{"id":"e056e410-f3fa-4177-8618-e7f2974a67f1","arxiv_id":"1908.07646","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors propose a learned shared representation (communal domain learning) for registering images across modalities and report improved affine registration on T1-T2 and T1-CT brain data, but the derivation contains invalid steps.","lead":"Medical scans of the same anatomy from different machines look so different that computers struggle to line them up. This paper proposes training a network to map both scan types into a shared space first, reporting accuracy gains, but the math behind the method has serious flaws.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (3) is not mutual information and Supplementary B does not establish Gaussian preservation; the CDL training objective and search direction are therefore unsupported, and the empirical table cannot repair this.","rationale":"As a second-pass reviewer, I focused on the condition under which the central claim would hold: the CDL network must actually maximize mutual information between source and target in the learned subspace, and the experiments must demonstrate the claimed improvement. The first condition is the least secure. The paper's own supplementary derivation contradicts the Gaussian-preservation premise, and the surrogate in eq (3) is algebraically not a mutual information estimate even under the assumed Gaussianity. This is not a disagreement with a consensus prior; it is an internal mathematical inconsistency that can be exhibited with a simple Gaussian example. The empirical evidence is also not reliable: impossible standard deviations in Table 1 mean the reported p < 0.001 cannot be accepted as reported. No formal verification or released code offsets these problems. I agree with the reader that the theoretical justification collapses; I would maintain the REJECT verdict. The check I propose is cheap and decisive: if eq (3) fails on simulated Gaussians and the transformed densities are non-Gaussian, the paper's cost function is not what the central claim requires.","tokens_in":8234,"tokens_out":6119,"duration_ms":144087,"concrete_test":"Run a two-part numerical check. (1) For jointly Gaussian variables with correlation rho = 0.9, evaluate eq (3); if the surrogate equals -0.05 instead of the true MI 0.83, the equation is not mutual information. (2) Draw x ~ N(0,1), set y = tanh(x), and compare the empirical density of y with eq (14) and with a best-fit Gaussian; the nonconstant Jacobian will make the exact density non-Gaussian. If both checks fail, the CDL cost and gradients are not optimizing the stated information-theoretic objective.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim that CDL maximizes shared information and thereby improves registration rests on eq (3), which replaces mutual information with a cross-correlation surrogate. This step fails for two independent reasons. First, even if the activations preserved Gaussianity, the mutual information of jointly Gaussian variables is I = -(1/2) ln(1 - rho^2), not -(1/2)(1 - rho). Eq (3) can be negative (e.g., rho = 0.9 yields -0.05 while the true MI is about 0.83), so it is not a valid MI estimate, not a lower bound, and maximizing it is not maximizing shared information. Second, Supplementary B claims tanh and sigmoid activations preserve Gaussianity, but the densities in eqs (14) and (16) are not Gaussian: the Jacobian factors 1/(1-y^2) and 1/(y(1-y)) are nonconstant and the inverse activations are nonlinear. The Taylor expansions are local approximations and do not make the transformed variable Gaussian; the statement that the expansion demonstrates Gaussianity is a non sequitur. Because the CDL training objective (2), the gradients (9)-(10), and the registration search direction (11) all differentiate this surrogate, the theoretical foundation of the central claim collapses. The experiments do not supply independent support: Table 1 lists impossible standard deviations for a Dice score bounded in [0,1] (e.g., 0.76 +/- 0.54), and the Wilcoxon test statistic is not reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Communal Domain Learning (CDL), a neural-network representation learned jointly from aligned source/target image pairs, followed by affine registration performed in the learned representation. The training objective (Eq. 2) combines a purported mutual-information term (Eq. 3), a distribution-difference term (Eq. 6), and weight regularization. The authors claim that a communal subspace is guaranteed to exist (Proposition 1), that hierarchical tanh/sigmoid transforms preserve Gaussianity (Supplementary B), and that CDL yields statistically significant Dice and Hausdorff distance improvements over MI baselines for T1-T2 and T1-CT registration (Table 1).","tokens_in":8617,"tokens_out":8133,"duration_ms":492474,"significance":"If the claims were valid, the paper would offer a generic learned representation for multi-sequence and multi-modal registration, a practically important goal. The authors are to be credited for evaluating on separate training and test sets and for comparing against several MI-based masking variants. However, the central theoretical mechanism is invalid: Eq. (3) is not mutual information, Supplementary B does not establish Gaussianity preservation, Eq. (6) is not the stated maximum mean discrepancy, and Proposition 1 is a near-tautological subspace statement. The empirical evidence is also internally questionable, with impossible standard deviations reported for some