{"id":"4e96d618-8829-4752-8daf-7f795300dba7","arxiv_id":"2412.20822","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding gradient correlation to the TransMorph loss and fine-tuning with FAdam yields smoother deformations for interpatient brain MRI registration, with small gains in Dice and HdDist95.","lead":"This challenge paper fine-tunes the TransMorph brain MRI registration model with the FAdam optimizer and a gradient-correlation similarity term. The combination slightly improves Dice and boundary distance scores and reduces non-diffeomorphic volume, a measure of deformation smoothness.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The NDV improvement is confounded: λ=2 doubles the diffusion-regularization weight relative to baseline, and checkpoints are selected by lowest weighted evaluation loss, so the smoother deformations may be a regularization-weight/selection artifact rather than an effect of the GC term.","rationale":"The reader's conditional verdict is appropriate: the paper reports a plausible but under-supported improvement, and the absence of uncertainty quantification and an unspecified evaluation split are real weaknesses. My stress-test sharpens the reader's concern in two ways. First, the loss-weight choice λ=2 is not merely a hyperparameter; it doubles the diffusion regularizer relative to the baseline, providing a direct alternative explanation for the NDV reduction that does not involve GC at all. Second, the checkpoint selection rule ('lowest evaluation loss') is likely to select for the same weighted objective, compounding the bias. The reader mentioned hand-set weights and checkpoint selection, so there is partial agreement, but the specific mechanism linking λ=2 to NDV is not made explicit in the reader's weakest_assumption. The proposed ablation is a single, cheap experiment that would settle whether the GC term itself contributes to smoothness. If the ablation shows no NDV benefit at λ=1, the central claim of the paper would be unsupported; if the benefit persists, the current conditional verdict could be upgraded. For now, UNCHANGED reflects that the reader's CONDITIONAL verdict already accommodates this unresolved confound.","tokens_in":4264,"tokens_out":3304,"duration_ms":34842,"concrete_test":"Run a controlled ablation on the same evaluation split: (a) FAdam+GC with the baseline regularizer weight λ=1 instead of λ=2, (b) FAdam without GC at λ=2, and (c) FAdam+GC at λ=2 but with checkpoints selected by validation Dice/TRE instead of lowest weighted evaluation loss. If FAdam+GC at λ=1 does not retain most of the NDV reduction relative to Baseline, the smoother-deformation claim is a regularization-weight or checkpoint-selection artifact, not evidence for GC. Report NDV with per-subject bootstrap confidence intervals or paired significance tests.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that adding gradient correlation (GC) improves deformation smoothness is not identifiable from the reported comparison. In Eq. 5, L = Lsim + λLreg, and Section 2.2 states λ = 2 'to account for the similarity loss being now about double compared to the baseline.' This implies the baseline used λ = 1 (or at least a smaller λ), so FAdam+GC is trained with twice the diffusion-regularization weight as Baseline. Moreover, both fine-tuned models are selected by 'the lowest evaluation loss' at epochs 123 and 170, and that evaluation loss presumably includes the same weighted regularizer. Thus the NDV drop from 0.3509% to 0.2690% could be caused by the stronger regularization pressure or by checkpoint selection favoring low-NDV solutions, independently of the GC term. The FAdam-only row (NDV 0.3549) does not control for this because its λ and checkpoint rule are not stated to match the FAdam+GC run. Additionally, no variance or significance is reported for NDV, TRE, or HdDist95, so the 0.08 percentage-point NDV gap cannot be separated from noise. A decisive ablation separating λ, GC, and checkpoint selection is required before the smoothness gain can be attributed to GC.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fine-tuning strategy for the TransMorph deformable registration model on the Learn2Reg 2024 LUMIR brain MRI dataset. Two modifications are introduced: replacing the Adam optimizer with FAdam, and adding a gradient correlation (GC) term to the image similarity loss. The authors compare a baseline TransMorph model, a model fine-tuned with FAdam only, and a model fine-tuned with FAdam plus GC. They report Dice, TRE, HdDist95, and NDV scores, finding slight Dice/HdDist95 improvements and a notable reduction in NDV for the FAdam+GC model, and they illustrate qualitative differences on a single slice. The central claim is that the GC term improves anatomical alignment and deformation smoothness.","tokens_in":4565,"tokens_out":3078,"duration_ms":31148,"significance":"If the observed NDV reduction is real and attributable to the GC term, the paper offers a simple, low-cost way to improve deformation plausibility in a strong pretrained registration model, which would be a useful contribution to the Learn2Reg challenge community. The paper is concise and the idea is easy to replicate, and the use of a