{"id":"777577b9-89f8-4960-baed-ebaa10b1f3c8","arxiv_id":"2509.08442","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SBDM applies a Brownian bridge diffusion process on registered cortical surfaces to forecast vertex-level cortical thickness, outperforming several baselines on ADNI and OASIS.","lead":"The paper introduces SBDM, a Brownian bridge diffusion model that forecasts vertex-level cortical thickness changes from a baseline scan and tabular covariates, using a new spherical U-Net denoiser. On ADNI and OASIS longitudinal data it reports lower mean absolute errors than prior regression and diffusion baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2 recursion is theoretically correct but underived and has an undefined division by zero at β=0; text contradicts Algorithm 1.","rationale":"The reader's weakest assumption was that the reverse-sampling recursion in Algorithm 2 might be mis-specified for the modified loss target. Re-deriving the conditional Brownian bridge posterior shows the recursion is actually correct: the coefficients ζ3 and δ̃ correspond exactly to the posterior mean and variance when f_θ predicts x_β-Δτ_t. Thus the core mathematical objection is not substantiated. Nevertheless, the paper's presentation is deficient: no derivation is given, the pseudocode contains a division by zero at β=0 (the ratios δβ+1/δβ and δβ|β+1/δβ are 0/0), and the prose in §3.1 contradicts Algorithm 1 regarding the direction of the forward process. These are not fatal to the central claim—the empirical results presumably come from a correct implementation—but they make the method unreproducible from the manuscript alone and weaken the assurance that the reported accuracy is not an artifact of an unspecified implementation choice. The reader's conditional verdict remains appropriate: the paper should be accepted only after the derivation is supplied, the β=0 case is clarified, the text contradiction is fixed, and code is made available. Therefore I recommend no change to the reader's verdict.","tokens_in":10177,"tokens_out":24752,"duration_ms":241673,"concrete_test":"Run a unit test of Algorithm 2 with a perfect predictor f_θ(x_β,β,t,c)=x_β-Δτ_t for a known Δτ_t, starting at β=0. With the correct limiting coefficients, x_B should equal Δτ_t up to the stated variance; a naive implementation will encounter 0/0. Additionally, independently re-derive ζ1, ζ2, ζ3, δ̃ from Eq. (1) to verify they match the derived α and variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (SBDM significantly reduces prediction errors) assumes Algorithm 2's recursion correctly samples the Brownian bridge of Eq. (1) under the training target of Eq. (3). I derived the bridge posterior: for mβ<mβ+1, q(x_{β+1}|x_β,τ0,Δτ_t) is Gaussian with mean x_β - α(x_β-Δτ_t), α=(mβ+1-mβ)/(1-mβ), and variance 2α(1-mβ+1)I. Substituting the model estimate Δτ_t ≈ x_β - f_θ gives x_{β+1}=x_β - α f_θ + sqrt(2α(1-mβ+1))η. Simplifying Algorithm 2's ζ1, ζ2, ζ3, δ̃ yields ζ1=1, ζ2=0, ζ3=α, δ̃=2α(1-mβ+1), exactly matching. So the recursion is mathematically correct. However, the paper never derives this, and the pseudocode is not executable as written: at β=0, δ0=0, making the ratios δβ+1/δβ and δβ|β+1/δβ 0/0; the limiting coefficients are not provided. Moreover, §3.1's prose states the forward process maps Δτ_t to τ0, opposite to Algorithm 1, which interpolates from τ0 to Δτ_t. These gaps make the method irreproducible from the paper and leave the possibility that the reported gains rely on an unspecified implementation detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SBDM, a Brownian bridge diffusion model defined on spherical cortical surface representations, for forecasting vertex-level cortical thickness (CTh) changes from a baseline CTh map and tabular covariates (age, sex, diagnoses, time interval). The forward model interpolates between baseline CTh and a future relative change, with a newly introduced conditional spherical U-Net (CoS-UNet) as denoiser. Experiments on ADNI (subject-level split) and an external OASIS test set report mean absolute errors around 0.097–0.100 mm, about 10% better than the strongest baseline (SiT), with significant paired Wilcoxon tests (p < 10^-3), plus an ablation of the denoising architecture and counterfactual trajectory demonstrations.","tokens_in":10614,"tokens_out":8978,"duration_ms":78923,"significance":"If the method and empirical results hold, SBDM is a practically useful, transferable vertex-wise CTh forecaster, and the CoS-UNet architecture with dense cross-attention is a sensible design for fusing spherical surface data with tabular conditions. The paper deserves credit for a subject-level split, an external dataset without fine-tuning, paired statistical tests, and an ablation that isolates the denoiser contribution. The main weakness is that the method section is not fully reproducible: the direction of the Brownian