{"id":"f0052876-d270-4a8d-b4f1-0dc59dff32c3","arxiv_id":"2412.13811","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A lightweight smoothness-regularized optimization of 3D tumor concentration predicts glioblastoma recurrence better than the 15 mm standard margin and runs in under a minute.","lead":"The authors show that a simple optimization approach, fitting a 3D tumor-cell concentration map to MRI segmentations while enforcing smoothness, predicts glioblastoma recurrence better than the standard 15 mm radiotherapy margin and several complex simulation baselines. It cuts runtime from about 30 minutes to under one minute, making patient-specific radiotherapy planning more practical in clinics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparison uses 'Ours Best Thresholds' selected on the test set; when thresholds are treated as fixed or chosen without seeing recurrence labels, the claimed advantage over Static Grid may shrink or disappear.","rationale":"The paper is a useful empirical contribution: it is simple, fast, has a public code link, and reports comparisons on two datasets. The reader's CONDITIONAL verdict is appropriate. The most load-bearing concern I see is not the recurrence-correlation assumption itself, but the way the headline comparison is selected. The 'Ours Best Thresholds' row is the best of 20 threshold combinations evaluated on the test data. On GliODIL the best-vs-static-grid gap is about 4.7 points, but the worst-vs-best spread is about 2.8 points, and on RHUH the spread (2.9 points) exceeds the best-vs-baseline advantage (2.2 points). This means the ordering of methods can flip depending on which threshold is used, so the claimed superiority is not a property of the framework alone; it is a property of the framework plus a favorable test-set threshold choice. The reader's rationale notes this issue, although the stated weakest assumption is the recurrence-correlation assumption; hence partial agreement. A pre-registered threshold selection or a corrected multiple-comparison test would settle whether the central claim survives. If the advantage survives such a test, the paper's main message is preserved; if not, the abstract should be weakened to 'comparable or sometimes better under favorable threshold settings'. This is a conditional concern, not a rejection: the code and speed are real, and the evaluation protocol is otherwise paired and reasonable. No ad hominem or fatal language is intended; the concern is about the statistical procedure, not the authors' integrity.","tokens_in":7457,"tokens_out":8772,"duration_ms":90043,"concrete_test":"Pre-register the thresholds: split GliODIL into training (e.g., 2/3) and test (1/3); choose tauCore and tauEdema maximizing any-recurrence coverage on the training split only; evaluate the fixed model on the held-out split and on RHUH. Repeat across multiple splits or bootstrap samples and compare the mean difference vs Static Grid with a paired test. In parallel, compute a max-statistic permutation test over the 20 thresholds on GliODIL: permute patient-level recurrence outcomes, recompute the maximum over thresholds of the Wilcoxon statistic against Static Grid, and derive a corrected p-value. If the adjusted p-value for Ours Best vs Static Grid is not <0.05, or if the fixed-threshold gap on the held-out split is less than the reported 4.68 points, the headline superiority is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that the proposed optimization significantly improves recurrence coverage over Static Grid Discretization. The number used for the headline, 'Ours Best Thresholds' (72.48% vs 67.80% for any recurrence on GliODIL), is the maximum over a 20-point sweep of tauCore x tauEdema evaluated on the same patients used to report the result. Table 1 also reports 'Ours Worst Thresholds' (69.72%) and 'Ours Median Thresholds' (70.93%), so the sweep is summarized by test-set order statistics. If the thresholds were chosen after looking at recurrence labels, the best-of-20 value overstates the expected performance of a fixed model. On the independent RHUH dataset, 'Ours Worst Thresholds' is 70.27% vs Static Grid 70.95%, so the advantage over the strongest baseline is not robust to threshold choice; only 'Ours Best' (73.18%) exceeds it. The paper does not state whether 'Ours Best' on RHUH is the threshold that was optimal on GliODIL or whether it was re-selected on RHUH; if the latter, the independent validation is itself test-set selected. Moreover, the enhancing-recurrence column in Table 1 shows no significant improvement for the default optimizer (Ours Best 85.19 vs Static Grid 84.42, no dagger), and the text says the wave-loss variant is needed to obtain a significant enhancing-core improvement. The large spread between worst and best thresholds on RHUH (2.91 points) exceeds the gap between Ours Best and Static Grid (2.23 points), so the threshold choice materially