{"id":"ac918d5e-7a71-4f81-8fdc-9d4d563adcf0","arxiv_id":"2509.09869","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Surrogate supervision applies the registration loss to clean surrogate images rather than raw inputs, improving robustness to artifacts, masks, and modality differences without extra inference cost.","lead":"Surrogate supervision trains a deformable registration network on raw, messy image pairs while computing the loss on clean surrogate copies, so the network learns robust registration without needing clean inputs at test time. The paper demonstrates this for MRI artifacts, inconsistent lung masking, and multi-modal MR across three network architectures.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Artifact experiment omits the natural w/IC test-time-preprocessing baseline, so the robustness gain over standard practice is not established.","rationale":"The reader's weakest_assumption is that surrogates are spatially congruent with originals. That assumption is indeed necessary for the framework, but in the three applications the surrogates are derived from the same subject and space (N4-corrected images, masked versions of the same images, pre-aligned same-subject T1/T2), so the assumption is satisfied in the experiments. A more immediate gap is the missing natural baseline in the artifact experiment, which directly affects the central claim about avoiding preprocessing. The reader correctly flagged that the w/IC baseline is not deployed in its natural configuration, and I agree this is a serious concern. I do not think it changes the overall verdict from CONDITIONAL, because the w/oIC comparison and the mask/multi-modal experiments still provide some support, but the artifact evidence is incomplete as written. A single new experiment would resolve the concern.","tokens_in":12984,"tokens_out":8271,"duration_ms":101288,"concrete_test":"Reproduce the artifact-robustness experiment (Section IV-B) with a fourth training/evaluation configuration: train the model on inhomogeneity-corrected images (w/IC) exactly as in the paper, and at test time apply N4 correction to the raw fixed and moving inputs before feeding them to the network. Evaluate DSC at inhomogeneity levels 1x, 2x, and 3x using the same test pairs and paired Wilcoxon signed-rank tests (alpha=0.01) against the surrogate-supervised model (Ours). If w/IC+N4 at test achieves statistically equivalent DSC to Ours, the claim that surrogate supervision improves robustness over standard practice is not supported; if Ours still outperforms, the central claim is strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that surrogate supervision makes registration robust to artifacts without requiring inference-time preprocessing (Section III-A, Eq. 3, and the abstract). The artifact experiment (Section IV-B) is the primary evidence for this. However, the comparison is asymmetric: the w/IC model is trained on N4-corrected images but evaluated on raw uncorrected inputs at test time. This is not the natural deployment configuration for that model—a user would apply N4 correction at test time, as the training preprocessing implies. The w/oIC baseline is evaluated in its natural raw-deployment mode, but the stronger baseline—w/IC with N4 applied at test time—is missing. Without this arm, the experiment only demonstrates that surrogate supervision helps when test-time preprocessing is unavailable; it does not show that surrogate supervision is preferable to the standard 'correct at test time' pipeline. The abstract's claim of robustness 'without increasing complexity' hinges on outperforming or matching standard preprocessing, which is untested. This is load-bearing because if w/IC+N4 at test time matches Ours, the contribution reduces to 'you can skip preprocessing,' which is weaker than the stated robustness advantage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes surrogate supervision, a training framework for deformable registration formalized in Eq. (3): the network estimates a deformation field φ from raw inputs (I_f, I_m), but the similarity loss is evaluated on surrogate images (eI_f, eI_m) that correspond to the original pair, after warping eI_m by φ. This decouples the input domain from the supervision domain, allowing standard similarity losses (MSE, NCC) to be used when the raw inputs are corrupted by artifacts, inconsistently masked, or of different modalities. The framework is instantiated for three applications: artifact-robust brain MR registration using N4-corrected surrogates, mask-agnostic lung CT registration using lung-masked surrogates, and T2-to-T1 multi-modal MR registration using same-subject T1 surrogates. Experiments cover three architectures (VoxelMorph, TransMorph, VFA) and report Dice or TRE improvements over the compared baselines in most settings, with the notable exception that VoxelMorph shows no clear artifact-robustness gain at low inhomogeneity levels.","tokens_in":13272,"tokens_out":4903,"duration_ms":53060,"significance":"If the result holds, the framework is a simple and broadly applicable training