{"id":"7da4ec52-88d3-482b-a873-4fbeb40c2432","arxiv_id":"2607.07401","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":10,"one_line_summary":"HA-DSB uses a diffusion Schrödinger bridge with vision-language model region embeddings and PET-guided noise modulation to translate whole-body MRI while preserving lesion fidelity.","lead":"This paper presents HA-DSB, a diffusion Schrödinger bridge model that translates between MRI sequences across the whole body, using PET scans and AI-generated anatomical labels to preserve lesion details. It matters because faster MRI translation could cut long PET-MR scan times, making clinical workflows more efficient and improving pathology visualization.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"PET-guided noise modulation (Eq. 7) changes the forward marginal distribution without adjusting the training objective (Eq. 3) or reverse process, creating a train-inference inconsistency whose impact on the bridge validity is unverified.","rationale":"The reader correctly identifies the most load-bearing concern: the PET-guided noise modulation in Eq. 7 modifies the Schrödinger bridge forward process without proving that the training objective (Eq. 3) or the reverse process remains consistent with this modification. My analysis confirms this: the modulated forward process changes the marginal distribution to have spatially varying variance, but the training objective does not account for the spatial scaling S. This is a genuine theoretical gap, not merely a missing experiment.\n\nI considered whether the lack of ablation between noise modulation and PET attention (reader's issue 2) might be more load-bearing. However, the bridge validity concern is more fundamental: it questions whether the core methodological contribution is theoretically sound, whereas the ablation concern is about attribution. Both matter, but the bridge validity issue is correctly prioritized.\n\nThe concern is addressable. The endpoints of the bridge are preserved (noise vanishes at t=0 and t=1), so the method may work empirically even if the intermediate dynamics are inconsistent. The paper could resolve this by either (a) deriving the correct objective and showing it matches Eq. 3, (b) showing empirically that the corrected objective yields similar results, or (c) reframing the noise modulation as a training-time augmentation rather than a bridge modification. Until one of these is done, CONDITIONAL is the appropriate verdict.\n\nThe reader's other concerns (no statistical significance tests, single-institution dataset, VLM label validation underspecified) are valid but secondary. The code availability and consistent improvements across regions provide reasonable empirical support, but the theoretical gap in the bridge formulation prevents full confidence in the central claim.","tokens_in":9349,"tokens_out":6121,"duration_ms":377070,"concrete_test":"Analytically derive the correct denoising objective for the modulated forward process in Eq. 7. If q(x_t|x_0,x_1) = N(μ_t, Σ_t diag(S²)), the score is ∇log q = −(x_t−μ_t)/(Σ_t S²), so the DSM target should involve 1/S scaling. Check whether Eq. 3's target (X_t−X_0)/σ_t reduces to S⊙ε under Eq. 7; if it does not, the objective is inconsistent. Then train two variants on the lesion cohort: (a) Eq. 3 as written, (b) the corrected objective with S-aware targets. If variant (b) outperforms (a) by more than the current PET gain margin (+0.80 PSNR), the inconsistency is material and the current results understate the modulation's effect. If (a) and (b) perform identically, the modulation acts as benign augmentation and the concern is less severe. Additionally, run an ablation with noise modulation ON but PET attention OFF (and vice versa) to isolate which component drives the lesion improvement","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly identifies the core issue. Eq. 7 replaces the standard bridge forward process x_t = μ_t + Σ_t^{1/2} ε with x_t = μ_t + Σ_t^{1/2}(S⊙ε), where S ∈ [0.2, 2.0]^{H×W} is predicted from PET. This changes the marginal distribution from N(μ_t, Σ_t I) to N(μ_t, Σ_t diag(S²)). The correct denoising score-matching objective for this modified marginal requires the network to predict S⊙ε (or equivalently, the score must be scaled by 1/S² spatially). However, Eq. 3 uses the target (X_t − X_0)/σ_t with no S-dependent correction, which is the objective appropriate for the unmodulated bridge. During inference, if the reverse process uses the standard bridge SDE (without spatially varying diffusion coefficient g·S), there is a train-test mismatch: the model is