affects the conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a lightweight optimization framework that estimates a 3D tumor-cell concentration field by fitting a scalar concentration to MRI-derived tumor segmentations (core, edema) while penalizing the Dirichlet energy of the field. The estimated concentration is then thresholded to a fixed volume to define a predicted clinical target volume, and the method is evaluated by recurrence coverage on two public glioblastoma datasets (GliODIL, 152 patients; RHUH, 40 patients). The authors report significant improvements over the clinical standard plan and over a static-grid discretization baseline for 'any recurrence' coverage, with runtime reduced from about 30 minutes to under one minute. The paper also demonstrates extensibility to PET imaging and to an additional wave-loss constraint derived from a traveling-wave solution of the Fisher-Kolmogorov equation.","tokens_in":7801,"tokens_out":6969,"duration_ms":61571,"significance":"If the result holds, the framework is practically attractive because it offers a simple, fast, and flexible alternative to PDE-based and learning-based tumor-growth models, with genuine clinical potential for individualizing radiotherapy volumes. Strengths of the paper include the use of two public datasets, paired nonparametric statistical testing, an external validation cohort, ablations for PET and wave losses, and publicly available code. The runtime reduction from 30 minutes to under one minute is a substantive practical contribution. However, the headline evaluation is currently weakened by test-set-dependent threshold selection, and some performance claims are broader than the reported results support.","major_comments":[{"comment":"The headline results in Table 1 are reported as 'Ours Best Thresholds', 'Ours Median Thresholds', and 'Ours Worst Thresholds' over the 20-point grid of (τCore, τEdema) shown in Figure 2c. Because the thresholds are selected after examining the recurrence-coverage outcome on the same 152 GliODIL patients, the 'Best' value (72.48% any recurrence) is an optimistically biased estimate of a fixed model's performance, and the associated significance marks (‡, ∗∗) are not valid for that row. The same issue applies to Table 2a for RHUH: it is not stated whether the reported Best/Median/Worst thresholds were fixed in advance from GliODIL or re-selected on RHUH; if re-selected, the independent validation inherits the same bias. Please provide a prespecified threshold-selection rule (e.g., choose thresholds on a training subset or fix them a priori at clinically motivated values) and report test-set performance for that fixed configuration, in addition to the sweep statistics.","section":"Section 3, Table 2b, Table 1"},{"comment":"The abstract and Section 3 claim that the proposed framework 'outperforms all existing methods' and achieves 'superior tumor recurrence prediction'. This claim is too broad. In Table 2b (PET subset, 58 patients), Dynamic Grid Discretization [2] achieves higher contrast-enhancing recurrence coverage (89.9%) than both Ours (88.2%) and Ours with additional PET loss (88.4%), and the text in Section 3 acknowledges this. On GliODIL (Table 1), the enhancing-core improvement of Ours Best (85.19%) over Static Grid (84.42%) is not statistically significant (no dagger). Please restrict the performance claim to the 'any recurrence' metric and to settings where the difference is significant, or provide a discussion of why the enhancing-core deficit is acceptable for the intended clinical use.","section":"Section 3, Table 2b, Table 1"},{"comment":"The paper uses paired Wilcoxon tests across many threshold configurations and multiple outcome metrics (any recurrence, enhancing core; GliODIL, RHUH, PET subset) without any correction for multiple comparisons. The selection of the 'Best' threshold among 20 configurations makes the problem worse. The p-values reported in Table 1 and Table 2 are therefore unadjusted and potentially anti-conservative. Please state the total number of comparisons performed and either apply a multiplicity correction (e.g., Benjamini-Hochberg) or clearly argue why the comparisons are considered pre-planned with a single primary outcome.","section":"Section 3 (Evaluation/Statistics)"},{"comment":"The evaluation is built on the assumption that 'post-operative tumor recurrence is correlated with preoperative tumor cell concentration'. This assumption is stated in Section 2 but is not supported by a citation or a sensitivity analysis in the paper. Since all clinical-utility conclusions follow from this correlation, the paper should either cite direct evidence for the correlation in the context of pre-treatment imaging or analyze how the reported coverage results change under alternative assumptions (e.g., using only enhancing recurrence as the outcome, or varying the volume-match protocol). At minimum, the dependence