Pith. sign in

REVIEW 4 major objections 5 minor 17 references

A Lightweight Optimization Framework for Estimating 3D Brain Tumor Infiltration

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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…

desk verdict 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. read the letter →

arxiv 2412.13811 v2 pith:6KXUJFT4 submitted 2024-12-18 physics.med-ph cs.CV

Jonas Weidner , Michal Balcerak , Ivan Ezhov , André Datchev , Laurin Lux , Lucas Zimmer , Daniel Rueckert , Björn Menze
show 1 more author
Benedikt Wiestler
This is my paper · ORCID
classification physics.med-phcs.CV
keywords glioblastomatumorinfiltrationrecurrencepredictionradiotherapytargetvolumeDirichletenergyregularizationreaction-diffusionmodelMRIsegmentationoptimizationframework
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Section 3, Table 2b, Table 1] 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.
  2. [Section 3, Table 2b, Table 1] 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.
  3. [Section 3 (Evaluation/Statistics)] 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.
  4. [Section 2, Evaluation] 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.
minor comments (5)
  1. [Table 2c] 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.
  2. [Section 2, Eq. (2)] 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.
  3. [Figure 2c] 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.
  4. [Table 1 and Table 2] 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.
  5. [References] Reference [2] is cited as 'a recent NeurIPS publication'; please add the year (2024) and the conference abbreviation for completeness.

Circularity Check

1 steps flagged · score 3.0 of 10

No derivation-level circularity; one evaluation-level circularity: the headline 'Ours Best Thresholds' is selected on the recurrence labels it then predicts.

  1. fitted input called prediction [Section 3 Results, Figure 2c and Table 1]
    "As it is unclear which tumor concentrations are visible in MR images, we tested an extensive, medically plausible range of thresholds τCore and τEdema. ... For the "any recurrence" prediction, we find an optimal outcome at (τCore = 0.6, τEdema = 0.2)."

    The operating point in the headline result is chosen by maximizing the same recurrence-coverage metric on which the method is then evaluated; "Ours Best Thresholds" is the maximum over the 20-point threshold sweep, not the performance of a fixed, pre-specified model. On the independent RHUH dataset, the worst threshold choice (70.27%) falls below Static Grid (70.95%), so the claimed advantage exists only for the test-set-selected optimum; the reported "best" is an upper-envelope statistic rather than an independent prediction.

full rationale

The core concentration estimation is not circular: LTotal optimizes a concentration field against preoperative MRI segmentations using Dice losses plus a Dirichlet-energy smoothness regularizer, and recurrence labels appear only in evaluation, not in the optimized objective. The wave-loss constraint is derived from a standard sigmoid traveling-wave solution of the Fisher-Kolmogorov equation with an external citation, so it is not a self-referential input. The comparison baselines [2,3,16] are largely prior work by the same group, but they are run as concrete algorithms on public datasets, so the self-citations are not load-bearing. The one genuine circular element is threshold selection: the paper sweeps 20 (τCore, τEdema) pairs and reports "Ours Best Thresholds" as the headline, with the optimum chosen using the same recurrence labels used for the comparison. This makes the headline number a best-of-sweep test-set statistic. The reporting of worst and median thresholds is transparent and partly mitigates the issue, and on GliODIL even the worst and median thresholds beat the Static Grid baseline, so the central claim retains independent content. On RHUH the worst threshold set falls below Static Grid, showing the headline advantage is partly a selection artifact. Score 3 reflects one evaluation-level circularity without derivation-level circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim is empirical, so the ledger consists mainly of domain assumptions and hand-tuned hyperparameters. No new physical entities or dimensions are introduced.

free parameters (5)
  • τCore = 0.6 (best for any recurrence)
    Core visibility threshold; sampled from {0.6, 0.7, 0.8, 0.9} and best selected on test data.
  • τEdema = 0.2 (best for any recurrence)
    Edema visibility threshold; sampled from {0.1, 0.2, 0.3, 0.4, 0.5} and best selected on test data.
  • Loss weights (λData, λDE, λWave, λPET, αCore, αEdema) = 1, 1000, 1000, 1, 1, 1
    Chosen empirically so loss terms are of similar order of magnitude; not tuned on a separate validation set.
  • k (wave slope) = 0.1 mm^-1
    Initial slope for the wave loss; not fitted to recurrence data.
  • Initial concentration offsets = 0.01
    Small initialization values for core, edema, and brain tissue to avoid zero gradients.
assumptions (6)
  • domain assumption Post-operative tumor recurrence is correlated with preoperative tumor cell concentration.
    Stated in Sec. 2 Evaluation; the entire recurrence-coverage metric depends on this correlation.
  • domain assumption BraTS Toolkit segmentations of enhancing tumor, necrosis, and edema are reliable proxies for visible tumor.
    Used to define the data loss and the standard plan; segmentation errors propagate directly.
  • domain assumption Deformable registration with ANTs and BraTSReg settings accurately aligns recurrence MRI to preoperative space.
    Required for measuring recurrence coverage; registration errors bias the evaluation.
  • domain assumption The Fisher-Kolmogorov equation and its traveling-wave sigmoid solution describe tumor growth locally.
    Used to derive the optional wave loss; the assumption of local sigmoid behavior is asserted without independent validation.
  • domain assumption A smooth concentration field, enforced by Dirichlet energy regularization, is physically plausible for tumor infiltration.
    The core regularization is motivated by reaction-diffusion PDEs, but the equivalence is heuristic.
  • domain assumption The synthetic standard plan (tumor core volume plus 15 mm margin) approximates the clinical radiotherapy protocol.
    Used as the baseline; any mismatch with real clinical margins affects the comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Lightweight Optimization Framework for Estimating 3D Brain Tumor Infiltration." pith.science (2026). https://pith.science/paper/6KXUJFT4