Dice scores. The central claims are therefore not supported by the manuscript as written.","major_comments":[{"comment":"Equation (3) is not mutual information and cannot justify the equality asserted immediately after it. For jointly Gaussian variables the correct mutual information is I = -1/2 ln(1 - rho^2), not -1/2(1 - rho); the two disagree numerically (e.g., rho = 0.9 gives I ≈ 0.83 while Eq. (3) gives -0.05). The inequality I ≥ -1/2(1 - rho) is vacuous because the right-hand side is non-positive for all |rho| ≤ 1. Since the CDL training objective, its gradients (4)-(5), and the registration search direction (11) all differentiate this surrogate, maximizing the cost in Eq. (2) is not maximizing shared information.","section":"§2.1, Eq. (3)"},{"comment":"The densities in Eqs. (14) and (16) are not Gaussian densities: the Jacobian factors 1/(1-y^2) and 1/(y(1-y)) are nonconstant and the exponential arguments contain the nonlinear inverse activations. The Taylor expansions in Eqs. (15) and (17) are local approximations of the inverse functions, not a proof that the transformed variables are Gaussian; the statement that the expansions demonstrate Gaussianity is a non sequitur. The Gaussianity preservation premise is therefore unsupported, and with it the derivation of Eq. (3).","section":"Supplementary B, Eqs. (14) and (16)"},{"comment":"Equation (6) is not the maximum mean discrepancy of Gretton et al.; MMD is defined as the RKHS norm of the difference between mean embeddings and requires a kernel. At best, Eq. (6) is a squared MMD for a linear kernel, which matches only first moments and says very little about the probability differences between the drifted domains. The distribution-difference term in the cost is therefore not what the paper claims it is.","section":"§2.1, Eq. (6)"},{"comment":"Proposition 1 is a near-tautological statement about non-orthogonal vector spaces sharing a non-empty intersection; it does not establish that a drift-invariant subspace exists, nor that registration performed in any such subspace will be accurate. In addition, Eqs. (2) and (12) are the same objective, so the training procedure maximizes exactly the criterion used at registration time; this explains the favorable comparison to MI but does not validate the proposed subspace theory.","section":"§2.1-§2.2, Proposition 1 and Eq. (12)"},{"comment":"Some reported Dice values are impossible for a score bounded on [0,1] if the ± entries are standard deviations (e.g., 0.76 ± 0.54 and 0.96 ± 0.26); at minimum, the table does not state what the error bars represent. The statistical claim p<0.001 is also incomplete: the test compares paired registration outcomes on the same image pairs, so a paired test such as Wilcoxon signed-rank would be appropriate, yet neither the test statistic nor the number of pairs per cell is reported. These issues prevent the reader from verifying the headline empirical claim.","section":"Table 1"}],"minor_comments":[{"comment":"The text refers to Fig. 3(a) and Fig. 3(b), but the displayed figure is captioned Fig. 2; figure numbering should be made consistent.","section":"Section 3.1"},{"comment":"The free parameters α, β, λ, and M are introduced but no selection procedure or sensitivity analysis is reported, which is important because the method is advertised as generic.","section":"Section 2, Algorithm 1"},{"comment":"The statement that p(0) = d = 3 for volumetric images conflicts with the use of scalar image volumes as network inputs; the input dimension should be the number of voxels in the representation, not 3.","section":"Section 2"},{"comment":"The reference to Katyal et al. [7] is to a MICCAI grand challenge paper on fluid-tissue categorization; it is not sufficient support for the general claim that joint and marginal distributions of multi-sequence/multi-modal intensities are Gaussian.","section":"Section 2.1"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the core idea—learn a shared subspace for multimodal registration—is reasonable, but the theoretical scaffold collapses and the reported statistics are impossible. As written, it should not be trusted.\n\nWhat's new: the CDL setup—training a small network to maximize shared information and minimize distribution distance, then registering in that learned space—is not in the cited registration literature. The T1-T2 and T1-CT experiments are clinically relevant, and the paper is clearly organized. The authors even acknowledge in a footnote that the Gaussian assumption is domain-dependent and a topic of future research, which is honest. That said, the paper doesn't cite the domain adaptation literature where learning invariant representations is standard, so the novelty is overstated.\n\nThe problems: eq (3) replaces mutual information with -(1/2)(1 - Corr). For Gaussians, MI = -(1/2)log(1 - Corr^2). Their formula is not a lower bound or an approximation; it can be negative at high correlation. So the training objective does not maximize shared information. The supplementary proof that tanh/sigmoid preserve Gaussianity is wrong: eqs (14) and (16) are not Gaussian densities—they include Jacobian factors 1/(1-y^2) and 1/(y(1-y)), and the exponential contains atanh(y), not y. The Taylor expansion only shows approximate linearity in a small neighborhood; it does not make the transformed variable Gaussian. So the theoretical justification for using cross-correlation is unsupported.\n\nThe MMD term (eq 6) is just the squared norm of the difference of sample means, not a kernel MMD. That's a misnomer, and the distribution matching is only first-order.