{"id":"c7006167-aaa6-4fc7-aef9-72703d3c9aa3","arxiv_id":"2509.10824","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A multitask diffusion model predicts future FLAIR images and probabilistic tumor-evolution maps from just two prior glioblastoma MRI scans at arbitrary future timepoints.","lead":"GliomaDiff is a diffusion-based model that, given two earlier MRI scans of a glioma patient, predicts what the next FLAIR scan will look like at any chosen future date and produces a pixel-by-pixel map of where the tumor is likely to grow. It is an attempt to give radiation oncologists an uncertainty-aware forecast for treatment planning, tested on a public glioblastoma dataset and a five-patient local set.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hand-set SDF soft labels make the calibrated-uncertainty claim circular; ECE should be recomputed against binary GTVs before the probability maps are accepted.","rationale":"The reader's weakest assumption is indeed the most load-bearing one. Equation (5) sets the logistic-normalized SDF parameters without an empirical calibration check, and all uncertainty metrics are computed against this same transform, making the calibration claim circular. The central promise of the paper is not merely a future segmentation but a pixel-wise probability map with calibrated uncertainty; if the soft-label target is arbitrary, the ECE, BCE, and KL numbers do not measure clinical calibration. The DSC threshold at P>0.8 is less affected by the transform choice, which is why the spatial prediction results may still be meaningful, but the uncertainty quantification claims cannot be accepted at face value. A secondary concern is that the augmentation-count selection (25 synthetic sequences) is performed on the same 7-patient test cohort used for the headline metrics, which risks optimistic estimates. This does not change the overall verdict: the method is plausible and the private-set evaluation is a useful independent check, but the strong claims about calibrated uncertainty and robust performance should be conditional on an empirical calibration test and a clean validation split. Hence the verdict remains CONDITIONAL, with no adjustment beyond what the reader already recommended.","tokens_in":14740,"tokens_out":5970,"duration_ms":61817,"concrete_test":"Recompute ECE and the reliability diagrams in Table 1 and Fig. 4 using the binary GTV masks at the target timepoint t3 as ground truth (pixel-level observed tumor presence), with patient-level clustering and the same probability bins. If ECE rises materially above the reported 0.010 (e.g., above 0.05) or the reliability curve deviates significantly from the diagonal, the calibrated-uncertainty claim is not established; then re-fit β and μ on a validation subset and repeat the Table 2 ablation to see whether the conclusions survive an unbiased calibration check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines the ground-truth probability map via Eq. (5), a hand-set logistic transform of the SDF with β=0.1 and μ=15, justified only by 'empirical observations and clinical practice' with no quantitative fit shown. Every uncertainty metric then measures agreement against this self-defined target: ECE, BCE, and KL in Tables 1, 2, and 4 and the reliability diagrams in Fig. 4 compare predictions to the logistic-normalized SDF, not to any independent clinical outcome. A model that merely outputs a distance-decayed function of the input mask would appear well calibrated under this protocol. Table 1's ECE drop from 0.016 to 0.010 also conflates a change in evaluation target with a change in model quality: the binary baseline is scored against hard masks while the proposed method is scored against its own soft labels. This is load-bearing because the central claim includes pixel-wise probability maps and calibrated uncertainty; DSC at threshold P>0.8 is comparatively robust to the transform choice, but the uncertainty numbers and the 'likelihood of tumor presence' interpretation are not yet supported. A separate, secondary issue is that the best augmentation count (25 synthetic sequences) is selected on the same 7-patient test cohort used for the headline results, further inflating confidence in the reported numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GliomaDiff, a multi-task diffusion framework that takes two prior longitudinal MRI scans (S_t1, S_t2) with acquisition dates and predicts a future FLAIR scan S_t3 at an arbitrary target timepoint, together with a pixel-wise probability map of tumor evolution. Tumor labels are converted to soft probability maps through a logistic-normalized signed distance field (Eq. 5) with hand-set parameters β=0.1, μ=15, and the framework also integrates a learned deformation module, a diffusion-based data augmentation pipeline, and a radiotherapy-weighted focal loss. The