@misc{pith2026241213811,
  author       = {Pith},
  title        = {Pith review of: A Lightweight Optimization Framework for Estimating 3D Brain Tumor Infiltration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6KXUJFT4}},
  note         = {Machine review of arXiv:2412.13811}
}
read the original abstract

Glioblastoma, the most aggressive primary brain tumor, poses a severe clinical challenge due to its diffuse microscopic infiltration, which remains largely undetected on standard MRI. As a result, current radiotherapy planning employs a uniform 15 mm margin around the resection cavity, failing to capture patient-specific tumor spread. Tumor growth modeling offers a promising approach to reveal this hidden infiltration. However, methods based on partial differential equations or physics-informed neural networks tend to be computationally intensive or overly constrained, limiting their clinical adaptability to individual patients. In this work, we propose a lightweight, rapid, and robust optimization framework that estimates the 3D tumor concentration by fitting it to MRI tumor segmentations while enforcing a smooth concentration landscape. This approach achieves superior tumor recurrence prediction on 192 brain tumor patients across two public datasets, outperforming state-of-the-art baselines while reducing runtime from 30 minutes to less than one minute. Furthermore, we demonstrate the framework's versatility and adaptability by showing its ability to seamlessly integrate additional imaging modalities or physical constraints.

Figures

Figures reproduced from arXiv: 2412.13811 by the authors.

Figure 1
Figure 1. We optimize (blue) a 3D scalar tumor concentration estimation (yellow) by simultaneously fitting the data while ensuring a smooth concentration landscape by minimizing the Dirichlet energy. Using this predicted tumor concentration (orange), we propose a radiotherapy plan (Clinical Target Volume (CTV), orange). We evaluate (green) our method’s ability to capture areas of subsequent tumor recurrence. Tumor growth mode… view at source ↗
Figure 2
Figure 2. a.) Demonstration of our method on example patients. In the first row, we show the two input MR images with the tumor and the recurrence that should be covered. Edema is shown in blue, enhancing tumor in green, and necrotic in red. Our method predicts a continuous estimation of tumor cells, as shown in the second row. This continuous concentration is thresholded to have the same volume as the standard plan (grey) to… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 15 canonical work pages

  1. [2]

    In: The Thirty-eighth Annual Conference on Neural Information Processing Systems

    Balcerak, M., Amiranashvili, T., Wagner, A., Weidner, J., Karnakov, P., Paet- zold, J.C., Ezhov, I., Koumoutsakos, P., Wiestler, B., et al.: Physics-regularized multi-modal image assimilation for brain tumor localization. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page

  2. [1]

    The Brain Tumor Sequence Registration (BraTS-Reg) Challenge: Establishing Correspondence Between Pre-Operative and Follow-up MRI Scans of Diffuse Glioma Patients

    Baheti,B.,Waldmannstetter,D.,Chakrabarty,S.,Akbari,H.,Bilello,M.,Wiestler, B., Schwarting, J., Calabrese, E., Rudie, J., Abidi, S., et al.: The brain tumor sequence registration challenge: establishing correspondence between pre-operative and follow-up mri scans of diffuse glioma patients. arXiv preprint arXiv:2112.06979 (2021)

    work page Pith review arXiv 2021

  3. [3]

    arXiv preprint arXiv:2312.05063 (2023)

    Balcerak, M., Ezhov, I., Karnakov, P., Litvinov, S., Koumoutsakos, P., Weidner, J., Zhang, R.Z., Lowengrub, J.S., Wiestler, B., Menze, B.: Individualizing glioma radiotherapy planning by optimization of a data and physics informed discrete loss. arXiv preprint arXiv:2312.05063 (2023)

    work page arXiv 2023

  4. [4]

    Physics in Medicine & Biology67(15), 155014 (2022)

    Bortfeld, T., Buti, G.: Modeling the propagation of tumor fronts with shortest path and diffusion models—implications for the definition of the clinical target volume. Physics in Medicine & Biology67(15), 155014 (2022)

    work page 2022

  5. [5]

    Data in Brief50, 109617 (2023)