\n\nTable 1 has standard deviations exceeding the theoretical maximum for Dice scores (e.g., 0.76 ± 0.54; the maximum possible for a [0,1] variable with mean 0.76 is about 0.41). The Wilcoxon test statistic is not reported, so the p<0.001 claim cannot be checked. Baselines are only standard MI; the authors say learning-based implementations were unavailable, which is understandable but leaves the comparison thin.\n\nThe same cost appears in training and registration (eqs 2 and 12), so the reported gains may not transfer to an independent similarity metric. That is a design choice, but it weakens the claim that the learned representation is generically useful.\n\nWho it's for: someone who wants a quick look at a plausible idea might skim Section 2, but the theory and numbers are unreliable. The paper needs a major rewrite: correct the information-theoretic formulas, provide a valid proof or an explicitly empirical justification, and redo the evaluation. In its current form, I would not send it to peer review; the math errors are clear enough for a desk reject.","headline":"Reasonable high-level idea, but the math is wrong in load-bearing places and the reported statistics are impossible; as written, not credible.","tokens_in":9062,"tokens_out":8427,"would_cite":false,"duration_ms":169603,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that registering drifted image pairs inside a learned communal subspace—shared by source and target domains—improves affine alignment of T1/T2 MRI and MR/CT over mutual-information baselines, with reported p<0.001 gains…","keywords":["image registration","multimodal registration","multisequence MRI","domain adaptation","communal subspace","mutual information","maximum mean discrepancy","deep representation learning"],"falsifier":"Compute the exact probability density of tanh(x) and sigmoid(x) for Gaussian inputs with the mean and variance actually seen in the training volumes, and measure how far the result is from Gaussian, for example by KL divergence over the intensity range. If the divergence is large, or if replacing the true mutual information by the cross-correlation surrogate in eq. (3) changes the affine parameters found by the optimizer, the claimed guarantee that CDL maximizes shared information is refuted.","tokens_in":8033,"feed_emoji":"🧠","tokens_out":7327,"duration_ms":76496,"temperature":0.7,"pith_summary":"Image pairs that see the same anatomy but come from different MRI sequences or modalities do not share an intensity distribution, which makes standard similarity metrics struggle. This paper proposes communal domain learning (CDL): train a small neural network to map source and target images into a shared 'communal' subspace, then run the registration similarity and optimizer inside that subspace instead of in the raw intensities. The CDL objective simultaneously maximizes shared information between domains and penalizes distribution differences, and the mutual-information term is replaced by a cross-correlation surrogate under a Gaussianity assumption. On affine registration of T1-T2 and T1-CT brain images, CDL reports statistically significant gains over mutual-information baselines (Wilcoxon rank sum, p<0.001), with higher Dice scores and lower Hausdorff distances. If the claim holds, one learned representation could serve as a generic front end for registering images across modalities, sequences, and acquisition settings.","feed_headline":"Shared image subspace beats mutual information for MRI/CT registration","feed_subtitle":"A neural network maps drifted image pairs into one common space, improving affine alignment over mutual information with p<0.001.","key_machinery":"The load-bearing object is the CDL network, a fully connected network with sigmoid activations trained to learn the communal subspace W. Its loss combines three terms: an information term I(h_t^(M); h_s^(M)) approximated by the cross-correlation formula in eq. (3), a distribution-difference term D based on maximum mean discrepancy, and Frobenius and L2 regularization on weights and biases. The same cost function is reused at registration time, and its derivative with respect to the transformation parameters gives the search direction (eq. 11), so the learned subspace is not a preprocessing step but the space in which the alignment is decided. The paper's justification that eq. (3) is a valid stand-in for mutual information is the assumption that hierarchical tanh/sigmoid transforms preserve the Gaussianity of intensities.","core_discovery":"The central claim is that whenever two domains image the same field of view they share a non-empty set of drift-invariant features, called the communal subspace W, and that a hierarchical nonlinear network can learn W from aligned training pairs. During training, the network output h(M) is optimized with the cost C = I - alpha*D - beta*regularization, where I is the mutual information approximated by the cross-correlation lower bound of eq. (3), and D is the maximum mean discrepancy between source and target outputs. At registration time, the same network is used as a feature extractor: the similarity term and the search direction for affine parameters are computed from h(M) and the same loss C. The paper reports that this beats MI-based registration for multisequence T1-T2 (Dice from 0.80 to 0.93 for the no-mask variant) and multimodal T1-CT (Dice from 0.83 to 0.96), with p<0.001.","pith_inferences":["Inference: The need for pre-aligned