method is trained on the public BGPD cohort (105 patients) and tested on 7 BGPD patients plus a 5-patient private dataset, reporting RMSE, BCE, KL divergence, ECE, Dice, and generative image quality metrics. The central claims are accurate time-agnostic prediction of glioma progression, calibrated uncertainty estimates, and generalization to an independent private set.","tokens_in":14985,"tokens_out":4280,"duration_ms":38340,"significance":"If the claims hold, the paper addresses a clinically relevant and technically challenging problem: forecasting tumor progression from sparse, irregularly collected longitudinal MRI. The multi-task formulation, the use of deformation fields as temporal priors, the augmentation pipeline for data-scarce settings, and the external private evaluation are all valuable contributions. The public BGPD training with a held-out private test set is a genuine strength, and the ablations in Table 2 give a reasonable picture of which components help. However, the current evaluation protocol does not yet support the calibrated-uncertainty claim, because all uncertainty metrics are computed against soft labels manufactured by the authors' own hand-set transform, and the augmentation budget is selected on the same test cohort that produces the headline numbers. These issues are load-bearing for the central claims, so the paper needs a revised validation protocol before it can be accepted.","major_comments":[{"comment":"The ECE, BCE, and KL metrics are all computed against the logistic-normalized signed distance field defined in Eq. (5) with hand-set β=0.1 and μ=15, justified only by 'empirical observations' with no quantitative fit. Because this transform is used as the probabilistic ground truth, the reported calibration improvement (ECE dropping from 0.0163 to 0.0100) measures agreement with a self-defined target rather than with any independent clinical probability. Under this protocol, a model that simply emits a distance-decayed function of the input mask would appear well calibrated. Please recompute ECE and reliability diagrams against binary GTV masks or another independently defined target, and report sensitivity of the uncertainty metrics to β and μ.","section":"Section 4.1, Eq. (5); Tables 1, 2, 4; Fig. 4"},{"comment":"The best synthetic-augmentation budget (25 sequences) is selected on the same 7-patient test cohort used to report the headline results (RMSE 0.048, DSC 0.736, ECE 0.010). This post-hoc selection on the test set inflates the reported gains and means the improvements in Table 2 are not a valid estimate of performance on unseen data. Please move augmentation-budget selection to a held-out validation split or use nested cross-validation, and report final test-set results only for a configuration fixed before looking at the test cohort.","section":"Section 5.4, Table 2"},{"comment":"The generalization claim rests on 7 test patients from BGPD and 5 private patients, and the paper provides no re-implemented baselines trained and evaluated under the same protocol; the only external comparison to [19] is cross-dataset and uses a different number of input scans. Please add at least one re-implemented baseline (for example, a standard conditional diffusion model without the proposed deformation, augmentation, and RT-weighted modules) on the same splits, and report per-patient or bootstrapped confidence intervals for the aggregate metrics instead of treating all evaluation slices as independent samples.","section":"Section 5.4, Table 4"},{"comment":"The binary baseline is scored against hard GTV masks while the probabilistic model is scored against its own logistic-normalized soft labels, so the ECE and CE improvements in Table 1 conflate a change in evaluation target with a change in model quality. Reporting DSC at a threshold P>0.8 chosen to mirror Eq. (5) does not resolve this, because the threshold is derived from the same transform being evaluated. Please report both models on both targets, or at least report DSC at a threshold that is not selected from the soft-label definition.","section":"Section 5.2 and Table 1"}],"minor_comments":[{"comment":"The abstract and Section 5.4 claim forecasting 'up to 12 months ahead,' while the Conclusion states 'reliable forecasts up to six months ahead'; Table 3 shows DSC=0.699 for the >365-day bin. Please align these statements with the actual evidence.","section":"Abstract vs. Conclusion"},{"comment":"There are several typographical errors and inconsistent notations, e.g., 'Comarison' in Table 1, 'boundery' in Eq. (4), 'descrease' in Section 5.4, and 'the the' in Section 4.4. Please proofread carefully.","section":"Throughout"},{"comment":"The table with per-patient results is referenced as 'Table 5.4' in the text but appears to be Table 4; please fix the cross-reference. Also, Figure 9 duplicates Figure 