{"id":"ce125e7a-ca8b-4f60-9559-3b784c13be89","arxiv_id":"2501.13370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A modality-agnostic deep learning model reconstructs healthy brain anatomy from pathological CT and MRI scans, trained with fluid-dynamics-based synthetic anomaly generation and contralateral brain symmetry.","lead":"UNA is a new AI method that reconstructs healthy brain anatomy from medical scans, including CT and several MRI contrasts, even when the scan contains stroke lesions or other abnormalities. It generates realistic synthetic pathologies on the fly to train the model, so it can be applied directly to new patients without fine-tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-pathology reconstruction is never directly measured: real stroke evaluation only scores healthy regions, while lesion-region numbers come from the model's own synthetic generator with hyperparameters tuned on the same test protocol.","rationale":"The paper has real strengths: the method is clearly specified, code is released, ablations isolate the fluid generator and the contralateral/contrastive components, and the healthy-region results on real data are sensible. None of these criticisms question UNA's utility as a healthy-anatomy synthesis tool. The concern is specifically about the scope of the headline: 'handle both healthy scans and cases with pathology' and 'state-of-the-art performance in all tasks and modalities.' For that scope, the decisive test is whether the model fills in lesion-occluded anatomy correctly on real clinical images. The authors do not and, with current datasets, cannot provide voxel-level ground truth for that quantity, so they instead report healthy-region scores on real stroke cases, which are near-copy tasks because Eq. 7 supervises healthy regions toward the input. The only lesion-region numbers come from synthetic images generated by the same fluid-driven pipeline used in training, and the hyperparameters were selected on that same simulated protocol. This creates a real risk that the lesion-region results are optimistic. The failure case in Sec. 5.2 shows the failure mode is not hypothetical. These issues are addressable with the proposed protocol checks; they do not invalidate the architecture or the training idea. Hence the correct verdict remains conditional. The reader's stated weakest assumption, brain symmetry, is related but secondary: it explains one failure mode, whereas the unmeasured real-pathology reconstruction and the circular synthetic evaluation are the load-bearing gap for the central claim.","tokens_in":20803,"tokens_out":7416,"duration_ms":762110,"concrete_test":"Re-run the Tab. 1 evaluation with a strict protocol: (i) train on lesion profiles from ATLAS/ISLES training subjects only, and generate the simulated test set from lesion masks held out from those subjects; (ii) choose lambda_p and lambda_contrast on a separate validation split, not on the test set; (iii) retrain with at least 3 seeds and report lesion-region L1/PSNR/SSIM with confidence intervals. If UNA's D-column advantage over Brain-ID/PEPSI collapses or overlaps under this protocol, the claim that UNA reconstructs pathology-occluded anatomy is not established by the current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing part of the central claim is that UNA correctly reconstructs anatomy occluded by real pathology, but the paper never directly measures this. On real stroke data (ATLAS, ISLES), Sec. 5.2 states that ground-truth healthy anatomy is unavailable, so Tab. 2 reports reconstruction only in healthy regions. Since the training loss for real images (Eq. 7) masks out lesions and supervises healthy regions toward the input, strong healthy-region scores on ATLAS are largely a copy-the-input result and do not speak to lesion-region restoration. The only quantitative lesion-region scores (Tab. 1, D columns) are produced by UNA's own fluid-driven generator, and the hyperparameters in Tab. C.2 are tuned on the same synthetic test protocol used in Tab. 1, so the reported D-region margins are vulnerable to generator/test overlap. The documented failure mode (Fig. 4 bottom, Sec. 5.2) is precisely a lesion-region error: UNA 'over-corrects' when pathology heavily occludes anatomy. Thus the strongest evidence for the headline claim is either circular (synthetic test from the training generator) or absent (real-data diseased regions unmeasured), independent of whether the symmetry prior in Eq. 8 is biologically reasonable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"UNA, the manuscript under review, proposes a contrast- and