{"id":"5ee4168d-d156-4d94-b63f-6cacaeedc62f","arxiv_id":"2412.13294","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GeoReg uses position-aware cross-attention as continuous deformation functions in a multi-resolution decoder, achieving state-of-the-art OCT registration and on-par brain MRI registration with 741k parameters.","lead":"This paper introduces GeoReg, a deformable image registration method that models image alignment with continuous geometric deep learning operations and refines deformations in a coarse-to-fine manner without feature resampling. It reports the best accuracy on longitudinal retinal OCT registration and competitive results on brain MRI, using a model far smaller than transformer baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's 'significant improvement' for brain registration is contradicted by the paper's own 'on par' conclusion and Table 1; no significance tests support the claim.","rationale":"The reader's weakest_assumption about integer rounding is a legitimate secondary issue, but it is not the most load-bearing problem. Rounding the target neighborhood center to the nearest integer grid index (Appendix 10) means sub-voxel displacements do not change the set of sampled neighbor features; however, the architecture still evaluates positional embeddings at continuous relative coordinates, so the claim of avoiding explicit resampling to a regular grid is weakened but not outright falsified. The more decisive issue is the abstract's assertion of 'significant improvement' for brain registration, which is directly contradicted by the paper's own Section 4.2 and Conclusion stating 'on par' performance. Table 1 shows overlapping standard deviations, mixed metric direction (worse HD95 on T1w, worse folding), and no statistical tests. This is an internal inconsistency, not a disagreement with external consensus, and it bears directly on the central claim's empirical validity. The proposed paired permutation test would settle whether the brain improvement is real or only noise. If the test shows significance, my objection is resolved; if not, the abstract must be weakened. Since the OCT results remain strong and the reader already issued CONDITIONAL based on addressable concerns, my analysis does not change the verdict.","tokens_in":22149,"tokens_out":5852,"duration_ms":57653,"concrete_test":"Using the released GeoReg code and the standard CamCAN split, compute per-subject DSC and HD95 for GeoReg, Transmorph, and LapIRN on the T1w-T1w and T1w-T2w test sets, then run paired permutation tests (or Wilcoxon signed-rank tests) between GeoReg and each baseline, reporting p-values and 95% confidence intervals. If p≥0.05 on both brain tasks, the abstract's 'significant improvement over state-of-the-art' for brain registration is unsupported, and the paper should be revised to claim only on-par brain performance plus OCT improvements.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—'significant improvement over state-of-the-art approaches for both mono- and multi-modal inter-subject brain registration' (Abstract)—is not supported by the paper's own results. Section 4.2 states that on both brain tasks the method 'shows on par performance', and the Conclusion repeats this. Table 1 shows GeoReg DSC 0.838±0.06 vs Transmorph 0.822±0.01 on T1w brain, and DSC 0.788±0.05 vs 0.788±0.02 on T1w-T2w, with HD95 2.05±0.82 vs 1.92±0.52 (GeoReg worse on T1w HD95) and higher folding (0.09 vs 0.02). No paired significance tests, confidence intervals, or effect sizes are reported, and standard deviations overlap. In contrast to the reader's weakest assumption, the integer-grid rounding described in Appendix 10 does not by itself falsify the 'no resampling' claim, because the position-aware embeddings still use continuous relative coordinates; the unsupported significance claim is more load-bearing. If it fails, the central claim reduces to OCT-only, and the abstract's headline claim must be revised to match the paper's own 'on par' wording.