trained on samples from a modulated forward process but evaluated with an unmodulated reverse process. The endpoints are preserved (noise vanishes at t=0 and t=1), so the method may still work empirically as a form of spatially adaptive data augmentation. But the paper frames this as a modification to the Schrödinger bridge dynamics, and without proving consistency between the modified forward process, the training objective, and the reverse process, the theoretical claim that this yields a valid transport plan is unsupported. The lesion improvements in Table 2 (PSNR +0.80, SSIM +1.7) could then stem from the multi-scale PET attention module (Eq. 8) alone, since no ablation isolates the noise modulation from the attention mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes HA-DSB, a Diffusion Schrödinger Bridge framework for whole-body MRI translation in integrated PET/MR settings. The method addresses two sources of heterogeneity: (1) cross-region anatomical variability, handled via region context embeddings derived from a vision-language model (PubMedBERT + Gemini), and (2) pathological tissue fidelity, addressed through a dual-stage PET guidance mechanism comprising forward-process noise modulation (Eq. 7) and reverse-process multi-scale attention (Eq. 8). Experiments on 246 whole-body PET/MR subjects across five anatomical regions show improvements over five baselines (Table 1), and a lesion-specific ablation (Table 2) isolates the PET guidance effect on pathological slices. Code is stated to be available.","tokens_in":9881,"tokens_out":1301,"duration_ms":233350,"significance":"The clinical motivation is strong: whole-body PET/MR acquisition times of 60–90 minutes are a real bottleneck, and MR-to-MR translation from fast sequences is a practical solution. The paper provides falsifiable quantitative predictions: per-region SSIM/PSNR gains (Table 1) and a lesion-subset ablation (Table 2) showing that PET guidance yields +0.80 PSNR / +1.7 SSIM on lesion slices versus +0.16 / +0.3 on overall data. The region-conditioning design via VLM-generated labels is a reasonable engineering contribution. The code availability statement is a positive for reproducibility. However, the theoretical framing of the PET-guided noise modulation as a modification to the Schrödinger bridge dynamics is not rigorously justified, which weakens the contribution's depth.","major_comments":[{"comment":"§2.3, Eq. (7): The PET-guided noise modulation replaces the standard bridge forward process x_t = μ_t + Σ_t^{1/2} ε with x_t = μ_t + Σ_t^{1/2}(S⊙ε), changing the marginal from N(μ_t, Σ_t I) to N(μ_t, Σ_t diag(S²)). However, the training objective in Eq. (3) uses the target (X_t − X_0)/σ_t with no S-dependent correction. For the modulated marginal, the correct denoising target should account for the spatially varying diffusion coefficient (e.g., predict S⊙ε or adjust the score by 1/S²). The paper does not prove that the modified forward process, the training objective, and the reverse process remain mutually consistent. This is load-bearing for the claim that the noise modulation modifies 'bridge dynamics' rather than functioning as a form of spatially adaptive data augmentation. The authors should either (a) derive the consistent objective for the modulated bridge and show that Eq. (3)近似","section":null},{"comment":"§2.3–2.4, Table 2: The lesion improvement (PSNR +0.80, SSIM +1.7) is attributed to both the forward noise modulation (Eq. 7) and the reverse multi-scale PET attention (Eq. 8), but no ablation isolates these two components. Since the attention module (Eq. 8) with zero-initialized W_out is a standard conditioning mechanism whose lesion benefit is plausible on its own, the marginal contribution of the noise modulation cannot be determined from the current experiments. An ablation with attention-only (no noise modulation) vs. full PET guidance would clarify whether the forward-process modification is necessary or whether the attention module suffices.","section":null},{"comment":"§3, Table 1: The 'HA-DSB (no PET)' variant already outperforms all baselines substantially (e.g., Head/Neck SSIM 94.0 vs. ResViT 91.0). However, the paper does not describe what 'no PET' entails—whether the noise modulator S is set to uniform (S=1), whether the PET attention module is removed, or whether both are disabled. This ambiguity makes it difficult to interpret the PET contribution. The variant configuration should be explicitly stated.","section":null}],"minor_comments":[{"comment":"§2.2: The VLM used is 'Google Gemini 3 Pro'—this should be cited properly if a reference exists, or