bridge is described inconsistently, the inference recursion is stated without derivation, and the pseudocode contains a division-by-zero at the initial step. These are fixable, but they are load-bearing for the central claim, because the empirical gains rest on the correctness of the sampling recursion and its alignment with the training target.","major_comments":[{"comment":"The prose states: “by definition, our forward process maps the relative change Δτ_t … and maps it to the baseline thickness τ_0 … Conversely, the reverse process reconstructs Δτ_t from τ_0.” But Eq. (1) with x_0=τ_0 and x_B=Δτ_t, and Algorithm 1 line 6 (x_β ← (1−m_β)τ_0 + m_β Δτ_t + √δ_β ϵ), define the opposite direction: the forward process maps τ_0 to Δτ_t. The direction convention is load-bearing because Algorithm 2 reverses this process; the text must be corrected.","section":"§3.1, Eq. (1)–(2) and Algorithm 1, line 6"},{"comment":"The inference recursion is introduced “in analogy to the original BBDM” but no derivation is given for the modified setting, where the training target in Eq. (3) is (1−m_β)(τ_0−Δτ_t)+√δ_β ϵ (equivalently x_β−Δτ_t), not a pure noise term, and where τ_0 is a fixed condition. The paper should derive the posterior q(x_{β+1} | x_β, τ_0, Δτ_t), state the substitution Δτ_t ≈ x_β − f_θ(x_β), and show that the coefficients ζ_1, ζ_2, ζ_3, δ̃ follow. Without this, the reader cannot verify that Algorithm 2 samples the conditional bridge.","section":"§3.1, Eq. (4)–(5) and Algorithm 2"},{"comment":"The pseudocode is not executable at β=0: δ_0=0 and m_0=0, so the expressions for δ_{β|β+1}, ζ_1, ζ_3, and δ̃ involve division by zero. The limiting values exist (e.g., ζ_1→1, ζ_2→0, ζ_3→m_1, δ̃→2m_1(1−m_1)), but they are not given. Because the loop starts at β=0, the implementation is underspecified. Please provide the limiting coefficients or start the recursion at β=1 with the appropriate initialization.","section":"Algorithm 2, lines 4–8"}],"minor_comments":[{"comment":"The text says the network recovers the “noise” from x_β, but the target is (1−m_β)(τ_0−Δτ_t)+√δ_β ϵ, which equals x_β−Δτ_t. Clarify this target and why it is not the pure diffusion noise.","section":"Eq. (3)"},{"comment":"The non-Markovian sampling strategy is mentioned but not specified: how are the 200 sampled stages chosen, and how do the recursion coefficients adapt to skipped steps? This is needed for reproducibility.","section":"§4.1, Implementation Details"},{"comment":"OASIS has no MCI column (shown as “–”). The phrase “across all diagnostic groups and datasets” should be qualified as “all available diagnostic groups” to avoid overstatement.","section":"Table 1"},{"comment":"The mean improvement over SiT is about 0.01 mm, which is 10% relative but small in absolute terms. Given the reported standard deviations (e.g., 0.097±0.031 vs. 0.108±0.028), please comment on the practical/clinical significance of this difference, beyond statistical significance.","section":"§4.2"},{"comment":"The code URL is given as “will be made available”; for review, a working repository or detailed hyperparameter/schedule listing would strengthen reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound in its empirical design, and I verified that the Algorithm 2 coefficients are consistent with the correct Brownian-bridge posterior if the paper's target is interpreted as a prediction of x_β−Δτ_t. The issues are therefore not fatal, but the presentation of the method must be corrected and made fully reproducible before the paper can be accepted. The contribution is somewhat incremental over BBDM and existing spherical architectures, but the application domain and the thorough comparison justify a major-revision route."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on arXiv:2509.08442. The paper does something straightforward but useful: it takes Brownian bridge diffusion (BBDM) and applies it to vertex-level cortical thickness forecasting, with a new spherical U-Net (CoS-UNet) that conditions on demographics, diagnosis, and follow-up time. The empirical work is honestly done: subject-level splits, external OASIS test, several baselines, ablations, and Wilcoxon tests. The reported ~10% MAE improvement over SiT is modest but consistent, and the external generalization is a nice plus.\n\nThe novelty is not in the stochastic process—that's BBDM unchanged. The new bits are the domain application, the CoS-UNet architecture, and the counterfactual trajectory generation. That's fine for an applied paper.\n\nSoft spots: Section 3.1 states the forward process maps Δτ_t to τ0, while Algorithm 1 interpolates from τ0 to Δτ_t. That's a direct contradiction and needs fixing. More importantly, the reverse-sampling recursion in Algorithm 2 is given \"in analogy to BBDM\" without derivation. I checked the math: the coefficients do match the correct posterior when you substitute the training target, so the recursion is theoretically sound. But the pseudocode as written has division by zero at β=0 (δ0=0), and the limiting coefficients aren't provided. So the paper is not fully reproducible from the text. These are fixable in revision, but they're real gaps.