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"GeoReg is a deformable image registration framework built on geometric deep learning principles. It uses a dual-stream encoder to extract independent source/target feature pyramids, then refines deformations in a coarse-to-fine manner with a position-aware cross-attention deformation function τ evaluated on continuous coordinates, together with a learned cross-resolution interpolation function δ. The paper claims that this spatially continuous formulation avoids resampling to a regular grid between successive refinements, and it evaluates the method on CamCAN T1w/T1w-T2w inter-subject brain registration, longitudinal retinal OCT registration, and synthetic large-deformation tasks, with code released publicly.","tokens_in":22367,"tokens_out":4579,"duration_ms":43579,"significance":"The geometric-deep-learning framing is timely and the architecture is genuinely parameter-efficient (741k parameters versus 46.8M for Transmorph). The qualitative scale-separation analysis in Figures 3-4 is a useful interpretability contribution, and the synthetic large-deformation experiments with ground-truth transformations in Table 3 are a valuable stress test in which GeoReg is consistently strong and nearly folding-free. If the empirical claims are properly supported, the paper would make a meaningful contribution to interpretable and data-efficient registration. However, as written, the headline claim in the abstract is not supported by the paper's own reported brain results, and the continuous-neighborhood claim is only approximately realized in the implementation; both need to be addressed before publication.","major_comments":[{"comment":"The abstract claims \"significant improvement in performance metrics over state-of-the-art approaches for both mono- and multi-modal inter-subject brain registration,\" but Section 4.2 states that on both brain tasks the method \"shows on par performance,\" and Table 1 shows overlapping standard deviations on every brain metric. For T1w brain, GeoReg's DSC is 0.838±0.06 versus Transmorph's 0.822±0.01, its HD95 is 2.05±0.82 versus 1.92±0.52, and its folding is 0.09 versus 0.02; on T1w-T2w the DSC values are identical at 0.788. No paired significance tests, confidence intervals, or effect sizes are reported. Please either add a proper statistical comparison (paired tests across subjects, with test-set sizes and confidence intervals) or revise the abstract and conclusion to state that brain performance is on par with, not significantly better than, the state of the art.","section":"Abstract, Section 4.2, Table 1"},{"comment":"The \"spatially continuous\" claim is weakened by the implementation detail in Appendix 10: target neighborhoods are obtained by \"mapping the current source coordinates into the index space of the target grid and rounding to the closest integer.\" This means that sub-voxel updates of the transformed position φ_n(x) do not change the set of target features seen by the deformation function; the continuous relative-coordinate embedding does not eliminate the discretization introduced by nearest-neighbor rounding. Please clarify how much of the claimed advantage over feature warping remains under this rounding, and ideally provide an ablation comparing against trilinear feature interpolation or soft/continuous neighborhood weighting.","section":"Appendix 10, Section 3.2"},{"comment":"The OCT improvement is also not statistically supported. GeoReg's DSC is 0.581±0.09 versus Transmorph's 0.577±0.12 and HD95 is 1.72±1.72 versus 1.95±2.54, all from a single 80/10/10 split with no significance tests. Given the large standard deviations, the reported differences may be within noise. Please report the number of test volumes, paired significance tests, and confidence intervals for all datasets.","section":"Section 4.2, Table 1 (OCT)"},{"comment":"The architectural axiom that features need not be propagated across decoder levels is not isolated by the \"feat. warp\" ablation. Replacing δ with bilinear feature warping changes several factors at once: how features are transported, whether neighborhoods are integer-rounded, and what information is available to the interpolation function. To support the claim in Section 3.3 that local features at each resolution are sufficient, please add an ablation that propagates or concatenates feature vectors across decoder levels while keeping the rest of the design fixed.","section":"Section 3.3, Table 1 (feat. warp ablation)"},{"comment":"LapIRN results are entirely missing from Table 3 (all entries are \"− ± −\") with no explanation, despite LapIRN being a key multi-resolution baseline in Table 1. Without these entries, or an explicit statement of why LapIRN could not be evaluated, the claim that GeoReg \"consistently outperforms other baselines\" on the synthetic large-deformation tasks is not fully supported.","section":"Table 3"}],"minor_comments":[{"comment":"The word \"correspondances\" should be \"correspondences\".","section":"Section 3.4"},{"comment":"The phrase \"utilize bending energy as a regularize\" should read \"as a regularizer\".","section":"Appendix 10"},{"comment":"Table 1 contains stray spaces in several numerical entries (e.g., \"5 .46\", \"2 .55\", \"2 .05\") and inconsistent decimal formatting; please clean up the table.","section":"Table 1"},{"comment":"The