of the conclusion on this assumption should be explicitly acknowledged in the Conclusion.","section":"Section 2, Evaluation"}],"minor_comments":[{"comment":"The row label 'Ours with Wave Loss Thresholds' is missing the word 'Best'; it should read 'Ours with Wave Loss Best Thresholds' for consistency with the other rows.","section":"Table 2c"},{"comment":"The traveling-wave solution is a one-dimensional object; the paper should state explicitly that the wave loss assumes the concentration gradient is dominated by a single direction of propagation, which is an approximation in 3D.","section":"Section 2, Eq. (2)"},{"comment":"The caption should clarify that the color plots show improvement in coverage over the standard plan (not absolute coverage), and the axes currently read 'Edema Threshold' and 'Core Threshold' without units; adding the τ ranges would improve readability.","section":"Figure 2c"},{"comment":"The captions do not define the symbols ∗, ∗∗, †, ‡; the definitions appear only in the text of Section 3. Please repeat them in the captions or provide a common notation table.","section":"Table 1 and Table 2"},{"comment":"Reference [2] is cited as 'a recent NeurIPS publication'; please add the year (2024) and the conference abbreviation for completeness.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The baseline methods compared in this paper (Static Grid Discretization, Dynamic Grid Discretization, NN unconstrained/constrained) originate largely from the same research group as the current submission. This is not a reason to reject, but the editor may wish to consider whether the novelty claims relative to [2,3] are sufficiently differentiated. The central methodological concern is the test-set-dependent threshold selection; if the authors can provide a prespecified threshold protocol or show that the qualitative conclusions survive under worst-case threshold choice on both datasets, the paper would be considerably stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this is a genuinely simple and fast method that gets recurrence coverage roughly on par with or somewhat better than a 30-minute PDE-constrained baseline, but the abstract oversells it. The 'Best Thresholds' numbers are best-of-grid on the test set, and the independent validation is only significantly positive for the best threshold.\n\nWhat's new: they replace the full reaction-diffusion simulation with a static Dice fit plus a Dirichlet energy smoothness penalty. That's a legitimate simplification, and it cuts runtime from 30 minutes to under a minute. On 152 GliODIL patients the median-threshold result (70.93% any recurrence) beats Static Grid (67.80%), and the worst-threshold result (69.72%) still beats it, so the main comparison is not entirely an artifact of cherry-picking. The wave-loss extension gives a significant enhancing-core improvement, and the PET extension is a nice demonstration of modularity.\n\nThe soft spots are real. First, the abstract says 'outperforming state-of-the-art baselines' without acknowledging that the headline numbers are selected from a threshold sweep on the same patients used for evaluation. On RHUH, the worst threshold (70.27%) is actually below Static Grid (70.95%), and only the best threshold (73.18%) carries a significance mark. So the external validation is thinner than it looks. Second, in the PET subset, the enhancing-core coverage is still below Dynamic Grid (88.4 vs 89.9), which the paper admits but the abstract's blanket claim does not. Third, the entire evaluation rests on the assumption that post-op recurrence correlates with pre-op concentration; that's clinically reasonable but unvalidated.\n\nThe paper is worth a serious referee. The threshold-selection problem is fixable: either fix thresholds a priori, use a validation split, or report the full sweep and interpret the median as the expected default. As written, it's a conditional accept with a requested revision to tone down the abstract and to clarify threshold selection on RHUH. I'd bring it to a reading group and would cite the method if I needed a fast baseline for concentration estimation.\n\nRecommendation: send to peer review, with a note flagging the threshold-selection issue.","headline":"A fast, simple tumor-concentration estimator that beats a 30-minute PDE baseline on median thresholds, but the abstract's 'best thresholds' are test-set selected and the external validation is thinner than it looks.","tokens_in":8339,"tokens_out":2862,"would_cite":true,"duration_ms":26650,"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":"This paper claims that a lightweight optimization of a 3D tumor-cell concentration field, regularized for smoothness, predicts glioblastoma recurrence better than the clinical 15 mm margin and state-of-the-art simulation baselines while…","keywords":["glioblastoma","tumor infiltration","tumor recurrence prediction","radiotherapy target