{"id":"fcb97c2f-83f2-4fd7-b1c3-880aa439dd75","arxiv_id":"1908.06194","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Conv2Warp uses deformable convolutions and a learnable Catmull-Rom spline resampler in an unsupervised network to match or improve on traditional deformable registration accuracy on lung CT and brain MRI datasets.","lead":"This paper introduces Conv2Warp, a deep learning method for deformable medical image registration that combines deformable convolutions with a learnable Catmull-Rom spline resampler to compute smooth deformation fields. The method matches the accuracy of established tools like SimpleElastix on lung CT and brain MRI while running much faster.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The training loss in Eq. (2) is not equivalent to NCC because it uses absolute values, so the central unsupervised objective rests on a false mathematical claim.","rationale":"The reader's weakest assumption is exactly the loss function issue: the paper claims a derived SSD that 'guarantees maximization of NCC' but provides no derivation, and the absolute-value term is not the standard centered z-score. I agree this is the most load-bearing technical concern because it directly affects the training objective, and therefore every reported accuracy number. A wrong loss formula does not necessarily invalidate the empirical results if the implementation used a different loss, but as written it makes the method unverifiable. The concern is internal to the manuscript: the stated mathematical equivalence is false, which is a correctness risk rather than a disagreement with consensus. The reader's conditional verdict already requires clarifying or correcting this point, so my read does not change the verdict. I did not identify a stronger concern: the missing VoxelMorph baseline and TRE are important evaluation gaps, but they are completeness issues, whereas the loss equation, if wrong, undermines the theoretical foundation of the proposed unsupervised objective. The proposed analytical counterexample can settle the mathematical claim without requiring the authors' code or retraining.","tokens_in":6223,"tokens_out":9207,"duration_ms":102572,"concrete_test":"Evaluate the paper's Eq. (2) on normalized vectors zs = [1,−1] and zt = [−1,1]: the quoted loss gives 0, while NCC gives −1, disproving the claimed equivalence. Repeating this check on random signed z vectors will confirm that the identity fails; if the authors' implementation instead uses signed z-scores, the printed formula must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 defines the data term as D(Is,It) = (1/2N) Σ ( |Is−μs|/√(σs²+ε²) − |It−μt|/√(σt²+ε²) )² and states that this 'derived SSD guarantees maximization of NCC metric [4]'. The standard SSD equivalent of NCC is E[(zs−zt)²] = 2 − 2E[zszt], where z = (I−μ)/σ, with no absolute values. Replacing z by |z| discards the sign of the normalized intensity deviation. The equivalence fails: for zs = [1,−1] and zt = [−1,1], both have zero mean and unit norm, but the paper's loss is 0 while NCC is −1, so the loss rewards contrast inversion. Therefore the written objective is not a proxy for NCC, and the claim that training 'guarantees maximization of NCC' is unsupported. All reported Dice results are obtained with a loss that may optimize a different similarity measure, so the accuracy comparisons rest on an unverified premise. If the absolute value is a typographical error, the method as printed is not reproducible; if it is not, the training objective is not the stated one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes Conv2Warp, an unsupervised deformable image registration network that concatenates linear and deformable convolutional layers with a Catmull-Rom spline resampler and a multi-scale continuous warping scheme. The training loss combines a normalized intensity-difference data term with a clamped L2 regularizer on the deformation field. The method is evaluated on lung CT (POPI) and brain MRI (MGH10, CUMC12) benchmarks, reporting Dice and Jaccard overlaps and CPU runtime against SimpleElastix, ANTS-SyN, and a linear-convolution-only ablation. The authors claim improved accuracy over classical methods and a 141x runtime speedup relative to SimpleElastix.","tokens_in":6501,"tokens_out":6972,"duration_ms":68959,"significance":"The idea of a differentiable continuous resampler and pyramid warping for unsupervised deep registration is timely, and the proposed architecture is lightweight and evaluated on public datasets. The internal ablation comparing the full model with a linear-convolution baseline is a useful control, and the speed advantage over classical tools is a genuine practical contribution. However, the