resolution-agnostic deep learning framework to reconstruct healthy brain anatomy from images that may contain pathology. The method has three main components: (i) fluid-driven anomaly randomization, which evolves a small set of real lesion segmentations through an advection-diffusion PDE to synthesize diverse pathology profiles; (ii) a 3D UNet trained on a combination of synthetic and real data with a reconstruction loss and a contralateral-paired input that exploits brain symmetry; and (iii) an intra-subject self-contrastive loss that encourages the reconstruction to match the healthy contralateral tissue while remaining distinct from the diseased region. The authors report state-of-the-art results on simulated and real data from CT, T1w, T2w, and FLAIR MRI, and demonstrate a direct application to anomaly detection.","tokens_in":21061,"tokens_out":8903,"duration_ms":71255,"significance":"If the central claim is correct, UNA would be a useful tool for enabling general-purpose healthy-brain analysis models (e.g., SynthSeg, FreeSurfer) to be applied to clinical scans with pathology, without requiring paired annotations or fine-tuning. The paper's strengths include a clearly specified PDE-based generative model with well-defined boundary conditions, a transparent implementation with code release, extensive experiments across eight public datasets and multiple modalities, and an honest discussion of the method's limitations, notably the failure case in Fig. 4. However, the evidence for the load-bearing part of the claim—accurate reconstruction of anatomy occluded by real pathology—is currently indirect: the only quantitative lesion-region scores come from the authors' own synthetic generator, and real stroke data are evaluated only in healthy regions, where the loss trivially encourages copying the input. The contribution is therefore real but not yet fully demonstrated.","major_comments":[{"comment":"The only quantitative evaluation on real stroke data is limited to healthy regions, where the training loss for real images (Eq. 7) supervises the reconstruction toward the input itself; these scores are thus largely a copy-the-input measure and do not attest to the central claim of restoring anatomy occluded by real pathology. The failure case in Fig. 4 (bottom row) shows exactly the lesion-region error that remains unmeasured. Please provide a quantitative assessment of lesion-region reconstruction on real data, for instance by evaluating on a dataset with pre-stroke or post-resolution healthy ground truth, or by a manual expert scoring study, and if none is feasible, state this restriction explicitly and temper the corresponding claims.","section":"Sec. 5.2 and Table 2"},{"comment":"The synthetic test set used in Table 1 is produced by the same fluid-driven anomaly randomization used for training, and the hyperparameters (λ_p, λ_contrast) are selected on the same synthetic test protocol described in Table C.2 as 'same as first-row group in Tab. 1'. Consequently, the diseased-region (D) scores in Table 1 are vulnerable to generator/test overlap and do not provide independent evidence of generalization to real pathology. Please validate on a synthetic test set generated with a different generator or with substantially different parameters, and ensure that hyperparameter selection is performed on a validation set disjoint from the test set.","section":"Sec. 5.1 and Table C.2"},{"comment":"The intra-subject contrastive loss relies on the assumption that the contralateral hemisphere is a reliable healthy reference. This assumption can fail in stroke patients with bilateral pathology, mid-line shift, or mass effect, and the paper's own failure case (Sec. 5.2, Fig. 4 bottom row) is a direct instance of the resulting over-correction. Please provide an analysis of the prevalence and severity of this failure mode on the real datasets, and discuss possible mitigations, such as restricting the contralateral reference to confidently healthy tissue.","section":"Sec. 4, Eq. (8)"},{"comment":"The definition of the voxel weight k(x) is inconsistent with the accompanying text. As printed, for d=1 (real images) the weight is 1−p(x) inside the lesion region and 0 outside it, which contradicts the statement that the loss is computed 'exclusively for the healthy regions'; for d=0 (synthetic images) the weight is 1 inside the lesion and 0 outside, again contradicting the statement that the loss spans the whole brain. Please correct the equation and provide the intended weighting.","section":"Sec. 4, Eq. (7)"}],"minor_comments":[{"comment":"The phrase 'state-of-the-art performance in all tasks