{"id":"7c0d125c-85db-4eda-9a87-31cdede30bb9","arxiv_id":"2412.05330","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A POD-neural-network pipeline estimates six glioblastoma growth parameters from two synthetic tumor snapshots and forecasts tumor volume with 96% accuracy at 150x speedup, but only on synthetic data from a single brain geometry.","lead":"This paper combines a brain-tumor growth model, reduced-order math, and neural networks to estimate a glioblastoma patient's growth parameters from two MRI snapshots and then predict tumor evolution in seconds. It is a proof of concept on synthetic data built from one real patient's brain anatomy, not yet a validated clinical tool.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inverse network's test split likely shares parameter sets between training and test, so the reported 96% volume accuracy and near-exact parameter recovery may reflect memorization rather than patient-specific generalization.","rationale":"The reader's weakest assumption correctly identifies synthetic validation and weak identifiability of delta and kappa, but the more load-bearing problem is internal to the synthetic evaluation: the inverse-network test set is not parameter-disjoint. Because each parameter set contributes 20 highly correlated input-output pairs, a random pair-level split lets the network see other pairs from the same parameter set during training. That can inflate the reported inverse error and the Fig. 5 proof-of-concept if Eq. (10) was among the training sets. The forward POD-NN is less affected by shared parameter sets because its input contains the parameters; the inverse NN, which outputs parameters, is directly compromised. The proposed group-wise split is a simple check that does not require real clinical data. If the check confirms the concern, the paper's central quantitative claim is not currently supported. Until such a parameter-disjoint evaluation is reported, I would move the verdict from CONDITIONAL to UNVERDICTED, since the evidence for patient-specific generalization is not verifiable from the manuscript as written. If the check refutes the concern, the conditional verdict can stand with the usual real-data caveats.","tokens_in":17066,"tokens_out":7289,"duration_ms":77307,"concrete_test":"Redo the inverse experiment with a group-wise split: partition the 750 parameter sets into 600 training and 150 held-out sets; train NNinv only on pairs from the 600 sets and evaluate only on pairs from the 150 held-out sets. Report normalized per-parameter MSE and tumor-volume prediction error on held-out sets, and state explicitly whether the Fig. 5 parameter set (Eq. 10) was included in training. If held-out volume error remains approximately 4% and parameter MSE approximately 15%, the leakage concern is resolved; if errors degrade substantially, the 96% claim is an artifact of the split.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim is that NNinv recovers six patient-specific parameters from two tumor maps. The only quantitative support is the inverse-network test error and the Fig. 5 proof-of-concept. Section 3.2 says: 'we extract twenty pairs of tumor distributions ... for each of the 750 parameter sets. This results in a total of 15000 input-output pairs. These are then split into a training set containing 11000 elements and a test set with 4000 elements.' That is a random split of pairs, not of parameter sets. Because all 20 pairs for one parameter set come from the same simulation (different 20-day windows of the same trajectory), a test pair almost certainly shares its parameter set with training pairs. The network can therefore memorize parameter-set-specific signatures and report artificially low error. The paper never states that the parameter set in Eq. (10) used for Fig. 5 was held out of training. Without a parameter-disjoint split, the reported ~15% parameter error and 96% volume accuracy do not demonstrate generalization to a new patient. This is an internal validity problem, distinct from the acknowledged lack of real-data validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a computational pipeline for glioblastoma growth prediction and patient-specific parameter identification. A diffuse-interface PDE model (Eq. 9) is discretized with finite elements, reduced via POD, and replaced by a neural network surrogate (POD-NN) for the forward problem. A second neural network (NNinv) maps the POD coefficients of two tumor distributions 20 days apart to six model parameters. All training and validation are performed on synthetic data generated from a single patient's brain anatomy (MRI/DTI). The authors report a ~150x speedup for the forward surrogate, a ~15% parameter estimation error, and \"96% accuracy\" in tumor volume forecasting, and they support the results with global Morris and local Monte Carlo sensitivity analyses.","tokens_in":17233,"tokens_out":6125,"duration_ms":55968,"significance":"If the inverse-network generalization were properly established, the pipeline would offer a fast, interpretable tool for estimating GBM growth parameters from longitudinal imaging, with potential clinical value. The use of a physics-based forward model with DTI-derived anisotropic tensors and the inclusion of global and local sensitivity analyses are strengths. However, the central claim of patient-specific generalization is not yet supported because the inverse-network test split is not parameter-disjoint and the accuracy metrics are not quantitatively defined. The work is a promising proof-of-concept, but its current validation is insufficient for the stated claims.","major_comments":[{"comment":"The inverse-network data set of 15,000 input-output pairs is split randomly into 11,000 training and 4,000 test pairs. Since all 20 pairs for a given parameter set are generated from the same simulation trajectory, a test pair almost certainly comes from a parameter set that also appears in the training set. Consequently, the reported 15% parameter error (Fig. 4) and the 96% volume accuracy claimed in Section 4 do not demonstrate generalization to a new