\n\nThere's also no code released yet, so the empirical claims rest on the numbers as reported. The ablation study is useful—it shows CoS-UNet outperforms MLP, plain S-UNet, and SiT within the same bridge framework.\n\nOverall, this is a solid applied diffusion paper with a credible but incremental advance. The central method holds up after derivation; the presentation needs a careful pass. I'd send it to peer review with the expectation of major revision on the method description. The authors should be asked to reconcile the forward/reverse direction, add a proper derivation of the recursion, and handle the β=0 edge case explicitly.\n\nWho's it for: people working on longitudinal cortical morphology or conditional diffusion for medical imaging. I'd cite it if I were in that area.\n\nRecommendation: accept for peer review, conditional on those revisions.","headline":"Solid applied diffusion paper; the math checks out after derivation, but the method section has contradictions and an under-specified recursion that need fixing before the results are fully reproducible.","tokens_in":11046,"tokens_out":3513,"would_cite":true,"duration_ms":35797,"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":"A spherical Brownian bridge diffusion model forecasts vertex-level cortical thickness from a baseline scan and tabular covariates, cutting prediction error by about 10%.","keywords":["longitudinal prediction","cortical thickness","diffusion models","Brownian bridge","spherical U-Net","cross-attention","Alzheimer's disease","counterfactual trajectories"],"falsifier":"Run Algorithm 2 on held-out training pairs with known future changes Delta_tau_t, starting from the true tau0, and compare the generated x_B against the true Delta_tau_t, or compare the empirical distribution against q(x_beta | tau0, Delta_tau_t). If the samples are systematically biased, the reported accuracy does not come from the Brownian bridge posterior. A quicker check: verify that the recursion's posterior mean for q(x_beta+1 | x_beta, tau0) equals the training loss target when the denoiser is perfect.","tokens_in":10117,"feed_emoji":"🧠","tokens_out":7672,"duration_ms":84813,"temperature":0.7,"pith_summary":"This paper introduces SBDM, a diffusion model that forecasts how an individual's cortical thickness will change at every vertex of the cortical surface. The central claim is that a conditional Brownian bridge—a stochastic path whose endpoints are the baseline thickness map and the future change map—can be learned directly on the spherical surface, and that it outperforms prior methods for this task. Reported mean absolute errors are 0.097 mm on ADNI and 0.100 mm on OASIS, roughly 10% below the best competing method, with Wilcoxon p < 10^-3 over both datasets. The same model can be conditioned on a target diagnosis to generate factual and counterfactual thickness trajectories. If the result holds, it offers a high-resolution, single-scan based tool for tracking neurodegeneration and exploring hypothetical disease courses.","feed_headline":"Bridge diffusion cuts cortical-thickness forecast error ~10%","feed_subtitle":"Model predicts vertex-level thinning from a single baseline scan and transfers to a new cohort without retraining.","key_machinery":"The central object is the conditional Brownian bridge on the sphere: q(x_beta | tau0, Delta_tau_t) = N(x_beta; (1 - m_beta) tau0 + m_beta Delta_tau_t, delta_beta I), with m_beta = beta/B and delta_beta = 2(m_beta - m_beta^2). Baseline thickness is one endpoint and the future change is the other; training asks the denoiser to reproduce the combined signal (1 - m_beta)(tau0 - Delta_tau_t) + sqrt(delta_beta) epsilon from x_beta. The companion architecture, CoS-UNet, lets the bridge live directly on the cortical surface: spherical convolutions preserve local geometry, cross-attention injects the tabular covariates and time interval at every stage, and the bridge step is added as an embedding.","core_discovery":"In the paper's own terms, SBDM is the first Brownian bridge diffusion model for conditional vertex-level predictions on spherical surfaces. It maps between baseline cortical thickness tau0 and a future relative change Delta_tau_t via q(x_beta | tau0, Delta_tau_t), training a spherical U-Net with cross-attention to denoise the bridge at every step. Inference starts at tau0 and walks a non-Markovian recursion to produce Delta_tau_t. Across all diagnostic groups and both datasets, SBDM reports the lowest errors—0.097 mm on ADNI and 0.100 mm on OASIS—and statistically significant improvements over the compared baselines, and it produces plausible counterfactual trajectories when