noted as a proprietary model with version and access date.","section":null},{"comment":"§2.2: The 96% accuracy for body-location labels is mentioned but the validation protocol is vague (how many samples, which rater, inter-rater agreement). A brief clarification would help.","section":null},{"comment":"Table 1: The dagger (†) notation is defined as 'second-best result' but is applied to HA-DSB (no PET) rows, which is confusing since it is the authors' own ablation, not a competing method. Consider relabeling.","section":null},{"comment":"Fig. 2 caption: The figure shows translated T2w results and error maps but does not label which method each column corresponds to. Adding method labels would improve readability.","section":null},{"comment":"§3, Dataset: The text mentions 'three co-registered modalities (LAVA, T2, and PET)' but the translation direction (LAVA→T2) is only clear from Fig. 1. Stating this explicitly in the Dataset paragraph would help.","section":null},{"comment":"References [9, 10, 14, 15] appear to be from 2025, which is fine, but [10] is an arXiv preprint with a 2507.xxxx number suggesting July 2025—verify this is the correct and citable version.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core issue raised by the reader and stress-test regarding Eq. (7) is legitimate and lands: the train-inference consistency of the modulated bridge is not verified. However, I would frame this as a fixable theoretical gap rather than a fatal flaw—the endpoints are preserved, so the method likely works empirically as adaptive augmentation even if the bridge formulation is not strictly valid. The authors should be given the opportunity to either justify the formulation rigorously or reframe the contribution honestly. The lesion ablation gap (noise modulation vs. attention) is also important but straightforward to address with an additional ablation. I would not reject on these grounds."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for the careful reading and constructive feedback. The referee raises three points: (1) theoretical consistency of the PET-guided noise modulation with the DSB training objective, (2) lack of an ablation isolating the forward noise modulation from the reverse attention module, and (3) ambiguity in the 'no PET' variant configuration. We address each below and describe the revisions we will make.","responses":[{"response":"The referee is correct that the modulated forward process in Eq. (7) changes the marginal covariance to N(μ_t, Σ_t diag(S²)), and that the training objective in Eq. (3) does not include an S-dependent correction. We acknowledge this as a genuine gap in the current presentation. In practice, our implementation trains the network to predict the standard target (X_t − X_0)/σ_t while the forward process uses modulated noise, which means the model implicitly learns to compensate for the spatially varying perturbation through the conditioning signal c (which includes region context and, during reverse sampling, PET features via the attention module). However, we agree that this falls short of a rigorous derivation showing mutual consistency of the modified forward process, training objective, and reverse process. We will revise the manuscript in two ways: (1) We will add a derivation showing that when the network is trained to predict S⊙ε (or equivalently, when the score is adjusted by the spatially varying coefficient), the objective remains a valid ELBO for the modulated marginal, and we will clarify that our implementation approximates this by absorbing the S-dependence into the conditioning. (2) We will temper the language from 'modifying bridge dynamics' to 'spatially adaptive forward corruption with approximate reverse consistency,' and explicitly discuss the theoretical limitation that the reverse process does not use a corrected score. We believe this is the honest characterization: the noise modulation functions as a principled form of spatially adaptive perturbation that is empirically effective, but the full theoretical consistency proof is not provided and we will not overclaim it.","revision_made":"partial","referee_comment":"§2.3, Eq. (7): The PET-guided noise modulation replaces the standard bridge forward process x_t = μ_t + Σ_t^{1/2} ε with x_t = μ_t + Σ_t^{1/2}(S⊙ε), changing the marginal from N(μ_t, Σ_t I) to N(μ_t, Σ_t diag(S²)). However, the training objective in Eq. (3) uses the target (X_t − X_0)/σ_t with no S-dependent correction. For the modulated marginal, the correct denoising target should account for the spatially varying diffusion coefficient (e.g., predict S⊙ε or adjust the score by 1/S²). The paper does not prove that the modified forward process, the training objective, and the reverse process remain mutually consistent."