{"id":"3930e89f-f362-4119-bbb3-6bc94b675bd6","arxiv_id":"2608.08173","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DisMorph decomposes longitudinal brain MRI deformations into separate distortion and atrophy fields, reducing atrophy estimation error in the presence of scanner distortion and attributing distortion-only image pairs to the distortion field.","lead":"A brain MRI registration model trained purely on synthetic images learns to separate each scan-to-scan difference into two parts: scanner distortion and true tissue change. The goal is to keep small biological changes, such as early Alzheimer's atrophy, from being masked or mimicked by hardware artifacts in longitudinal studies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No experiment validates the two-field split on real images with both distortion and atrophy present; fBIRN lacks atrophy and ADNI lacks ground truth, so the central claim rests on an untested synthetic prior.","rationale":"The paper's design is strong in several respects: the fBIRN distortion-only specificity test is a clever use of a real technical artifact, and the simulated experiment with an external biophysical atrophy model provides a controlled check of the anatomical stream. However, the central claim of disentangling real technical and anatomical deformation is only tested when one of the two effects is absent, or when the distortion is drawn from the same generative model used for training. The ADNI analysis, while suggestive, cannot detect misattribution because both predicted fields are free to vary. The proposed mixed experiment would close the gap by using real GNL distortion while providing independent ground truth for atrophy. Because the reader already assigned CONDITIONAL and this concern is consistent with that verdict, no change in verdict is needed. This is an untested-assumption objection rather than evidence that the method is wrong, so it does not warrant rejection.","tokens_in":7866,"tokens_out":9791,"duration_ms":102323,"concrete_test":"Create mixed real/simulated pairs by applying known biophysical atrophy (as in Sec. 3.1) to the 18 fBIRN distortion-corrected images, then adding the real mri_gradunwarp distortion; run DisMorph and verify that the atrophy field recovers the prescribed volume changes within the simulation error bounds and that the distortion field matches the known mri_gradunwarp displacement field (e.g., mean endpoint error below 1 mm across the brain). If either check fails, the decomposition is not validated when both effects are present.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that, for real longitudinal pairs containing both GNL distortion and biological change, DisMorph's predicted fields isolate the two sources. The only real-data experiment with both effects present (ADNI, Sec. 3.2) has no ground truth for either field, and the fBIRN experiment (Sec. 3.2) has real distortion but no atrophy, while never comparing the predicted distortion field to the known mri_gradunwarp field. The simulated experiment (Sec. 3.1) uses distortion generated by the same elastic transform as training (Sec. 2.1), so it cannot detect a miscalibrated synthetic prior. If real residual GNL differs from the training distribution in smoothness, magnitude, or spectral content, the quarter-resolution distortion stream can absorb smooth components of atrophy or miss high-frequency components of distortion, biasing both fields. The authors' own skull-stripping example shows that out-of-distribution technical artifacts are misassigned, demonstrating that the split is not robust beyond the training prior. Without a real-data mixed test, the claimed separation is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DisMorph, a learning-based deformable registration framework trained entirely on synthetic data that predicts two dense deformation fields—one intended to capture scanner-induced technical distortion (notably gradient non-linearity, GNL) and one intended to capture anatomical change. Training uses a procedural generator that applies a smooth random elastic transform to model distortion and label-boundary reassignment to model atrophy, with domain randomization over image intensities and corruptions. The method is evaluated in three settings: synthetic data with known atrophy generated by an external biophysical model; real fBIRN image pairs that differ only by GNL distortion; and real ADNI longitudinal pairs from Alzheimer's disease subjects. The results suggest that using only the anatomical field reduces atrophy estimation error relative to a single composite field in simulation, and that most distortion-only real pairs are attributed to the distortion field, while ADNI pairs show plausible atrophy in subcortical structures.","tokens_in":8019,"tokens_out":4699,"duration_ms":51702,"significance":"If