pretrained baseline is a practical, environmentally motivated choice. The main weakness is that the evidence for the central claim is not statistically grounded: there are no variance or significance measures for NDV, TRE, or HdDist95, and the comparison is confounded by a doubled diffusion-regularization weight and an unspecified checkpoint-selection rule. The qualitative analysis is limited to one slice, and the GC improvement is partly circular, since GC is part of the optimized loss.","major_comments":[{"comment":"The central claim that FAdam+GC achieves a notable NDV reduction from 0.3509% to 0.2690% is not supported by any measure of variance or statistical significance. NDV is reported as a single scalar without a standard deviation, and the same is true for TRE and HdDist95. The difference is 0.082 percentage points, which may fall within the noise envelope of the evaluation set. Please provide per-subject distributions for all metrics and a paired significance test (e.g., Wilcoxon signed-rank test) or confidence intervals across the evaluation set for the FAdam+GC versus Baseline and FAdam+GC versus FAdam comparisons. Without this, the 'notable reduction' asserted in the abstract and the conclusion is not established.","section":"Section 3.1, Table 1"},{"comment":"The NDV improvement is confounded by the change in the diffusion regularization weight lambda. The text states lambda = 2 'to account for the similarity loss being now about double compared to the baseline,' implying the baseline used a smaller lambda (likely 1). Since the diffusion regularizer directly penalizes non-smooth deformations, the lower NDV of FAdam+GC could be caused by the stronger regularization pressure rather than by the GC term. The FAdam-only row does not control for this because its lambda is not stated and its checkpoint selection follows the same loss-based rule. A decisive ablation is required: train the baseline and FAdam-only models with lambda = 2 and the same checkpoint-selection rule, and train FAdam+GC with lambda = 1, so that the effect of GC can be isolated from the regularization weight.","section":"Section 2.2, Eq. (5) and Section 3.1"},{"comment":"The checkpoint selection criterion is described only as 'the version with the lowest evaluation loss at 123 and 170 epochs, respectively.' It is not specified which split (training, validation, or test) was used for this loss, nor whether the evaluation loss includes the full objective L = L_sim + lambda * L_reg. If the selection loss includes the lambda-weighted regularizer and the GC term, then the selection procedure itself may favor low-NDV and high-GC solutions, independently of any genuine improvement. Please state the split, the selection metric, and the epoch range searched, and ideally select checkpoints by a metric independent of the training objective (e.g., NDV or Dice on a held-out set).","section":"Section 3.1, checkpoint selection"},{"comment":"The qualitative evidence is based on a single slice (slice 70) of a single image pair (3457 and 3456). The reported GC values of 0.521 (baseline) and 0.528 (FAdam+GC) are used as supporting evidence, but GC is part of the training objective for FAdam+GC, so improving GC on the training distribution is expected and does not independently validate the method. Please report GC on the full evaluation set or a held-out set, and quantify the artifact reduction beyond the one highlighted slice. A single-slice visual inspection is too anecdotal to support the claim of improved boundary alignment.","section":"Section 3.2, Figure 3"}],"minor_comments":[{"comment":"The table header shows 'HdDist95 ↑', but HdDist95 is a distance measure where lower is better, and the text in Section 3.1 states 'HdDist95 ↓'. Please correct the arrow direction.","section":"Section 3.1, Table 1 header"},{"comment":"The term 'Normalized Gradient Cross Correlation' is used for GC, but the formula is an average of NCC of gradients. Consider renaming to 'Gradient Correlation' (as in the title) and defining NCC clearly; the current notation may be confused with 'normalized cross-correlation' of the images themselves.","section":"Section 2.2, Eq. (1)"},{"comment":"The caption contains 'cfg. figure 2', which appears to be a typo for 'cf. Figure 2'. Please fix.","section":"Figure 2 caption"},{"comment":"The description of FAdam as incorporating 'principles from natural gradient descent and Riemannian geometry' is vague. Consider adding a brief technical detail or equation from the FAdam paper to make the modification concrete, since the paper is otherwise short.","section":"Section 1/2.1, FAdam description"},{"comment":"The paper states the LUMIR dataset contains over 4,000 scans but does not specify how many were used for fine-tuning versus evaluation, nor whether the evaluation pairs are the same across the three models. Please clarify the exact data split and the number of test cases for each metric.","section":"Section 3.1, dataset split"}],"recommendation":"major_revision","confidential_remarks":"This is a short, well-written challenge paper with a simple and practical idea. The main scientific issue is the missing causal attribution of the NDV improvement to the GC term: the