mathematical justification of the loss is questionable, the experimental comparison omits the most relevant deep-learning baselines, and the evaluation relies solely on segmentation overlap without landmark-based error or statistical tests. If the loss issue is clarified and the comparisons are completed, the paper could make a useful contribution to fast deformable registration.","major_comments":[{"comment":"The data term D(Is,It) = 1/(2N) Σ ( |Is−μs|/√(σs²+ε²) − |It−μt|/√(σt²+ε²) )² is stated to 'guarantee maximization of NCC metric [4]', but for the standard definition of normalized cross-correlation this is not correct. Since E[(|zs|−|zt|)²] = 2 − 2 E[|zs zt|], the loss maximizes the expected product of absolute normalized deviations, not E[zs zt]. For example, with zs = [1,−1] and zt = [−1,1], the loss is 0 while NCC is −1, so the objective rewards contrast inversion. If the absolute values are intentional, the loss is not a valid reformulation of NCC and the theoretical justification is unsupported; if they are a typographical error, the equation must be corrected and the experiments rerun. Because every reported accuracy result depends on this loss, the issue is load-bearing.","section":"Section 2, Loss function"},{"comment":"The experimental comparison omits the most relevant deep-learning baselines for this task, namely VoxelMorph (ref. [1]) and the unsupervised framework of de Vos et al. (ref. [10]), even though both are cited in the introduction. Without these comparisons, the claim that Conv2Warp advances the state of the art in deep deformable registration is not established. Please add these baselines or explicitly restrict the claims to the classical methods that were actually compared.","section":"Tables 1 and 2"},{"comment":"The text in Section 3.2 states that Conv2Warp 'outperforms all other state-of-the-art methods and ConvNet for almost all considered pairs', but Table 1 reports the same mean Dice for SimpleElastix and Conv2Warp (0.90) and a slightly lower mean Jaccard for Conv2Warp (0.84 vs 0.85). This statement overstates the numerical results and should be corrected.","section":"Table 1"},{"comment":"No statistical significance tests or confidence intervals are provided, and the sample sizes are small (7 lung pairs, 9 and 11 brain pairs). For example, in Table 2 the MGH10 mean Dice values for SimpleElastix and Conv2Warp are both 0.95 with overlapping standard deviations, so the claimed advantage is not substantiated. In addition, no target registration error (TRE) is reported for POPI, which has publicly available landmark annotations; segmentation Dice alone is an indirect accuracy measure and may be insensitive to clinically relevant misalignments. Paired significance tests and a landmark-based evaluation are needed.","section":"Tables 1 and 2"},{"comment":"The phrase 'learnable bicubic Catmull-Rom spline resampler' is ambiguous because Catmull-Rom interpolation coefficients are fixed by the interpolation conditions. The only learnable component appears to be a convolutional filter applied after resampling. Please specify exactly which parameters are learned in this block and how; otherwise the contribution of the resampler cannot be reproduced.","section":"Section 2, NL-DVF-R"},{"comment":"The multi-scale warping scheme is described only briefly. In particular, it is not stated how the image pyramid and the deformation fields are downsampled and upsampled between levels, which interpolation is used for the upsampled DVF, or how the concatenated input images are handled at each pyramid level. These details are necessary for reproducibility.","section":"Section 2, Continuous Warping Stage"}],"minor_comments":[{"comment":"The 'Pre-align.' column is not defined; please state what transformation is applied to obtain these Dice values.","section":"Table 1"},{"comment":"The regularization weight λ=0.001 and the clamp [0, 0.25] are presented without a sensitivity analysis; a small ablation would help justify these choices.","section":"Section 2, Loss function"},{"comment":"The description of the lung training data is confusing: 'only breathing cycles 00-50, 10-80 and 30-90 were considered from 10 different sets (a total of 120 volumes)' needs clarification on how the 120 volumes are formed.","section":"Section 3.1"},{"comment":"The paper refers to supplementary figures (e.g., Fig. 2 and Fig. 3 in the supplement) that are not included in the submission; the supplementary material should be provided.","section":"References to supplementary material"},{"comment":"The runtime comparison should report the CPU model, the number of threads, and whether the classical methods are run single-threaded; the 141x speedup factor depends on these details.","section":"Section 3.2"},{"comment":"The loss is described as 'multi-modal NCC', but all experiments are mono-modal; please clarify whether the method is intended for multi-modal registration and whether the loss supports it.