{"id":"73bab00d-5bdb-4efd-bb0e-b3d163b49c4a","arxiv_id":"2505.00133","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"BlindHarmonyDiff converts MR images from unseen scanners to a target scanner style using a 3D edge-to-image rectified flow trained only on target data, plus a refinement step that preserves source structure.","lead":"This paper presents BlindHarmonyDiff, a 3D method that harmonizes MRI scans from different scanners by detecting edges in a source scan and generating a new scan that matches a target scanner's appearance. The goal is to make multi-scanner medical imaging datasets usable for machine learning without needing paired source-domain data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No quantitative test of the edge-invariance assumption (Eq. 2); the paper's own T2 conversion shows imperfect edges, and the refinement module is only a partial, unquantified safeguard.","rationale":"The reader's weakest_assumption identifies the same load-bearing risk: Eq. (2) may fail, and the paper provides only qualitative acknowledgment plus a refinement module that is not quantitatively validated for structural recovery. My stress-test agrees with that assessment and adds that the paper's own Supplementary S1 and T2-conversion results are internal evidence that the assumption can fail. A direct edge-map overlap measurement on the existing traveling-subject data would settle whether the assumption holds for the four T1 domains central to the headline claim. Since the reader's verdict is already CONDITIONAL and this concern does not change the assessment, the verdict should remain UNCHANGED. I do not find a separate, more damaging flaw: the ablations show the conditional edge-to-image model contributes beyond refinement, the 3D vs 2D comparison supports the inter-slice consistency claim, and the downstream task improvements are consistent with the method's mechanism.","tokens_in":17498,"tokens_out":5246,"duration_ms":61068,"concrete_test":"For all traveling-subject pairs used in Table 1, compute the Dice coefficient between the Canny edge maps D(xtar) and D(xsrc) using the authors' subject-level 8% threshold, separately for each source domain (Sonata, BioGraph, Vision, Vida). Then correlate per-subject edge Dice with per-subject PSNR/SSIM of the final BlindHarmonyDiff output. If any source domain has median edge Dice below, say, 0.9, or if low edge Dice is associated with low output fidelity, Eq. (2) fails in exactly the large-domain-gap regime the paper claims to handle. As a second check, repeat the T2-to-T1 experiment with this metric to quantify how far the admitted imperfect edge detection is from the assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that BlindHarmonyDiff can harmonize unseen source domains rests on Eq. (2): D(xtar) = D(xsrc) = e, with a subject-level Canny threshold fixed to 8% edge voxels. If this equality fails for a realistic source domain, the edge-to-image model receives structural information that is not the subject's true edge map, and the generated image inherits wrong structure. The paper explicitly acknowledges in Supplementary S1 that thresholding 'might occasionally result in inconsistent edge detection and produce poor edge-to-image model output' and in the Discussion that T2-to-T1 conversion exhibits 'imperfect edge detection' and 'limited' performance. Yet no experiment measures how often or how severely Eq. (2) is violated for the four T1 source domains used in Tables 1 and 2, nor how much the refinement module actually recovers when it is violated. The refinement module maximizes global normalized cross-correlation between the generated image and the source image; this is a whole-volume statistic and is not shown to restore a locally misplaced or missing edge. The lesion example in Fig. 6 is a single anecdote, not a quantitative test. Because the edge-invariance assumption is the load-bearing premise that distinguishes this method from unconditional generation or simple histogram matching, leaving it unmeasured leaves the central claim conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BlindHarmonyDiff, a 3D blind harmonization framework that trains a rectified-flow-based edge-to-image model on target-domain images only. At inference, it detects edges in a source-domain image and generates a harmonized image, followed by a refinement module that maximizes normalized cross-correlation with the source image to reduce hallucination. The authors also introduce multi-stride patch training for efficient 3D training. Experiments on OASIS3 with four source