doubled lambda and the loss-based checkpoint selection are strong confounds. The requested ablation (varying lambda and checkpoint rule) is well within the scope of the work and can be done with the existing codebase and data. The lack of variance measures is also easily fixable by reporting per-subject results. I would be comfortable accepting a revised version that resolves these points, but the current evidence does not support the abstract's 'notable reduction' claim with the necessary rigor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read of the six-pager: the new thing is an empirical result—fine-tuning a pre-trained TransMorph with FAdam and a gradient-correlation term on the LUMIR challenge data gives a modest Dice/HdDist95 bump and an NDV drop from 0.351% to 0.269%. The methods section is clear, the metrics are standard, and the authors are honest that the Dice/TRE gains are tiny. That honesty matters.\n\nThe soft spot is exactly where the stress-test lands. Eq. 5 sets λ=2, and the text says this was done to account for the similarity loss being 'about double compared to the baseline.' So the FAdam+GC run is trained with twice the diffusion-regularization weight as the baseline (and, presumably, as the FAdam-only run, though that row's λ isn't stated). The NDV improvement is therefore a regularization-weight/selection artifact candidate, not clean evidence for GC. On top of that, both fine-tuned checkpoints were chosen by 'lowest evaluation loss' at epochs 123 and 170, with no variance or significance reported for NDV, TRE, or HdDist95. The 0.08-percentage-point NDV gap is within the noise envelope as far as the paper shows. A single-slice qualitative figure doesn't close that gap.\n\nThat said, I don't think this is a dishonest paper. It reads like a competent challenge report, the kind that usually ships with a clear ablation. The missing piece is exactly one: a λ-matched control (GC at λ=1, or λ=2 without GC) plus confidence intervals. Without that, the central claim is not identifiable from the reported comparison.\n\nWho gets value: readers tracking the Learn2Reg LUMIR leaderboard or experimenting with TransMorph fine-tuning recipes. It won't change how I set up registration experiments, but it's a useful data point if someone wants to chase smoother deformations. I'd send it to review—not because it's a breakthrough, but because the confound is fixable and the result, once cleaned up, is worth having on record. I would not cite it in its current form.","headline":"Short challenge report claiming FAdam+GC fine-tuning smooths TransMorph deformations; the NDV gain is real but confounded by a doubled regularization weight and checkpoint selection, so the headline claim isn't identifiable.","tokens_in":5060,"tokens_out":2840,"would_cite":false,"duration_ms":27589,"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":"Fine-tuning TransMorph with FAdam and a gradient-correlation similarity term improves anatomical alignment and reduces non-diffeomorphic volume in unsupervised brain MRI registration.","keywords":["deformable image registration","TransMorph","gradient correlation","FAdam","brain MRI","unsupervised registration","non-diffeomorphic volume","Learn2Reg LUMIR"],"falsifier":"Run the same fine-tuning protocol multiple times with different random seeds, and for each seed select the checkpoint by lowest evaluation loss; if the distributions of NDV for baseline and FAdam+GC overlap substantially, the claimed smoothness improvement is not established.","tokens_in":4059,"feed_emoji":"🧠","tokens_out":7239,"duration_ms":64381,"temperature":0.7,"pith_summary":"Unsupervised brain MRI registration must balance segmentation accuracy against deformation smoothness. This paper proposes fine-tuning a pre-trained TransMorph model with two modifications: the FAdam optimizer and a gradient-correlation (GC) term added to the similarity loss. On Learn2Reg 2024 LUMIR data, the GC-augmented model matches the baseline in landmark accuracy, improves Dice and 95% Hausdorff distance slightly, and cuts non-diffeomorphic volume from 0.3509% to 0.2690%. The authors read the NDV reduction as evidence of smoother, more anatomically plausible deformations, and a qualitative slice comparison supports better alignment at tissue boundaries with fewer artifacts. The contribution is a targeted fine-tuning recipe that improves deformation quality without retraining from scratch.","feed_headline":"Adding gradient correlation cuts folding in brain MRI registration","feed_subtitle":"Fine-tuning TransMorph with FAdam plus GC lowers non-diffeomorphic volume from 0.35% to 0.27%.","key_machinery":"The load-bearing object is the gradient-correlation term $L_{GC}=1-\\mathrm{GC}(I_t, I_s \\circ \\phi)$, where $\\mathrm{GC}$ is the average normalized cross-correlation of the $x$-, $y$-, and $z$-gradients of the fixed and warped images. This is added to the local normalized cross-correlation intensity loss as $L_{sim}=L_{IC}+\\gamma L_{GC}$ with $\\gamma=0.5$, and the total objective becomes $L=L_{sim}+\\lambda L_{reg}$ with $\\lambda=2$, where $L_{reg}$ is the diffusion regularizer from TransMorph. The other component is FAdam, an Adam variant that uses