{"id":"b8fdfea3-fa10-48fb-8c26-c2a30ef54c81","arxiv_id":"2511.12853","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A ControlNet-guided diffusion model reconstructs pseudo-healthy brain MRIs from tumor scans by inpainting with healthy prompts and mirrored contralateral edge maps, improving FID, SSIM, and tumor false-positive rate on BraTS2020.","lead":"BrainNormalizer is a machine-learning method that creates a 'pseudo-healthy' brain MRI by digitally removing the tumor and filling the gap using the patient's own healthy side as a guide. It reports better realism, symmetry, and fewer leftover tumor signals than two comparison methods on the BraTS2020 brain-tumor dataset, offering a way to picture what a patient's brain looked like before the tumor.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mirrored-contralateral conditioning and the SSIM metric share the same symmetry assumption, so the reported structural-consistency gains may reflect symmetry enforcement rather than accurate pseudo-healthy reconstruction.","rationale":"The reader's weakest_assumption correctly identifies the mirrored-contralateral edge map as the fragile link in the method. My stress-test sharpens this into an epistemic problem: the paper's principal structural metric, SSIM, is computed against the same contralateral region from which the inference condition is constructed. This makes the quantitative evidence partly circular. Even in unilateral cases without midline involvement, mass effect and natural brain asymmetry mean the contralateral hemisphere is not a ground-truth template; the paper itself acknowledges the failure modes in Section 5. The central quantitative claim in Table 3 is therefore not enough to establish that BrainNormalizer recovers the subject's true pseudo-healthy anatomy rather than a more aggressively symmetrized plausibility. I do not reject the work: the two-stage training and ControlNet conditioning are coherent, the qualitative examples are consistent with the design, and a synthetic-counterfactual benchmark could settle the concern. The appropriate level remains conditional acceptance, so I keep the reader's verdict unchanged.","tokens_in":15471,"tokens_out":7807,"duration_ms":78837,"concrete_test":"Build a synthetic-counterfactual benchmark with known ground truth. Take non-tumorous T1CE slices from an independent healthy dataset (or use BraTS2020 slices with no tumor), add a realistic tumor patch and optionally a mass-effect warp to surrounding tissue, and run Wolleb et al., SD Inpaint, and BrainNormalizer using the same tumor mask. Compare each reconstruction to the unmodified original slice in the masked region using LPIPS, PSNR, SSIM, and ventricle/cortical-surface overlap, stratified by tumor location (unilateral peripheral, unilateral with mass effect, central/midline-crossing) and size. If BrainNormalizer does not consistently beat the baselines on these ground-truth-anchored metrics, the Table 3 gains are likely artifacts of the symmetry-based SSIM and FP measures rather than evidence of accurate pseudo-healthy reconstruction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To establish the central claim, the method must show that a mirrored contralateral edge map is a valid template for the patient's missing pre-tumor anatomy. The evidence offered in §4.3 does not test this. SSIM is computed between the inpainted region and the contralateral non-tumorous region of the same slice, i.e., against the exact structure that c_mirrored is derived from. A reconstruction that copies/symmetrizes the contralateral hemisphere will be rewarded by this SSIM even if the tumor has displaced or effaced the true anatomy. The FP metric rewards removal of tumor-like signal, which the same symmetrization achieves; FID only checks that generated images look like non-tumorous brains in aggregate. Thus the three headline numbers are consistent with a model that is better at imposing symmetry, not at recovering the subject's true counterfactual anatomy. Section 5 concedes the scheme fails for large central or bilateral tumors, and healthy brains are not perfectly symmetric (the paper's own ref [48]); for unilateral tumors with mass effect, the contralateral side is not a clean template either. The paper also does not stratify Table 3 by tumor location/size, so the averages mix regimes where the conditioning is and is not informative. Without paired ground truth or an independent metric, the 'subject-specific' and 'anatomically consistent' claims are not adequately supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"BrainNormalizer proposes a two-stage diffusion framework for pseudo-healthy brain MRI reconstruction from tumorous scans without paired pre-tumor data. The first stage fine-tunes Stable Diffusion v1.5 with an inpainting objective on tumorous and non-tumorous BraTS2020 slices; the second stage trains a ControlNet branch on Canny edge maps. At inference, the model receives the tumorous masked slice, a non-tumorous text prompt, and a mirrored contralateral edge map, deliberately misaligning the input to induce counterfactual healthy generation. The paper reports FID=27.1, SSIM=0.76, and FP rate=7.1% on a 35-subject test set (Table 3), claiming improved realism, structural