{"id":"b90ef708-55d3-4f7e-9337-1c41845b0db6","arxiv_id":"2506.19975","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A discrete optimization registration method uses voxel-wise displacement entropy to guide adaptive message passing, matching supervised methods on abdominal CT without training.","lead":"VoxelOpt is a training-free deformable image registration method that uses per-voxel displacement entropy to adaptively smooth displacement fields, reaching 58.5% Dice on abdominal CT without labels. It runs in under a second and matches supervised deep learning methods, which could make accurate registration available where training data or labels are scarce.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'no label supervision' claim is not established: VoxelOpt's accuracy gain comes mostly from a label-supervised foundation segmentation model (12.8 Dice in Table 2), and the paper never checks whether that model's pretraining data overlaps the 30-scan evaluation cohort.","rationale":"The reader's weakest assumption is Eq. (6); I partially agree but believe feature provenance is the more load-bearing issue. The paper's own ablation ranks the contributions: foundation features account for 12.84 Dice over raw (45.67 vs 58.51) and 8.53 over MIND (49.98 vs 58.51), whereas adaptive message passing accounts for 2.11 Dice (56.40 vs 58.51). Therefore, even if Eq. (6) is suboptimal and alpha is arbitrary, the worst-case blow to the central claim is bounded by a few Dice points. In contrast, the reliance on a label-trained foundation segmentation model makes the entire 'no label supervision' framing questionable, and if the evaluation cohort appears in the foundation model's training set, the result is confounded at its root. The absence of a provenance check is not a stylistic omission: public pretrained models trained on public medical datasets create exactly this overlap risk. A single manifest check plus a disjoint-backbone rerun would settle the matter. If the concern is resolved, the engineering contribution stands and conditional acceptance is appropriate; if not, the central claim needs to be substantially revised. I therefore keep the reader's CONDITIONAL verdict unchanged, with the explicit condition that this provenance analysis be provided.","tokens_in":8879,"tokens_out":7599,"duration_ms":84137,"concrete_test":"Obtain the pretraining data manifest for the foundation model [19] (from the authors, repository, or paper) and check whether any of the 30 patient scans in [28] appear among its training images. If no manifest is available, rerun the Table 2 row #3 configuration with a feature backbone that is provably disjoint from [28] (e.g., retrain [19] after holding out that cohort, or use a self-supervised backbone trained only on non-abdominal data) and compare Dice/HD95 on the 42 test pairs. A drop of more than ~2 Dice points relative to 58.51 would indicate that the 'without label supervision' accuracy is largely inherited from labeled pretraining; a maintained result with a disjoint backbone would clear the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4/Table 2 shows that the feature extractor, not the adaptive message passing, is the dominant contributor to the headline result: replacing raw image features with foundation-model features raises Dice from 45.67 (#1) to 58.51 (#3), a 12.84-point gain, while the proposed voxel-adaptive term contributes only 2.11 Dice (#3 vs #4). The conclusion claims VoxelOpt achieves SOTA accuracy 'without relying on label supervision or complex training.' But the feature map comes from [19], a foundation model trained with dense organ-segmentation labels on large abdominal CT corpora, and the evaluation set [28] is a public abdominal CT dataset with 13-organ segmentations. The paper provides no analysis of whether [19]'s pretraining data includes scans from [28] or their patient cohort. If overlap exists, the reported accuracy reflects memorized segmentation knowledge leaking into the cost volume rather than the registration algorithm. Even absent exact overlap, relying on a target-domain label-trained segmentation model is label supervision in substance, which conflicts with the advertised 'without label supervision' framing. This is a confound that must be removed or quantified before the central claim can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VoxelOpt, a training-free discrete optimization framework for deformable abdominal CT registration. The method builds a 27-neighbor cost volume from feature maps, computes per-voxel displacement entropy from that cost volume, and uses the entropy to set a per-voxel Gaussian blur width that controls adaptive message passing. Large deformations are handled with a multi-level Laplacian pyramid, and the feature extractor is a pretrained foundational segmentation model. Experiments on 