and modalities' is broader than the evidence: the real-data evaluation on stroke datasets covers only healthy regions, and ISLES is evaluated only qualitatively. Please make the claim more precise, e.g., 'on simulated lesion regions and healthy regions of real scans'.","section":"Abstract and Sec. 5"},{"comment":"The novelty relative to PEPSI is not fully clear because PEPSI also handles pathology in a contrast-agnostic manner; the Introduction lists several differences, but a direct empirical comparison on the same evaluation protocol would strengthen the positioning.","section":"Introduction vs. Related Work"},{"comment":"References [44] and [45] are the same paper (Modat et al., NiftyReg), and references [16] and [17] are the same paper (Fischl et al.); please merge the duplicate entries.","section":"References"},{"comment":"The claim of 'approximately 10%' improvement over the next best method on ATLAS is not tied to a specific metric; please specify the metric and report confidence intervals or statistical significance.","section":"Table 2"},{"comment":"The anomaly detection Dice scores for UNA are in the range 0.31–0.37; a brief discussion of the operating point and the clinical or practical utility of these scores would help the reader interpret the results.","section":"Table 3"},{"comment":"The Perlin noise code seeds the random number generator with the current time, which hurts reproducibility; please fix the seed using a constant or a configurable parameter.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-written and technically sound in its formulation, but the evaluation of the core claim—reconstruction of anatomy occluded by real pathology—is currently indirect. I recommend major revision rather than rejection because the issues are fixable within the manuscript's scope: the authors can add an independent real-data evaluation of lesion regions, remove the hyperparameter/test overlap by tuning on a disjoint validation set, and compare against a synthetic generator not from their own method. The duplicated references and minor presentation issues are easy to fix. The skeptical assessment raised in the stress-test is largely on target: the paper never directly measures the headline claim on real lesion regions, and the synthetic lesion-region scores are not independent of the training generator."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"UNA proposes a modality-agnostic model to reconstruct healthy brain anatomy from CT and MRI containing pathology. The two genuinely new pieces are the fluid-driven anomaly randomization (advection-diffusion PDEs starting from a few pathology segmentations or Perlin noise) and the intra-subject contralateral contrastive loss. Both are cleanly specified, the code is available, and the synthetic-data evaluation is thorough: UNA beats SynthSR, Brain-ID, and PEPSI across T1w, T2w, FLAIR, and CT on full-brain, healthy-region, and diseased-region metrics. The real-data results on healthy regions are consistently better, and the paper honestly reports a failure case where the model over-corrects when pathology heavily occludes anatomy.\n\nThe stress-test note is right: the load-bearing claim—reconstructing anatomy occluded by real pathology—is never directly measured. On ATLAS and ISLES, the paper only reports healthy-region reconstruction because ground truth healthy anatomy is unavailable. Those strong healthy-region scores partly reflect copy-the-input behavior, not lesion restoration. The only quantitative lesion-region numbers (the D columns in Table 1) come from UNA's own synthetic generator, and the hyperparameters in Table C.2 are tuned on the same synthetic protocol, so those margins are optimistic. This is a real gap, not a manufactured one. The paper is transparent about it in Sec 5.2, yet the abstract and conclusion still say \"state-of-the-art performance in all tasks and modalities,\" which overstates what is shown.\n\nAlso, there are no error bars or significance tests anywhere. That is a moderate issue given the large claimed margins, but the consistent direction of results helps.\n\nWhat holds up: the PDE formulation is well-posed, the training losses are explicit, the ablations show each component contributing, and the method is reproducible from the code. The brain-symmetry assumption is a genuine limitation, and the paper says so.