{"id":"9283d9d4-ad48-4e24-9bad-c1a8046f7978","arxiv_id":"2508.03073","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Nexus-INR combines arbitrary-scale implicit-neural super-resolution, cross-modal knowledge distillation, and joint segmentation to improve multimodal brain MRI reconstruction and tumor segmentation.","lead":"This preprint describes Nexus-INR, a system for sharpening medical MRI images at any requested zoom level using several scan types together. It also trains a tumor-segmentation task in the same network, and reports better results than existing methods on the BraTS2020 benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The submitted full text is a different paper (SSFMamba, arXiv 2508.03069v2), so the Nexus-INR abstract's claims have no supporting equations, architecture details, or experiments in the reviewed artifact; the central claim is therefore unverified rather than validated.","rationale":"The supplied artifact is internally mismatched: the metadata and abstract describe Nexus-INR, an arbitrary-scale multimodal super-resolution framework, while the full text is SSFMamba, a 3D segmentation paper with different authors, no super-resolution component, and a different loss function. Treating the full text as in-scope evidence, it contains no Nexus-INR equations, architecture diagrams, training details, or experimental tables, so the paper's headline result cannot be checked. This is the most load-bearing problem because every downstream question about disentanglement or distillation presupposes that such a method is actually specified. The reader also flagged the mismatch and assigned UNVERDICTED; I agree that the appropriate outcome is unchanged, since the defect is insufficiency of evidence rather than an established error. My proposed check is to retrieve the correct PDF from arXiv and verify that it supplies the missing architecture and experiments; if it does not, no scientific verdict beyond UNVERDICTED is possible. I do not attribute the mismatch to the authors, and the concern is not about the plausibility of the abstract's mechanism but about the absence of any verifiable specification in the reviewed manuscript.","tokens_in":17872,"tokens_out":2467,"duration_ms":30362,"concrete_test":"Retrieve the full PDF for arXiv:2508.03073 directly from arXiv or from the authors and verify that its main body describes Nexus-INR rather than SSFMamba. Specifically, check for: (1) a loss expression combining reconstruction, self-supervised consistency, classification, and segmentation terms; (2) an experimental table comparing Nexus-INR against state-of-the-art arbitrary-scale super-resolution methods on BraTS2020 with metrics such as PSNR and SSIM; (3) downstream segmentation results obtained from the reconstructed images. If the retrieved PDF matches the supplied SSFMamba text or lacks these components, the central claim remains unverified and the verdict should stay UNVERDICTED. If the correct body is supplied and contains these items, the concern dissolves and a normal technical review can proceed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Nexus-INR outperforms state-of-the-art methods on BraTS2020 for arbitrary-scale super-resolution and downstream segmentation. For that claim to hold, the paper must contain a coherent specification of the dual-branch encoder with auxiliary classification, the cross-modal attention distillation with self-supervised consistency loss, the integrated segmentation module, and the experimental protocol with metric tables. The reviewed full text contains none of these: it is SSFMamba, a Mamba-based segmentation method with a different author list and no super-resolution component. The only Nexus-INR content in the artifact is the abstract. This is not a disagreement with consensus; it is a missing-evidence problem. The abstract's assertion that experiments demonstrate superiority cannot be checked, and the disentanglement and distillation premises cannot be derived from the supplied text. The manuscript body provides no figure, equation, loss definition, dataset split, or comparison table for Nexus-INR. Therefore the strongest claim is unsupported by the artifact as reviewed. If the correct PDF exists elsewhere, the submitted artifact is defective; if the supplied body is the actual submission, the claim is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The submitted manuscript is titled 'Nexus-INR: Diverse Knowledge-guided Arbitrary-Scale Multimodal Medical Image Super-Resolution' and its abstract describes a three-component framework for arbitrary-scale super-resolution of multimodal brain MRI, combining a dual-branch encoder with an auxiliary classification task, a cross-modal attention distillation module with a self-supervised consistency loss, and an integrated segmentation module. The abstract claims state-of-the-art performance on the BraTS2020 dataset for both