volume","Dirichlet energy regularization","reaction-diffusion model","MRI segmentation","optimization framework"],"falsifier":"On a held-out cohort with registered recurrence masks, compute the recurrence coverage of a same-volume thresholded concentration map against the 15 mm standard plan; if the paired difference is not significant under the Wilcoxon test the paper itself uses, or if predicted concentration does not correlate with recurrence voxels, the core claim is refuted.","tokens_in":7270,"feed_emoji":"🧠","tokens_out":4988,"duration_ms":43336,"temperature":0.7,"pith_summary":"This paper tries to establish that the hidden microscopic spread of glioblastoma can be estimated directly, patient by patient, by optimizing a 3D tumor-cell concentration field against standard MRI tumor segmentations while penalizing roughness of the field. The authors argue that this simple, physics-inspired smoothing replaces the need for expensive reaction-diffusion simulations or learned inverse solvers. On 192 patients across two public datasets, the method predicts where tumor will recur better than the current clinical standard of a uniform 15 mm margin and better than state-of-the-art baselines, while cutting runtime from about 30 minutes to under one minute. If correct, this makes personalized radiotherapy target volumes clinically practical without specialized compute or complex biophysical models.","feed_headline":"Optimizer predicts tumor recurrence better than 30-minute models","feed_subtitle":"A concentration-mapping method beats the standard 15 mm margin and fastest baselines on 192 patients, in under a minute.","key_machinery":"The central object is a 3D scalar field of tumor-cell concentration over the brain, optimized directly with the Adam optimizer. The objective combines a data term that uses Dice scores between thresholded concentration and BraTS.Toolkit segmentations of core and edema with the Dirichlet energy $L_{\\mathrm{DE}} = \\frac{1}{N}\\sum_{n=1}^N |\\nabla \\hat{c}_n|^2$, which penalizes gradient magnitude and enforces a smooth, physically plausible concentration landscape. The optional wave loss enforces the gradient-concentration relation $|\\nabla \\hat{c}| = k\\,\\hat{c}(1-\\hat{c})$ derived from the traveling-wave solution of the Fisher-Kolmogorov equation, and the optional PET loss adds correlation with amino-acid PET intensity. The optimizer runs 500 steps in under a minute on a GPU.","core_discovery":"The paper's central claim is that the hidden infiltration boundary needed for radiotherapy can be estimated by direct optimization of a per-voxel tumor concentration against MRI segmentations, without simulating tumor growth over time. On the GliODIL dataset (152 patients), the best threshold configuration reaches 72.48% ± 1.99% coverage of any recurrence, compared with 67.80% ± 2.09% for the best baseline (static grid discretization) and 63.59% ± 2.26% for the standard plan, all at matched irradiated volume; on the independent RHUH dataset (40 patients), the method again outperforms both. The same framework accepts extra loss terms: adding PET correlation raises any-recurrence coverage to 77.4% ± 3.4% on the 58-patient PET subset, and adding a Fisher-Kolmogorov wave constraint improves robustness across threshold choices. The paper concludes that simpler, efficient optimization can effectively address tumor recurrence prediction, challenging the demand for complex simulations.","pith_inferences":["The success of a pure smoothness penalty suggests that much of the predictive information for recurrence may be contained in the spatial envelope of the preoperative MRI abnormality, not in the detailed growth dynamics; this is an implication the paper does not fully explore.","A direct comparison of dosimetric consequences, such as dose to healthy tissue under a concentration-derived target versus the 15 mm margin, would be a natural next test; the paper only measures geometric coverage, not delivered dose.","If the recurrence-concentration correlation holds generally, the same optimization might apply to other infiltrating tumors, but the assumed growth law and visibility thresholds would need re-validation.","The threshold sweep shows results are fairly stable, which could indicate robustness to segmentation errors; this could be tested directly by perturbing input segmentations."],"forward_implications":["If correct, individualized radiotherapy target volumes can be derived from a routine preoperative MRI in under a minute, replacing the uniform 15 mm margin.","At equal irradiated volume, the method covers more of the actual recurrence region on both datasets, meaning the same planned dose is placed where recurrence actually happens more often.","Because the framework accepts additional loss terms, centers with PET or other modalities can incorporate them without re-architecting the method; the PET extension shows further gains.","The