consistency, and clinical plausibility over a DDIM baseline (Wolleb et al.) and a first-stage-only inpainting baseline.","tokens_in":15859,"tokens_out":4425,"duration_ms":43944,"significance":"If the central claim is established, the method would be valuable: it addresses a real clinical need for subject-specific pseudo-healthy references, uses only public data, and explicitly avoids paired data requirements. The combination of inpainting fine-tuning and ControlNet edge conditioning is sensible, and the paper honestly states limitations in Section 5, including failure on large central and bilateral tumors. However, the current evidence does not separate 'enforcing symmetry' from 'recovering true subject-specific anatomy.' The SSIM metric is computed against the same contralateral structure that is used as the inference condition, so the reported structural-consistency gains are partly circular. FID and FP rate check aggregate realism and tumor-signal removal, not anatomical correctness. Without an independent validation protocol—e.g., synthetic paired data or external anatomical metrics—the 'subject-specific' and 'anatomically consistent' claims are not adequately supported. The contribution is promising but needs substantially stronger evidence.","major_comments":[{"comment":"The SSIM metric is circular with the inference design. Section 4.3 states SSIM is computed between the inpainted region and the contralateral non-tumorous region, while Section 3.4 conditions generation on a mirrored contralateral edge map. A model that copies the contralateral anatomy will be rewarded by this metric by construction, even if the tumor has displaced or effaced the true midline anatomy. Thus the SSIM improvement over baselines does not establish that the reconstruction matches the subject's missing pre-tumor anatomy. Please add an independent validation, e.g., take healthy slices from non-tumor subjects, artificially simulate tumorous inputs by masking and/or deforming them, and compare reconstructions against the known healthy slice. Report SSIM or PSNR against that ground truth, with confidence intervals.","section":"§4.3 (SSIM protocol; Table 3)"},{"comment":"The load-bearing premise of the method is that the mirrored contralateral edge map is a valid structural template for the tumor-affected region. Section 5 concedes the scheme fails for large central or bilateral tumors, but even unilateral tumors with mass effect deform the contralateral hemisphere, and the paper itself cites evidence (ref. [48]) that healthy brains are not perfectly symmetric. This is not a minor edge case; it directly affects the central claim. The manuscript should quantify when the assumption is violated: stratify Table 3 by tumor location, size, and midline shift, and show failures. Without this stratification, the reported averages mix regimes where the conditioning is informative with regimes where it is misleading.","section":"§3.4 and §5 (mirrored-contralateral assumption)"},{"comment":"The quantitative results are point estimates with no error bars, significance tests, or multiple-seed variance. The differences—SSIM 0.76 vs. 0.68 and FP 7.1% vs. 10.4%—may be within stochastic variation, especially on a test set of 35 subjects. Please report per-subject or bootstrap confidence intervals, the number of seeds, and the exact test-set size for each metric. Also specify the FID feature extractor and whether it was adapted to MRI; features from ImageNet-trained Inception networks may not capture clinically relevant anatomical fidelity.","section":"§4.4 (Table 3 statistics)"},{"comment":"The comparison isolates the effect of adding inpainting and the effect of adding ControlNet, but it does not isolate the deliberate misalignment strategy. To attribute the gains to mirrored contralateral conditioning, the authors should include controls at inference: (i) same model with a non-mirrored edge map, (ii) same model with the mirrored edge map but a tumorous prompt, and (iii) same model with no edge map. The current design cannot distinguish the contribution of the mirrored edge condition from generic inpainting or the text prompt. Additionally, Wolleb et al. [28] is a non-inpainting method; comparing against it conflates task formulation with model quality.","section":"§4.3 (experimental design / ablations)"}],"minor_comments":[{"comment":"For non-tumorous slices, masks are 'randomly borrowed from a tumorous slice with the same slice index.' Please clarify whether this is from the same subject or a different subject, and discuss whether this artificial pairing biases the learned inpaint distribution. The assumption that this pairing provides valid training targets is plausible but currently unexamined.","section":"§3.2 (training data construction)"},{"comment":"The tumor size categories in Table 2 (small, mild, medium, moderate, large) are given approximate pixel counts, while Section 4.1 states slices with 1,000–3,000 tumor pixels are used. The mapping between the categorical prompts and the numerical range is unclear. Please define the bins explicitly.","section":"§3.2 / §4.1 (tumor size categories)"},{"comment":"The figure descriptions contain repeated placeholder/watermark text ('274+ Thousand Frozen Icon...') that should be removed. Figures