{"id":"3b0fcb59-34ed-45db-bcf0-ceca6ed46549","arxiv_id":"2509.09610","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"Combining a fitted ODE tumor-growth model with gradient-guided DDIM produces future MRI slices and tumor-growth probability maps for pediatric DMG patients.","lead":"This paper combines a mathematical model of tumor growth with an AI image generator to create predicted follow-up MRI scans for children with brain tumors. The goal is to help doctors anticipate how and where a diffuse midline glioma will grow from only a few earlier scans.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HD95 directionality claim is confounded by size effect; no size-matched baseline in evaluation.","rationale":"The reader's weakest assumption exactly identifies that the size-only gradient may not encode direction, making HD95 improvement a size artifact. My analysis agrees and sharpens the test: any larger mask reduces HD95 toward a larger target, so a size-matched isotropic dilation baseline is the minimal control needed. The paper's own evaluation lacks such a baseline; the static setup (true size) and dynamic setup (bootstrap sizes) both vary size while holding only the scalar target. The HD95 comparison to the initial mask is therefore confounded. Other issues (Eq. 6 vs Algorithm 1 inconsistency, tuning overlap, static evaluation) are real but secondary; the directionality claim is load-bearing for the clinical utility statement. A single concrete control experiment can settle whether the concern lands. Since the reader already flagged this and recommended a conditional verdict, my read does not change the verdict.","tokens_in":9210,"tokens_out":1704,"duration_ms":21387,"concrete_test":"Add a baseline where the initial tumor segmentation is expanded isotropically (e.g., morphological dilation with a spherical structuring element) to match the target tumor area, then compute HD95 between this dilated mask and the target. If the dilated baseline achieves HD95 comparable to the generated dynamic/static probability maps (no significant difference), the directionality claim collapses to a size effect. Alternatively, compute HD95 for generated images conditioned on the true target size but with the regressor gradient replaced by a random uniform expansion direction (or no gradient) while matching target size; if HD95 does not degrade, direction is not being learned. Report the difference and significance for at least the 60 slices used in Fig. 4.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the framework captures both extent and directionality of tumor growth, supported by HD95 between thresholded generated probability maps and target segmentations being significantly lower than HD95 between initial and target (Fig. 4B). However, the guidance signal in Eq. (2) is the gradient of a scalar regressor R that predicts tumor size relative to brain area. This gradient carries no explicit spatial-direction information; it only pushes the generated image toward larger tumor area. Consequently, the diffusion model can satisfy the guidance by enlarging the tumor in almost any direction, as long as the area increases. The HD95 improvement is then expected even for directionally incorrect growth, because any mask that is larger and overlaps the target will reduce the 95th-percentile Hausdorff distance relative to the initial smaller mask. The paper's own static evaluation (target set to true size) further removes the ODE's role and isolates the size-guidance effect; the dynamic evaluation adds bootstrap size variation but still no spatial baseline. Without a control that expands the initial tumor mask isotropically (or via a non-directional baseline) to the same target size, the reported HD95 reduction cannot be attributed to correct growth directionality. In addition, Eq. (6) defines lambda'(t) with a tanh form, while Algorithm 1 uses lambda_eff = -lambda*tanh(...); these are inconsistent, and the sign convention may affect the fitted dynamics. However, the most load-bearing issue is the missing size-matched baseline for the directionality claim, because it directly undermines the headline contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid 'mechanistic learning' framework for spatio-temporal brain tumor growth prediction. A mechanistic ODE model, including exponential pre-RT growth, compartmental survival/death fractions at RT onset, and a tanh-delayed decay term, is fitted per patient and per 2D axial slice to longitudinal tumor area measurements. The fitted model extrapolates future tumor burden, which is then used as a target for a gradient-guided DDIM: a regressor trained to predict tumor size relative to brain area steers the denoising process toward images with the desired tumor size. Follow-up FLAIR slices are synthesized, and tumor growth probability maps are computed by aggregating difference maps from multiple generated images with varying target sizes (static: true target size; dynamic: ODE bootstrap estimates). The method is trained on BraTS adult/pediatric data and evaluated on 60 axial slices from 8 in-house pediatric DMG patients. The main reported results are region-specific SSIM for image quality and a significant reduction in HD95 between generated thresholded probability maps and the true target segmentation compared with initial-to-target