principle: rather than designing task-specific similarity losses, one chooses a trustworthy surrogate domain and applies standard losses. The paper has several concrete strengths: it tests three architectures across three distinct tasks, uses a deliberately mismatched test regime in the lung experiment, and the surrogate construction is transparent and does not introduce fitted constants beyond regularization weights inherited from prior work. The inference-time pipeline is unchanged, which is a practical advantage. However, the artifact-robustness experiment is incomplete as evidence for the strongest claim: the w/IC baseline is evaluated in an unnatural deployment mode, and the standard 'correct at test time' baseline is missing. The conceptual contribution is also somewhat incremental relative to prior label-map supervision and Cao et al.'s intra-modality supervision, although the paper does position these as special cases. Overall, the central mechanism is credible, but one load-bearing comparison needs to be repaired before the stated conclusions are fully supported.","major_comments":[{"comment":"The artifact experiment compares Ours (trained on raw inputs with N4-corrected surrogates, tested on raw corrupted inputs) against w/IC (trained on N4-corrected images, tested on raw corrupted inputs) and w/oIC (trained on raw, tested on raw). The w/IC arm is not evaluated in its natural deployment configuration. A user who trained on N4-corrected images would apply N4 correction at test time; without that arm, the experiment only shows that surrogate supervision helps when test-time preprocessing is unavailable. It does not establish the abstract's claim that robustness is achieved 'without increasing complexity' relative to standard preprocessing pipelines. Please add the w/IC model with N4 applied at test time, or explicitly reframe the claim as robustness under missing/incomplete test-time preprocessing.","section":"Section IV-B, Figure 2"},{"comment":"The formulation does not state sufficient conditions on the surrogates (eI_f, eI_m) relative to (I_f, I_m). It says they 'correspond to' the original pair, but the loss is only meaningful if the surrogates share the coordinate frame of the original images; otherwise eI_m ∘ φ supervises the wrong transformation. The three applications satisfy this by construction (N4 correction, masking, and rigidly aligned same-subject T1), but a precise definition should be added, e.g., 'surrogates are obtained by intensity-only transformations of the input pair, or by paired acquisitions already aligned to the input geometry.' Without this, a user applying the framework to surrogates with different acquisition geometry could silently learn a biased deformation. This is a correctness-risk concern for the general formulation, not for the specific experiments.","section":"Section III-A, Eq. (3)"},{"comment":"The text states that for VoxelMorph, surrogate supervision yields 'similar performance' to w/oIC at inhomogeneity levels 1× and 2×, and attributes this to architecture capacity. This is a partial counterexample to the claim that surrogate supervision makes any registration network artifact-robust. The paper should either report the paired test statistics for this comparison, or temper the cross-architecture claim. As written, the evidence supports architecture-dependent benefits, not a general property of the framework.","section":"Section IV-B, VoxelMorph results"}],"minor_comments":[{"comment":"The multi-modal experiment does not explicitly describe how the moving surrogate is constructed. It should state that the surrogate is the T1-weighted image of the moving subject, pre-aligned with the T2 input, and that this paired image is available only during training.","section":"Section IV-D"},{"comment":"Typo: 'VoxelMorph, surrogate supervision yield similar performance' should be 'yields.' Also, Figure 2 is dense with significance brackets across three architectures and three inhomogeneity levels; consider separate panels or reporting exact p-values.","section":"Section IV-B"},{"comment":"The NDV values for NCC-based configurations (including Ours) are substantially larger than for CR/MI in VoxelMorph and TransMorph. The text says this is 'due to our choice of λ values,' but λ for NCC was taken from prior work rather than tuned. Please clarify that λ was not re-tuned for Ours/NCC in this comparison.","section":"Table II"},{"comment":"Reference [22] has formatting errors ('Me. Ren' and 'P. E. Grant'); Reference [3] lists 'L Tian' with inconsistent capitalization. Check the bibliography for consistency.","section":"References"},{"comment":"The 'Random' training setting is described as choosing both inputs either masked or unmasked per sample. It may be worth clarifying whether mixed masked/unmasked pairs (one masked, one not) were ever presented during training; this is the setting that appears at test time as 'Mixed.'","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The framework is a clean unifying perspective, but the novelty relative to prior intra-modality surrogate supervision and label-map supervision should be sharpened; currently the paper's contribution is more of a systematic framing than a new mechanism. The missing artifact-test baseline is fixable with additional experiments and should be requested. The paper is within scope for the journal, but the experimental reporting needs to match the generality of the stated claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a useful, clearly written paper that formalizes a known trick—supervising on surrogate images rather than raw inputs—and shows it works for two genuinely new applications: artifact-robust and mask-agnostic training. The lung CT experiments are the strongest part; the artifact experiment has a missing baseline that weakens the headline claim.