should be cleaned before submission.","section":"Figures 1–4"},{"comment":"The Abstract appears twice in the manuscript (once at the top and again in the full text), and the conclusion states 'clinically reliable anatomical references' even though no clinical validation was performed. Please soften the claim or add a clear caveat that clinical utility has not yet been assessed.","section":"Abstract and §5"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and practically important problem, and the proposed architecture is reasonable. The main issue is evidentiary: the headline structural metric is aligned with the conditioning mechanism, so the central claim of subject-specific anatomical recovery is not yet demonstrated. I would like the revision to include an independent validation protocol (e.g., synthetic paired evaluation), stratified results by tumor geometry, and proper uncertainty quantification. I do not see this as a reject, because the problems are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The core idea is a genuine new combination: use a ControlNet branch conditioned on Canny edge maps, and at inference feed a mirrored contralateral edge map along with a healthy prompt to inpaint the tumor region. That is a sensible way to inject subject-specific structure without paired data, and I don't see it in the cited pseudo-healthy work (Wolleb, Durrer, Kwark, Ferreira). The two-stage training—inpainting fine-tuning then ControlNet—is straightforward and well documented. The mask robustness experiment is a nice touch, and the paper is honest about limitations (large or bilateral tumors).\n\nThe soft spots are mostly about evaluation. The SSIM is computed between the inpainted region and the contralateral non-tumorous region of the same slice. Since the model is explicitly conditioned on a mirrored contralateral edge map, a reconstruction that simply symmetrizes the hemisphere will score well even if the tumor has displaced the actual anatomy. So the 0.76 SSIM supports \"the output is symmetric,\" not \"the output matches the patient's true pre-tumor anatomy.\" That is not a fatal flaw—symmetry is a reasonable prior—but the paper overstates what the metric demonstrates. The FP rate and FID are less circular but still aggregate; FP rewards removing tumor-like signal, which symmetrization also achieves. There are no error bars or significance tests, no comparison to the recent pseudo-healthy baselines cited in the related work, no stratification by tumor location/size, and no code or weights. The single-modality (T1CE) choice limits generality.\n\nThe central claim—that this method gives a subject-specific pseudo-healthy reference—is plausible but not proven. The paper deserves a serious referee; the approach is new enough and the experiments, despite their gaps, are a reasonable start. I would ask for confidence intervals, an external structural metric that does not rely on intact symmetry (e.g., agreement with known non-tumorous tissue labels or deformation field plausibility), and comparisons to Durrer et al. and Kwark et al. before accepting. The writing is clear and the thinking is honest, so I'd engage with it.","headline":"BrainNormalizer combines ControlNet edge conditioning with mirrored contralateral edge maps for pseudo-healthy brain MRI reconstruction; the combination is new, but the main structural metric is aligned with the method's own symmetry prior, so the reported gains may reflect symmetry enforcement rather than accurate counterfactual reconstruction.","tokens_in":16275,"tokens_out":2758,"would_cite":true,"duration_ms":25244,"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 tries to establish that a tumorous brain MRI can be converted into a subject-specific healthy-looking reference without paired pre-tumor scans, by guiding diffusion inpainting with the patient's mirrored contralateral edge map and","keywords":["pseudo-healthy brain MRI","diffusion inpainting","edge-guided conditioning","contralateral symmetry","counterfactual medical imaging","tumor removal","brain MRI reconstruction"],"falsifier":"Take healthy brain MRIs, digitally insert realistic tumors with surrounding edema and deformation, run the pipeline, and compare each reconstruction against the original healthy image; if the reconstruction error is no better than simply copying the mirrored healthy hemisphere, the contralateral edge prior is not carrying the claimed benefit.","tokens_in":15409,"feed_emoji":"🧠","tokens_out":7855,"duration_ms":61641,"temperature":0.7,"pith_summary":"BrainNormalizer tries to establish that subject-specific pseudo-healthy brain MRIs can be reconstructed directly from tumorous scans, with no paired pre-tumor images and no healthy reference used in training. The core move is to keep the diffusion model's inpainting objective but feed it deliberately mismatched guidance at inference: a healthy text prompt and an edge map mirrored from the patient's healthy hemisphere. If correct, this gives clinicians a per-patient counterfactual baseline for planning surgery and studying tumor-induced deformation, something that does not exist in practice today. On a public brain-tumor MRI dataset, the reconstructed images score a perceptual distribution distance of 27.1, a