HD95, which the authors interpret as evidence that the framework captures both extent and directionality of tumor growth.","tokens_in":9549,"tokens_out":5067,"duration_ms":64429,"significance":"If the claims hold, the paper would make a useful contribution: it combines a biologically motivated ODE model with a modern generative model in a data-sparse clinical setting, and it reports an in-house longitudinal pediatric DMG evaluation rather than relying only on public datasets. The public code availability and the explicit use of bootstrap uncertainty quantification are strengths. The mechanistic module is evaluated separately from the generative module, which is good practice. However, the central claim about growth 'directionality' is currently supported only by a metric that is confounded by tumor-size increase, and the ODE model has an internal mathematical inconsistency. Because these issues are load-bearing for the main conclusions, the paper requires substantive revision before the claims can be accepted.","major_comments":[{"comment":"The delayed decay rate is defined inconsistently. Eq. (6) states lambda'(t) = lambda_decay * tanh((t - t_RTstart - delta) * slope), while Algorithm 1 line 7 sets lambda_eff = -lambda * tanh((t - t_RTstart - delay)/slope). These differ in sign, in the magnitude parameter (lambda_decay vs lambda), in the delay symbol, and in the slope scaling. More importantly, both Eqs. (6) and Algorithm 1 write the dying compartment as A_d(t) = A_d(t_RTstart) * exp(lambda_eff(t) * (t - t_RTstart)) with a time-dependent rate. This is not the solution of dA_d/dt = lambda_eff(t) * A_d; the correct exponential factor is exp(integral lambda_eff(s) ds). Since the ODE predictions are the conditioning targets for image generation, this inconsistency and integration error must be corrected and the resulting fits re-evaluated.","section":"Section 2.2, Eq. (6) and Algorithm 1 line 7"},{"comment":"The claim that the framework captures growth directionality is confounded by tumor-size increase. The HD95 of the generated (thresholded probability) mask against the true target segmentation is compared with the HD95 of the initial (smaller) segmentation against the target. Because the guidance signal is a scalar regressor of tumor size relative to brain area, the generated image is explicitly pushed toward larger tumor area; any mask that enlarges the tumor and overlaps the target will reduce HD95 relative to the smaller initial mask, even if the direction of growth is wrong. A size-matched baseline is required: for example, isotropically dilating the initial tumor mask to the target area, or using a non-spatial growth baseline, and comparing HD95 against that baseline. Without such a control, the reported HD95 improvement does not demonstrate correct directionality.","section":"Section 3.3, Fig. 4B (HD95 evaluation)"},{"comment":"The static probability-map evaluation conditions on the true tumor size at the next imaging session and then compares the generated mask with the true segmentation at that same session. This is a valid oracle test of the generative model's ability to produce a target-sized tumor, but it removes the mechanistic ODE from the spatial prediction path. The dynamic evaluation uses ODE-estimated sizes, but it still lacks a size-matched spatial control. Consequently, the specific contribution of the mechanistic model to spatial accuracy is not identified. Please report the results separately for the static and dynamic settings with the control baseline, and state explicitly which claim each comparison is intended to support.","section":"Section 3.3 (static vs. dynamic evaluation)"}],"minor_comments":[{"comment":"Typographical errors: 'cnetral' should be 'central' (Section 3.2), 'Institue' should be 'Institute' (affiliations), and 'bootstrap' capitalization is inconsistent.","section":"Throughout"},{"comment":"The notation epsilon_theta^{(l)} is not defined clearly; it should distinguish the network's noise prediction at diffusion step l from the noise terms in Eq. (1).","section":"Eq. (2)"},{"comment":"The dynamic guidance scaling s_R(dyn) is described verbally but no update rule is given. Please specify the formula used to adapt the gradient scale during sampling.","section":"Section 2.3"},{"comment":"The y-axis label 'HD95 Distance' lacks units and the figure does not report the number of paired samples, exact p-values, or effect sizes. Please include these details in the text or caption.","section":"Figure 4B"},{"comment":"The criterion for using 'tumor sizes below 90th percentile of the bootstrap-estimated values' as targets is not justified. Please clarify how the target sizes were sampled and why the upper tail was excluded.","section":"Section 3.3"},{"comment":"The labels 'all' and 'train' are used without a definition in the figure caption. Define them in the caption or in the main text for clarity.","section":"Figure 3B and Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"I concur with the reader's assessment: the HD95 directionality claim is the main obstacle. The size-match confound is severe because the regressor gradient is a scalar size signal, and the reported comparison to the initial segmentation cannot isolate spatial direction. The ODE inconsistency in Eq. (6)/Algorithm 1 is also real and should be fixed before