\n\nThis paper is for anyone working on pathology-robust analysis of clinical brain images or synthetic-data training in medical imaging. It deserves a serious referee. My recommendation: send it to review, and ask the authors to add a real-data evaluation of diseased regions—expert rating or a downstream surrogate like lesion segmentation agreement—and to validate hyperparameters on a held-out synthetic protocol, with confidence intervals.","headline":"UNA's fluid-driven anomaly randomization and contralateral contrastive loss are genuinely new and the evaluation is thorough, but the central claim about real-pathology reconstruction is only supported indirectly, so the paper deserves a serious referee with requests for direct validation.","tokens_in":21638,"tokens_out":2315,"would_cite":true,"duration_ms":20630,"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":"UNA is a single modality-agnostic deep network that reconstructs healthy brain anatomy from CT and T1w/T2w/FLAIR MRI scans containing stroke lesions, without needing fine-tuning on the input image.","keywords":["normal brain anatomy reconstruction","fluid-driven anomaly randomization","advection-diffusion PDE","modality-agnostic learning","contralateral symmetry prior","stroke lesion synthesis","anomaly detection","synthetic medical image generation"],"falsifier":"Apply UNA to stroke cases where the lesion substantially involves both hemispheres, so the mirrored counterpart of the diseased tissue is itself diseased; if reconstruction error in the affected regions does not clearly exceed that for unilateral lesions, the symmetry assumption is not in fact load-bearing, whereas a sharp degradation would confirm the mechanism. A complementary check the paper does not report is to compare the shape and intensity distribution of the fluid-generated anomaly profiles against held-out real lesion segmentations with a distributional metric, which would test whether the simulated pathologies are quantitatively realistic rather than merely visually plausible.","tokens_in":20571,"feed_emoji":"🧠","tokens_out":13642,"duration_ms":105852,"temperature":0.7,"pith_summary":"UNA is a single deep-learning network that takes a brain scan, CT or any of the major MRI contrasts, and reconstructs the same brain as it would appear without pathology, working on both healthy scans and scans with stroke lesions and requiring no fine-tuning on the image being analyzed. This matters because the standard toolkit of brain analysis, segmentation, parcellation, registration, and super-resolution, is built and trained for healthy, high-resolution T1w MRI, and degrades in clinical settings where both the acquisition and the anatomy differ. To train such a model without expensive pathology annotations, the paper introduces fluid-driven anomaly randomization, which evolves a handful of real stroke segmentations through an advection-diffusion equation inside the brain boundary to produce an unlimited supply of realistic lesion shapes and intensities on the fly. Exploiting the brain's near-mirror symmetry, the network borrows healthy tissue from the contralateral hemisphere through a self-contrastive loss, and the authors report state-of-the-art reconstruction across CT and T1w/T2w/FLAIR MRI on both simulated and real stroke data, with direct application to anomaly detection.","feed_headline":"One network erases stroke lesions from CT and MRI brain scans","feed_subtitle":"A fluid-flow simulator invents endless realistic lesions, so a single model handles CT and every major MRI contrast.","key_machinery":"The load-bearing mechanism is fluid-driven anomaly randomization: a pathology probability map $P(x,t)$ is treated as mass evolving under the advection-diffusion PDE $\\frac{\\partial P}{\\partial t} = -\\nabla\\times\\Psi(x)\\cdot\\nabla P + \\nabla\\cdot(\\Phi^2(x)\\,\\nabla P)$, where the curl potential $\\Psi$ guarantees incompressible transport and the squared field $\\Phi^2$ guarantees non-negative diffusion, with a zero-Neumann boundary condition keeping the anomaly inside the brain. This turns a few gold-standard stroke segmentations into an unlimited stream of training anomalies whose shapes and intensities obey realistic constraints by construction. The second mechanism is the contralateral-paired input: each image is fed together with its axial-flipped version, registered back to the original frame, so the model can borrow healthy anatomy from the mirror hemisphere; an intra-subject contrastive loss (Eq. 8) enforces similarity to that healthy mirrored tissue and distinctiveness from the diseased input. Together these let a standard 3D U-Net, trained half on synthetic and half on real data, map any input modality to clean anatomy.","core_discovery":"The paper's central claim is that healthy brain anatomy can be reconstructed from pathological scans of any common modality by a single