structural similarity of 0.76 against the mirrored healthy side, and a 7.1 percent false-positive rate from a downstream tumor detector, all better than the compared baselines.","feed_headline":"Mirrored brain edges guide pseudo-healthy MRI reconstruction","feed_subtitle":"No pre-tumor scan needed: a diffusion model fills the tumor site using the patient's healthy side.","key_machinery":"The load-bearing mechanism is the mirrored contralateral edge map combined with the deliberate misalignment of inputs. An edge detector extracts structural boundaries from the patient's own MRI; mirroring the hemisphere opposite the tumor supplies a structural template for the missing healthy tissue, while a text prompt describing a healthy brain steers semantic interpretation. A trainable control branch attached to the frozen diffusion decoder via zero-initialized convolutions injects those edges, so generation follows stable anatomical contours instead of unconstrained inpainting. This lets the model use the patient's own anatomy as the reference, which is what makes the reconstruction sub","core_discovery":"The central claim is that a tumor region can be inpainted into healthy-looking tissue that preserves the subject's own anatomy by conditioning generation on boundary contours taken from the patient's contralateral hemisphere. The model is built in two stages: first, a pretrained latent diffusion model is fine-tuned on MRI slices to reconstruct masked regions from surrounding context, using text prompts that describe either healthy or tumorous scans; second, a separately trainable control branch learns to inject edge maps into the frozen decoder through zero-initialized convolutions. At inference the paper pairs a tumorous slice with a non-tumorous prompt and a mirrored edge map, a deliberate","pith_inferences":["The reported structural similarity compares the reconstructed region to the mirrored healthy side, so it rewards symmetry; the method may favor over-smoothed or symmetric-looking tissue in naturally asymmetric brain areas.","A direct validation would use healthy MRIs with digitally inserted tumors and known deformations, then measure reconstruction error against the original scan—something the unpaired-data setting cannot provide.","Because the reference is the patient's own anatomy, the same pipeline could be reused at multiple time points to track how deformation progresses, an extension the paper does not test.","The mirrored-edge prior is trustworthy mainly for unilateral tumors; for midline or bilateral disease, the model may hallucinate rather than recover structure, so clinical deployment should be limited until tested on such cases."],"forward_implications":["One scan becomes a source of a personalized healthy-reference image, removing the need for historically unavailable pre-tumor MRI.","Generated pseudo-healthy images can serve as counterfactual pairs with the tumor scan for studying displacement and infiltration, once clinically validated.","Rough tumor masks are enough: expanded masks produce similar reconstructions, which lowers the annotation burden in practice.","A downstream tumor detector marks only 7.1 percent of reconstructions as tumorous, indicating that tumor-specific signal is largely removed while surrounding tissue is preserved.","The method's clinical value is bounded to mostly unilateral tumors; the paper itself notes degradation for large central or bilateral tumors."],"fun_headline_variants":["No pre-tumor scan: mirrored edges reconstruct healthy brain","Diffusion model fills tumors using mirrored healthy tissue","Pseudo-healthy brain from mirrored edges: no paired data needed","Flip MRI edges to reconstruct the brain without a pre-scan","Mirrored hemisphere edges guide tumor-free brain reconstruction"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The mirrored contralateral edge map must be a valid structural template for the missing non-tumorous tissue; as the paper itself notes in its limitations, this breaks down for tumors in large central regions or affecting both hemispheres, or when mass effect deforms the opposite side.","fun_headline_variants_meta":{"raw":{"variants":["No pre-tumor scan: mirrored edges reconstruct healthy brain","Diffusion model fills tumors using mirrored healthy tissue","Pseudo-healthy brain from mirrored edges: no paired data needed","Flip MRI edges to reconstruct the brain without a pre-scan","Mirrored hemisphere edges guide tumor-free brain reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000987,"raw_usage":{"total_tokens":4030,"prompt_tokens":762,"completion_tokens":3268,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":3188}},"tokens_in":506,"tokens_out":3268,"duration_ms":18446,"temperature":1.0,"reasoning_tokens":3188,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:56:42.313212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take healthy brain MRIs, digitally insert realistic tumors with surrounding edema and deformation, run the pipeline, and compare each reconstruction against the original healthy image; if the reconstruction error is no better than simply copying the mirrored healthy hemisphere, the contralateral edge prior is not carrying the claimed benefit.","supporting_citations":[],"review_version":1}