super-resolution and downstream segmentation. However, the full text of the submitted artifact is a different paper, 'SSFMamba: Learning Symmetry-driven Spatial-Frequency Modeling for Physically Consistent 3D Medical Image Segmentation' (arXiv:2508.03069v2), with a different title, author list, and no super-resolution content. The body of the submission contains no equations, architecture diagrams, loss definitions, or experimental results for Nexus-INR. The only Nexus-INR content is the abstract, so the central claims of the submission cannot be verified or evaluated from the provided artifact.","tokens_in":18013,"tokens_out":2409,"duration_ms":28550,"significance":"If Nexus-INR works as described, the proposed combination of classification-guided disentanglement, cross-modal knowledge distillation, and integrated segmentation for arbitrary-scale super-resolution would be a potentially valuable contribution to multimodal medical image analysis, addressing a real limitation of INR-based methods in handling multi-modal inputs with varying resolutions. However, the submitted manuscript provides no evidence toward this claim: the full text is an unrelated segmentation paper, and the Nexus-INR abstract alone cannot support any assessment of soundness, novelty, or experimental validity. The SSFMamba paper that constitutes the body is a separate work with its own experiments and contributions, but it is not the paper under review and does not address super-resolution. Consequently, the significance of the submitted manuscript as it stands is unassessable.","major_comments":[{"comment":"The submitted full text is the SSFMamba paper (arXiv:2508.03069v2), not Nexus-INR. The title, author list, abstract, method, and experiments all describe a frequency-domain Mamba segmentation framework, with no mention of arbitrary-scale super-resolution, implicit neural representations, or the Nexus-INR components. The central claim that Nexus-INR outperforms state-of-the-art methods on BraTS2020 for super-resolution and downstream segmentation is therefore completely unsupported by the artifact as submitted.","section":"Full text (all sections, I–V)"},{"comment":"The abstract describes three specific components: a dual-branch encoder with an auxiliary classification task for disentanglement, a knowledge distillation module using cross-modal attention with a self-supervised consistency loss, and an integrated segmentation module. None of these components appear anywhere in the full text. The SSFMamba body contains a dual-branch encoder, but it separates spatial and frequency features for segmentation and includes no classification task, no cross-modal attention, no consistency loss, and no super-resolution module. The claimed mechanism is therefore unverifiable from the submitted manuscript.","section":"Abstract, 'three key components'"},{"comment":"The abstract says experiments on BraTS2020 demonstrate superiority in both super-resolution and downstream segmentation. The full text reports only Dice and HD95 segmentation results on BraTS2020, BraTS2023, and BTCV for SSFMamba, with no super-resolution metrics (e.g., PSNR, SSIM), no arbitrary-scale evaluation, and no comparison for the super-resolution task. The reported experiments are for a different method and task, so the empirical claims in the abstract are not backed by any table, figure, or statistical analysis in this submission.","section":"Experimental sections (Sections IV.A–IV.C)"}],"minor_comments":[{"comment":"The arXiv identifier in the reader's materials is 2508.03073 (eess.IV) but the full text carries the arXiv identifier 2508.03069v2 and a different title. This appears to be a packaging error, but it means the submitted artifact is internally inconsistent between the abstract and the body.","section":"Manuscript metadata"},{"comment":"The SSFMamba body contains corrupted mathematical expressions and garbled symbols in several equations and tables (e.g., the loss function in Section III.C and various numbers in Tables I–IV). Since this text is not the submitted paper, I do not assess it substantively, but the rendering issues would need to be fixed before any resubmission.","section":"Formatting of the supplied full text"}],"recommendation":"reject","confidential_remarks":"This submission appears to be the result of an uploading error: the abstract describes Nexus-INR while the full text is the SSFMamba segmentation paper. As a referee I can only evaluate the artifact as provided, and by that standard the manuscript does not contain the work it claims to present. Resubmission of a correct Nexus-INR manuscript would be required before any soundness assessment is possible. I would not treat the SSFMamba paper's content as evidence for Nexus-INR, even though that paper may itself be a valid submission elsewhere."