},{"response":"This is a fair and important point. The current experiments do not isolate the individual contributions of the forward noise modulation (Eq. 7) and the reverse multi-scale PET attention (Eq. 8). We will conduct the requested ablation on the lesion-containing subset (Table 2 cohort) with three configurations: (a) attention-only (S = 1, PET attention active), (b) noise modulation only (S active, PET attention removed), and (c) full dual-stage PET guidance. This will allow us to report the marginal contribution of each component. Based on our understanding of the method, we expect both components to contribute—the noise modulation redistributes perturbation strength to emphasize pathological regions during training, while the attention module provides explicit PET-conditioned feature recovery during inference—but we will report whatever the data shows. We will add these results to a revised Table 2 (or an additional table) and update the discussion accordingly.","revision_made":"yes","referee_comment":"§2.3–2.4, Table 2: The lesion improvement (PSNR +0.80, SSIM +1.7) is attributed to both the forward noise modulation (Eq. 7) and the reverse multi-scale PET attention (Eq. 8), but no ablation isolates these two components. Since the attention module (Eq. 8) with zero-initialized W_out is a standard conditioning mechanism whose lesion benefit is plausible on its own, the marginal contribution of the noise modulation cannot be determined from the current experiments. An ablation with attention-only (no noise modulation) vs. full PET guidance would clarify whether the forward-process modification is necessary or whether the attention module suffices."},{"response":"The referee is correct that the manuscript does not explicitly define the 'no PET' configuration. To clarify: in the 'HA-DSB (no PET)' variant, both PET-dependent components are disabled—the noise modulator S is set to uniform (S = 1 everywhere, equivalent to the standard bridge forward process), and the multi-scale PET attention modules are removed from the UNet. The model retains the region context embeddings and all other architectural components. We will add this explicit description to Section 3 (Experiments) where the variant is introduced, so readers can unambiguously interpret the PET contribution shown in Tables 1 and 2.","revision_made":"yes","referee_comment":"§3, Table 1: The 'HA-DSB (no PET)' variant already outperforms all baselines substantially (e.g., Head/Neck SSIM 94.0 vs. ResViT 91.0). However, the paper does not describe what 'no PET' entails—whether the noise modulator S is set to uniform (S=1), whether the PET attention module is removed, or whether both are disabled. This ambiguity makes it difficult to interpret the PET contribution. The variant configuration should be explicitly stated."}],"tokens_in":9276,"tokens_out":1210,"duration_ms":357205,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper combines three known ingredients — I2SB Schrödinger bridges, VLM-generated region labels, and PET-guided attention — into a whole-body MRI translation pipeline that shows consistent quantitative improvements over five baselines. The clinical motivation is real (60–90 minute scan times), and the experimental gains are substantial, not marginal. But there's a theoretical gap in the PET-guided noise modulation that needs to be addressed before the central claim is fully supported. The reader and stress-test both flag it correctly, and I agree it's the load-bearing issue. Eq. 7 replaces the standard bridge forward process with a spatially modulated version where the scaling map S is predicted from PET. This changes the marginal distribution from N(μ_t, Σ_t I) to N(μ_t, Σ_t diag(S²)). But the training objective in Eq. 3 uses the standard target (X_t − X_0)/σ_t with no S-dependent correction. For the modified forward process, the correct denoising target should account for the spatially varying noise. The paper doesn't prove the bridge remains valid under this modification, and there's a potential train-test mismatch if the reverse process uses the standard bridge SDE without the spatially varying diffusion coefficient. The endpoints are preserved (noise vanishes at t=0 and t=1), so the method may work empirically as a form of spatially adaptive data augmentation — but the theoretical framing as a modified Schrödinger bridge is unsupported as written. What's genuinely new: the combination of