\n\nWhat's actually new: the authors don't pretend label-map supervision or paired multimodal supervision are new. They unify these as \"surrogate supervision\" and add two applications that aren't in the cited prior work. The mask-agnostic result is impressive: across VoxelMorph, TransMorph, and VFA, the surrogate-supervised model matches or beats the oracle models on masked, unmasked, and mixed test settings, with no test-time mask requirement. That's a clean, practical win. The multi-modal revisit is a useful head-to-head against modern MI and CR losses, though it's closer to a replication of Cao et al.\n\nThe soft spots: the artifact experiment compares \"Ours\" (raw at test) to a w/IC model trained on corrected images but evaluated on raw uncorrected inputs. No one would deploy that model that way. The missing arm is w/IC with N4 applied at test time—the standard pipeline. Without that, the experiment only shows you can skip preprocessing, not that surrogate supervision is better than doing it correctly. The VoxelMorph gain is not significant at low inhomogeneity, which the authors attribute to capacity; plausible, but worth noting. In the multi-modal experiment, lambda is tuned to 5 for CR/MI and left at 1 for NCC, and the resulting NDVs are higher for Ours. The authors say lowering lambda for CR/MI doesn't help validation accuracy, so this is a minor concern, but the comparison isn't perfectly clean. No code or proprietary data is released, which limits reproducibility.\n\nThe framework and the mask experiments deserve serious referee time. The artifact claim needs the missing baseline before it can support the abstract's \"without increasing complexity\" framing. Who this is for: anyone training deformable registration networks in medical imaging, especially with artifacts or masks. It's a practical, architecture-agnostic recipe. My recommendation: send it to peer review, and ask the authors to add the test-time preprocessing arm and to soften the robustness claim if they don't.","headline":"A solid training-recipe paper: the lung mask-agnostic results are strong and the framing is useful, but the artifact experiment omits the natural test-time-preprocessing baseline, so the headline robustness claim is overextended.","tokens_in":13745,"tokens_out":2392,"would_cite":true,"duration_ms":27736,"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":"Surrogate supervision trains registration networks on cleaned stand-ins, giving robustness to artifacts and modality gaps with no extra test-time preprocessing.","keywords":["surrogate supervision","deformable image registration","robustness","artifact robustness","mask-agnostic registration","multi-modal registration","similarity loss","deep learning"],"falsifier":"Train a model on surrogates known to be spatially shifted relative to the input anatomy (e.g., a bias-corrected image that was resampled, or a distortion-corrected diffusion volume) and show that registration accuracy collapses or the deformation field becomes systematically biased; alternatively, measure the correlation between surrogate geometric error and registration error across a test set with known spatial distortions.","tokens_in":12927,"feed_emoji":"🩻","tokens_out":7473,"duration_ms":77947,"temperature":0.7,"pith_summary":"The paper proposes surrogate supervision, a training paradigm for deformable image registration. The network still receives raw image pairs and predicts a deformation field, but the similarity loss is computed on surrogate images—bias-corrected, masked, or same-subject other-modality versions—that correspond to the same anatomy. This decouples the input domain from the supervision domain, so standard losses like mean squared error or normalized cross-correlation can train a model even when raw inputs are corrupted or multimodal. If correct, any registration architecture can be made artifact-robust, mask-agnostic, and multi-modal simply by choosing a trustworthy surrogate domain, with no added cost at deployment. Experiments across three tasks and three architectures show the expected gains while keeping accuracy on clean data.","feed_headline":"Surrogate images make registration robust to artifacts and modality gaps","feed_subtitle":"A training trick applies deformations to cleaned stand-ins, so models handle raw, masked, or multi-modal