any further claims about mechanistic predictions. The paper fits the scope of the journal, and the in-house longitudinal dataset is valuable, but the central claim is not yet supported. A revision that adds a size-matched control and corrects the ODE formulation could change my recommendation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a plausible integration of a per-slice ODE tumor-growth model with gradient-guided DDIM, plus bootstrap-aggregated growth probability maps, evaluated on a small in-house longitudinal DMG cohort. The combination is new, the code link appears, and the idea of using a mechanistic size forecast as a guidance target is reasonable. The probability-map aggregation is a nice touch. Credit where due: the ODE module is simple but evaluated with bootstrap and nRMSE on held-out time points, and the diffusion training on BraTS is standard.\n\nThe headline claim about directionality, though, is not yet supported. The guidance signal is the gradient of a scalar regressor that predicts tumor area relative to brain area. That gradient carries no explicit spatial-direction information. The diffusion model can satisfy it by expanding the tumor in almost any direction, as long as the area increases. The reported HD95 improvement over the initial mask is then expected even for directionally wrong growth, because a larger overlapping mask reduces HD95 purely by size. Without a size-matched baseline—e.g., isotropically dilating the initial mask to the same target area—the HD95 comparison cannot isolate directionality. That is the load-bearing flaw.\n\nThere is also a concrete inconsistency: Eq. (6) defines lambda'(t) = lambda_decay * tanh(...), while Algorithm 1 uses lambda_eff = -lambda * tanh(...). Those are different models, and the sign conventions differ. The paper never reconciles them. The static evaluation in Sec. 3.3 conditions on the true target size, which removes the ODE's forecast role; the dynamic maps use ODE sizes but still suffer the same size-effect issue. The hyperparameter grid search was done on the same Zurich cohort; the paper says the 8 evaluation patients are 'additional,' but it should state explicitly that they are disjoint from the 29 used for tuning.\n\nThese are fixable. The right fix is to add a non-directional size-control baseline (isotropic dilation, or a diffusion run with the same target size but no guidance beyond area) and show the generated direction beats that baseline. Also fix the lambda inconsistency and report error bars over slices. The writing is honest about limitations, and the work is serious, not a toy.\n\nWho is this for: researchers working on mechanistic generative models for longitudinal medical imaging. As written, I would not let the directionality claim stand, but the framework deserves referee time. Recommendation: send to peer review with a request for the size-matched baseline and the equation fix.","headline":"A sensible hybrid of ODE tumor-burden forecasting and guided DDIM, but the directionality claim needs a size-matched baseline before it stands.","tokens_in":10053,"tokens_out":2391,"would_cite":false,"duration_ms":25773,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A mechanistic ODE forecast of tumor area, used as a conditioning target for a gradient-guided DDIM, generates follow-up brain MRIs whose tumor growth probability maps match observed growth direction, as measured by 95th percentile Hausdorff","keywords":["diffusion models","tumor growth prediction","mechanistic learning","pediatric diffuse midline glioma","longitudinal MRI synthesis","tumor growth probability maps","ordinary differential equations","Hausdorff distance"],"falsifier":"A direct test is to replace the ODE-extrapolated target size with a size-matched but direction-neutral target (e.g., the true final area, or a shuffled size from another patient) and measure whether thresholded growth probability maps still reduce HD95 relative to the initial segmentation. If an isotropic enlargement of the initial tumor mask to the same area achieves equal or better HD95, then the reported directionality reduces to a size effect.","tokens_in":9096,"feed_emoji":"🧠","tokens_out":4062,"duration_ms":47233,"temperature":0.7,"pith_summary":"This paper tries to show that a hybrid pipeline can predict how a brain tumor will grow in space and time from a recent MRI and a mechanistic model of tumor dynamics. It combines an ordinary differential equation model, which forecasts tumor area after radiotherapy, with a denoising diffusion implicit model steered by a regressor gradient toward that forecasted size. The authors test the approach on pediatric diffuse midline glioma, a rare and fatal tumor with sparse longitudinal imaging data. They report that synthesized follow-up scans preserve anatomy, and that tumor growth probability maps produced by aggregating bootstrap-conditioned generations are significantly closer to the true follow-up segmentation than the starting scan is, by 95th percentile Hausdorff distance. The value, if correct, is that biologically informed image generation becomes possible in data-limited clinical settings.","feed_headline":"Mechanistic diffusion model grows brain tumors in the right direction","feed_subtitle":"A tumor-size ODE forecast guides image diffusion, yielding follow-up MRIs that match where