{"id":"277513b2-a491-4349-8959-6c1d4ab97b7f","arxiv_id":"1908.05926","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A variational Gaussian mixture model with missing-data handling predicts missing MRI contrasts and CT images from observed brain scans using small training sets.","lead":"This paper shows that a probabilistic model that already segments brain scans can also fill in missing medical images, such as missing MRI contrasts or CT scans. It offers an interpretable alternative to deep learning methods and works with small training datasets, which matters for clinical applications with limited data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported PSNRs are computed over all voxels, including large background/air regions; this may inflate the evidence for clinical usefulness and should be checked by re-scoring inside a brain mask.","rationale":"The reader identified missing-at-random as the load-bearing assumption. That concern is legitimate for the differing-fields-of-view experiment, where missingness is spatially structured and may correlate with anatomy. However, the paper's central claim is about predicting entirely missing modalities from one or a few MR contrasts; in those experiments the missing modality is absent for all voxels by acquisition design, so missingness is effectively MAR/MCAR and does not bias the conditional predictions. The stronger threat to the evidence base is that the sole quantitative metric, PSNR, is computed over the entire volume including background. Because head images are dominated by air/background, a model that learns a background class can achieve misleadingly high PSNR while still synthesizing tissue poorly. This concern applies uniformly to all three validation tables, making it more load-bearing for the headline claim than the MAR issue. The paper is otherwise a coherent extension of a known generative model, and the held-out-subject design is appropriate; no circularity is apparent. The verification step of recomputing masked PSNRs would settle whether the quantitative claims survive. Since the reader's verdict is already CONDITIONAL, my read does not change that verdict, but it does redirect attention to a more direct threat. A masked-PSNR reanalysis is a cheap, decisive check that should be part of any revision.","tokens_in":8673,"tokens_out":13006,"duration_ms":145815,"concrete_test":"Recompute Tables 1-3 with PSNR restricted to voxels within a brain/body mask (e.g., from the learned template at tissue probability >0.5, or from a thresholded PDw/T1w intensity). Report masked vs unmasked values; also report PSNR for tissue subclasses. If the masked values drop by more than about 2-3 dB, the conclusions about clinical usefulness and 'good job' need to be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. 3 the only quantitative evidence for the central claim is PSNR, and the paper explicitly states 'no voxels are excluded when computing the PSNR'. Head MR/CT volumes contain a large fraction of background/air voxels with near-constant intensity. A Gaussian component with a small-variance background class can score these almost perfectly, so a substantial fraction of the reported MSE can be trivial. For example, if 60-70% of voxels are background and are predicted with negligible error, a whole-volume PSNR of 28 dB corresponds to roughly 22-24 dB inside the brain. The claims that the model does 'a good job' at filling missing fields of view and yields clinically useful synthesized images therefore rest on a metric that may be inflated by non-tissue voxels. The same issue affects the missing-contrast and MR-to-CT tables, since CT volumes also include air and surrounding material. This is not an objection to the generative model itself, but to the strength of the quantitative support for the paper's headline claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a generative model for medical image translation based on an empirical Bayesian Gaussian mixture model. The unified segmentation model of Ashburner and Friston is extended to handle missing data, so that a multivariate Gaussian mixture with a deformable tissue template can be trained on partially observed multi-channel images and then used to predict missing modalities (MR contrasts or CT) for unseen subjects. The training procedure is a variational EM algorithm in which subject-specific posteriors are computed with a mean-field approximation and population prior hyper-parameters are estimated from the data. Experiments are reported on three tasks: inferring missing spatial voxels in MR volumes with different fields of view (IXI dataset), inferring entirely missing MR contrasts (IXI), and translating T1w/T2w MRI into CT (RIRE). All quantitative results are given as PSNR computed on held-out subjects. The paper's central claim is that image translation is a by-product of learning a joint intensity distribution, and that the model works with a small number of training subjects.","tokens_in":8788,"tokens_out":4332,"duration_ms":44948,"significance":"If the findings are supported, the proposed model would be a valuable interpretable alternative to deep-learning-based image translation, with the same model providing segmentation, spatial normalization, and missing-data synthesis in a single probabilistic framework. The derivation of the variational updates (Eqs. 12–23) is standard and internally consistent, and the use of held-out subjects for all reported PSNR values is a clear strength. The paper also honestly acknowledges several limitations, including the qualitative comparison with deep learning and the failure to capture detailed bone structure in CT synthesis. However, the quantitative evidence for the