training pairs is the clear practical bottleneck; a self-supervised variant that generates pseudo-alignments from initial registrations would be the natural next test.","Inference: The Gaussianity-preservation step is the theoretical hinge; a direct diagnostic on the actual distribution of network outputs would show whether eq. (3) is maximizing true mutual information or only a proxy.","Inference: Since CDL is a learned representation rather than a metric, it could complement other learned or classical metrics; the question whether the benefits come from the subspace or from the cross-correlation surrogate could be tested by ablating each term in eq. (2)."],"forward_implications":["If the central claim is correct, registration across modalities and sequences reduces to a training problem: one CDL network per domain pair replaces hand-tuned similarity metrics and the choice of masking.","Because the same loss and optimizer are used for both learning and registration, CDL should transfer to other transformation models, including nonrigid registration, without changing the metric or the subspace.","Reported robustness to masking suggests the learned subspace does not depend on accurate background or brain extraction, which would make the method easier to deploy in clinical pipelines.","The faster convergence shown in Fig. 3(b) implies that besides a better final alignment, registration in the communal subspace can reduce the number of optimizer iterations needed."],"supporting_citations":[{"why":"Supplies the human-visual-system theory that a drift-invariant communal space exists, motivating learning W instead of mapping one domain to the other.","marker":"[6]"},{"why":"Provides the empirical Gaussian intensity model for same-field-of-view multisequence MR images that lets the paper model mutual information parametrically.","marker":"[7]"},{"why":"Gives the lower bound connecting mutual information to cross-correlation for Gaussian variables, from which eq. (3) is derived.","marker":"[8]"},{"why":"Supplies the maximum mean discrepancy criterion used as the distribution-difference penalty in the CDL cost function.","marker":"[9]"},{"why":"Underwrites the constrained parameter search and gradient-descent optimizer used for registration in the learned subspace and for the MI baselines.","marker":"[5]"}],"fun_headline_variants":["Drift-proof image registration via learned communal subspace","Neural network learns shared subspace for robust MRI/CT alignment","Cross-modal registration with a single shared subspace","One subspace to align them all: domain-invariant registration","Beyond MI: communal subspace lifts MRI/CT registration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the assumption that the network's nonlinear activations leave Gaussian-distributed intensities Gaussian enough that a simple cross-correlation formula can stand in for mutual information; if that assumption fails, the training objective is not actually maximizing shared information.","fun_headline_variants_meta":{"raw":{"variants":["Drift-proof image registration via learned communal subspace","Neural network learns shared subspace for robust MRI/CT alignment","Cross-modal registration with a single shared subspace","One subspace to align them all: domain-invariant registration","Beyond MI: communal subspace lifts MRI/CT registration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001299,"raw_usage":{"total_tokens":5339,"prompt_tokens":1024,"completion_tokens":4315,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":4239}},"tokens_in":640,"tokens_out":4315,"duration_ms":34116,"temperature":1.0,"reasoning_tokens":4239,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:00:42.072826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact probability density of tanh(x) and sigmoid(x) for Gaussian inputs with the mean and variance actually seen in the training volumes, and measure how far the result is from Gaussian, for example by KL divergence over the intensity range. If the divergence is large, or if replacing the true mutual information by the cross-correlation surrogate in eq. (3) changes the affine parameters found by the optimizer, the claimed guarantee that CDL maximizes shared information is refuted.","supporting_citations":[{"cited_title":"Proceedings of the National Academy of Sciences 106(21) (2009) 8748--8753","cited_arxiv_id":null,"evidence_quote":"Supplies the human-visual-system theory that a drift-invariant communal space exists, motivating learning W instead of mapping one domain to the other."},{"cited_title":"Proceedings of the MICCAI Grand Challenge on MR Brain Image Segmentation (2013)","cited_arxiv_id":null,"evidence_quote":"Provides the empirical Gaussian intensity model for same-field-of-view multisequence MR images that lets the paper model mutual information parametrically."},{"cited_title":"Entropy 13(3) (2011) 612--649","cited_arxiv_id":null,"evidence_quote":"Gives the lower bound connecting mutual information to cross-correlation for Gaussian variables, from which eq. (3) is derived."},{"cited_title":"In: Advances in Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum mean discrepancy criterion used as the distribution-difference penalty in the CDL cost function."},{"cited_title":"IEEE Transactions on Image Processing 16(12) (2007) 2879--2890","cited_arxiv_id":null,"evidence_quote":"Underwrites the constrained parameter search and gradient-descent optimizer used for registration in the learned subspace and for the MI baselines."}],"review_version":1}