the two-field decomposition is reliable, DisMorph would be a practically valuable tool for longitudinal morphometry in multi-site, retrospective, or protocol-changing studies where residual GNL distortion confounds atrophy estimates. The paper has notable strengths: the simulated atrophy ground truth comes from an independent biophysical model rather than the training generator; the fBIRN distortion-only experiment is a genuine out-of-distribution specificity test on real data; and the synthetic training strategy with domain randomization is a reasonable way to obtain dense supervision for two deformation fields. However, the central claim—that DisMorph separates true biological change from technical distortion on real images containing both effects—is not yet established. The current real-data evaluation lacks ground truth for the decomposition, and the simulation uses a distortion generator that appears to match the training distribution. The contribution is promising but needs additional validation before the claimed separation can be considered supported.","major_comments":[{"comment":"The central claim that DisMorph separates technical distortion from biological change in real longitudinal pairs containing both is not established. The ADNI experiment (Fig. 5B) has no ground truth for either the distortion field or the atrophy field, so the reported decomposition cannot be verified. The fBIRN experiment (Fig. 5A) uses real pairs that differ only by GNL distortion, but because no atrophy is present, it cannot test whether the model misattributes real anatomical change to the distortion field when both effects coexist. A real-data mixed test is needed, for example by comparing the predicted distortion field against a known mri_gradunwarp correction on fBIRN pairs while simultaneously measuring atrophy on images with known disease-related changes, or by constructing a synthetic mixed test with a held-out distortion generator.","section":"Sec. 3.2, Fig. 5"},{"comment":"The simulated experiment in Sec. 3.1 is likely in-distribution for the distortion prior. The training data in Sec. 2.1 synthesize GNL distortion with smooth random elastic transforms sampled on a coarse control lattice, and the simulation in Sec. 3.1 is described only as 'adding distortion' without stating whether it uses a different generator. If the same generator is used, the experiment cannot detect a miscalibrated synthetic prior: the model may have memorized the specific smoothness and magnitude distribution of the training distortions. The paper should explicitly state the relationship between the training and simulation distortion generators and should add out-of-distribution tests, such as distortion fields derived from a measured or biophysical GNL model, or with smoothness and magnitude ranges outside the training support, to demonstrate that the decomposition generalizes beyond the training distribution.","section":"Sec. 2.1 vs. Sec. 3.1"},{"comment":"The training objective supervises the distortion field through L_d and the composed deformation through L_a, but it does not supervise the anatomical field independently. Because the only label losses are on the intermediate distorted label map and the final distorted-and-atrophied label map, any pair (phi_d, phi_a) that preserves the composition phi_da is equally consistent with the label losses. The smoothness regularizer is applied to the composed displacement u, not to the individual fields, and the quarter-resolution distortion stream versus half-resolution anatomical stream is an architectural prior rather than a validated physical constraint. As a result, the split between v_d and v_a may be driven by inductive biases rather than by the data. The paper should provide a sensitivity analysis varying distortion smoothness and magnitude and should compare the predicted distortion field against a known reference such as mri_gradunwarp output to demonstrate that the decomposition is not arbitrary.","section":"Sec. 2.2, Eq. (2)"},{"comment":"The skull-stripping inconsistency in the fBIRN example is explicitly assigned to the atrophy field and described as 'apparent atrophy.' This is a concrete case of an out-of-distribution technical artifact being misattributed to biological change, which undermines the claim of specificity beyond the training prior. The manuscript should quantify how frequently such misassignments occur across the fBIRN pairs and discuss the implications for analyses of cortical or boundary-adjacent structures where preprocessing inconsistencies are common. Without such quantification, the example suggests a known failure mode of the proposed decomposition rather than an isolated anecdote.","section":"Sec. 3.2, Fig. 6 (left)"}],"minor_comments":[{"comment":"The notation in Eq. (1) is slightly confusing: the