VLM-generated region context embeddings (body-location labels from Gemini, encoded by PubMedBERT, fused via cross-attention with organ labels) with PET-guided dual-stage guidance for whole-body translation. The region conditioning is well-motivated and the per-region results in Table 1 show consistent gains across all five anatomical regions. The lesion-specific evaluation (Table 2) is a good experimental design — isolating 526 lesion-containing slices shows PET guidance yields +0.80 PSNR / +1.7 SSIM on pathology versus +0.16 / +0.3 overall. That's a meaningful signal. The ablation between HA-DSB(no PET) and HA-DSB(PET) does isolate the combined effect of noise modulation plus attention, but it does not separate the two. So we can't tell whether the lesion improvements come from the noise modulation (Eq. 7) or the multi-scale PET attention (Eq. 8). Given the theoretical concern about the noise modulation, this matters — if the bridge modification is invalid, the gains might be entirely from the attention module, which is architecturally sound. Minor issues: no statistical significance tests, single-institution dataset, VLM label validation is described as ~96% accuracy under professional review but the methodology is underspecified. These are standard limitations for a medical imaging paper at this stage. The core contribution — region-conditioned bridge translation with PET guidance — is sound and the experimental results are credible. The paper deserves a serious referee who can push on the bridge validity question and request the missing ablation. If the authors can either prove consistency of the modified forward process or show that the noise modulation contributes independently of the attention module, this is a solid contribution to the medical image translation literature.","headline":"Whole-body MRI translation via Schrödinger bridge with VLM region conditioning and PET-guided noise modulation; bridge validity under spatially modulated forward process is unproven.","tokens_in":10478,"tokens_out":755,"would_cite":true,"duration_ms":171451,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"PET-guided diffusion bridge translates whole-body MRI across regions and lesions","keywords":[],"falsifier":"If an ablation removing only the PET-guided noise modulation (but keeping PET attention) shows the same lesion-region improvements, the noise modulation mechanism would not be contributing the claimed effect.","tokens_in":9373,"feed_emoji":"🧲","tokens_out":1353,"duration_ms":75020,"temperature":0.7,"pith_summary":"This paper proposes a method called HA-DSB for translating one type of MRI scan into another across the entire body, a task complicated by the fact that different body regions (head, thorax, abdomen, pelvis, thighs) have very different image characteristics and lesions look different from healthy tissue. The authors frame translation as a Diffusion Schrödinger Bridge — a stochastic process that transports images from a source distribution to a target distribution — and make two key modifications. First, they inject region context embeddings generated by a vision-language model into the bridge's conditioning, so the transport dynamics adapt to which body region is being processed. Second, they use co-registered PET scans as a metabolic prior: during the forward (noise-injection) phase, PET uptake spatially modulates how much corruption each pixel receives, and during the reverse (denoising) phase, multi-scale PET-guided attention amplifies lesion-relevant features. The result is that the method achieves the best average structural similarity (SSIM 91.1%) and signal-to-noise ratio (PSNR 26.25 dB) across five body regions compared to five baselines, with PET guidance yielding substantially larger gains on lesion-containing slices than on healthy tissue.","feed_headline":"PET-guided diffusion bridge translates whole-body MRI, boosts lesion fidelity","feed_subtitle":"A single model handles five body regions by routing metabolic PET signals into the noise schedule of a Schrödinger bridge, with gains five-","key_machinery":"The core object is the Diffusion Schrödinger Bridge (DSB), specifically the I2SB instantiation, which defines a stochastic transport between two endpoint image distributions via an analytic Gaussian posterior q(X_t|X_0,X_1). The authors modify the standard bridge in two ways: (1) they replace the spatially uniform forward noise with a PET-guided spatial scaling map S that modulates per-pixel perturbation strength, and (2) they condition the reverse denoising network on region context embeddings derived from a vision-language