resolution-specific weights α_r and the regularization weight λ are not specified in the paper; please report their values for reproducibility.","section":"Equation (10), Appendix 10"},{"comment":"The caption of Figure 16 lists method names without a clear correspondence to panels; please label the subfigures and describe the layout.","section":"Figure 16"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The manuscript presents a solid technical idea with a useful interpretability analysis and an impressively parameter-efficient architecture. The main problems are claim calibration and statistical support: the abstract overstates the brain results relative to the paper's own conclusions, and the continuous-neighborhood claim is only approximately realized in the implementation. These issues are fixable within the scope of the paper, so I recommend major revision rather than rejection. Please also verify that the anonymous code repository and any hyperparameter settings are available at the camera-ready stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read this one. The architecture is worth your time: 741k parameters, a coherent geometric-deep-learning framing for deformable registration, and genuinely strong results on large synthetic deformations. But the abstract claims 'significant improvement' on brain registration, and the paper's own results say 'on par' — twice, in Section 4.2 and the Conclusion. No significance tests appear anywhere. Read the body, not the abstract.\n\nWhat is actually new: the combination of dual-stream encoders, coarse-to-fine refinement via position-aware cross-attention, and a learned cross-resolution interpolation δ that replaces feature warping. Each piece is known, but the assembly is new for volumetric registration. The parameter efficiency is real — roughly 60x smaller than Transmorph while staying in the same performance band on brain MRI. The synthetic deformation experiments are the strongest evidence: at ±45° rotation, GeoReg's AEE is about 2.0 versus 26 for Transmorph, with zero folding. That is a substantial result, and the MNIST scale-separation figures are a useful qualitative addition.\n\nThe soft spots, in order. First, the abstract/body contradiction. Table 1 backs the body: GeoReg is better on T1w DSC (0.838 vs 0.822) but worse on HD95 and folding, and T1w-T2w is a tie. With no paired tests and overlapping standard deviations, 'significant improvement' is not defensible. Either the abstract gets rewritten to match the body, or the claim needs statistical support. Second, the 'spatially continuous' framing is approximate: Appendix 10 is candid that target neighborhoods are fetched by rounding the float coordinate to the nearest integer grid index. That is a real discretization step. I part ways with the reader's weakest-assumption here — the position embeddings still encode continuous relative coordinates, so the operation is more continuous than feature warping, and the rounding does not by itself collapse the no-resampling claim. But the paper should say 'approximately continuous' and justify the rounding. Third, the OCT outperformance rests on thin means (DSC 0.581 vs 0.577, overlapping SDs); the qualitative figures and synthetic experiments carry that claim more than the numbers do. Minor items: LapIRN is missing from the synthetic table, some hyperparameters live only in the anonymous repo, and the Appendix 6 'proof' is just distributivity of the dot product.\n\nTo the paper's credit, it discloses its costs — the Conclusion admits the memory footprint, and Table 2 confirms GeoReg uses more VRAM than Transmorph.\n\nWho this is for: people working on parameter-efficient registration or GDL-based medical imaging. My recommendation: send it to a serious referee. The fixes — abstract rewrite, significance tests, qualifying 'continuous' — are honest revisions, not a rebuild.","headline":"Genuinely novel, parameter-efficient registration architecture with strong synthetic-deformation results, but the abstract overclaims brain improvements that the paper's own body reports as 'on par' and no significance tests support.","tokens_in":22926,"tokens_out":6760,"would_cite":true,"duration_ms":59518,"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":"GeoReg models image registration as continuous-coordinate cross-attention, avoiding feature resampling between refinement steps.","keywords":["deformable image registration","geometric deep learning","position-aware cross-attention","coarse-to-fine refinement","continuous coordinates","multi-modal registration","longitudinal retinal OCT","interpretable registration"],"falsifier":"A controlled ablation would settle the central claim: replace the