trained as a pair through a shared time-dependent critic, and at sampling their predicted denoised images are combined by averaging the posterior means. An uncertainty-aware mask loss derived from critic attention maps reweights the synthesis loss toward high-uncertainty regions. In the reported comparisons, the full D2Diff system outperforms the previous state-of-the-art baselines on all tested contrasts, and its synthetic images support downstream tumour segmentation with Dice scores comparable to those obtained from real contrasts.","pith_inferences":["The fixed DCT kernel sizes of 3, 5, and 7 are an arbitrary choice; learning the kernels or using a learned frequency transform might yield further gains, especially for 3D volume synthesis.","Because frequency features encode global intensity shifts, the frequency branch may also make the model more robust to scanner-to-scanner intensity variation; the paper's multi-site data hints at this but does not test it explicitly.","The same dual-domain conditioning could plausibly transfer to other synthesis settings, such as CT from MRI or missing-modality PET, where global intensity differences between modalities are also prominent."],"forward_implications":["If D2Diff works as reported, multi-contrast MRI synthesis improves across standard fidelity metrics, with the largest visible gains on challenging tumour-enhancement contrasts such as T1CE.","Downstream tumour segmentation using synthetic contrasts tracks segmentation on real contrasts, suggesting the synthetic images carry diagnostically relevant information rather than looking plausible only.","The mutual-learning design means each domain branch alone is weaker than the pair; ablation results show combined dual-domain training outperforms either individual branch.","The uncertainty mask loss and adaptive frequency aggregation each add measurable benefit, so the full architecture, not just the diffusion backbone, drives the result."],"supporting_citations":[{"why":"It defines the discrete cosine transform used for the frequency branch.","marker":"[2]"},{"why":"It supplies the forward and reverse diffusion formulations and the noise schedule.","marker":"[4]"},{"why":"It provides the Hi-Net baseline that the method must outperform.","marker":"[9]"},{"why":"It provides the multi-modal GAN baseline for comparison.","marker":"[12]"},{"why":"It supplies the denoising-diffusion-GAN formulation and the shared critic mechanism.","marker":"[13]"},{"why":"It is the adversarial-diffusion baseline compared against.","marker":"[16]"},{"why":"It supplies the brain tumour MRI images used for training and for the segmentation test.","marker":"[17]"},{"why":"It supplies the segmentation network used to measure the diagnostic value of synthetic images.","marker":"[22]"}],"fun_headline_variants":["Dual-domain diffusion boosts MRI synthesis accuracy","Spatial and DCT features improve synthetic MRI contrasts","Two-domain diffusion refines multi-contrast MRI synthesis","Frequency-spatial diffusion sharpens MRI contrast synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-chosen DCT frequency features are genuinely complementary to spatial features; if replacing the frequency branch with another spatial branch of the same size changes nothing, the dual-domain design has no real advantage.","fun_headline_variants_meta":{"raw":{"variants":["Dual-domain diffusion boosts MRI synthesis accuracy","Spatial and DCT features improve synthetic MRI contrasts","Two-domain diffusion refines multi-contrast MRI synthesis","Frequency-spatial diffusion sharpens MRI contrast synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1525,"prompt_tokens":863,"completion_tokens":662,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":600}},"tokens_in":479,"tokens_out":662,"duration_ms":6925,"temperature":1.0,"reasoning_tokens":600,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:41:45.936075+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training pipeline on the T1CE task with the frequency branch replaced by a second spatial branch of matched parameter count. If synthesis metrics do not drop meaningfully, the claim that frequency-domain features carry complementary information is not supported.","supporting_citations":[{"cited_title":"& Rao, K","cited_arxiv_id":null,"evidence_quote":"It defines the discrete cosine transform used for the frequency branch."},{"cited_title":"& Abbeel, P","cited_arxiv_id":null,"evidence_quote":"It supplies the forward and reverse diffusion formulations and the noise schedule."},{"cited_title":"& Shao, L","cited_arxiv_id":null,"evidence_quote":"It provides the Hi-Net baseline that the method must outperform."},{"cited_title":"& Çukur, T","cited_arxiv_id":null,"evidence_quote":"It is the adversarial-diffusion baseline compared against."}],"review_version":2}