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This appears to be an early arXiv version of a workshop-oriented paper. The central idea is plausible and the empirical results, if correct, could be valuable, but the mathematical claim around Eq. (2) is not acceptable as written and the experiments need the standard deep baseline and statistical support. I am not recommending rejection; the issues appear addressable with additional analysis and a careful re-derivation of the loss. The absence of TRE and the small sample sizes should be weighed carefully by the editor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, this paper reports a genuinely fast unsupervised deformable registration pipeline, and the Dice/Jaccard numbers on POPI, MGH10, and CUMC12 are plausibly competitive with SimpleElastix. Second, the training loss as printed is not NCC, and the claim that it guarantees maximization of NCC is wrong.\n\nWhat is actually new: the integration of deformable convolutions into a registration network, a learnable Catmull-Rom spline resampler, and a multi-scale continuous warping scheme. Each component is published, but the combination is not in prior work, and the authors show the spline choice matters by comparing training convergence across interpolants. The speed advantage is real and important—about 141x faster than SimpleElastix on lung CT—and the comparison against their own linear-only ConvNet is a useful ablation.\n\nThe load-bearing problem is Eq. (2). NCC uses z = (I − μ)/σ, without absolute value. The paper replaces z with |z|. That changes the objective: for a source with centered intensities [1, −1] and a target with [−1, 1], the paper's loss is 0 while NCC is −1, so the loss rewards contrast inversion and is not a proxy for NCC. If the bars are a typographical error, the paper as printed is not reproducible; if they are not, the objective is not the one stated. All reported Dice results come from this loss, so the accuracy comparisons rest on an unverified premise. I treat this as a major-revision issue, not a desk-reject killer, because the architecture and experiments are otherwise coherent.\n\nThe evaluation is also thinner than it should be: no VoxelMorph, which is the natural deep-learning baseline; no landmark-based target registration error; and no significance tests. VoxelMorph in particular would tell you whether the spline resampler and deformable convolutions actually beat a simpler unsupervised learner.\n\nBottom line: this is a useful contribution for people building fast clinical registration workflows, and the speed win is clear. But the loss-function claim needs fixing, and the evaluation needs the missing baselines and error metrics before the accuracy comparisons can be trusted. It deserves a serious referee—major revision, not rejection.","headline":"A fast unsupervised DIR pipeline with real speed gains, but the printed loss function is not NCC and the evaluation misses the key deep learning baseline.","tokens_in":6968,"tokens_out":2959,"would_cite":false,"duration_ms":29102,"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":"Conv2Warp learns continuous warp fields for deformable registration with deformable convolutions and a learnable Catmull-Rom spline, matching SimpleElastix's Dice while running 141× faster.","keywords":["unsupervised deformable image registration","deformable convolution","Catmull-Rom spline","continuous warping","normalized cross-correlation","lung CT","brain MRI","medical image registration"],"falsifier":"Train Conv2Warp on a fixed pair set with the proposed D loss, and in parallel train the same architecture with a standard negative NCC or SSD loss. If the paper's equivalence claim is correct, the proposed loss should track NCC (lower D should correspond to higher NCC) and produce comparable or better Dice; if the D-trained model underperforms the NCC-trained one, or if NCC does not improve as D decreases, the loss-equivalence premise fails. A second check: on POPI, which has landmark point pairs, compare the point-wise target registration error of Conv2Warp against SimpleElastix; Dice gains without corresponding TRE gains would indicate that accuracy is not truly in the deformation fields.","tokens_in":6028,"feed_emoji":"🧠","tokens_out":7914,"duration_ms":69033,"temperature":0.7,"pith_summary":"Conv2Warp is an unsupervised network for deformable medical image registration that aims to capture large non-linear deformations while keeping computation fast. The paper claims that by