headline claim is weakened by the choice of evaluation metric (whole-volume PSNR without excluding background), the absence of baseline comparisons on the same data, and an unstated missingness assumption. These issues are addressable, but they need to be resolved before the central claim is fully convincing.","major_comments":[{"comment":"The paper explicitly states that \"no voxels are excluded when computing the PSNR,\" and this is the only quantitative metric used. Head MR and CT volumes contain a large fraction of air/background voxels with near-constant intensities, so a Gaussian mixture component with a small-variance background class can score those voxels almost perfectly. This can substantially inflate the mean PSNR values reported in Tables 1 and 2 and the MR-to-CT result in Section 3.2. Please recompute PSNR within a brain mask (or a tissue mask) and report both whole-volume and masked values, so that the claim that the model \"does a good job at filling in missing fields of view\" is supported by tissue-only accuracy.","section":"Section 3, Eq. (24)"},{"comment":"No baseline comparison is provided on the same held-out data. The PSNR values are reported only for the proposed model, and the reference to prior work in Section 3.1 is a vague statement that results are \"close to those previously reported\" without citing specific numbers. Similarly, the CT experiment in Section 3.2 states the result is \"not on pair with deep learning based techniques\" but does not give any quantitative comparison. Please add at least one baseline on the same data splits (for example, a conditional-mean imputation from a complete-case Gaussian mixture, or a published synthesis method) so that the performance gain of the proposed model is demonstrated rather than assumed.","section":"Section 3, Tables 1 and 2"},{"comment":"The derivation of the observed-channel likelihood and the conditional prediction of missing channels assumes that missingness is ignorable, i.e., the probability that a channel is missing does not depend on the missing intensity value. This assumption is never stated or tested in the manuscript. In clinical data, missingness may be associated with anatomy, pathology, or field-of-view choices, in which case the predicted values from Eq. (9) would be biased. Please state the missingness assumption explicitly and discuss its validity; a sensitivity analysis with simulated non-ignorable missingness would strengthen the claim that the model has clinical utility.","section":"Section 2, Eqs. (6)–(7)"}],"minor_comments":[{"comment":"The equation for the categorical prior is not typeset clearly: the denominator is missing a separating space and the definition of a_dk(φ) is deferred to later text. Please revise for readability.","section":"Eq. (2)"},{"comment":"The subscripts o and m in the expectations E[z_dk x_d]_o, E[z_dk x_d]_m, etc., are not self-explanatory. Define these subscripts (observed and missing parts) before presenting the update equations.","section":"Eq. (17)"},{"comment":"The text says the model is trained on \"50 subjects from the IXI dataset\" and then specifies \"IXI[064-118]\" and \"IXI[002-063]\"; please verify that these indices each correspond to 50 subjects and state the counts explicitly.","section":"Section 3.1"},{"comment":"The comparison with Chartsias et al. is only qualitative; please include the exact PSNR value from that reference when you say the results are \"close\", or remove the comparison.","section":"Section 3.1"},{"comment":"There is a typo in \"can then be drived to zero\" (should be \"driven\"). Please correct this and any other grammatical errors.","section":"Conclusions"},{"comment":"The paper does not mention code or data availability. For a methodological paper, please include a statement about whether the implementation will be released.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The variational derivation and the three proof-of-concept experiments are solid, and the authors are appropriately cautious in their claims. However, the evaluation is currently too weak to support the abstract's assertion that the model yields clinically useful synthesized images. The masking and baseline issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. I would also encourage the editor to consider whether the journal expects a code/data availability statement for this type of paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful news: this paper shows a classical generative model can do medical image translation with small training sets and no black-box machinery. The core idea is natural but genuinely new in combination: a variational empirical-Bayes Gaussian mixture, with deformable templates, extended to handle entirely missing channels. The derivations in Eqs. (12)–(23) are standard and consistent, and the experiments use held-out subjects, so the central claim is not circular. I think the reader's verdict is right: conditional accept, not reject.\n\nWhat the paper does well: it frames image translation as a by-product of joint intensity modeling, which is clean and interpretable. The three experiments are clinically relevant, and the approach is credible even with eight training subjects for MR-to-CT. The qualitative result in Fig. 4 is honest—it admits bone detail and meninges are not captured well.