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know upfront: the full text attached to arXiv 2508.03073 is not Nexus-INR. It is a separate paper, SSFMamba, with a different title, a mostly different author list, and no super-resolution component. The only Nexus-INR content in the artifact is the abstract. That is a load-bearing defect, not a quibble: we literally cannot check the architecture, losses, or experiments the abstract claims.\n\nFor what it is worth, the abstract describes a sensible research direction: a dual-branch INR encoder for arbitrary-scale multimodal super-resolution, using classification to disentangle shared anatomy from modality-specific detail, cross-modal attention distillation from a high-resolution reference, and an integrated segmentation head. The pieces are all known in the literature, so the contribution is a specific combination plus an empirical claim on BraTS2020. I can see why someone might want to read that paper.\n\nBut on the evidence in front of us, the claim is unverified. There are no equations, no tables, no error bars, no ablations. The abstract states SOTA without statistical support, and the disentanglement premise—that the auxiliary classification task really separates structure from modality detail—is exactly the kind of thing that needs experimental validation. The stress-test note is right: the central claim is unsupported by the artifact as reviewed.\n\nThe SSFMamba text itself is not garbage, but it is not the paper under review. Reviewing it would be reviewing the wrong manuscript. I would not spend referee time on this artifact, and I would not cite Nexus-INR based on what I have seen. If the correct PDF exists, the authors should resubmit it; if this PDF is what they intended, then the submission is fundamentally broken.\n\nMy recommendation: desk reject this artifact and ask the authors to check their upload. The underlying idea may deserve a serious referee, but only when we can actually read it.","headline":"The submitted PDF is not the Nexus-INR paper—it is a different Mamba segmentation paper (SSFMamba) with different authors—so the artifact cannot be reviewed as submitted.","tokens_in":18589,"tokens_out":1315,"would_cite":false,"duration_ms":17411,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Nexus-INR is an arbitrary-scale multimodal medical image super-resolution framework that, according to the paper, outperforms state-of-the-art methods on the BraTS2020 dataset for both reconstruction and downstream tumor segmentation.","keywords":["arbitrary-scale super-resolution","implicit neural representation","multimodal medical imaging","cross-modal knowledge distillation","brain MRI","BraTS2020","tumor segmentation","classification-guided disentanglement"],"falsifier":"Run Nexus-INR on BraTS2020 while withholding or randomly permuting the registration of the high-resolution reference modality. If super-resolution quality stays high, the cross-modal attention is not the source of the gain, contradicting the claimed mechanism. A second check: compare against a single-modality INR with identical capacity and same training budget; if the multimodal version does not beat it clearly, the claimed benefit of cross-modal distillation is not demonstrated.","tokens_in":17621,"feed_emoji":"🧠","tokens_out":4743,"duration_ms":51521,"temperature":0.7,"pith_summary":"Nexus-INR claims to solve arbitrary-scale super-resolution for multimodal medical images. The paper proposes a three-component framework on an implicit neural representation (INR) backbone: a dual-branch encoder that uses an auxiliary classification task to split features into shared anatomical structure and modality-specific detail, a cross-modal attention distillation module that transfers high-frequency information from a high-resolution reference modality to a low-resolution one, and an integrated segmentation module that injects anatomical semantics. The claimed payoff is that a single model can reconstruct a brain MRI at any requested scale while borrowing detail from companion modalities, and the same features improve downstream tumor segmentation. The paper reports that on the BraTS2020 dataset, Nexus-INR outperforms state-of-the-art methods on both super-resolution and segmentation metrics.","feed_headline":"Nexus-INR super-resolves brain MRI at any scale, beating prior methods","feed_subtitle":"Combines anatomy-disentangled encoders, cross-modal distillation, and joint segmentation on BraTS2020.","key_machinery":"The load-bearing mechanism is the dual-branch encoder whose auxiliary classification task forces the latent space to separate shared anatomical structure from modality-specific features; on top of it, the knowledge distillation module uses cross-modal attention to copy high-frequency detail from the high-resolution reference modality into the low-resolution modality's reconstruction, stabilised by a self-supervised consistency loss. A third, integrated segmentation module embeds anatomical semantics so that reconstruction