model, without fine-tuning, provided the training set is enriched with synthetic pathologies generated by a fluid-dynamics process. Starting from a real stroke segmentation or a random noise shape, the anomaly map is transported forward in time under an incompressible velocity field and a non-negative diffusion field confined by the brain boundary, yielding endless variations of lesion geometry and intensity; these are encoded onto synthetic healthy scans of randomized contrast and resolution. The reconstruction network is trained with a voxel-wise loss whose weighting focuses on the diseased regions, plus an intra-subject contrastive loss that pulls the output toward the registered mirror-image healthy anatomy and pushes it away from the abnormal input. Evaluated on simulated lesions with known ground truth, on two real stroke datasets, and in a downstream anomaly-detection task, the authors report that UNA outperforms the closest contrast-agnostic synthesis and anomaly-detection methods in full-brain, healthy-region, and diseased-region metrics.","pith_inferences":["Because the anomaly generator is initialized from stroke segmentations but also from pure noise, the same training recipe probably transfers to other lesion types such as tumors, white-matter disease, or hemorrhage, provided the intensity-encoding rule (Eqs. 5-6) is adapted; the paper does not test this.","The over-correction failure mode implies a ceiling on any contralateral-borrowing method: pathologies that are bilateral or that nearly replace an entire hemisphere lack a healthy mirror reference, so clinical deployment would need to detect and flag such cases rather than trust the reconstruction.","A cheap, decisive experiment is to ablate the PDE-based generator against pure noise augmentation while matching lesion volume and intensity statistics; the paper's UNA-(b) ablation is close but does not match statistics, so it does not isolate the value of fluid realism.","UNA's reconstructed healthy anatomy is itself a candidate covariate for stroke outcome prediction or for pathology-aware registration, directions the paper lists as future work but does not quantify."],"forward_implications":["General-purpose analysis tools that expect healthy, high-resolution T1w MRI, including segmentation, parcellation, registration, and super-resolution, can be applied to clinical CT and MRI scans with stroke lesions because UNA first restores the underlying healthy anatomy.","One trained UNA model spans CT, T1w, T2w, and FLAIR MRI at variable resolution without retraining, which would let a single pipeline handle heterogeneous clinical archives.","Anomaly detection follows directly from the reconstruction: the voxel-wise absolute difference between the diseased input and UNA's healthy output locates lesions, and the paper reports state-of-the-art Dice scores against dedicated unsupervised anomaly detectors.","Training requires only a few gold-standard pathology segmentations alongside healthy anatomy labels, since the fluid-driven generator multiplies those few examples into an unlimited set of anomaly profiles.","On healthy inputs, the same network performs anatomy reconstruction and super-resolution, so it can serve as a standard normalizer upstream of existing tools rather than only as a pathology-specific fix."],"supporting_citations":[{"why":"It is the closest prior contrast-agnostic MRI method that handles pathology, and it supplies both the limitations UNA overcomes and the main comparison baseline.","marker":"[39]"},{"why":"It supplies the random-modality, contrast-agnostic synthesis recipe that UNA builds on and serves as a primary reconstruction baseline.","marker":"[38]"},{"why":"It supplies a modality-agnostic super-resolution and T1w synthesis baseline, and UNA uses it in preprocessing to estimate T1w counterparts for the contralateral-paired registration.","marker":"[26]"},{"why":"It establishes the synthetic contrast-agnostic generation paradigm whose label maps and corruption styles UNA reuses to synthesize random-modality healthy images.","marker":"[7]"},{"why":"It provides the gold-standard stroke lesion segmentations that seed the anomaly generator as initial conditions and serve as real evaluation data.","marker":"[36]"},{"why":"It provides a second source of gold-standard stroke lesion segmentations for initializing anomaly profiles and for real-data evaluation.","marker":"[22]"},{"why":"It supplies the regularization-free incompressible-flow and non-negative-diffusion representation of the advection-diffusion PDE that keeps the anomaly transport well posed.","marker":"[40]"},{"why":"It