the tumor actually spreads.","key_machinery":"The load-bearing object is the gradient of a trained regressor R(x_l, l) that predicts tumor size relative to brain area from a noisy image. This gradient is added to the DDIM denoising step (Eq. 2), scaled by a dynamic factor s_R(dyn) that depends on the gap between the regressor's current output and the ODE-extrapolated target size, times a constant scale. The ODE model itself partitions the tumor at radiotherapy onset into surviving and dying compartments, with delayed exponential decay for the dying fraction, and is fitted per patient and per slice with bootstrap resampling to produce a distribution of target sizes. The machinery's job is to convert a scalar forecast (tumor area at follo","core_discovery":"The central claim is that mechanistic model estimates of future tumor burden can serve as conditioning targets for a gradient-guided DDIM, and that doing so yields realistic follow-up MRI scans in which the direction of tumor growth is clinically relevant. The authors introduce tumor growth probability maps, computed by averaging binarized difference maps across generations using bootstrap-sampled target sizes, and show that thresholded versions of these maps have a significantly lower HD95 to the true target segmentation than the initial tumor mask does. They interpret this as evidence that the generated growth is both anatomically feasible and directionally aligned with observed progressio","pith_inferences":["If the regressor gradient is indeed the carrier of directional information, a testable extension is to visualize the gradient field on held-out images and check whether its support concentrates on tumor-adjacent tissue; diffuse support would predict limited transfer to 3D volumes or tumors with less distinct margins.","The bootstrap-to-probability-map aggregation is a general recipe: any scalar biomarker forecast from a mechanistic model could be converted into a spatial probability map over image changes, potentially useful for other diseases with sparse longitudinal imaging.","A natural next experiment is to compare the generated growth maps against a purely data-driven spatio-temporal diffusion model trained on the same longitudinal pairs; the mechanistic prior should improve data efficiency and directional consistency, and that comparison would isolate the contribution of the ODE.","Because the ODE parameters are per patient and per slice, the framework yields a quantitative, time-continuous tumor response curve that could be correlated with clinical endpoints beyond imaging, such as progression-free survival."],"forward_implications":["Follow-up MRIs can be generated from a recent scan plus an ODE-based area forecast, without needing a large longitudinal training set for the target population.","Tumor growth probability maps can be thresholded to produce predicted future tumor segmentations for treatment planning.","The per-slice, per-patient ODE fits give time-continuous forecasts of tumor response to radiotherapy, including the timing of delayed cell death.","Bootstrapped target sizes yield dynamic probability maps with lower HD95 variability than static maps produced with a single target size.","The framework offers generative-space-time predictions that incorporate mechanistic priors, which is relevant for data-scarce pediatric cancers like DMG."],"fun_headline_variants":["Guide diffusion with tumor growth equations for faithful MRI forecasts","ODE-driven diffusion predicts where brain tumors spread","Mechanistic prior steers diffusion for spatio-temporal tumor prediction","Growth probability maps reveal tumor spread direction from AI MRIs","Hybrid model uses tumor ODE to guide image diffusion for follow-up MRIs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole directionality claim rests on the regressor's gradient, trained only on static single-time-point images, encoding spatial growth direction when pushed toward a larger tumor-area scalar; if that gradient encodes size increase without anatomical direction, the HD95 improvement is just a size artifact.","fun_headline_variants_meta":{"raw":{"variants":["Guide diffusion with tumor growth equations for faithful MRI forecasts","ODE-driven diffusion predicts where brain tumors spread","Mechanistic prior steers diffusion for spatio-temporal tumor prediction","Growth probability maps reveal tumor spread direction from AI MRIs","Hybrid model uses tumor ODE to guide image diffusion for follow-up MRIs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1408,"prompt_tokens":694,"completion_tokens":714,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":630}},"tokens_in":438,"tokens_out":714,"duration_ms":7916,"temperature":1.0,"reasoning_tokens":630,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:44:51.610699+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to replace the ODE-extrapolated target size with a size-matched but direction-neutral target (e.g., the true final area, or a shuffled size from another patient) and measure whether thresholded growth probability maps still reduce HD95 relative to the initial segmentation. If an isotropic enlargement of the initial tumor mask to the same area achieves equal or better HD95, then the reported directionality reduces to a size effect.","supporting_citations":[],"review_version":1}