42 test pairs from a public abdominal CT dataset report 58.51% Dice with sub-second runtime, matching a semi-supervised RDP baseline and outperforming iterative and unsupervised learning baselines. Ablations attribute 12.8 Dice points to the foundation-model features and 3.7 Dice points to the adaptive message-passing component.","tokens_in":9129,"tokens_out":4081,"duration_ms":48122,"significance":"If the results hold, the paper demonstrates a useful point: a non-learning discrete optimizer with entropy-guided adaptive smoothing and features from a pretrained segmentation model can reach accuracy comparable to semi-supervised registration networks while running in under a second. The multi-level pyramid with fixed 27-neighbor cost volumes is a sensible way to avoid exponential search-space growth, and the entropy-based interpretation of feature quality for registration is an interesting diagnostic idea. The paper includes a concrete held-out test split and ablation studies for feature type, kernel size, adaptive message passing, and cost-volume pre-filtering. However, the central claim of label-supervision-free performance depends on a label-trained foundation model whose pretraining data overlap with the evaluation set is never analyzed, and the core adaptive mapping is introduced heuristically without sensitivity analysis. The empirical evidence is also based on a single small test set without significance testing.","major_comments":[{"comment":"The claim that VoxelOpt achieves state-of-the-art accuracy 'without relying on label supervision' is not established as stated. The dominant accuracy gain comes from features extracted by the foundational segmentation model [19], which was trained with dense organ-segmentation labels on large CT corpora, and Table 2 shows that this feature source contributes 12.8 Dice points (#1 vs #3). The evaluation dataset [28] is a public abdominal CT dataset with 13-organ segmentations, but the paper never checks whether [19]'s pretraining data includes scans from [28] or from the same patient cohort. If such overlap exists, the reported 58.51 Dice partly reflects memorized segmentation knowledge leaking into the cost volume. The authors should either demonstrate non-overlap of the pretraining and evaluation data and quantify the sensitivity to that overlap, or reframe the contribution as registration without training a dedicated network and provide results with features that are not label-trained.","section":"Sections 3.1, 3.2, and 4"},{"comment":"The voxel-adaptive message-passing contribution rests entirely on the mapping sigma(x) = alpha * log(E(x)/max(E) + 1), but the paper provides no derivation, optimality argument, or sensitivity analysis for this mapping. The ablation in Table 2 (#3 vs #4) reports only 3.7 Dice points for the adaptive term, and it is unclear whether this gain is robust to the choices of alpha = 1.5, beta, and the theta schedule used in Eq. (5). Please report results for a range of alpha and beta values, and ideally compare the log mapping against a simpler alternative such as a linear or rank-based mapping, so readers can assess whether the specific functional form is load-bearing for the claimed improvement.","section":"Equation (6) and Section 3.4"},{"comment":"All accuracy comparisons are based on 42 test pairs from a single dataset, and no statistical significance tests, confidence intervals, or per-pair variance measures are reported. Several headline comparisons are close (VoxelOpt at 58.51 Dice vs. RDP semi-supervised at 58.77 Dice), and the claimed parity/outperformance should be supported by paired tests (e.g., Wilcoxon signed-rank) and confidence intervals for Dice and HD95. Without this, the differences between methods in Table 1 may be within noise, and the conclusion that VoxelOpt 'matches' or 'outperforms' specific baselines is not yet quantitatively supported.","section":"Tables 1 and 2, Section 3.3"},{"comment":"The implementation of the proposed spatially varying Gaussian filtering is underspecified. Equation (6) defines a per-voxel sigma field, but the paper does not state how a space-variant Gaussian is applied to the 6D cost volume or to the displacement field; the statement that adaptive 3D Gaussian filtering is implemented as three separable 1D filters is insufficient because space-variant filtering is not separable in general. Since this filtering is the core mechanism of the method, please provide the exact discretization, including how per-voxel sigma is converted to filter kernels and how border conditions are handled, or make the code available in the review version.","section":"Sections 2.4, 2.5, and 3.2"}],"minor_comments":[{"comment":"There are several spacing artifacts such as 'V oxelOpt' and 'T able 1' that should be corrected during production.","section":"Throughout"},{"comment":"The text says