supplies the first-order upwind numerical scheme used to discretize and solve the forward advection-diffusion process.","marker":"[33]"},{"why":"They compute the deformations between each scan and its axial-flipped counterpart that make the contralateral-paired input and the symmetry loss possible.","marker":"[44,50]"}],"fun_headline_variants":["Fluid simulation teaches AI to erase brain lesions","One model reconstructs healthy anatomy from any brain scan","UNA erases pathology from CT and MRI with synthetic fluid","Fluid-driven anomaly randomization normalizes brain scans","Single network erases stroke lesions across modalities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the brain is close enough to mirror-symmetric that the flipped healthy hemisphere supplies a valid reference for the diseased side; the paper's own failure case shows what happens when that premise runs out, when a lesion heavily occludes the underlying anatomy, UNA over-corrects because no healthy mirror tissue remains to borrow from.","fun_headline_variants_meta":{"raw":{"variants":["Fluid simulation teaches AI to erase brain lesions","One model reconstructs healthy anatomy from any brain scan","UNA erases pathology from CT and MRI with synthetic fluid","Fluid-driven anomaly randomization normalizes brain scans","Single network erases stroke lesions across modalities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000372,"raw_usage":{"total_tokens":2011,"prompt_tokens":992,"completion_tokens":1019,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":945}},"tokens_in":608,"tokens_out":1019,"duration_ms":22929,"temperature":1.0,"reasoning_tokens":945,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:00:49.388539+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply UNA to stroke cases where the lesion substantially involves both hemispheres, so the mirrored counterpart of the diseased tissue is itself diseased; if reconstruction error in the affected regions does not clearly exceed that for unilateral lesions, the symmetry assumption is not in fact load-bearing, whereas a sharp degradation would confirm the mechanism. A complementary check the paper does not report is to compare the shape and intensity distribution of the fluid-generated anomaly profiles against held-out real lesion segmentations with a distributional metric, which would test whether the simulated pathologies are quantitatively realistic rather than merely visually plausible.","supporting_citations":[{"cited_title":"PEPSI: Pathology-enhanced pulse-sequence-invariant representations for brain MRI","cited_arxiv_id":null,"evidence_quote":"It is the closest prior contrast-agnostic MRI method that handles pathology, and it supplies both the limitations UNA overcomes and the main comparison baseline."},{"cited_title":"Alexander, and Juan E","cited_arxiv_id":null,"evidence_quote":"It supplies the random-modality, contrast-agnostic synthesis recipe that UNA builds on and serves as a primary reconstruction baseline."},{"cited_title":"Magdamo, Steve Arnold, Sudeshna Das, et al","cited_arxiv_id":null,"evidence_quote":"It supplies a modality-agnostic super-resolution and T1w synthesis baseline, and UNA uses it in preprocessing to estimate T1w counterparts for the contralateral-paired registration."},{"cited_title":"Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce R","cited_arxiv_id":null,"evidence_quote":"It establishes the synthetic contrast-agnostic generation paradigm whose label maps and corruption styles UNA reuses to synthesize random-modality healthy images."},{"cited_title":"A large, open source dataset of stroke anatomical brain images and manual lesion segmentations","cited_arxiv_id":null,"evidence_quote":"It provides the gold-standard stroke lesion segmentations that seed the anomaly generator as initial conditions and serve as real evaluation data."},{"cited_title":"ISLES 2022: A multi-center magnetic resonance imag- ing stroke lesion segmentation dataset","cited_arxiv_id":null,"evidence_quote":"It provides a second source of gold-standard stroke lesion segmentations for initializing anomaly profiles and for real-data evaluation."},{"cited_title":"Discovering hidden physics be- hind transport dynamics","cited_arxiv_id":null,"evidence_quote":"It supplies the regularization-free incompressible-flow and non-negative-diffusion representation of the advection-diffusion PDE that keeps the anomaly transport well posed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the first-order upwind numerical scheme used to discretize and solve the forward advection-diffusion process."}],"review_version":1}