{"id":"3894d013-2388-4588-ba09-893b65b4b0ec","arxiv_id":"2608.08401","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A T1c-guided, three-view fusion pipeline improves downstream tumor segmentation Dice over anisotropic T2w input, but the gain is confounded by cross-contrast information and a questionable smoothing objective.","lead":"VIPP-SR tries to make blurry, side-view brain MRI images sharper by using a patient's high-resolution T1 scan as a guide, then combines three viewing directions into one volume. The authors report that tumor segmentation improves when using their reconstructed T2-like image instead of the original anisotropic T2 image, but the evaluation may mostly show the benefit of adding T1 information rather than true T2 super-resolution.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VIPP-SR never uses the patient's anisotropic T2w at inference: Eq. (12)/Algorithm 1 optimize only T1c-derived views, so the reported segmentation gains are confounded by T1c anatomy; the method is T1c-to-T2w synthesis, not super-resolution.","rationale":"The reader's weakest assumption was the validity of L_proj as an anatomical consistency prior, which is a real issue but not the most fundamental. The decisive flaw is that the patient's anisotropic T2w never enters the optimization objective in the main method: L_dir (Eq. 11) is fidelity to T1c-derived views, and L_proj (Eq. 10) is a flatness prior on the fused volume. The acquired Y appears only in the optional L_aux used in the v3 ablation. Therefore, the output is a synthesis from T1c, and the comparison to the T2w baseline cannot establish super-resolution. The missing T1c-only baseline is the concrete experimental gap that would settle this. I agree with the reader's rejection but identify a different load-bearing concern, hence 'partial' agreement.","tokens_in":18009,"tokens_out":8116,"duration_ms":69478,"concrete_test":"Train nnU-Net with the same protocol (3d_fullres, 250 epochs, fold 0, same labels and split) using the patient's T1c alone as the single input channel on BraTS-MET, and also using T1c+T2w as two input channels. Compare mean-label Dice on the held-out 73-subject test set to Table 3's VIPP-SR row (0.465) and T2w-orig row (0.330). If T1c-only Dice reaches or exceeds 0.465, the reported gain over the anisotropic baseline is attributable to T1c anatomy, not to super-resolution; if T1c+T2w does not exceed T1c-only, the synthetic T2w adds no information beyond the T1c input.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper frames VIPP-SR as super-resolving the patient's anisotropic T2w (Sec. 1), but at inference the acquired T2w Y is never used. Algorithm 1 Stage 2 optimizes L_star(P) = alpha*L_proj + beta*L_dir (Eq. 12), where L_dir (Eq. 11) measures fidelity to the three T1c-derived view volumes P*_hd, P*_wd, P*_hw produced by applying the frozen generator to the T1c volume (Sec. 3.4). The only term containing Y is the optional L_aux (Sec. 3.5.3), used solely in ablation v3; the main VIPP-SR objective and its reported rows do not include it. Thus the final output is a pure function of X (T1c), the brain mask, and the trained GAN; Y imposes no constraint on the reconstruction. The method is therefore T1c-to-T2w cross-contrast synthesis, not super-resolution of the acquired T2w, despite the claim that the patient's T2w anchors the result. Consequently, the headline Dice improvement over the T2w baseline (0.330 to 0.465) may simply reflect that isotropic T1c provides richer anatomical structure to the downstream nnU-Net. The paper reports no T1c-only segmentation baseline, so this confound is uncontrolled. The Canny-Dice metric (Eq. 13) compares the synthetic T2w edges to the T1c edge map from which the output was derived, making high scores expected rather than validating T2w fidelity. The ablation claim that self-consistency is the main source of gain is also contradicted by Table 3, where v1 (fidelity-only) has higher mean-label Dice (0.484) than VIPP-SR (0.465).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VIPP-SR, a two-stage framework for producing an isotropic T2-weighted (T2w) volume from an anisotropic T2w acquisition using a co-registered isotropic T1c volume. Stage 1 trains a view-independent GAN (VIP-GAN) on axial T1c-to-T2w patch pairs. Stage 2 applies the frozen generator along axial, coronal, and sagittal views of the T1c volume and fuses the three view-specific outputs through a projection-based optimization that balances a slice-mean self-consistency penalty (Eq. 10) against a view-fidelity penalty (Eq. 11). The method is evaluated on BraTS-MET (held-out test set) and BraTS-GLI (zero-shot cross-cohort) using image-quality metrics (Canny-Dice, SSIM, Lap-RMSE, HFEN) and downstream nnU-Net segmentation. The paper claims that VIPP-SR improves mean-label Dice over the real anisotropic T2w baseline (from 0.330 to 0.465 on BraTS-MET and from 0.473 to 0.563 on BraTS-GLI) and that ablation studies identify inter-plane self-consistency as the main source of the gain.","tokens_in":18341,"tokens_out":5174,"duration_ms":49797,"significance":"If the central claims were supported, the paper would offer a practically valuable approach for retrospective neuro-oncology cohorts where isotropic T2w ground truth is unavailable. The cross-cohort, zero-shot evaluation is a strength, and the intention to release code is commendable. However, the paper's framing as super-resolution is not supported by the algorithm, which does not use the acquired T2w volume at inference. The reported ablation contradicts the headline claim, and the primary structural-fidelity metric is circular. These issues are load-bearing, so the significance of the contribution as stated cannot be accepted without fundamental revision.","major_comments":[{"comment":"The patient's anisotropic T2w volume Y is never used in the main VIPP-SR objective. Eq. (12) optimizes L*(P) = alpha*L_proj + beta*L_dir, where L_dir depends only on the T1c-derived view volumes P*_hd, P*_wd, P*_hw. The only term containing Y, L_aux, is introduced in Sec. 3.5.3 and used solely in the v3 ablation. Algorithm 1 confirms that Stage 2 operates on X, M, and the frozen generator, with no reference to Y. Consequently, the final output is a pure function of the T1c volume, making the method a T1c-to-T2w synthesis rather than a super-resolution of the patient's acquired T2w. This directly contradicts the framing in Sec. 1 and the abstract, and it confounds the reported segmentation gains because the output carries all of the T1c anatomy.","section":"Sec. 3.4-3.5, Algorithm 1, Eq. (12)"},{"comment":"The abstract and Sec. 5.2 claim that ablation studies identify inter-plane self-consistency as the main source of the downstream gain. Table 3 shows the opposite: the fidelity-only ablation v1 achieves mean-label Dice 0.484, higher than VIPP-SR's 0.465, and v1 also gives the highest edema Dice (0.529). The text explicitly acknowledges that 'v1 gives the highest mean-label Dice (0.484) and edema Dice (0.529)' but still concludes that self-consistency is the main contributor. This is a direct contradiction between the reported numbers and the paper's central attribution claim, so the ablation conclusion is not supported.","section":"Table 3, Sec. 5.2, Abstract"},{"comment":"The structural-fidelity metric Canny-Dice computes the Dice coefficient between the synthetic T2w edge map and the T1c edge map. Since T1c is the generator's input, high scores can be obtained by copying T1c edges into the output, and the metric is circular as evidence of T2w fidelity. The paper's claims in Sec. 5.1 that the synthetic volumes 'agree with the subject's own T1c anatomy' and that high Canny-Dice indicates recovered inter-plane structure are therefore not established; an explicit ablation or a non-circular reference (e.g., an independent modality or manual annotations) is needed.","section":"Eq. (13), Sec. 4.2, Sec. 5.1"},{"comment":"The projection term L_proj penalizes the squared distance of each voxel from its slice mean along each of the three axes. The paper itself notes that a perfectly self-consistent volume is constant along each axis, i.e., constant over the whole brain mask. This is a strong spatial-flatness prior, not a physically motivated model of anatomical inter-plane consistency for a real brain T2w volume, which has substantial structure along all axes. Consequently, the interpretation of the fusion stage as enforcing 'anatomical consistency' is not supported, and the mechanism by which it might help segmentation remains unclear.","section":"Sec. 3.5.1, Eq. (10)"},{"comment":"No baseline is reported in which the isotropic T1c volume alone is used as the input to the downstream nnU-Net segmentation. Because VIPP-SR's output is a function of T1c and the trained generator, the observed improvement over the real T2w baseline could be due to the T1c information alone rather than to any property of the proposed reconstruction. Without a T1c-only channel baseline, the central segmentation claim is confounded and cannot be interpreted as evidence that the reconstructed T2w is superior to the acquired anisotropic T2w.","section":"Sec. 5.2, Tables 3 and 4"}],"minor_comments":[{"comment":"There are several apparent OCR artifacts in the typesetting, for example 'bY  M=1 ≈Y iso M=1' in Eq. (1) and 'P ∗ ∗hd' in Sec. 3.4; these should be corrected before publication.","section":"Eq. (1) and Sec. 