scanners report PSNR/SSIM improvements over prior blind methods and favorable downstream segmentation and age-prediction results.","tokens_in":17760,"tokens_out":5620,"duration_ms":52442,"significance":"If the central edge-invariance assumption holds, the method is a meaningful advance: it provides a 3D, low-cost blind harmonization approach that does not require source-domain training data, with clear ablations (multi-stride ratio, refinement, conditional vs unconditional, 3D vs 2D) and downstream validation. The paper's strengths include systematic ablations and the novel use of edge-to-image generation for harmonization. However, the load-bearing assumption that edge maps are scanner-invariant is not directly measured, and the quantitative claims lack variance analysis.","major_comments":[{"comment":"The central assumption D(xtar)=D(xsrc)=e is load-bearing but is not directly verified for the four T1-weighted source domains in Tables 1 and 2. The paper itself notes in Supplementary S1 that the 8% edge-voxel threshold 'might occasionally result in inconsistent edge detection and produce poor edge-to-image model output' and in the Discussion that the T2-to-T1 conversion exhibits 'imperfect edge detection' and 'limited' performance. Because the harmonized image is generated solely from D(xsrc), any failure of Eq. (2) propagates into the output, and the refinement module is only a partial safeguard. I request a quantitative evaluation of the assumption: report edge-map agreement (e.g., Dice or Jaccard) between D(xsrc) and D(xtar) on registered traveling subjects per source domain, and show how harmonization fidelity varies with that agreement. Without this, the claim that the method generalizes to unseen source domains remains conditional.","section":"§3.1, Eq. (2); Supplementary S1"},{"comment":"All quantitative results are reported as point estimates without error bars, confidence intervals, or statistical tests. The number of image-level test subjects varies from 7 to 30 across source domains (Table S1), so the reported average differences (e.g., Table 1: BlindHarmonyDiff PSNR 23.4 vs. histogram matching 21.8 on Sonata) may not be significant. The downstream results in Table 2 likewise lack per-subject variance. To support the claim of 'superior performance across scanners', please provide per-subject distributions and paired significance tests (e.g., Wilcoxon signed-rank) against the strongest baseline for each metric, and report the number of subjects in each table.","section":"Tables 1-4, Table S1"},{"comment":"The refinement module is an inference-time gradient ascent on the normalized cross-correlation between the harmonized image and the source image. The authors do not report how the step size (0.02) and the number of iterations (6) were chosen, nor how sensitive the results are to them. More importantly, the objective is defined with respect to the source domain, and the paper argues that source and target images are highly correlated in low frequencies; the reported PSNR/SSIM gains in Table 4 show empirical benefit, but a mechanistic analysis (e.g., comparing the frequency spectra of the refined output against target-domain statistics) would clarify why maximizing source NCC improves target fidelity. Without this, the refinement module risks being seen as a source-domain post-processing rather than a target-domain harmonization step.","section":"§3.5, Table 4"}],"minor_comments":[{"comment":"The threshold adjustment procedure is described only informally ('if the ratio ... exceeds the predetermined value, the threshold is decremented'); please specify the iterative algorithm and convergence criterion.","section":"§3.2"},{"comment":"The text says 'rectified flow [29]', but reference [29] is 'Flow matching for generative modeling' (Lipman et al.); the rectified flow method appears in reference [31] (Liu et al.). Please correct the citation.","section":"§3.3"},{"comment":"The curve has no error bars; please add inter-subject variability. Also clarify the definition of 'multi-stride patch ratio' in the caption.","section":"Figure 2"},{"comment":"The use of E2I(D(xsrc), θ*) omits the coordinate conditions (i,j,k) that are part of the model input in Eq. (6); please align notation.","section":"Equation (4)"},{"comment":"Add footnotes indicating the number of test subjects per domain and the metric computation mask (brain mask) for reproducibility.","section":"Tables 1 and 2"},{"comment":"The caption contains a typo: 'BlindHarmonDiff' should be 'BlindHarmonyDiff'.