model (body-location and organ labels encoded by PubMedBERT) fused with diffusion timestep embeddings.","core_discovery":"The central discovery is that spatially modulating the noise schedule of a Schrödinger bridge using PET uptake — so that regions with high metabolic activity receive different perturbation strengths than surrounding tissue — combined with region-aware conditioning from a vision-language model, produces a single model that handles whole-body MRI translation with improved fidelity in both normal anatomy and pathological regions. The PET guidance effect is concentrated where it matters: lesion-containing slices see a PSNR improvement of 0.80 dB from PET guidance, versus only 0.16 dB on the overall test set, indicating the noise modulation and attention mechanisms are selectively helping diseas-","pith_inferences":["The modest overall gains from PET guidance (PSNR +0.16) combined with large lesion-specific gains (PSNR +0.80) imply that the noise modulation is doing something localized rather than globally transformative — which is consistent with the design intent but also means the method's value proposition is narrow: it helps mainly when lesions are present.","The paper does not prove that spatially modulating the forward noise preserves the Schrödinger bridge's theoretical guarantees. If the bridge formulation is broken, the method may functionally reduce to a conditional diffusion model with PET attention — still potentially useful, but the bridge framing would be incidental rather than load-bearing.","The reliance on a specific VLM (Google Gemini 3 Pro) for label generation introduces an external dependency whose failure modes on unusual anatomy (e.g., post-surgical changes, congenital anomalies) are not characterized."],"forward_implications":["If the spatial noise modulation is valid, the same principle could extend to other paired-modalities settings where an auxiliary image provides a spatial prior for where translation difficulty is concentrated (e.g., CT-guided MR translation, or ultrasound-guided CT synthesis).","The region context embedding approach suggests that vision-language models can serve as lightweight anatomical priors for any medical imaging task with body-region heterogeneity, potentially replacing hand-crafted region labels.","If PET-guided noise modulation genuinely helps lesion fidelity, it could reduce the need for full multi-sequence MRI acquisition in PET/MR protocols, shortening clinical scan times.","The concentration of PET-guidance gains in lesion regions (5x the overall improvement) suggests that evaluation on whole-image metrics underestimates clinically relevant gains, and pathology-specific evaluation protocols should become standard."],"fun_headline_variants":["PET-modulated noise schedule lifts lesion fidelity in whole-body MRI translation","Single Schrödinger bridge model translates five body regions with PET-guided noise","Metabolic PET signals steer diffusion bridge noise for sharper lesion MRI","PET uptake modulates bridge noise schedule, improving lesion PSNR by 0.80 dB","Vision-language region context plus PET guidance enables whole-body MRI translation"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The PET-guided noise modulation modifies the forward noise schedule of the Schrödinger bridge without a corresponding proof that the resulting process still defines a valid bridge with a consistent training objective. If the spatial modulation breaks the bridge formulation, the lesion improvements could come from the PET attention module alone rather than from the noise modulation.","fun_headline_variants_meta":{"raw":{"variants":["PET-modulated noise schedule lifts lesion fidelity in whole-body MRI translation","Single Schrödinger bridge model translates five body regions with PET-guided noise","Metabolic PET signals steer diffusion bridge noise for sharper lesion MRI","PET uptake modulates bridge noise schedule, improving lesion PSNR by 0.80 dB","Vision-language region context plus PET guidance enables whole-body MRI translation"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":703,"prompt_tokens":607,"completion_tokens":96,"prompt_tokens_details":null},"tokens_in":607,"tokens_out":96,"duration_ms":31990,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T11:55:16.394339+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If an ablation removing only the PET-guided noise modulation (but keeping PET attention) shows the same lesion-region improvements, the noise modulation mechanism would not be contributing the claimed effect.","supporting_citations":[],"review_version":1}