3.4"},{"comment":"When H or W is not a multiple of P, the stride formulas produce non-integer values; the implementation detail of how non-integer strides are rounded or otherwise handled should be stated explicitly.","section":"Sec. 3.2.2, Eqs. (2)-(3)"},{"comment":"The image-quality metrics are computed on the single largest-mask-area slice from each anatomical plane rather than over the full volume; this limited sampling should be stated as a potential source of bias.","section":"Sec. 4.2"},{"comment":"The statement that the segmentation results are 'single-fold nnU-Net results without cross-validation, we treat the gains as indicative rather than formally significant' is an important caveat that should also appear in the abstract, which currently states the results 'validate' the method.","section":"Sec. 5.2"},{"comment":"References [14] and [22] appear to be the same paper (identical title and venue); this duplication should be resolved.","section":"References [14] and [22]"}],"recommendation":"reject","confidential_remarks":"The core issue is that the method does not use the acquired T2w volume at inference, so the paper's central framing as super-resolution is not supported by the algorithm; it is a cross-contrast synthesis method. The self-reported ablation data directly contradict the abstract's claim about the main source of gain, and the Canny-Dice metric is circular. These are load-bearing problems that would require substantial rework (including actually anchoring the reconstruction to Y, adding a T1c-only baseline, and re-examining the ablation claim), rather than minor revisions. For this reason, I recommend rejection. The paper does address a clinically relevant problem and includes a useful cross-cohort evaluation, so a fundamentally revised version might warrant reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper you're asking about is worth reading, but not for the reason it thinks. The reported method never uses the patient's anisotropic T2w at inference. Algorithm 1's Stage 2 optimizes only a weighted sum of the projection term and fidelity to three T1c-derived views; the acquired T2w enters only in the v3 ablation. So the final volume is a pure function of T1c, the mask, and the frozen generator. That makes it cross-contrast synthesis, not super-resolution, despite the title and abstract.\n\nWhat's genuinely new here is the combination: a view-independent patched GAN that transfers zero-shot from axial to coronal and sagittal, plus a projection-based fusion step. The shape-preserving patching with mask conditioning and deepest-skip removal is a smart way to reduce orientation-specific shortcuts. The cross-cohort evaluation on BraTS-GLI is a real strength, and the paper is clearly written, with code and a candid note that single-fold nnU-Net results are indicative only.\n\nThe soft spots are not minor. The central claim that 'inter-plane self-consistency is the main source of the gain' is contradicted by the paper's own Table 3: on BraTS-MET, the fidelity-only variant v1 achieves mean-label Dice 0.484, while VIPP-SR gets 0.465. The projection objective L_proj penalizes deviation from the slice mean along each axis; as the authors state, perfect self-consistency means the volume is constant along that axis. That is a flatness prior, not an anatomical consistency model. The structural metric Canny-Dice is computed against the T1c edge map, which is the generator's input, so high scores are partly circular.\n\nBecause the output never uses the T2w, the reported segmentation gains over the real anisotropic T2w are confounded. Without a T1c-only baseline and a T1c+T2w baseline, there is no way to know whether the improvement comes from recovering T2w structure or simply from injecting isotropic T1c anatomy. That is a load-bearing issue for the paper's framing.\n\nThat said, I would still send it to review, not desk-reject it. The underlying idea—training on axial pairs and transferring across planes with a patched generator—is worth exploring, and the empirical work is substantial enough to demand a careful referee. The authors need to either reframe the paper as T1c-guided T2w synthesis for downstream segmentation, or actually use the acquired T2w in the fusion objective, and add the missing baselines. As written, the claims do not hold, but the potential is there.","headline":"The paper has a load-bearing framing problem: the acquired T2w is never used at inference, so it is really T1c-to-T2w synthesis, and the ablation claim is contradicted by its own table; still worth a careful referee.","tokens_in":18966,"tokens_out":3860,"would_cite":false,"duration_ms":37665,"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":"The