\n\nThe soft spots, in order of importance. First, the stress-test note is correct: the PSNR is computed over all voxels, including a large background/air region. A near-constant background class can inflate the numbers substantially. The authors even state that no voxels are excluded. This is a real weakness in the quantitative support, especially because the abstract says \"results appear promising\" without a brain-mask or tissue-specific metric. It would be easy to re-score inside a mask, and they should. Second, there is no same-data baseline against a deep learning method or a simpler GMM. The comparison to published numbers on other datasets is suggestive but not rigorous. Third, the missing-at-random assumption is never mentioned, let alone tested. In clinical data, missingness often correlates with anatomy or acquisition artifacts, so the predicted images could be biased. At minimum they should state the assumption and discuss its limits. Minor: no code is released, and the template update is only cited, which hurts reproducibility. The self-citations are fine here; they point to the prior segmentation framework the work directly extends.\n\nThe math and the model are solid enough for a serious referee. The paper deserves peer review with requests for revision rather than desk rejection. It will be of real value to researchers who need small-data, interpretable synthesis—especially for PET attenuation correction and multi-modal registration. I would not cite it in my own work in the next year unless I were working in that niche, but I would send it to a capable referee with a request to push on the metric and baseline issues.","headline":"A solid, interpretable generative alternative to deep learning for medical image translation, but the quantitative evidence is inflated by whole-volume PSNR and the missing-at-random assumption goes untested.","tokens_in":9378,"tokens_out":1124,"would_cite":false,"duration_ms":13381,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A probabilistic model of joint brain-scan intensities can predict entirely missing MR contrasts and CT images from one or a few observed contrasts.","keywords":["medical image translation","Gaussian mixture model","missing data","empirical Bayes","variational inference","MRI synthesis","CT synthesis","unified segmentation"],"falsifier":"Take a cohort with naturally truncated fields of view, infer the missing voxels with the trained model, and compare them against the same anatomy imaged in a separate full-coverage scan; if inferred intensities diverge systematically in exactly the regions where coverage is most often cut off, such as the posterior fossa, the missing-at-random assumption fails.","tokens_in":8407,"feed_emoji":"🧠","tokens_out":8218,"duration_ms":66450,"temperature":0.7,"pith_summary":"This paper tries to establish that medical image translation — synthesizing one imaging modality from another — can be achieved as a by-product of learning a joint probability distribution over multi-channel intensities, rather than by training a dedicated mapping. The authors extend the unified segmentation model, a generative Gaussian mixture with a deformable tissue template, to handle voxels and entire channels that are missing. Once fitted to a population, the model can predict an entirely unacquired MR contrast from one or a few observed contrasts, or a CT image from MR scans, using the conditional Gaussian of the missing channels given the observed ones. The practical payoff is that translation inherits the benefits of a principled likelihood model: it works with small training sets, produces probabilistic predictions, and comes bundled with segmentation, bias correction, and spatial normalization.","feed_headline":"One joint-intensity model predicts missing MR contrasts and CTs","feed_subtitle":"Synthesis falls out of learning one probabilistic distribution, so small training sets and per-voxel uncertainty come along.","key_machinery":"The workhorse is the unified segmentation model recast as a variational Gaussian mixture with Gauss-Wishart priors on each class's mean and precision, and a deformable template defining voxel-wise class probabilities. The key move is treating unobserved channels as missing data: the observed-channel likelihood marginalizes the Gaussian over missing dimensions (Eq. 6), and the expected marginal posterior over missing values, $E[h_d] = \\sum_k \\tilde{z}_{dk} \\tilde{h}_{dk}$ from Eq. (9), is the synthetic image. Population parameters (template and prior hyperparameters) are learned by variational expectation-maximisation across training subjects, which is what makes the model empirical Bayesian.","core_discovery":"The central claim is that if a group-wise generative model of brain scans — a variational Gaussian mixture with class labels driven by a deformable template — is allowed to treat unobserved channels as missing data, then the same fitted model that normalizes and segments a subject also outputs the missing images. For each voxel, the posterior over the missing contrasts is a mixture of Gaussians, and its expected value $E[h_d] = \\sum_k \\tilde{z}_{dk} \\tilde{h}_{dk}$ gives the predicted intensities. The authors demonstrate this on three tasks: filling voxels absent because of differing fields of view, predicting entirely missing MR contrasts in all permutations of observed and missing channels, and synthesizing CT from T1w and T2w MR images. They report that the model produces realistic images, that T1w is the most predictive contrast, and that MR-to-CT translation from only eight training subjects reaches a mean PSNR of 25.5 dB.","pith_inferences":["The same machinery could be pushed to scanner harmonization: fitting the joint model on multi-site data and generating a template-contrast image for each subject would isolate site-specific intensity shifts from biological variability.","The missing-at-random assumption, which the paper leaves implicit, can be tested directly by comparing synthetically masked voxels (where the assumption holds by construction) with naturally missing