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's main novelty is the application of edge-to-image rectified flow to blind 3D harmonization with a refinement module. The scientific contribution is incremental but potentially solid if the edge-invariance assumption is validated. The citation mismatch for rectified flow and the absence of any code/data release are additional concerns for reproducibility. The paper seems within scope for a medical imaging journal; the main risk is the untested central assumption."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this is a solid applied paper that deserves a proper review. What's new is the specific combination—a 3D edge-to-image rectified flow trained only on target data, with multi-stride patch training and an NCC-based refinement module—applied to blind harmonization. The ablations are informative and mostly support the design choices. The paper is also unusually honest about its limitations, especially for T2-to-T1 conversion.\n\nWhat works: the 3D approach removes the inter-slice heterogeneity you get with 2D slice-wise harmonization. The multi-stride patch trick is simple and appears to give a real gain over plain patch training (PSNR 27.5 vs 26.9 on the target domain reconstruction). The refinement module, maximizing normalized cross-correlation between the harmonized image and the source, clearly helps preserve lesions and improves quantitative metrics. The downstream task tests—tissue segmentation and age prediction—are a good addition, because PSNR/SSIM alone are weak proxies for clinical utility.\n\nThe soft spots are real but not disqualifying. The tables report point estimates only. No error bars, no significance tests. Some of the reported gains are modest—for BioGraph, the Dice improvement over no harmonization is 0.872 vs 0.870, and the age-prediction MAE drops from 5.02 to 4.91. Those numbers could easily be noise. The central assumption, that Canny edges at an 8% threshold are stable across scanners for the same subject, is not directly validated. The authors acknowledge in the supplement that thresholding can occasionally give inconsistent edges, but they never measure how often that happens or how much the refinement module actually recovers. The refinement module is a global NCC match; it's not shown to restore a locally missing or misplaced edge. The T2-to-T1 example is a single anecdote, and the authors themselves say performance there is limited.\n\nThe citation pattern looks fine; the relationship to the authors' prior BlindHarmony work is clearly stated. No code or model is released, which is a limitation for reproducibility but not a fatal one.\n\nWho is this for? Anyone working on multi-site MRI pooling or scanner-invariant analysis. It gives a practical, credible recipe that outperforms previous blind harmonization methods on the tested T1 datasets. It deserves a serious referee. My recommendation: send it to review, but ask the authors for confidence intervals, a code release or model weights, and a direct test of the edge-invariance assumption—ideally on a held-out contrast or a set of subjects with more than one scan.","headline":"Solid 3D blind harmonization paper with a load-bearing edge-invariance assumption that is acknowledged but not directly tested; deserves review with requests for statistical rigor.","tokens_in":18283,"tokens_out":2704,"would_cite":true,"duration_ms":26077,"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":"Edges alone can bridge large MRI domain gaps in 3D, using only target-domain training data and a correlation-based refinement step.","keywords":["blind harmonization","MRI harmonization","edge-to-image model","rectified flow","3D medical imaging","domain gap","Canny edge detection","hallucination suppression"],"falsifier":"Compute the Dice coefficient between subject-level 8% Canny edge maps from the same subject scanned on two different scanners; if the average Dice is substantially below 1 in fine-structure or lesion regions, the edge-invariance identity that the pipeline rests on fails for a meaningful share of inputs, and the harmonized image inherits the discrepancy.","tokens_in":17282,"feed_emoji":"🧠","tokens_out":6319,"duration_ms":55567,"temperature":0.7,"pith_summary":"BlindHarmonyDiff claims that an MRI scan from an unseen scanner can be harmonized to a target scanner's style using only a subject-level edge map, without ever seeing source-domain data during training. The paper trains a 3D rectified-flow model on target-domain volumes to reconstruct a full image from its Canny edge map, then applies that model to the edges of source-domain