round-to-nearest neighborhood lookup with true continuous sampling, or conversely replace the learned interpolation $\\delta$ with bilinear feature warping while holding everything else fixed, and measure the same Dice and Hausdorff metrics on the brain and OCT splits. If accuracy does not change, the continuous-coordinate mechanism is not the source of the gains.","tokens_in":21912,"feed_emoji":"🧠","tokens_out":7621,"duration_ms":67713,"temperature":0.7,"pith_summary":"This paper argues that deformable image registration should be built from operations that explicitly reason about source and target coordinate systems, rather than treating images as fixed grids and warping features back to a grid at every refinement. It introduces GeoReg, a dual-stream encoder with a coarse-to-fine decoder in which a deformation function predicts displacement for each point by cross-attending to neighbors in the source and target domains, with positions encoded as continuous relative coordinates. The central claim is that this spatially continuous formulation avoids interpolation errors from repeated feature resampling and yields registration that is more interpretable and parameter-efficient. On the paper's experiments, the 741k-parameter model reports improved or on-par accuracy across mono- and multi-modal brain MRI and longitudinal retinal OCT registration, with the largest gains on OCT and on large synthetic deformations.","feed_headline":"Continuous coords make a 741k-parameter model competitive","feed_subtitle":"Cross-attention on continuous coordinates skips feature resampling between refinement steps and wins on retinal OCT.","key_machinery":"The load-bearing object is the deformation function $\\tau$, implemented as position-aware cross-attention. For a source point with feature $f$, the query is $q = f W_Q$, and keys and values are built from neighboring source or target features $F_N$ added to a Fourier positional embedding $E(X_N - \\phi_n(x))$ of their coordinates relative to the point's current transformed position $\\phi_n(x)$; a softmax-weighted combination yields the displacement update. Because neighbor contributions are weighted by continuous relative coordinates rather than fixed kernel positions, the function can be applied repeatedly at floating-point locations without resampling to a regular grid. A second learned function $\\delta$ interpolates deformations across resolutions by letting child points cross-attend to parent control points in the coarser resolution, which the paper argues preserves sharp transformation boundaries that naive interpolation would smooth.","core_discovery":"On its own terms, the paper's discovery is that deformation refinement can be cast as repeated position-aware cross-attention over continuous coordinates, so that successive refinements do not require resampling features to a regular grid. The authors show that a small model can capture the majority of a transformation at coarse resolutions, with finer decoder levels acting largely as interpolation, and that this scale-separated design remains end-to-end trainable with supervision at every resolution. They report that this formulation matches or exceeds larger baselines on inter-subject brain registration and outperforms them on longitudinal retinal OCT, while also recovering large synthetic affine-plus-noise deformations that other learned methods fail on.","pith_inferences":["One consequence the authors leave implicit is that their implementation gathers target neighborhoods by rounding the current floating-point coordinate to the nearest grid index, so the 'continuous' advantage is only approximate; isolating true continuous sampling from round-to-nearest would clarify which design choice drives the reported gains.","A natural testable extension is to apply the same $\\tau$ and $\\delta$ formulation to non-grid domains such as cortical surfaces or point clouds, where rounding to a grid is impossible and the geometric deep learning framing is native.","The observed scale separation suggests the multi-resolution objective could be tuned with different similarity or regularization weights per resolution, a knob the paper discusses qualitatively but does not investigate quantitatively."],"forward_implications":["Registration models can refine transformations without warping feature vectors or re-encoding images between steps, removing a source of interpolation error and repeated encoding cost.","Feature vectors do not need to be carried across decoder levels; only transformation vectors are passed between resolutions, simplifying the architecture.","Because the majority of the deformation is assigned to coarse resolutions, finer decoder levels