the follow-up di","pith_inferences":["The reported ablations change the denoising network while keeping the bridge; a same-backbone comparison between DDPM and BBDM would isolate how much of the ~10% gain comes from the Brownian bridge rather than the cross-attention architecture.","The same bridge formulation could be applied to other surface-valued longitudinal biomarkers, such as cortical volume, curvature, or myelin maps, because the geometry handling is in the network, not in the diffusion process.","The counterfactual diagnosis switch is associative rather than causal: diagnosis and disease trajectory co-evolve in reality, so the switched trajectories should be validated against observed converters before being used for clinical inference.","A testable extension is to check whether the counterfactual trajectories predict faster conversion in subjects who later convert to AD; if they do, the model captures disease-relevant signal beyond average aging."],"forward_implications":["Vertex-level cortical thickness forecasts from a single baseline scan plus tabular covariates reach about 0.1 mm mean absolute error, around 10% lower than the best reported baseline.","Because the model is a direct map between baseline and future change, predicting a follow-up visit does not require multiple prior visits or region-level aggregation.","The model transfers to an unseen dataset without retraining, with OASIS error close to the ADNI test error.","Conditioning on follow-up diagnosis produces factual and counterfactual trajectories, so the same architecture supports 'what if the diagnosis had been AD' simulations at the individual level.","Explicit time conditioning makes predictions feasible at arbitrary follow-up intervals, matching the irregular visit structure of real longitudinal studies."],"supporting_citations":[{"why":"Supplies the Brownian bridge diffusion formulation and the reverse-sampling recursion that SBDM adapts to cortical surfaces.","marker":"[12]"},{"why":"Supplies the spherical convolution operator and U-Net backbone that CoS-UNet builds on, and serves as a baseline.","marker":"[30]"},{"why":"Supplies the surface vision transformer architecture used as a baseline and as an alternative denoising network.","marker":"[4]"},{"why":"Supplies the prior conditional diffusion model for cortical thickness trajectory prediction, re-implemented and compared against.","marker":"[27]"},{"why":"Supplies the base denoising diffusion training and sampling implementation used as a starting point for SBDM.","marker":"[10]"},{"why":"Supplies the FreeSurfer cortical surface reconstruction pipeline that generates vertex-level thickness measurements.","marker":"[6]"},{"why":"Supplies unbiased within-subject template estimation for longitudinal analysis, used to compute cortical thickness trajectories.","marker":"[16]"},{"why":"Supplies the non-Markovian accelerated sampling strategy used during SBDM inference.","marker":"[25]"}],"fun_headline_variants":["Spherical Brownian bridge forecasts cortical thickness curves","Brownian bridge diffusion predicts cortex thinning on sphere","Diffusion bridge model sharpens cortical thickness forecasts","First Brownian bridge diffusion for cortical thickness prediction","Bridge diffusion trims cortical thickness forecast error"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The benchmark gains rest on the assumption that Algorithm 2's recursion actually samples the conditional Brownian bridge that the model was trained on; the paper adapts the coefficients from the image-based BBDM 'in analogy' without deriving them for its modified denoising target, and the prose and Algorithm 1 disagree about the forward direction.","fun_headline_variants_meta":{"raw":{"variants":["Spherical Brownian bridge forecasts cortical thickness curves","Brownian bridge diffusion predicts cortex thinning on sphere","Diffusion bridge model sharpens cortical thickness forecasts","First Brownian bridge diffusion for cortical thickness prediction","Bridge diffusion trims cortical thickness forecast error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1115,"prompt_tokens":738,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":482,"tokens_out":377,"duration_ms":4605,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T20:39:04.847254+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 on held-out training pairs with known future changes Delta_tau_t, starting from the true tau0, and compare the generated x_B against the true Delta_tau_t, or compare the empirical distribution against q(x_beta | tau0, Delta_tau_t). If the samples are systematically biased, the reported accuracy does not come from the Brownian bridge posterior. A quicker check: verify that the recursion's posterior mean for q(x_beta+1 | x_beta, tau0) equals the training loss target when the denoiser is perfect.","supporting_citations":[],"review_version":1}