scans. A refinement step that raises normalized cross-correlation with the source scan is shown to be indispensable for keeping lesions and other small structures intact. On T1-weighted traveling subjects across four source scanners and one target scanner, the method reports higher PSNR and SSIM and better downstream segmentation and age-prediction performance than prior blind harmonization methods.","feed_headline":"Edges alone can bridge large MRI domain gaps in 3D","feed_subtitle":"A target-only flow model turns source edges into harmonized volumes, beating prior blind harmonization.","key_machinery":"The load-bearing object is the edge-to-image model: a 3D rectified flow (straight-line transport between Gaussian noise and image patches, Eq. 5) conditioned on a Canny edge map patch and normalized spatial coordinates, trained only on target-domain volumes. It operationalizes the edge-invariance identity $D(x_{\\mathrm{tar}})=D(x_{\\mathrm{src}})=e$ and generates $x_{\\mathrm{har}}=\\mathrm{E2I}(D(x_{\\mathrm{src}}),\\theta^*)$. Supporting machinery includes subject-level Canny thresholding that fixes 8% of voxels as edge voxels, multi-stride patch training that downsamples sub-volumes to $64^3$ patches to blend local and global context under memory constraints, and a refinement module that iteratively raises normalized cross-correlation between the harmonized and source images to suppress hallucination and preserve structures.","core_discovery":"The central discovery is that a single edge-to-image model trained only on target-domain images, together with a correlation-based refinement step, can harmonize images from multiple unseen scanners, including large domain gaps that break earlier blind methods. Concretely, the paper formulates the same subject scanned in two domains as sharing an edge map: $D(x_{\\mathrm{tar}})=D(x_{\\mathrm{src}})=e$, so generating an image from the source edges with a flow model trained on target edges yields a target-contrast image. BlindHarmonyDiff implements this with a 3D rectified flow trained on $64^3$ patches with multi-stride sampling to capture global context, and a refinement module that performs gradient ascent on the normalized cross-correlation between the harmonized volume and the source volume. The paper reports that this 3D processing removes inter-slice heterogeneity, that the refinement module is essential for reliable outputs, and that the method outperforms prior blind harmonization on PSNR and SSIM as well as downstream tissue segmentation and age prediction on four scanners.","pith_inferences":["Because the method's success rests on edges surviving across scanners, a natural stress test is to evaluate it on datasets with pathologies where Canny edges are ambiguous; the refinement module may not rescue all such cases.","The edge-invariance identity could be tested directly by measuring edge-map agreement across paired traveling-subject scans; if agreement is high only for large structures, the method's robustness may degrade at fine anatomical detail.","The same training recipe could extend to other contrasts or modalities, such as CT to MR or T2 to T1, if a contrast-invariant structural representation replaces Canny edges; the paper's own T2-to-T1 result suggests this is currently limited by edge detection.","Multi-stride patch training is a general scheme: since it outperforms whole-image training in the paper's 2D ablation, it may benefit other 3D diffusion tasks beyond harmonization."],"forward_implications":["Harmonization no longer requires paired or source-domain training data, so a model built once for a target scanner can be applied to any new scanner without retraining.","3D processing removes inter-slice discontinuities that plague slice-wise harmonization, so downstream volumetric analysis sees more homogeneous inputs.","The refinement module's correlation step prevents the generator from deleting or inventing small structures such as lesions, which is a precondition for clinical use.","For same-contrast multi-scanner harmonization, the reported PSNR, SSIM, and downstream-task results approach supervised methods trained on each source domain, suggesting blind harmonization is a practical substitute when source data are scarce.","The same edge-to-image formulation can be pointed at a different target domain by retraining only on that domain, making the method a reusable harmonization building block."],"supporting_citations":[{"why":"Establishes the blind-harmonization setting and the observation that same-subject MR images share edges while differing in