the cost volume grows from (2*1+1)^3 to (2*2+1)^3 and (3*1+1)^3; the last expression should be (2*3+1)^3 for kernel size k = 3.","section":"Section 3.4"},{"comment":"The paper uses both '27-neighbor' and '26-neighborhood' for the same k = 1 cost volume; please standardize the terminology.","section":"Sections 1 and 3.4"},{"comment":"The symbol L1 is used both for the label of the displacement neighborhood and for the L1 dissimilarity function mentioned in Section 3.2, which is confusing; please disambiguate the notation.","section":"Equations (2)-(4)"},{"comment":"The pre-filtering of the cost volume (the step that produces C-tilde from Ck) is mentioned in the framework description and ablation but is not defined by an equation; a precise definition would improve reproducibility.","section":"Section 2.5 and Figure 2"},{"comment":"Please clarify whether 'pre-softmax features' are taken from a single layer or from multiple layers of the foundation model, and which layer is used for each pyramid level.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely to receive strong scrutiny on the 'without label supervision' framing because the feature extractor is a label-trained foundation model. I recommend requiring the overlap analysis and a reframed claim, as well as sensitivity analysis for the entropy-to-smoothing mapping, before publication. The small test set without significance testing is also likely to be raised by other reviewers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the core mechanism in VoxelOpt—per-voxel entropy from a local cost volume controlling the spatial extent of message-passing smoothing—is genuinely new in the registration literature. The paper does a clean job of isolating it in the ablations: switching the entropy-adaptive smoothing off costs 3.7% Dice (Table 2 #3 vs #4), and the full system runs sub-second on abdominal CT, roughly matching the best semi-supervised network on Dice while beating it on HD95 (18.54 vs 20.07). That is a real engineering result.\n\nBut the paper's advertised framing—\"without relying on label supervision\"—does not survive contact with its own Table 2. Swapping raw image features for features from a foundation segmentation model (CLIP-driven universal model, trained with dense organ labels on large abdominal CT datasets) is worth 12.84 Dice (#1 vs #3). The proposed adaptive message passing is worth 2.11 Dice (#3 vs #4). In other words, the main driver of the headline accuracy is not the new algorithm; it is a label-trained feature extractor. The paper never checks whether that extractor's pretraining data overlaps the 30-scan evaluation cohort. That is a confound that must be addressed—at minimum by stating the possibility and arguing against it, or by re-running the experiments with a feature extractor guaranteed not to have seen the test scans.\n\nOther soft spots are minor relative to this. The sigma(x)=alpha*log(E/max(E)+1) mapping is justified by intuition only; no sensitivity analysis for alpha or beta. The evaluation uses 42 test pairs from a single dataset with no significance tests. These are common limitations in the field and fixable.\n\nWhat the paper does well: the multi-level pyramid with fixed 27-neighbor cost volumes avoids exponential search space growth; the ablations are honest and informative; the runtime comparison is fair; and the entropy-based analysis of features (Fig. 1) is a useful diagnostic. The citation pattern to prior registration work looks appropriate.\n\nBottom line: this deserves serious peer review. The mechanism is worth publishing, and the authors have the tools to fix the label-supervision framing. But the conclusion as written overclaims. Send it to reviewers who will push on the foundation model overlap and the feature-dependence of the claimed result.","headline":"Entropy-guided adaptive message passing is a real contribution, but the 'no label supervision' claim is undercut by reliance on a label-trained foundation model.","tokens_in":9663,"tokens_out":1953,"would_cite":true,"duration_ms":20290,"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":"A discrete optimizer with entropy-guided adaptive smoothing registers abdominal CT scans as accurately as semi-supervised networks, without any training.","keywords":["deformable image registration","discrete optimization","message passing","displacement entropy","abdominal CT","mean-field inference","foundation model features","adaptive filtering"],"falsifier":"Run the same six-iteration VoxelOpt pipeline on a second abdominal CT dataset from a different scanner or contrast protocol, comparing the adaptive version against the non-adaptive version (ablation #3 vs #4): if the adaptive component does not consistently improve Dice, the entropy-to-smoothing map is not a general mechanism. A second check would be to compare the list of images