paper claims that anisotropic T2w brain MRI can be super-resolved to near-isotropic quality using only the patient's isotropic T1c as anatomical reference, with no isotropic T2w ground truth, improving downstream tumor segmentation.","keywords":["MRI super-resolution","anisotropic reconstruction","cross-contrast guidance","projection optimization","generative adversarial networks","inter-plane consistency","tumor segmentation","zero-shot transfer"],"falsifier":"Acquire thin-slice isotropic T2w for a subset of the same subjects, run VIPP-SR on the thick-slice versions, and compare the fused output against the real isotropic T2w in inter-plane high-frequency bands; if the optimization step moves the fused volume away from the real isotropic T2w relative to the simple three-view average, the self-consistency prior is not recovering anatomy.","tokens_in":17717,"feed_emoji":"🧠","tokens_out":8924,"duration_ms":90710,"temperature":0.7,"pith_summary":"The paper tries to establish that the missing inter-plane detail in an anisotropic T2-weighted brain MRI can be recovered by borrowing anatomy from the same patient's isotropic T1c image, without ever needing an isotropic T2w scan as ground truth. This matters because clinical T2w is routinely acquired with thick slices, and the resulting blur and stair-stepping obscure small structures and weaken 3D analysis for tumor segmentation and radiotherapy planning. The proposed method trains a view-independent patch generator on axial T1c-to-T2w slices, applies it along all three anatomical axes, and fuses the three outputs by a projection-based optimization that balances per-view fidelity against inter-plane self-consistency. On a metastasis cohort the fused volume improved mean-label tumor segmentation Dice from 0.330 to 0.465, and on a glioma cohort without retraining from 0.473 to 0.563, supporting the claim.","feed_headline":"T2w MRI super-resolution lifts tumor Dice to 0.563","feed_subtitle":"No isotropic T2w ground truth needed: one frozen T1c-to-T2w generator plus cross-view projection fusion restores inter-plane detail.","key_machinery":"The load-bearing mechanism is the projection-based fusion objective. It combines a per-axis inter-plane variance penalty $L_{\\text{proj}}(P)=\\sum_a \\|(P-\\bar{P}^a)\\odot M\\|_F^2$ with a view-fidelity term $L_{\\text{dir}}(P)=\\sum_a \\|(P-P^*_a)\\odot M\\|_F^2$, optimized over the masked volume $P$ with Adam. Because $L_{\\text{proj}}$ vanishes only if the volume is constant along every axis, it acts as a strong 3D coherence prior rather than a physical model of slice acquisition. Around this sits the view-independent patched generator VIP-GAN: a U-Net with mask conditioning, coarsened skip connections, and no deepest skip, trained on 128x128 axial patches without resampling, which lets one frozen checkpoint serve all three anatomical views.","core_discovery":"VIPP-SR's central claim is that cross-view agreement can substitute for missing isotropic supervision. A single generator, trained only on high-resolution axial T1c-to-T2w pairs, is applied unchanged to axial, coronal, and sagittal slices of the T1c volume; each resulting volume is sharp in its own plane and unreliable elsewhere. The fused estimate is obtained by minimizing a weighted sum of a self-consistency term, which penalizes the squared distance of every masked voxel from its slice mean along each axis, and a data-fidelity term, which keeps the result close to the three view-specific volumes. The paper reports that this fusion, without any isotropic T2w ground truth, improves downstream single-channel tumor segmentation over the acquired anisotropic T2w baseline on both cohorts, and attributes the improvement mainly to the inter-plane self-consistency term.","pith_inferences":["Editor's inference: because the self-consistency term penalizes deviation from slice-constancy, it likely smooths genuine inter-plane high frequencies; testing an alpha/beta sweep on a dataset with true isotropic T2w would reveal whether very thin structures are over-smoothed.","Editor's inference: the reported ablation numbers leave room for another reading, since the fidelity-only variant reaches comparable or higher mean Dice on the training cohort; a systematic weight sweep would clarify whether self-consistency itself or multi-view averaging drives the gain.","Editor's inference: replacing the slice-constancy penalty with a slice-profile or motion-aware forward model could be tested within the same framework, and the paper's edge-consistency and downstream-segmentation metrics would quantify the change."],"forward_implications":["Retrospective cohorts could receive near-isotropic T2w without rescanning, as long as a co-registered