low-frequency contrast.","marker":"[23]"},{"why":"Provides the rectified flow framework used to learn the 3D edge-to-image transport.","marker":"[29]"},{"why":"Defines the Canny edge detector whose subject-level 8% threshold implements the edge-invariance assumption.","marker":"[8]"},{"why":"Supplies the coordinate-encoded patch training recipe adapted for memory-efficient 3D diffusion training.","marker":"[6]"},{"why":"Contributes the patch-diffusion strategy that multi-stride patch training extends to capture global context.","marker":"[49]"},{"why":"Provides the multi-scanner traveling-subject dataset used for training, validation, and the reported harmonization and downstream evaluations.","marker":"[27]"},{"why":"Serves as the supervised DeepHarmony baseline whose target-domain correspondence the blind method is compared against.","marker":"[12]"},{"why":"Provides the spectrum-swapping SSIMH baseline for image-level harmonization comparison.","marker":"[20]"}],"fun_headline_variants":["Target-only edges harmonize any MRI, even large 3D gaps","No source data: edge-flow harmonizes 3D MRI across large gaps","Edge-to-image flow: target-only model beats big MRI domain gaps","Edge maps from any scanner become target MRI in 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the same subject yields the same Canny edge map across scanners once the threshold is set to keep 8% of voxels as edges; if a source scan's edges differ from the target scan's edges, the harmonized image inherits wrong structure and the refinement module can only partially fix it.","fun_headline_variants_meta":{"raw":{"variants":["Target-only edges harmonize any MRI, even large 3D gaps","No source data: edge-flow harmonizes 3D MRI across large gaps","Edge-to-image flow: target-only model beats big MRI domain gaps","Edge maps from any scanner become target MRI in 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00159,"raw_usage":{"total_tokens":6350,"prompt_tokens":967,"completion_tokens":5383,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":5318}},"tokens_in":583,"tokens_out":5383,"duration_ms":36555,"temperature":1.0,"reasoning_tokens":5318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:50:12.911852+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the Dice coefficient between subject-level 8% Canny edge maps from the same subject scanned on two different scanners; if the average Dice is substantially below 1 in fine-structure or lesion regions, the edge-invariance identity that the pipeline rests on fails for a meaningful share of inputs, and the harmonized image inherits the discrepancy.","supporting_citations":[{"cited_title":"Blindharmony:” blind” harmonization for mr images via flow model","cited_arxiv_id":null,"evidence_quote":"Establishes the blind-harmonization setting and the observation that same-subject MR images share edges while differing in low-frequency contrast."},{"cited_title":"Flow matching for generative modeling","cited_arxiv_id":null,"evidence_quote":"Provides the rectified flow framework used to learn the 3D edge-to-image transport."},{"cited_title":"A computational approach to edge detection","cited_arxiv_id":null,"evidence_quote":"Defines the Canny edge detector whose subject-level 8% threshold implements the edge-invariance assumption."},{"cited_title":"Memory-efficient 3d de- noising diffusion models for medical image processing","cited_arxiv_id":null,"evidence_quote":"Supplies the coordinate-encoded patch training recipe adapted for memory-efficient 3D diffusion training."},{"cited_title":"Patch diffusion: Faster and more data- efficient training of diffusion models","cited_arxiv_id":null,"evidence_quote":"Contributes the patch-diffusion strategy that multi-stride patch training extends to capture global context."},{"cited_title":"Oasis-3: longitudinal neuroimaging, clin- ical, and cognitive dataset for normal aging and alzheimer disease","cited_arxiv_id":null,"evidence_quote":"Provides the multi-scanner traveling-subject dataset used for training, validation, and the reported harmonization and downstream evaluations."},{"cited_title":"Deepharmony: A deep learning approach to contrast harmonization across scanner changes","cited_arxiv_id":null,"evidence_quote":"Serves as the supervised DeepHarmony baseline whose target-domain correspondence the blind method is compared against."},{"cited_title":"Fast image-level mri harmonization via spec- trum analysis","cited_arxiv_id":null,"evidence_quote":"Provides the spectrum-swapping SSIMH baseline for image-level harmonization comparison."}],"review_version":1}