and downstream segmentation are optimised jointly. The INR backbone supplies arbitrary-scale coordinate querying, so the entire pipeline can upscale to any resolution in one pass.","core_discovery":"On its own terms, the paper's central discovery is that arbitrary-scale multimodal super-resolution can be made substantially better by explicitly separating what is common across modalities from what is specific to each modality, and then using the common anatomy plus a sharp reference modality to guide reconstruction of a blurry one. Concretely, Nexus-INR couples an INR, which maps continuous coordinates to intensities and so naturally handles any upsampling factor, with three task-level modules: a classification-guided dual-branch encoder for the disentanglement, a cross-modal attention distillation with a self-supervised consistency loss for detail transfer, and a segmentation module that ties the representation to anatomical labels. The paper argues that this combination yields state-of-the-art reconstruction quality and better tumor segmentation on BraTS2020 than existing super-resolution methods.","pith_inferences":["The provided full text in this file is a different manuscript (a 3D segmentation framework called SSFMamba); the Nexus-INR experiments, ablations, and exact numbers are not present here, so the abstract's reported superiority could not be checked against details.","A natural untested extension is zero-shot reconstruction of an entirely missing modality: the shared-anatomy branch could generate structure while the modality-specific branch is synthesised from the reference, though the paper does not claim this.","The auxiliary classification task focuses the disentanglement on tumor-relevant classes; one could test whether the same framework helps super-resolution of healthy tissue or other organs, where the classification signal would need to be redefined."],"forward_implications":["A single trained Nexus-INR can upsample a low-resolution MRI to any requested scale without retraining for each factor, since the INR queries continuous coordinates.","When several aligned modalities are available, the sharpest one (e.g., T1ce or T2) can be used as a reference to improve restoration of the blurriest, so the method degrades gracefully as reference quality drops.","Because segmentation semantics are injected during training, the same network can produce a tumor segmentation map alongside the super-resolved image, shortening clinical pipelines that currently run SR and segmentation separately.","If the disentanglement is real, the shared-anatomy branch should transfer across modality pairs, making the method applicable when the high-resolution reference is a different MRI sequence than the one being restored."],"supporting_citations":[],"fun_headline_variants":["Any-scale brain MRI upscaling that beats the state of the art","Nexus-INR: anatomy-guided multimodal MRI super-resolution","Cross-modal distillation lifts MRI resolution at any scale","One network for any resolution: brain MRI comes into focus","BraTS2020: anatomy-aware multimodal super-resolution tops baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premise that the auxiliary classification task actually separates each modality's features into shared anatomy and modality-specific detail, and that the cross-modal attention can then transfer useful high-frequency information from a high-resolution reference to a low-resolution modality, is what the reported gains stand on; if that separation is imperfect or the modalities are not well aligned, the distillation branch cannot improve reconstruction.","fun_headline_variants_meta":{"raw":{"variants":["Any-scale brain MRI upscaling that beats the state of the art","Nexus-INR: anatomy-guided multimodal MRI super-resolution","Cross-modal distillation lifts MRI resolution at any scale","One network for any resolution: brain MRI comes into focus","BraTS2020: anatomy-aware multimodal super-resolution tops baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000911,"raw_usage":{"total_tokens":3892,"prompt_tokens":904,"completion_tokens":2988,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":2903}},"tokens_in":520,"tokens_out":2988,"duration_ms":25745,"temperature":1.0,"reasoning_tokens":2903,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:40:50.546453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Nexus-INR on BraTS2020 while withholding or randomly permuting the registration of the high-resolution reference modality. If super-resolution quality stays high, the cross-modal attention is not the source of the gain, contradicting the claimed mechanism. A second check: compare against a single-modality INR with identical capacity and same training budget; if the multimodal version does not beat it clearly, the claimed benefit of cross-modal distillation is not demonstrated.","supporting_citations":[],"review_version":1}