mixing deformable convolution layers with linear convolutions and using a learnable bicubic Catmull-Rom spline to resample the predicted deformation field, the network produces smoother and more accurate alignments than linear-only ConvNets. On lung CT and brain MRI benchmarks it reports Dice scores on par with or above conventional optimization-based methods like SimpleElastix and ANTS, but with a roughly 141-fold speed-up (about 3 seconds per pair on CPU). The practical point is that accurate deformable registration could run quickly enough for routine clinical use without manually generated training deformations.","feed_headline":"Deformable registration drops from minutes to 3 seconds","feed_subtitle":"Deformable convolutions plus spline resampling aligns lung and brain images in seconds, not minutes.","key_machinery":"The network has three linked components. LD-ConvNet alternates five linear and three deformable convolution layers (with batch norm and ELU) to predict a raw deformation vector field; deformable convolution adds learnable spatial offsets to the sampling grid, relaxing the uniform-pixel sampling of standard filters. NL-DVF-R resamples that field with a learnable bicubic Catmull-Rom spline—a cubic interpolant that passes through its control points and is C1 continuous—rather than with linear interpolation, which is the paper's mechanism for smoother fields and smaller interpolation error. CWS applies the field at multiple pyramid levels, warping each coarser source level and summing the per-level data terms. The data term D(Is,It) is written as a 'derived SSD' from standardized intensities (with absolute values) that the paper says guarantees maximization of normalized cross-correlation, and an L2 regularizer (clamped to [0,0.25]) penalizes deviation between the final DVF and the upsampled previous-level DVF to keep the field smooth.","core_discovery":"The central claim is that a registration network built from linear plus deformable convolutions and a learnable Catmull-Rom spline resampler can learn smooth deformation fields that handle large non-linear deformations better than networks restricted to linear convolutions and linear resampling. Deformable convolutions relax the uniform spatial sampling of standard filters, the Catmull-Rom spline provides C1-continuous, locally supported interpolation that passes through control points, and a multi-scale continuous warping stage aggregates deformations across pyramid levels. On the POPI lung 4D-CT dataset the method reaches mean Dice 0.90, equalling SimpleElastix; on the MGH10 and CUMC12 brain MRI datasets it reaches mean Dice 0.95 and 0.97, matching or exceeding SimpleElastix and beating ANTS, while running in about 2.94 s on CPU and under 1 s GPU. The authors argue these results show that large non-linear deformations can be captured without ground-truth deformation fields and with drastically lower computation than iterative methods.","pith_inferences":["A caveat that follows from the paper's own comparison: Dice is computed on segmentation masks, not on the deformation fields themselves; mask overlap can improve even when the underlying voxel correspondence does not, so the claim of 'accurate DVFs' would be better tested against point-wise target registration error on datasets like POPI that provide landmarks.","The loss's equivalence to NCC is asserted without derivation, and the absolute-value standardization is not the standard z-score; if this loss does not actually behave as a correlation objective, a model trained on a conventional NCC or SSD loss might perform differently, so a direct ablation would strengthen the argument.","Catmull-Rom splines pass through their control points whereas B-splines do not; this may be why the paper observes lower training loss, but the comparison is on losses only, and final registration quality could be compared on DVF error metrics instead."],"forward_implications":["Pair-wise registration of a 256×256 2D image takes about 2.94 s on CPU and under 1 s on GPU, so the method could fit into clinical workflows that currently use minutes-long optimization.","Training is unsupervised: no ground-truth deformation fields are needed, only image pairs, which removes the bottleneck of generating synthetic DVFs for training.","The multi-scale continuous warping component is modular and could be combined with other data terms or regularizers without changing the architecture.","On the tested benchmarks, the model matches or exceeds SimpleElastix Dice while being roughly 141 times faster, suggesting the accuracy/speed trade-off of learning-based DIR can be resolved in favor of speed without sacrificing accuracy."],"supporting_citations":[{"why":"Supplies the deformable