wave-loss version is more stable across threshold choices, which the paper reads as a safety property for clinical deployment.","The speed and simplicity make the approach feasible for large retrospective validation and prospective planning studies."],"supporting_citations":[{"why":"Supplies the Static Grid Discretization baseline and the GliODIL dataset used for the main evaluation.","marker":"[3]"},{"why":"Supplies the Dynamic Grid Discretization method with PET, the comparison target for the PET extension.","marker":"[2]"},{"why":"Defines the clinical 15 mm margin standard plan that the method must beat.","marker":"[12]"},{"why":"Provides the visibility thresholds tau_Core and tau_Edema and the personalized fitting approach the thresholds are sampled from.","marker":"[11]"},{"why":"Supplies the physics-constrained neural network baseline.","marker":"[7]"},{"why":"Supplies the numerical physics simulations baseline with 2-hour runtime.","marker":"[16]"},{"why":"Supplies the independent RHUH dataset for external validation.","marker":"[5]"},{"why":"Provides the optimized registration settings used to align recurrence images into preoperative space.","marker":"[1]"},{"why":"Provides the Fisher-Kolmogorov traveling-wave solution used to derive the wave loss.","marker":"[15]"},{"why":"Provides the BraTS.Toolkit segmentations used to define core and edema data losses.","marker":"[8]"}],"fun_headline_variants":["One-minute optimizer beats standard margin for tumor recurrence","Fast optimizer maps hidden tumor spread in under a minute","Lightweight method beats 30-minute models for tumor recurrence","Optimizer cuts runtime to 1 minute, improves recurrence prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All conclusions about clinical utility depend on the assumption that postoperative tumor recurrence is correlated with preoperative tumor-cell concentration; the paper states this assumption but does not independently validate it.","fun_headline_variants_meta":{"raw":{"variants":["One-minute optimizer beats standard margin for tumor recurrence","Fast optimizer maps hidden tumor spread in under a minute","Lightweight method beats 30-minute models for tumor recurrence","Optimizer cuts runtime to 1 minute, improves recurrence prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000906,"raw_usage":{"total_tokens":3886,"prompt_tokens":927,"completion_tokens":2959,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":2894}},"tokens_in":543,"tokens_out":2959,"duration_ms":21477,"temperature":1.0,"reasoning_tokens":2894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:45:27.941373+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out cohort with registered recurrence masks, compute the recurrence coverage of a same-volume thresholded concentration map against the 15 mm standard plan; if the paired difference is not significant under the Wilcoxon test the paper itself uses, or if predicted concentration does not correlate with recurrence voxels, the core claim is refuted.","supporting_citations":[{"cited_title":"arXiv preprint arXiv:2312.05063 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the Static Grid Discretization baseline and the GliODIL dataset used for the main evaluation."},{"cited_title":"In: The Thirty-eighth Annual Conference on Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Supplies the Dynamic Grid Discretization method with PET, the comparison target for the PET extension."},{"cited_title":"Radiotherapy and Oncology184, 109663 (2023) Estimating 3D Brain Tumor Infiltration 11","cited_arxiv_id":null,"evidence_quote":"Defines the clinical 15 mm margin standard plan that the method must beat."},{"cited_title":"IEEE transactions on medical imaging38(8), 1875–1884 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the visibility thresholds tau_Core and tau_Edema and the personalized fitting approach the thresholds are sampled from."},{"cited_title":"Medical Image Analysis83, 102672 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the physics-constrained neural network baseline."},{"cited_title":"Data in Brief50, 109617 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the independent RHUH dataset for external validation."},{"cited_title":"The Brain Tumor Sequence Registration (BraTS-Reg) Challenge: Establishing Correspondence Between Pre-Operative and Follow-up MRI Scans of Diffuse Glioma Patients","cited_arxiv_id":"2112.06979","evidence_quote":"Provides the optimized registration settings used to align recurrence images into preoperative space."},{"cited_title":"British journal of cancer98(1), 113–119 (2008)","cited_arxiv_id":null,"evidence_quote":"Provides the Fisher-Kolmogorov traveling-wave solution used to derive the wave loss."},{"cited_title":"Frontiers in neuroscience p","cited_arxiv_id":null,"evidence_quote":"Provides the BraTS.Toolkit segmentations used to define core and edema data losses."}],"review_version":1}