the foundation model was pretrained on against the 30-scan evaluation set; any overlap would change the interpretation of the training-free claim.","tokens_in":8663,"feed_emoji":"🩻","tokens_out":6519,"duration_ms":65393,"temperature":0.7,"pith_summary":"VoxelOpt aims to close the gap between slow iterative registration and fast learning-based registration for deformable abdominal CT. The paper argues that the entropy of a local displacement cost volume measures how confident each voxel is in its own displacement, and that this signal should control how much each voxel listens to its neighbors: confident voxels keep their displacement, uncertain voxels borrow from neighbors. Using this entropy-adaptive message passing inside a multi-level discrete optimizer, with features from a pretrained segmentation model, VoxelOpt reports Dice scores on par with the best label-supervised learning method and better than the best iterative method, all in under one second and with no training on the target data. A sympathetic reader would take this as evidence that discrete optimization can be made both fast and accurate when the smoothing schedule is driven by per-voxel signal strength rather than a fixed isotropic regularizer.","feed_headline":"Discrete optimizer matches supervised CT registration without training","feed_subtitle":"Entropy-guided adaptive smoothing beats iterative methods and equals semi-supervised networks on abdominal CT.","key_machinery":"The load-bearing object is the per-voxel displacement entropy $E(x)$, computed by converting a 27-neighbor cost volume $C^1(x)$ into a probability distribution via softmax with temperature $\\beta$ and taking the entropy of that distribution. Entropy is then normalized and mapped by $\\sigma(x) = \\alpha \\log(E(x)/\\max(E) + 1)$ with $\\alpha = 1.5$ to a per-voxel Gaussian blur width used during message passing. This adaptive smoothing replaces the isotropic regularizer used in earlier discrete and learning-based registration, and it is what lets the optimizer propagate displacement signals from boundary and texture-rich voxels into uniform regions in only six iterations.","core_discovery":"The central claim is that a training-free discrete optimization framework can match semi-supervised deep registration accuracy on abdominal CT by making the smoothing step of mean-field inference voxel-adaptive. Specifically, VoxelOpt computes a probabilistic cost volume over 27 neighbor displacements at each pyramid level, derives per-voxel displacement entropy, and maps that entropy through a fixed logarithmic rule to a per-voxel Gaussian blur width. Voxels with strong displacement signals (low entropy) are smoothed less, preserving their information; voxels with weak signals (high entropy) are smoothed more, receiving more neighbor influence. Together with a five-level image pyramid, a 27-neighbor search per level, and feature maps from a pretrained foundational segmentation model, this reaches 58.5% Dice on the testing split, compared with 53.6% for the best iterative baseline and 58.8% for the best semi-supervised network, at sub-second runtime.","pith_inferences":["Inference beyond the paper: the same entropy-gated smoothing could be applied to other per-voxel discrete estimates, such as optical flow or stereo disparity, where confidence varies spatially.","Inference beyond the paper: the logarithmic mapping in Eq. (6) is one of many possible entropy-to-blur curves; scanning $\\alpha$ or fitting the mapping on a small validation set would show how much of the 3.7% Dice gain depends on that particular choice.","Inference beyond the paper: because the evaluation set is a public 30-scan dataset, the pretrained segmentation model's training-data overlap with it is a concrete risk; checking for overlap would determine whether the reported gains transfer to unseen scanners.","Inference beyond the paper: since the optimizer is differentiable and runs in a single forward pass, it could be unrolled and fine-tuned end-to-end, which the paper mentions as a possibility but does not test."],"forward_implications":["Sub-second deformable registration becomes achievable without any training or label supervision, which matters for low-data clinical settings where semi-supervised networks cannot be trained.","The entropy of a local cost volume provides a usable per-voxel confidence signal, so the same adaptive message-passing idea could be attached to other discrete optimization solvers.","Pretrained segmentation features can replace hand-crafted or contrastively learned features for registration, removing an entire training stage.","A fixed small search neighborhood (27 neighbors) with an image pyramid suffices; enlarging the kernel only adds runtime and degrades smoothness, so exponential search-space growth is avoidable."],"supporting_citations":[{"why":"Supplies the