voxels from truncated acquisitions; systematic divergence would indicate bias in real-world applications.","The posterior variance of the expected marginal could serve as a per-voxel confidence map, flagging regions where the synthesized modality is unreliable — a guardrail that adversarial image translation methods lack.","A natural extension is to optimize the template on the target modality as well as the intensity hyperparameters; the paper notes that the meninges and bone detail are missed when only intensities are retrained, so joint template-intensity fitting could close that gap."],"forward_implications":["Missing fields of view in routine clinical MR scans can be filled in rather than discarded, with mean PSNR above 36 dB when half the field of view is missing.","An entirely missing MR contrast can be predicted from any combination of remaining contrasts, with T1w the most informative; predicted contrasts are accurate enough to serve as inputs to deep-learning pipelines trained on a fixed contrast.","CT images can be synthesized from T1w and T2w MR scans after training on as few as eight subjects, offering a radiation-free route to attenuation correction in PET-MR.","The same fitted model yields segmentation, bias correction, and spatial normalization alongside translation, so downstream analysis needs no separate preprocessing."],"supporting_citations":[{"why":"Supplies the Gaussian missing-data machinery (marginalization and conditional prediction) used in Eqs. (6)-(7).","marker":"[15]"},{"why":"The unified segmentation model that the method extends to handle missing modalities.","marker":"[16]"},{"why":"Provides the group-wise template optimization / VEM framework used to learn population priors.","marker":"[17]"},{"why":"Defines the empirical Bayes paradigm that motivates marginalizing subject parameters and optimizing population priors.","marker":"[18]"},{"why":"The variational inference scheme for segmentation that the missing-data ELBO builds on.","marker":"[19]"},{"why":"The deep-learning MR contrast synthesis baseline whose reported PSNR the paper compares against.","marker":"[11]"},{"why":"The deep MR-to-CT synthesis baseline used as comparison for the CT experiment.","marker":"[13]"},{"why":"The dataset providing the MR and CT training/test subjects for the translation experiments.","marker":"[22]"}],"fun_headline_variants":["Bayesian mixture predicts missing brain scans from few MR","One model infers absent MR and CT from sparse inputs","Small-data model fills missing modalities in brain imaging","Predict missing CT and MR contrasts via Bayesian mixture","Eight subjects enough to synthesize CT from MR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole prediction rests on the assumption that missingness is ignorable: whether a voxel's channel is unobserved must be unrelated to what its value would have been, which the paper never states or tests.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian mixture predicts missing brain scans from few MR","One model infers absent MR and CT from sparse inputs","Small-data model fills missing modalities in brain imaging","Predict missing CT and MR contrasts via Bayesian mixture","Eight subjects enough to synthesize CT from MR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2655,"prompt_tokens":840,"completion_tokens":1815,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":1742}},"tokens_in":456,"tokens_out":1815,"duration_ms":11694,"temperature":1.0,"reasoning_tokens":1742,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:00:47.641884+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a cohort with naturally truncated fields of view, infer the missing voxels with the trained model, and compare them against the same anatomy imaged in a separate full-coverage scan; if inferred intensities diverge systematically in exactly the regions where coverage is most often cut off, such as the posterior fossa, the missing-at-random assumption fails.","supporting_citations":[{"cited_title":"Supervised learning from incomplete data via an EM approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian missing-data machinery (marginalization and conditional prediction) used in Eqs. (6)-(7)."},{"cited_title":"Unified segmentation,","cited_arxiv_id":null,"evidence_quote":"The unified segmentation model that the method extends to handle missing modalities."},{"cited_title":"Generative diffeomor- phic modelling of large MRI data sets for probabilistic template construction,","cited_arxiv_id":null,"evidence_quote":"Provides the group-wise template optimization / VEM framework used to learn population priors."},{"cited_title":"Empirical Bayes: Past, present and future,","cited_arxiv_id":null,"evidence_quote":"Defines the empirical Bayes paradigm that motivates marginalizing subject parameters and optimizing population priors."},{"cited_title":"Variational inference for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"The variational inference scheme for segmentation that the missing-data ELBO builds on."},{"cited_title":"Multimodal mr syn- thesis via modality-invariant latent representation,","cited_arxiv_id":null,"evidence_quote":"The deep-learning MR contrast synthesis baseline whose reported PSNR the paper compares against."},{"cited_title":"Deep MR to CT synthesis using unpaired data,","cited_arxiv_id":null,"evidence_quote":"The deep MR-to-CT synthesis baseline used as comparison for the CT experiment."},{"cited_title":"Comparison and evaluation of retrospective intermodality image registration techniques,","cited_arxiv_id":null,"evidence_quote":"The dataset providing the MR and CT training/test subjects for the translation experiments."}],"review_version":1}