convolution layers that give Conv2Warp its non-uniform sampling of features.","marker":"[3]"},{"why":"Source of the correlation-transform loss that the paper restates as a 'derived SSD' claimed to maximize NCC.","marker":"[4]"},{"why":"Earlier linear-sampling unsupervised registration model that Conv2Warp extends and compares against.","marker":"[7]"},{"why":"Earlier B-spline-based registration framework whose interpolation choice Conv2Warp challenges.","marker":"[10]"},{"why":"Canonical unsupervised linear-convolution registration baseline for deformable DIR.","marker":"[1]"},{"why":"Provides the POPI lung 4D-CT dataset with segmentation masks used for evaluation.","marker":"[9]"},{"why":"Provides the MGH10 and CUMC12 brain MRI datasets with masks used for evaluation.","marker":"[5]"},{"why":"Provides the LBPA40 brain dataset used for training.","marker":"[8]"},{"why":"Provides the DIR-LAB 4D CT dataset used for training.","marker":"[2]"}],"fun_headline_variants":["Continuous conv and spline warp cut registration to seconds","Unsupervised deformable registration in seconds with Conv2Warp","Spline resampling and deformable convs speed image alignment","Fast deformable registration with continuous warping and conv","Conv2Warp: smooth warps for lung and brain images in seconds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loss is stated to be a 'derived SSD' that guarantees maximization of normalized cross-correlation, but the derivation is not shown and the absolute-value standardized terms do not match the standard NCC form; if the loss does not truly optimize correlation, the reported accuracy rests on an unverified objective.","fun_headline_variants_meta":{"raw":{"variants":["Continuous conv and spline warp cut registration to seconds","Unsupervised deformable registration in seconds with Conv2Warp","Spline resampling and deformable convs speed image alignment","Fast deformable registration with continuous warping and conv","Conv2Warp: smooth warps for lung and brain images in seconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1398,"prompt_tokens":953,"completion_tokens":445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":359}},"tokens_in":569,"tokens_out":445,"duration_ms":4900,"temperature":1.0,"reasoning_tokens":359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:52:49.661221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Conv2Warp on a fixed pair set with the proposed D loss, and in parallel train the same architecture with a standard negative NCC or SSD loss. If the paper's equivalence claim is correct, the proposed loss should track NCC (lower D should correspond to higher NCC) and produce comparable or better Dice; if the D-trained model underperforms the NCC-trained one, or if NCC does not improve as D decreases, the loss-equivalence premise fails. A second check: on POPI, which has landmark point pairs, compare the point-wise target registration error of Conv2Warp against SimpleElastix; Dice gains without corresponding TRE gains would indicate that accuracy is not truly in the deformation fields.","supporting_citations":[{"cited_title":"In: 2017 IEEE International Conference on Computer Vision (ICCV)","cited_arxiv_id":null,"evidence_quote":"Supplies the deformable convolution layers that give Conv2Warp its non-uniform sampling of features."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Source of the correlation-transform loss that the paper restates as a 'derived SSD' claimed to maximize NCC."},{"cited_title":"In: 2018 IEEE 15th International Sympo- sium on Biomedical Imaging (ISBI 2018)","cited_arxiv_id":null,"evidence_quote":"Earlier linear-sampling unsupervised registration model that Conv2Warp extends and compares against."},{"cited_title":"Medical Image Analysis 52, 128–143 (2019)","cited_arxiv_id":null,"evidence_quote":"Earlier B-spline-based registration framework whose interpolation choice Conv2Warp challenges."},{"cited_title":"IEEE Transac- tions on Medical Imaging 38(8), 1788–1800 (Aug 2019)","cited_arxiv_id":null,"evidence_quote":"Canonical unsupervised linear-convolution registration baseline for deformable DIR."},{"cited_title":"Med Phys 38(1), 166–178 (2011)","cited_arxiv_id":null,"evidence_quote":"Provides the POPI lung 4D-CT dataset with segmentation masks used for evaluation."},{"cited_title":"NeuroImage 46(3), 786 – 802 (2009)","cited_arxiv_id":null,"evidence_quote":"Provides the MGH10 and CUMC12 brain MRI datasets with masks used for evaluation."},{"cited_title":"NeuroImage 39(3), 1064 – 1080 (2008)","cited_arxiv_id":null,"evidence_quote":"Provides the LBPA40 brain dataset used for training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DIR-LAB 4D CT dataset used for training."}],"review_version":1}