pre-softmax feature maps from a pretrained foundational segmentation model, replacing raw images and hand-crafted descriptors.","marker":"[19]"},{"why":"Supplies the discrete optimization-with-convex-relaxation formulation (Eqs. 2 and 5) that VoxelOpt's coordinate descent solves.","marker":"[13]"},{"why":"ConvexAdam is the closest iterative baseline; VoxelOpt compares directly against it and cites its exponential cost-volume growth as the problem to avoid.","marker":"[21]"},{"why":"Deeds is the best iterative baseline and the discrete MRF registration method VoxelOpt outperforms in Dice.","marker":"[12]"},{"why":"Cited as evidence that photometric dissimilarity gives limited anatomical correspondence, motivating the use of segmentation-derived features.","marker":"[16]"},{"why":"MIND features are the ablation alternative against which raw image and foundation-model features are compared.","marker":"[11]"},{"why":"Supplies the 30-scan abdominal CT dataset with 13-organ masks used for evaluation and the train/test split.","marker":"[28]"},{"why":"VoxelMorph is the canonical amortized learning baseline that VoxelOpt seeks to match in speed while beating in unsupervised accuracy.","marker":"[5]"},{"why":"Ants SyN is the classical iterative baseline representing the variational optimization approach.","marker":"[4]"},{"why":"Supplies the TV-regularized denoising step whose fixed-point iteration is approximated as Gaussian filtering in the u-subproblem.","marker":"[7]"}],"fun_headline_variants":["Training-free CT registration matches supervised accuracy","Voxel-adaptive entropy beats iterative registration","Discrete optimizer rivals learned CT registration","Entropy-guided smoothing speeds deformable CT alignment","Fast CT registration without training via entropy-aware smoothing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole accuracy gain from adaptivity rests on the untested assumption that the fixed logarithmic formula $\\sigma(x)=1.5\\log(E(x)/\\max(E)+1)$ turns cost-volume entropy into the right per-voxel blur strength for every anatomy and scanning protocol.","fun_headline_variants_meta":{"raw":{"variants":["Training-free CT registration matches supervised accuracy","Voxel-adaptive entropy beats iterative registration","Discrete optimizer rivals learned CT registration","Entropy-guided smoothing speeds deformable CT alignment","Fast CT registration without training via entropy-aware smoothing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000617,"raw_usage":{"total_tokens":2886,"prompt_tokens":986,"completion_tokens":1900,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":1833}},"tokens_in":602,"tokens_out":1900,"duration_ms":14628,"temperature":1.0,"reasoning_tokens":1833,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:59:27.903594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same six-iteration VoxelOpt pipeline on a second abdominal CT dataset from a different scanner or contrast protocol, comparing the adaptive version against the non-adaptive version (ablation #3 vs #4): if the adaptive component does not consistently improve Dice, the entropy-to-smoothing map is not a general mechanism. A second check would be to compare the list of images the foundation model was pretrained on against the 30-scan evaluation set; any overlap would change the interpretation of the training-free claim.","supporting_citations":[{"cited_title":"In: Biomedical Image Registration: 6th In- ternational Workshop, WBIR 2014, London, UK, July 7-8, 2014","cited_arxiv_id":null,"evidence_quote":"Supplies the discrete optimization-with-convex-relaxation formulation (Eqs. 2 and 5) that VoxelOpt's coordinate descent solves."},{"cited_title":"IEEE Transactions on Medical Imaging (2024)","cited_arxiv_id":null,"evidence_quote":"ConvexAdam is the closest iterative baseline; VoxelOpt compares directly against it and cites its exponential cost-volume growth as the problem to avoid."},{"cited_title":"IEEE transactions on medical imaging32(7), 1239–1248 (2013)","cited_arxiv_id":null,"evidence_quote":"Deeds is the best iterative baseline and the discrete MRF registration method VoxelOpt outperforms in Dice."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited as evidence that photometric dissimilarity gives limited anatomical correspondence, motivating the use of segmentation-derived features."},{"cited_title":"IEEE Transactions on Biomedical Engineering 63(8), 1563–1572 (2016)","cited_arxiv_id":null,"evidence_quote":"Supplies the 30-scan abdominal CT dataset with 13-organ masks used for evaluation and the train/test split."},{"cited_title":"Journal of Mathematical imaging and vision20, 89–97 (2004)","cited_arxiv_id":null,"evidence_quote":"Supplies the TV-regularized denoising step whose fixed-point iteration is approximated as Gaussian filtering in the u-subproblem."}],"review_version":1}