text introduces intensity images I_F and I_M, but the loss is defined on one-hot encoded label maps. Clarify that Eq. (1) is computed on label representations, not on intensity images.","section":"Sec. 2.2, Eq. (1)"},{"comment":"The results in Figs. 4 and 5 are presented without confidence intervals or statistical tests. Given the small sample sizes (18 fBIRN pairs and 40 ADNI pairs), error bars or a paired statistical comparison would strengthen the claims of consistent improvement and specificity.","section":"Fig. 4 and Fig. 5"},{"comment":"The metric AAE is used in the text and Fig. 4 but is never explicitly defined. Please provide the formula relating AAE to the prescribed and estimated ASPC values.","section":"Sec. 3.1, Metrics"},{"comment":"The phrase 'lattice resolutions between 2 and 10 control points for each dimension' is ambiguous; clarify whether the number of control points is sampled independently per dimension and whether this refers to the spacing or the count.","section":"Sec. 2.1, Geometry generator"},{"comment":"The claim that DisMorph is 'the first registration framework' that explicitly decomposes longitudinal deformation into technical and anatomical transforms is strong and is not substantiated by a literature survey. Soften the claim or provide a focused comparison with prior works that estimate separate deformation components.","section":"Conclusion, Sec. 4"},{"comment":"The fBIRN experiment uses mri_gradunwarp with the manufacturer's coefficient files, but no details are given about the scanner model, version of mri_gradunwarp, or whether the same coefficients were used across subjects. These details would help readers assess the reliability of the distortion reference.","section":"Sec. 3.2, Data"}],"recommendation":"major_revision","confidential_remarks":"The paper does not mention release of code, trained models, or the synthetic training pipeline. Given that the method is defined entirely by the synthetic data distribution, releasing the generator and training configuration would be important for reproducibility. The topic is well within the scope of an imaging journal, but the validation gap described in the major comments should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. DisMorph's core idea is genuinely new: a registration network trained on synthetic data where GNL distortion and atrophy are applied separately, with separate losses supervising two output fields, and the claim that using only the anatomical field improves atrophy estimation. The simulation with Khanal's biophysical model is a real strength — independent ground truth for atrophy, and the AAE reduction is consistent across atrophy levels. The fBIRN experiment is a nice specificity check: real distortion-only pairs, and the model puts most change into the distortion field. I also read the skull-stripping misassignment as a mark of honesty, not just a failure.\n\nNow the soft spots. There are no error bars or significance tests on any of the quantitative figures, so we don't know how stable the gains are. More fundamentally, the split into distortion and atrophy fields is only identifiable through an explicit smoothness/scale prior — distortion at quarter resolution, atrophy at half resolution — and there's no sensitivity analysis around that architectural choice. The simulated distorted pairs in Sec. 3.1 appear to use the same elastic transform distribution as training (Sec. 2.1), making the distortion-recovery test partly in-distribution. And on real data, fBIRN has no atrophy and ADNI has no ground truth for either field, so the central claim — that the split holds when both effects co-occur on real scans — is plausible but not demonstrated. The paper also doesn't compare its predicted distortion field to the known mri_gradunwarp field in fBIRN, which would be a cheap and strong validation.\n\nThe citation pattern is fine; the SynthMorph line of work is properly credited, and the external biophysical model is a good choice. No code is released, which limits reproducibility.\n\nOverall, this is a solid, honest paper with a real contribution to longitudinal registration. A serious referee could push for the evidence to match the claim: error bars, sensitivity analysis, the mri_gradunwarp comparison, and ideally a mixed real-data test with an independent reference. I'd send it out.","headline":"New disentanglement idea with a solid synthetic check; real-data evidence for the split is still thin, but it deserves refereeing.","tokens_in":8596,"tokens_out":3178,"would_cite":false,"duration_ms":34886,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DisMorph trains a registration network on synthetic images to factor each longitudinal brain scan pair into a scanner-distortion field and a biological-change field, and using only the anatomical field lowers