    Cepeda, S., García-García, S., Arrese, I., Herrero, F., Escudero, T., Zamora, T., Sarabia, R.: The río hortega university hospital glioblastoma dataset: A compre- hensive collection of preoperative, early postoperative and recurrence mri scans (rhuh-gbm). Data in Brief50, 109617 (2023)

    work page 2023

  6. [6]

    Dice,L.R.:Measuresoftheamountofecologicassociationbetweenspecies.Ecology 26(3), 297–302 (1945)

    1945

  7. [7]

    Medical Image Analysis83, 102672 (2023)

    Ezhov, I., Scibilia, K., Franitza, K., Steinbauer, F., Shit, S., Zimmer, L., Lipkova, J., Kofler, F., Paetzold, J.C., Canalini, L., et al.: Learn-morph-infer: a new way of solving the inverse problem for brain tumor modeling. Medical Image Analysis83, 102672 (2023)

    work page 2023

  8. [8]

    Frontiers in neuroscience p

    Kofler, F., Berger, C., Waldmannstetter, D., Lipkova, J., Ezhov, I., Tetteh, G., Kirschke, J., Zimmer, C., Wiestler, B., Menze, B.H.: Brats toolkit: translating brats brain tumor segmentation algorithms into clinical and scientific practice. Frontiers in neuroscience p. 125 (2020)

    work page 2020

Show all 17 references
  1. [9]

    In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2006: 9th International Conference, Copenhagen, Denmark, October 1-6, 2006

    Konukoğlu, E., Clatz, O., Bondiau, P.Y., Delingette, H., Ayache, N.: Extrapo- lating tumor invasion margins for physiologically determined radiotherapy regions. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2006: 9th International Conference, Copenhagen...

    2006

  2. [10]

    EJNMMI research11, 1–6 (2021)

    Liesche-Starnecker, F., Prokop, G., Yakushev, I., Preibisch, C., Delbridge, C., Meyer, H.S., Aftahy, K., Barz, M., Meyer, B., Zimmer, C., et al.: Visualizing cellularity and angiogenesis in newly-diagnosed glioblastoma with diffusion and perfusion mri and fet-pet imaging. EJNM...

    2021

  3. [11]

    IEEE transactions on medical imaging38(8), 1875–1884 (2019)

    Lipkova, J., Angelikopoulos, P., Wu, S., Alberts, E., Wiestler, B., Diehl, C., Preibisch, C., Pyka, T., Combs, S.E., Hadjidoukas, P., et al.: Personalized radio- therapy design for glioblastoma: integrating mathematical tumor models, multi- modal scans, and bayesian inference....

    2019

  4. [12]

    Radiotherapy and Oncology184, 109663 (2023) Estimating 3D Brain Tumor Infiltration 11

    Niyazi, M., Andratschke, N., Bendszus, M., Chalmers, A.J., Erridge, S.C., Galldiks, N., Lagerwaard, F.J., Navarria, P., af Rosenschöld, P.M., Ricardi, U., et al.: Estro- eano guideline on target delineation and radiotherapy details for glioblastoma. Radiotherapy and Oncology18...

    2023

  5. [13]

    Bulletin of mathematical biology76, 2306–2333 (2014)

    Saut, O., Lagaert, J.B., Colin, T., Fathallah-Shaykh, H.M.: A multilayer grow-or- go model for gbm: effects of invasive cells and anti-angiogenesis on growth. Bulletin of mathematical biology76, 2306–2333 (2014)

    2014

  6. [14]

    Journal of mathematical biology 79, 941–967 (2019)

    Subramanian, S., Gholami, A., Biros, G.: Simulation of glioblastoma growth using a 3d multispecies tumor model with mass effect. Journal of mathematical biology 79, 941–967 (2019)

    2019

  7. [15]

    British journal of cancer98(1), 113–119 (2008)

    Swanson, K.R., Rostomily, R.C., Alvord, E.: A mathematical modelling tool for predicting survival of individual patients following resection of glioblastoma: a proof of principle. British journal of cancer98(1), 113–119 (2008)

    2008

  8. [16]

    IEEE Transactions on Medical Imaging PP (Nov 2024)

    Weidner, J., Ezhov, I., Balcerak, M., Metz, M.C., Litvinov, S., Kaltenbach, S., Feiner, L., Lux, L., Kofler, F., Lipkova, J., et al.: A learnable prior improves inverse tumor growth modeling. IEEE Transactions on Medical Imaging PP (Nov 2024). https://doi.org/10.1109/TMI.2024....

    2024

  9. [17]

    Medical Image Analysis 101, 103423 (2025)

    Zhang, R.Z., Ezhov, I., Balcerak, M., Zhu, A., Wiestler, B., Menze, B., Lowengrub, J.S.: Personalized predictions of glioblastoma infiltration: Mathematical models, physics-informed neural networks and multimodal scans. Medical Image Analysis 101, 103423 (2025)

    2025

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.