natural-gradient and Riemannian-geometry corrections intended to improve convergence stability. The GC term carries the smoothness claim: by rewarding aligned image gradients rather than only intensities, it pushes the deformation to keep structural boundaries consistent.","core_discovery":"The central claim is that adding gradient correlation to the image similarity loss, together with FAdam-based fine-tuning, produces deformation fields that are substantially smoother than the TransMorph baseline while preserving segmentation and landmark accuracy. Quantitatively, Dice rises from 0.7594 to 0.7614, HdDist95 falls from 3.5074 to 3.4899, and NDV falls from 0.3509% to 0.2690%; TRE changes from 2.4225 mm to 2.4599 mm. The authors attribute the smoothness gain to the GC term, since the FAdam-only model has NDV 0.3549%, close to baseline.","pith_inferences":["An untested implication is that the gradient-correlation term behaves as an implicit regularizer on deformation topology: aligning image gradients may suppress folding directly, so sweeping $\\gamma$ would map a dose-response curve between the GC weight and NDV.","The same loss modification could transfer to other intensity-only registration networks, not only TransMorph; the paper only demonstrates it on the provided baseline.","A practical next step the paper leaves open is multi-seed fine-tuning with statistical comparison, since the magnitude of the NDV gain relative to run-to-run variation is not quantified."],"forward_implications":["On the LUMIR evaluation, FAdam+GC achieves Dice 0.7614 and HdDist95 3.4899, slightly better than baseline 0.7594 and 3.5074.","NDV drops from 0.3509% to 0.2690%, indicating that the deformation fields contain fewer non-diffeomorphic, folding voxels.","The improvement is concentrated at tissue boundaries where image gradients are high, which is consistent with the HdDist95 gain.","Landmark accuracy is not sacrificed: TRE is 2.4599 mm for FAdam+GC versus 2.4225 mm for baseline.","FAdam alone does not explain the smoothness effect, as the FAdam-only model has NDV 0.3549%, essentially unchanged from baseline."],"supporting_citations":[{"why":"Supplies the TransMorph architecture and the pre-trained weights used as the starting point for fine-tuning.","marker":"[1]"},{"why":"Defines the Learn2Reg 2024 LUMIR challenge and its evaluation setting.","marker":"[2]"},{"why":"Introduces the gradient consistency loss that the gradient-correlation term builds on.","marker":"[4]"},{"why":"Introduces the FAdam optimizer used for the fine-tuning.","marker":"[5]"},{"why":"Defines the non-diffeomorphic volume computation used as the smoothness metric.","marker":"[6]"},{"why":"Provides the original gradient-correlation similarity measure for registration.","marker":"[8]"},{"why":"Provides anatomical fiducial landmarks used for the TRE evaluation.","marker":"[9]"}],"fun_headline_variants":["Gradient correlation fine-tuning reduces brain MRI folding by 23%","Adding GC to TransMorph yields smoother deformations in brain registration","GC improves brain MRI alignment with less folding, better Dice","FAdam plus gradient correlation smooths TransMorph for brain MRI","Gradient correlation cuts non-diffeomorphic volume in brain MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the NDV decrease from 0.3509% to 0.2690% is a real effect of the gradient-correlation loss, not an artifact of the chosen checkpoint or of run-to-run variation; the paper reports no uncertainty or significance testing for NDV, TRE, or HdDist95.","fun_headline_variants_meta":{"raw":{"variants":["Gradient correlation fine-tuning reduces brain MRI folding by 23%","Adding GC to TransMorph yields smoother deformations in brain registration","GC improves brain MRI alignment with less folding, better Dice","FAdam plus gradient correlation smooths TransMorph for brain MRI","Gradient correlation cuts non-diffeomorphic volume in brain MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001054,"raw_usage":{"total_tokens":4364,"prompt_tokens":825,"completion_tokens":3539,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":3451}},"tokens_in":441,"tokens_out":3539,"duration_ms":25255,"temperature":1.0,"reasoning_tokens":3451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:08:29.958351+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same fine-tuning protocol multiple times with different random seeds, and for each seed select the checkpoint by lowest evaluation loss; if the distributions of NDV for baseline and FAdam+GC overlap substantially, the claimed smoothness improvement is not established.","supporting_citations":[{"cited_title":"doi:10.5281/ZENODO.10991880, https://zenodo.org/doi/10.5281/zenodo.10991880","cited_arxiv_id":null,"evidence_quote":"Defines the Learn2Reg 2024 LUMIR challenge and its evaluation setting."},{"cited_title":"International Journal of Computer Vision 132(9), 3678--3688 (Sep 2024)","cited_arxiv_id":null,"evidence_quote":"Defines the non-diffeomorphic volume computation used as the smoothness metric."},{"cited_title":"Scientific Data 10(1), 449 (Jul 2023)","cited_arxiv_id":null,"evidence_quote":"Provides anatomical fiducial landmarks used for the TRE evaluation."}],"review_version":1}