atrophy-estimation error.","keywords":["longitudinal morphometry","deformable image registration","gradient non-linearity distortion","disentanglement","synthetic training data","domain randomization","neuroimaging","Alzheimer's disease"],"falsifier":"Construct a distortion-only pair by applying mri_gradunwarp to one scan and leaving the other original, as in the paper's fBIRN setup; if the predicted anatomical field then produces statistically significant volume change in structures, the split is misattributing technical distortion as biology.","tokens_in":7604,"feed_emoji":"🧠","tokens_out":6826,"duration_ms":63473,"temperature":0.7,"pith_summary":"DisMorph is a registration framework that takes a longitudinal pair of brain MR scans and predicts two separate deformation fields: one for scanner-induced geometric distortion (primarily gradient non-linearity) and one for biological change such as atrophy. The model is trained entirely on synthetic images in which distortion and atrophy are generated independently, giving it per-field supervision without needing real ground-truth pairs. On simulated data with known atrophy, computing volume change from the predicted anatomical field alone reduces estimation error compared with a single composite deformation. On real distortion-only pairs, the model assigns most geometric change to the distortion field, and on longitudinal Alzheimer's disease pairs it isolates AD-related volume change in the anatomical field while capturing residual distortion. If the decomposition holds, it would make longitudinal morphometry more accurate in multi-site or retrospective studies where vendor distortion correction is incomplete.","feed_headline":"Registration model isolates scanner distortion from real brain change","feed_subtitle":"Trained only on synthetic images, DisMorph cuts atrophy-estimation error and flags residual distortion in real MRI.","key_machinery":"The central mechanism is the factored deformation model φ_da = φ_d ∘ φ_a, in which a dual-stream encoder with a pyramid-warping decoder predicts two stationary velocity fields that are integrated via scaling and squaring into diffeomorphic transforms. The network predicts the distortion field at quarter resolution and the anatomical field at half resolution, encoding the prior that GNL distortion is smoother than atrophy. Training supervises each field separately with the synthetic ground truth: the loss is the sum of a label-matching term for the distortion field, a label-matching term for the composed deformation, and a smoothness regularizer. The synthetic generator provides the disentanglement supervision: an elastic lattice deformation models GNL distortion, and iterative boundary reassignment between adjacent labels models local atrophy or growth, followed by an intensity generator that adds bias fields, blurring, noise, and contrast variation for domain randomization.","core_discovery":"The paper's central claim is that the longitudinal deformation between two scans can be factorized into a technical transform and an anatomical transform, and that a network trained purely on synthetic data can recover both components separately. DisMorph outputs two stationary velocity fields, integrated by scaling and squaring into diffeomorphisms φ_d and φ_a, whose composition φ_da = φ_d ∘ φ_a aligns the moving image with the fixed image. The synthetic training pipeline constructs each effect independently—smooth random elastic fields for GNL distortion and boundary-reassignment for atrophy—so each predicted field has a direct ground-truth supervision target. With this split, the anatomical field alone yields lower absolute atrophy estimation error than the composite field on simulated data across low, moderate, and high atrophy levels, and on real fBIRN pairs that differ only by known GNL distortion the model attributes the majority of geometric change to the distortion field. On two-year ADNI Alzheimer's disease pairs, the anatomical field captures the expected subcortical volume loss while the distortion field absorbs remaining scanner-induced misalignment.","pith_inferences":["If the scale-separation prior is correct, one testable extension is to compare DisMorph's distortion field against direct phantom-based GNL measurements on the same scans; agreement would confirm that the smooth-field prior captures the true physical distortion.","The factorization assumes distortion and atrophy compose sequentially; a natural extension would be to allow the distortion field to depend on tissue geometry, which the current generative model does not represent.","The method could be applied across sites as a harmonization step: estimating each site's distortion signature from pairs of repeated scans and removing it before pooling longitudinal data, a use the paper does not explicitly