6.","section":"Section 5.4, Table 4"},{"comment":"The augmentation pipeline relies on two diffusion networks from prior work, but no quantitative evaluation of the synthetic sequences is provided; please add at least a small quantitative or qualitative validation of the synthesized follow-ups.","section":"Section 4.4 and 5.3"},{"comment":"The deformation loss notation is unclear: the similarity term is written as a sum over φ1,2 with no explicit per-pixel definition, and λ_smooth appears only in the second term while the first term is unweighted. Please clarify the exact loss expression.","section":"Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant clinical problem and has a useful external evaluation, but the current validation of the uncertainty claims and the test-set-based selection of the augmentation budget are significant concerns. I believe the work is worth a major revision rather than rejection, provided the authors can re-validate the calibration against independent targets and fix the evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a real architectural contribution, but the headline uncertainty-calibration numbers are not measuring what they claim. The model combines a generative FLAIR branch with a predictive probability branch, uses logistic-normalized SDFs for soft labels, adds learned deformation fields between timepoints, and weights the loss with radiotherapy dose maps. That combination is new relative to GP-GAN and the treatment-aware diffusion model, and the ability to query predictions at arbitrary future timepoints is clinically sensible. They also test on an independent private set, which is good practice.\n\nThe soft spots are load-bearing. The ground-truth probability maps come from Eq. (5), a logistic transform of the SDF with β=0.1 and μ=15, justified only by 'empirical observations'. All of the calibration metrics—ECE, BCE, KL—are computed against these self-defined soft labels, not against any independent clinical probability. A model that simply outputs a distance-decayed function of the input mask would look well calibrated under this protocol. The ECE improvement in Table 1 is also apples-to-oranges: the binary baseline is scored against hard masks, the probabilistic method against its own soft labels. The 38% ECE drop is mostly an artifact of changing the target, not a better model. The DSC numbers are more robust because they threshold at the same soft-label boundary, but they still inherit the hand-set transform.\n\nSecondary issues: the best augmentation count (25 synthetic sequences) is selected on the same 7-patient test cohort used for the headline results, which inflates confidence. The test cohort is small—7 patients, 300 slice-level evaluation points—and the private set is only 5 patients. There are no re-implemented baselines; the only comparison is a reported DSC from another paper. Also, the conclusion says 'reliable forecasts up to six months ahead' while the results section claims 12 months; that inconsistency should be fixed.\n\nThat said, the central architectural idea is plausible, and the ablation study is honest about component contributions. The paper is worth reading for anyone working on longitudinal tumor prediction or medical diffusion models, but the uncertainty quantification needs to be re-done against binary GTVs or independent annotations, the augmentation budget needs a validation-based selection, and the test set should be bigger before strong clinical claims are made.\n\nThis deserves a serious referee. Send it to peer review, but expect major revision.","headline":"A genuinely new multitask diffusion architecture for time-agnostic glioma progression prediction, but the calibrated-uncertainty claim rests on hand-built soft labels and a 7-patient test set, so the numbers are less solid than they look.","tokens_in":15549,"tokens_out":2065,"would_cite":true,"duration_ms":19665,"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":"Two prior MRI scans are enough for a diffusion model to predict a glioma's future shape and a pixel-level risk map at any chosen day.","keywords":["Diffusion Models","Longitudinal Follow-up","Multimodal MRI","Tumor Progression Prediction","Uncertainty Quantification","Glioma","Signed Distance Fields","Data Augmentation"],"falsifier":"Compare GliomaDiff's predicted probability maps and expected calibration error against independently produced clinical ground truth, for example expert neuroradiologist delineations of future GTV or biopsy-verified recurrence regions, rather than against the logistic-normalized SDF. If the ECE computed against those independent labels is substantially worse than the reported 0.0100, the calibrated-uncertainty claim is not supported. A simpler check: recompute all metrics on the seven public test patients using hard tumor masks only and see whether the reported Dice gains