isotropic T1c exists.","The same frozen generator can serve a new tumor population at zero training cost; only the per-subject voxel-space fusion step is run at deployment.","Segmentation gains concentrate on small focal structures such as tumor, enhancing tissue, and resection cavity, so tasks that depend on those structures benefit most.","The method avoids iterative backpropagation through a large generative prior; fusion is voxel-space and linear-time per update, keeping per-case compute low.","Future super-resolution designs should treat explicit cross-view 3D consistency as a first-class objective rather than an implicit side effect of slice-wise generation."],"supporting_citations":[{"why":"Supplies the benchmark dataset that provides both evaluation cohorts and the tumor labels used for downstream segmentation.","marker":"[34]"},{"why":"Supplies the self-configuring segmentation model that converts each T2w input into the Dice scores grounding the downstream claim.","marker":"[37]"},{"why":"Defines the generative-prior baseline whose assumptions match the no-HR-T2w, no-degradation-model setting and anchors the comparison.","marker":"[17]"},{"why":"Cited for the risk that T1c-to-T2w synthesis alone can hallucinate pathology, motivating the anchored super-resolution rather than direct synthesis formulation.","marker":"[8]"},{"why":"Defines the SSIM metric used for intensity-fidelity evaluation in the image-quality tables.","marker":"[35]"},{"why":"Represents the cross-orientation self-supervision line this work extends by adding explicit projection-based 3D consistency.","marker":"[16]"}],"fun_headline_variants":["No isotropic T2w needed: T1c-guided super-res boosts tumor Dice","Cross-view fusion lifts T2w MRI segmentation without ground truth","VIPP-SR: Zero-shot T2w super-resolution sharpens tumor boundaries","T1c-to-T2w generator plus cross-view consistency improves Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that penalizing how far each voxel is from the average of its slice along every axis is a valid stand-in for anatomical consistency; a real brain is not constant from slice to slice, so in practice this term is a smoothing prior, not a measured relationship between anisotropic and isotropic anatomy.","fun_headline_variants_meta":{"raw":{"variants":["No isotropic T2w needed: T1c-guided super-res boosts tumor Dice","Cross-view fusion lifts T2w MRI segmentation without ground truth","VIPP-SR: Zero-shot T2w super-resolution sharpens tumor boundaries","T1c-to-T2w generator plus cross-view consistency improves Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000628,"raw_usage":{"total_tokens":2972,"prompt_tokens":1084,"completion_tokens":1888,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":1807}},"tokens_in":700,"tokens_out":1888,"duration_ms":16827,"temperature":1.0,"reasoning_tokens":1807,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:37:03.300080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire thin-slice isotropic T2w for a subset of the same subjects, run VIPP-SR on the thick-slice versions, and compare the fused output against the real isotropic T2w in inter-plane high-frequency bands; if the optimization step moves the fused volume away from the real isotropic T2w relative to the simple three-view average, the self-consistency prior is not recovering anatomy.","supporting_citations":[{"cited_title":"Jaeger, Simon A","cited_arxiv_id":null,"evidence_quote":"Supplies the self-configuring segmentation model that converts each T2w input into the Dice scores grounding the downstream claim."},{"cited_title":"Inversesr: 3d brain mri super-resolution using a latent diffusion model","cited_arxiv_id":null,"evidence_quote":"Defines the generative-prior baseline whose assumptions match the no-HR-T2w, no-degradation-model setting and anchors the comparison."},{"cited_title":"Greve, Koen Van Leemput, Bruce Fischl, and Adrian V","cited_arxiv_id":null,"evidence_quote":"Cited for the risk that T1c-to-T2w synthesis alone can hallucinate pathology, motivating the anchored super-resolution rather than direct synthesis formulation."},{"cited_title":"Image quality assessment: fromerrorvisibilitytostructuralsimilarity.IEEE Transactions on Image Processing, 13(4):600– 612, 2004","cited_arxiv_id":null,"evidence_quote":"Defines the SSIM metric used for intensity-fidelity evaluation in the image-quality tables."},{"cited_title":"SIMPLE: Simultaneous Multi-Plane Self-Supervised Learning for Isotropic MRI Restoration from Anisotropic Data","cited_arxiv_id":"2408.13065","evidence_quote":"Represents the cross-orientation self-supervision line this work extends by adding explicit projection-based 3D consistency."}],"review_version":1}