discuss."],"forward_implications":["Atrophy estimates can be computed from the predicted anatomical field alone, removing distortion-related bias without modifying the scanner or acquisition protocol.","The distortion field itself can serve as a quality-control signal, flagging residual GNL distortion in datasets that were assumed corrected, as the method does for ADNI.","Because training needs only synthetic data, the approach can be extended to other distortion sources such as B0 inhomogeneity or chemical shift by adding corresponding generative models.","The largest relative improvement appears at low atrophy levels, meaning early-disease or slow-progressing studies benefit most from the decomposition.","The two-field output preserves diffeomorphic smoothness for both components, so downstream morphometric measures remain well-defined."],"supporting_citations":[{"why":"Supplies the scaling-and-squaring integration used to convert predicted velocity fields into diffeomorphic deformations.","marker":"[3]"},{"why":"Documents GNL distortion as the dominant technical artifact and the standard correction approach this method targets.","marker":"[30]"},{"why":"Provides the synthesis-based registration paradigm and intensity generator that DisMorph builds on.","marker":"[22]"},{"why":"Provides the label-to-intensity synthesis workflow for creating MRI-like contrast from label maps.","marker":"[9]"},{"why":"Supplies the biophysical deformation model used to generate ground-truth longitudinal atrophy.","marker":"[32]"},{"why":"Defines the symmetric percent-change metric used to evaluate volume change.","marker":"[38]"},{"why":"Provides the ADNI cohort used for real longitudinal evaluation.","marker":"[28]"},{"why":"Provides the fBIRN dataset of real images differing only by distortion.","marker":"[31]"}],"fun_headline_variants":["DisMorph splits scan warps into distortion and true change","Two warps, one truth: DisMorph isolates real brain change","Synthetic-only AI untangles MRI distortion from brain atrophy","DisMorph separates scanner distortion from true brain change"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Real residual GNL distortion is smooth enough to be captured by the quarter-resolution random elastic fields used in training, and longitudinal deformation factorizes exactly as a distortion step followed by an anatomical step.","fun_headline_variants_meta":{"raw":{"variants":["DisMorph splits scan warps into distortion and true change","Two warps, one truth: DisMorph isolates real brain change","Synthetic-only AI untangles MRI distortion from brain atrophy","DisMorph separates scanner distortion from true brain change"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000595,"raw_usage":{"total_tokens":2829,"prompt_tokens":1030,"completion_tokens":1799,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":1729}},"tokens_in":646,"tokens_out":1799,"duration_ms":13787,"temperature":1.0,"reasoning_tokens":1729,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:18:23.193512+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a distortion-only pair by applying mri_gradunwarp to one scan and leaving the other original, as in the paper's fBIRN setup; if the predicted anatomical field then produces statistically significant volume change in structures, the split is misattributing technical distortion as biology.","supporting_citations":[{"cited_title":"In: MICCAI 2006","cited_arxiv_id":null,"evidence_quote":"Supplies the scaling-and-squaring integration used to convert predicted velocity fields into diffeomorphic deformations."},{"cited_title":"NeuroImage30(2), 436–443 (2006)","cited_arxiv_id":null,"evidence_quote":"Documents GNL distortion as the dominant technical artifact and the standard correction approach this method targets."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Provides the synthesis-based registration paradigm and intensity generator that DisMorph builds on."},{"cited_title":"Medical Image Analysis86, 102789 (2023)","cited_arxiv_id":null,"evidence_quote":"Provides the label-to-intensity synthesis workflow for creating MRI-like contrast from label maps."},{"cited_title":"NeuroImage134, 35–52 (2016)","cited_arxiv_id":null,"evidence_quote":"Supplies the biophysical deformation model used to generate ground-truth longitudinal atrophy."},{"cited_title":"NeuroImage61(4), 1402–1418 (2012)","cited_arxiv_id":null,"evidence_quote":"Defines the symmetric percent-change metric used to evaluate volume change."},{"cited_title":"J Magn Reson Imaging27(4), 685–691 (2008)","cited_arxiv_id":null,"evidence_quote":"Provides the ADNI cohort used for real longitudinal evaluation."},{"cited_title":"NeuroImage124, 1074–1079 (2016)","cited_arxiv_id":null,"evidence_quote":"Provides the fBIRN dataset of real images differing only by distortion."}],"review_version":1}