and calibration improvement persist.","tokens_in":14556,"feed_emoji":"🧠","tokens_out":7412,"duration_ms":60464,"temperature":0.7,"pith_summary":"GliomaDiff predicts how a glioma will evolve from just two earlier MRI scans: it outputs both a synthetic future FLAIR image and a pixel-by-pixel probability of tumor presence at any future date the clinician chooses. The paper's central claim is that this works because tumor boundaries are treated as soft, uncertain regions defined through signed distance fields rather than binary masks, and because a learned deformation field captures how the tumor changes shape between scans. If the claim holds, routine clinical follow-up data—two scans acquired months apart—would be enough to interrogate progression risk at arbitrary future milestones, complete with uncertainty estimates. On a held-out portion of a public glioblastoma dataset the model reaches an average Dice overlap of 0.736, and on an independent five-patient private set 0.759, while calibration error drops by 38% relative to a binary-mask baseline.","feed_headline":"Two MRI scans predict glioma growth up to a year ahead","feed_subtitle":"A multitask diffusion model turns two follow-up scans into a future FLAIR image and a pixel-level tumor risk map.","key_machinery":"The load-bearing object is the logistic-normalized signed distance field (SDF) probability map, $P_t(x)=\\sigma(\\phi(x),\\beta=0.1,\\mu=15)$, which converts a binary Gross Tumor Volume mask into a smooth pixel-wise likelihood of tumor presence. This single representation does three jobs: it is the regression target for the predictive branch, the input that the deformation module warps between timepoints, and the reference against which uncertainty metrics (ECE, BCE, KL) are computed. The second mechanism is the pre-trained tumor evolution module, a U-Net-style network that takes two SDF maps and a time gap and returns a dense 2D displacement field via a spatial transformer, giving the diffusion model a learned prior for how the tumor moves and grows. Around these sit a sinusoidal temporal embedding injected through feature-wise modulation layers, a two-stage diffusion augmentation pipeline (full-sequence synthesis plus missing-modality imputation), and a radiotherapy-weighted focal loss that up-weights pixels inside high-dose regions.","core_discovery":"On its own terms, the paper establishes that glioma progression can be framed as a conditional generation task: given scans $S_{t_1}$ and $S_{t_2}$ with their acquisition dates, GliomaDiff synthesizes the future FLAIR (a fluid-sensitive MRI sequence) at a requested date $d_{t_3}$ and simultaneously regresses a probability map $P_{t_3}$ of tumor presence. The key move is to replace hard tumor masks with logistic-normalized signed distance fields: each pixel's distance to the tumor boundary is pushed through a logistic function ($\\beta=0.1$, $\\mu=15$ mm), so the core stays above 0.8 probability, the edge decays to 0.5 at 15 mm, and beyond that tumor presence is considered unlikely. A pre-trained U-Net deformation module predicts a dense displacement field between the two input SDF maps, conditioned on the time gap, and supplies this as a deformation prior to a shared-encoder, two-decoder diffusion network; a radiotherapy-weighted focal loss biases learning toward high-dose regions. The authors report that the full system beats a probabilistic-mask diffusion baseline on RMSE, KL divergence, and Dice, keeps calibration error low (ECE 0.0100), and remains accurate for intervals up to about a year, with Dice falling from 0.781 at 0–60 days to 0.699 beyond 365 days.","pith_inferences":["The deformation module's SDF displacement field could be reused as a registration or shape-evolution prior for other longitudinal imaging tasks, since it is trained purely on tumor-distance maps and is time-gap conditioned.","The paper's uncertainty claim is only as strong as the SDF assumption; a natural next step is to validate $P_{t_3}$ against expert consensus or recurrence maps, and if that fails, the method may still generate good FLAIR images while its risk maps remain self-referential.","Because the model works on 2D axial slices and discards small-tumor slices, extending to full 3D volumes might change the error profile; testing on isotropic acquisitions would reveal whether the reported Dice holds in three dimensions.","The radiotherapy-weighted focal loss already uses dose maps, so the same architecture could be extended to predict response to a specific planned treatment, such as simulating growth under a proposed dose distribution, rather than only unconditional progression."],"forward_implications":["A clinician with two routine follow-up scans could query the model at any future date and receive a risk map, enabling adaptive surveillance intervals, for example shorter scans when the