inputs without preprocessing.","key_machinery":"The central mechanism is Eq. (3), the surrogate supervision objective. It modifies the standard registration loss by replacing the raw inputs in the similarity term with surrogates: LSim(eIf, eIm∘phi)+lambda LReg(phi). The deformation field is still estimated from the raw pair; the differentiable warping operator carries gradients from the surrogate-domain loss back into the network. This separates the question of what the network sees from the question of where similarity is well defined, letting a trainer use a domain (preprocessed, masked, or mono-modal) that is reliable even when the raw inputs are heterogeneous.","core_discovery":"Surrogate supervision claims that a deformation field predicted from a raw pair can be trained using a loss evaluated entirely in a separate, cleaner domain. Concretely, raw fixed and moving images are fed to the network to predict phi; instead of minimizing LSim(If, Im∘phi), one selects surrogate images (eIf, eIm)—bias-corrected, masked, or paired same-subject scans—and minimizes LSim(eIf, eIm∘phi)+lambda LReg(phi). Because the warp is differentiable, gradients flow back through the surrogate loss to the network. The paper reports that this yields resilience to simulated 3x MRI intensity inhomogeneity, makes lung CT registration indifferent to input masking, and lets T2-to-T1 MR registratio","pith_inferences":["Editorial: the same surrogate-supervision recipe should generalize to other dense-warp tasks such as optical flow, slice-to-volume alignment, or atlas construction—any setting with a differentiable warp and a trustworthy surrogate domain.","Editorial: the choice of surrogate makes the target invariance explicit and controllable; one could train a single architecture for multiple deployment conditions by swapping surrogate generators, effectively treating surrogate selection as a new hyperparameter.","Editorial: augmentation and surrogate supervision are complementary, so combining them could yield models that are simultaneously robust to a wide range of artifacts (input variability) and well-supervised (clean targets). A practical caveat the paper leaves open is that surrogates that alter geometry, such as distortion-corrected diffusion MRI, must be handled with care."],"forward_implications":["Registration models trained this way can run on raw, unprocessed images at test time, eliminating the need to reproduce the training-time preprocessing pipeline.","Standard intensity-based losses (MSE, NCC) become sufficient for artifact-robust and multi-modal registration, weakening the need for specialized similarity measures.","Established preprocessing algorithms (bias correction, ROI extraction) and paired data can be repurposed as surrogate generators, embedding their domain knowledge into the trained network.","Models become insensitive to whether inputs are pre-masked or not, performing consistently on masked, unmasked, and mixed inputs.","Because the framework imposes no constraints on architecture or loss form, it can be layered onto any existing registration network."],"fun_headline_variants":["Training on surrogate images boosts deformable registration robustness","Surrogate supervision: robust registration despite artifacts, masks, modality","Surrogate images decouple inputs for generalizable registration networks","Train on surrogates to make your registration model resilient","A simple training twist improves registration robustness across domains"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the surrogate images inhabit the same coordinate space as the raw inputs, so warping the surrogate is equivalent to warping the underlying anatomy; the paper itself also notes that success depends on the availability and quality of surrogates, shifting the central burden to domain knowledge.","fun_headline_variants_meta":{"raw":{"variants":["Training on surrogate images boosts deformable registration robustness","Surrogate supervision: robust registration despite artifacts, masks, modality","Surrogate images decouple inputs for generalizable registration networks","Train on surrogates to make your registration model resilient","A simple training twist improves registration robustness across domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1373,"prompt_tokens":737,"completion_tokens":636,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":558}},"tokens_in":481,"tokens_out":636,"duration_ms":6829,"temperature":1.0,"reasoning_tokens":558,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:32:37.338077+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a model on surrogates known to be spatially shifted relative to the input anatomy (e.g., a bias-corrected image that was resampled, or a distortion-corrected diffusion volume) and show that registration accuracy collapses or the deformation field becomes systematically biased; alternatively, measure the correlation between surrogate geometric error and registration error across a test set with known spatial distortions.","supporting_citations":[],"review_version":1}