can be cheap interpolation-only layers.","Supervising every refinement and interpolation step propagates meaningful gradients through the whole pipeline and acts as an implicit regularizer on featureless regions.","Encoding positions as continuous coordinates means the same formulation can in principle adapt to variable grid spacing, which the authors suggest as relevant to anisotropic data."],"supporting_citations":[{"why":"Supplies the continuous-convolution formulation that the paper's deformation function generalizes.","marker":"[4]"},{"why":"Is the single-stream baseline that motivates separating feature extraction from deformation modeling.","marker":"[3]"},{"why":"Is the transformer baseline with 46.8M parameters that GeoReg is compared against.","marker":"[5]"},{"why":"Provides the coarse-to-fine multi-resolution baseline (LapIRN) that the decoder schema refines.","marker":"[28]"},{"why":"Documents the curse of dimensionality in high-dimensional interpolation, the core reason the paper avoids feature warping.","marker":"[41]"},{"why":"Provides the Fourier feature positional embedding used to encode relative coordinates in the cross-attention.","marker":"[39]"},{"why":"Is the dual-stream pyramid baseline that warps features and that the paper's decoding scheme contrasts with.","marker":"[20]"},{"why":"Is the recursive cascaded baseline that motivates avoiding iterative re-encoding and warping between steps.","marker":"[44]"}],"fun_headline_variants":["Continuous coords skip resampling in deformable registration","Small model with cross-attention beats larger registration methods","Geometric deep learning makes registration interpretable and efficient","Coarse-to-fine refinement without resampling wins on brain and retina","Cross-attention on continuous coordinates improves deformable registration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that evaluating the deformation function at continuous coordinates actually provides the target information it is supposed to, since neighborhoods are gathered by rounding the current floating-point coordinate to the nearest grid index; if sub-voxel motion does not change the gathered neighbors, the claimed advantage over feature warping would rest on the learned cross-resolution interpolation and supervision scheme instead.","fun_headline_variants_meta":{"raw":{"variants":["Continuous coords skip resampling in deformable registration","Small model with cross-attention beats larger registration methods","Geometric deep learning makes registration interpretable and efficient","Coarse-to-fine refinement without resampling wins on brain and retina","Cross-attention on continuous coordinates improves deformable registration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000702,"raw_usage":{"total_tokens":3146,"prompt_tokens":898,"completion_tokens":2248,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":2168}},"tokens_in":514,"tokens_out":2248,"duration_ms":16028,"temperature":1.0,"reasoning_tokens":2168,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:15:54.680482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled ablation would settle the central claim: replace the round-to-nearest neighborhood lookup with true continuous sampling, or conversely replace the learned interpolation $\\delta$ with bilinear feature warping while holding everything else fixed, and measure the same Dice and Hausdorff metrics on the brain and OCT splits. If accuracy does not change, the continuous-coordinate mechanism is not the source of the gains.","supporting_citations":[{"cited_title":"Balakrishnan, A","cited_arxiv_id":null,"evidence_quote":"Is the single-stream baseline that motivates separating feature extraction from deformation modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the transformer baseline with 46.8M parameters that GeoReg is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the coarse-to-fine multi-resolution baseline (LapIRN) that the decoder schema refines."},{"cited_title":"Verleysen and D","cited_arxiv_id":null,"evidence_quote":"Documents the curse of dimensionality in high-dimensional interpolation, the core reason the paper avoids feature warping."},{"cited_title":"Tancik, P","cited_arxiv_id":null,"evidence_quote":"Provides the Fourier feature positional embedding used to encode relative coordinates in the cross-attention."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the dual-stream pyramid baseline that warps features and that the paper's decoding scheme contrasts with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the recursive cascaded baseline that motivates avoiding iterative re-encoding and warping between steps."}],"review_version":1}