predicted probability is high.","The soft-label SDF target is a reusable recipe for other tumors or organs where boundary uncertainty matters more than crisp segmentation.","Synthetic sequence augmentation helps only up to a point: 25 synthetic follow-ups improve Dice from 0.714 to 0.736 on the public test set, while 50 or 100 degrade it, so augmentation ratios must be tuned rather than maximized.","The temporal behavior of the model gives a practical operating envelope: reliable up to roughly a year, with Dice above 0.73 through 365 days and declining beyond."],"supporting_citations":[{"why":"Supplies the public training and test cohort of glioblastoma patients with T1CE and FLAIR follow-up scans.","marker":"[32]"},{"why":"Provides the diffusion forward/backward formulation and simplified denoising loss that the generative branch is built on.","marker":"[20]"},{"why":"Supplies the U-Net deformation architecture used for the tumor evolution module's dense displacement field prediction.","marker":"[27]"},{"why":"The closest comparable treatment-aware diffusion model, reported at DSC 0.719 with an extra input scan, serving as the comparison baseline.","marker":"[19]"},{"why":"Provides the automatic segmentation framework used to generate GTV masks when manual annotations are missing.","marker":"[33]"},{"why":"Clinical guideline defining the 15 mm CTV safety margin that justifies the $\\mu=15$ parameter in the logistic SDF.","marker":"[25]"},{"why":"Origin of the focal loss formulation adapted into the radiotherapy-weighted loss term.","marker":"[29]"}],"fun_headline_variants":["Two MRI scans yield a year of glioma risk maps","Multitask diffusion predicts glioma growth from two MRIs","Two scans, one diffusion model: future glioma maps","Pixel-level tumor risk from time-gapped MRI pairs","Glioma progression forecast via multitask diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The soft probability maps used as ground truth are generated by the authors' own logistic-normalized signed distance transform with hand-set parameters ($\\beta=0.1$, $\\mu=15$), and every calibration or uncertainty metric is measured against these synthetic labels; if that transform does not match real spatial uncertainty of tumor spread, the claim of calibrated uncertainty is only self-consistency.","fun_headline_variants_meta":{"raw":{"variants":["Two MRI scans yield a year of glioma risk maps","Multitask diffusion predicts glioma growth from two MRIs","Two scans, one diffusion model: future glioma maps","Pixel-level tumor risk from time-gapped MRI pairs","Glioma progression forecast via multitask diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000808,"raw_usage":{"total_tokens":3616,"prompt_tokens":1084,"completion_tokens":2532,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":2457}},"tokens_in":700,"tokens_out":2532,"duration_ms":15714,"temperature":1.0,"reasoning_tokens":2457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:52:30.970323+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare GliomaDiff's predicted probability maps and expected calibration error against independently produced clinical ground truth, for example expert neuroradiologist delineations of future GTV or biopsy-verified recurrence regions, rather than against the logistic-normalized SDF. If the ECE computed against those independent labels is substantially worse than the reported 0.0100, the calibrated-uncertainty claim is not supported. A simpler check: recompute all metrics on the seven public test patients using hard tumor masks only and see whether the reported Dice gains and calibration improvement persist.","supporting_citations":[{"cited_title":"Burdenko’s glioblastoma progression dataset (Burdenko- GBM-Progression), 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the public training and test cohort of glioblastoma patients with T1CE and FLAIR follow-up scans."},{"cited_title":"V oxelmorph: a learn- ing framework for deformable medical image registration","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net deformation architecture used for the tumor evolution module's dense displacement field prediction."},{"cited_title":"Treatment-aware diffusion probabilistic model for longitudinal mri generation and diffuse glioma growth prediction.IEEE Transactions on Medical Imaging, 2025","cited_arxiv_id":null,"evidence_quote":"The closest comparable treatment-aware diffusion model, reported at DSC 0.719 with an extra input scan, serving as the comparison baseline."},{"cited_title":"Estro-eano guideline on tar- get delineation and radiotherapy for idh-mutant who cns grade 2 and 3 diffuse glioma.Radiotherapy and Oncol- ogy, 202:110594, 2025","cited_arxiv_id":null,"evidence_quote":"Clinical guideline defining the 15 mm CTV safety margin that justifies the $\\mu=15$ parameter in the logistic SDF."}],"review_version":1}