patient's parameter set; they may reflect memorization of parameter-set-specific signatures. The authors should re-split the data by parameter set (e.g., hold out a subset of the 750 parameter sets entirely) and report test errors on unseen parameter sets. They should also state whether the parameter set in Eq. (10) used for the Fig. 5 proof-of-concept was held out during training.","section":"Section 3.2"},{"comment":"The conclusion states \"an accuracy of 96% in forecasting tumor volume\" but no metric is defined in the text or computed from the data shown in Fig. 5. The volume fraction curves appear close, but the paper should report a quantitative error measure (e.g., relative L2 error or relative volume error at t=30 days) computed on a properly held-out test set. As written, this claim is unverifiable and inconsistent with the vague \"well-tracked\" description in Section 3.2.","section":"Section 4 / Fig. 5"},{"comment":"The inverse network is trained and evaluated on synthetic data generated from a single brain anatomy and a single initial tumor condition (the Gaussian φ0 defined in Section 3.2). The proof-of-concept in Fig. 5 uses the same initial condition and anatomy. Thus the claim of \"patient-specific parameter identification\" is currently limited to the DTI-derived tensors; generalization to different tumor shapes, locations, and brain geometries is untested. Although the Conclusions acknowledge this, the Abstract and Introduction should be tempered to reflect that this is a single-anatomy, single-initial-condition proof-of-concept.","section":"Sections 3.1-3.2 and Conclusions"}],"minor_comments":[{"comment":"The paragraph beginning \"The computational demand of the POD-Galerkin solution is generally high...\" appears twice verbatim; remove one occurrence.","section":"Section 3.2"},{"comment":"The reported speed-up is quoted as \"approximately 150 times\" in Section 3.2 and Fig. 5, but the Conclusions say \"computational speed-up of approximately 99%\"; the latter is undefined and inconsistent, and should be corrected.","section":"Section 4"},{"comment":"The elapsed time for parameter estimation is described as \"of the order of seconds\"; please report the actual inference time (e.g., the mean over the test set).","section":"Section 3.2"},{"comment":"The model in Eq. (9) includes parameters ε (diffuse interface thickness) and r (anisotropy tuning factor) that are set a priori; these should be listed explicitly as fixed parameters, and the sensitivity analysis should state that their effects are not studied.","section":"Section 2.1"},{"comment":"The learning curves show training and test errors for the direct and inverse networks, but the panels are not labeled; please clarify which panel corresponds to which network and whether the plotted error is absolute or relative.","section":"Fig. 4"},{"comment":"In the normalization of the tensor T, the notation with hats (ˆT) is not clearly distinguished from T; please clarify the relationship between D, T, and ˆT.","section":"Appendix A.2"}],"recommendation":"major_revision","confidential_remarks":"The data-split issue is the critical barrier to publication. If the authors cannot demonstrate generalization on a parameter-disjoint split, the paper's central claim fails. Given the manuscript's framing as a proof-of-concept, a major revision with re-splitting and re-analysis is within scope. The paper's novelty is moderate (POD-NN is established; the application to GBM with sensitivity analysis is useful), but the fit with eess.IV is acceptable. I recommend the editor require a clear statement of what is held out during training and testing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read. This is a well-executed proof-of-concept that assembles a diffuse-interface GBM model, a POD-NN forward surrogate, and an inverse neural network into a single pipeline on a realistic brain mesh. That assembly is new, and the sensitivity analyses are a real strength: they correctly show that ν and δn dominate and that δ and κ are weakly identifiable. The authors are also honest about the synthetic validation and single geometry. The math and numerics look internally consistent.\n\nThe soft spot is bigger than the authors admit. The inverse network is trained on 20 tumor-distribution pairs per parameter set, then split randomly into 11,000 training and 4,000 test pairs. Because all 20 pairs from one parameter set come from the same simulation, a test pair almost certainly shares its parameter set with training pairs. The network can memorize parameter-set-specific signatures, so the reported ~15% parameter error and 96% volume accuracy don't demonstrate generalization to a new patient. The paper never states that the parameter set in Eq. (10) was held out. The forward POD-NN has a milder version of the same issue — temporal snapshots from the same trajectory are split across train/test — but there the function is smoother, so it's less damaging. Still, a parameter-disjoint split is required for both.\n\nThe lack of real-data validation and code/data release are also gaps, but not flaws for a proof-of-concept. The weak identifiability of δ and κ is acknowledged and consistent with the sensitivity analysis, so it's a limitation, not a mistake.\n\nIn short: the methodology is plausible and worth engaging with, but the central quantitative claims about inverse accuracy are not yet supported. I'd send it to peer review — it deserves referee time — but I'd make a parameter-disjoint split and a held-out patient test mandatory. For my own work, I wouldn't cite the accuracy numbers, and I'd probably only cite it as an example of the pipeline approach.","headline":"Useful pipeline, but the inverse network's validation split is leaky, so the reported accuracy numbers overstate patient-specific generalization.","tokens_in":17859,"tokens_out":2592,"would_cite":false,"duration_ms":24129,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C50","65M60","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural network predicts glioblastoma growth 150x faster from two MRIs.","keywords":["glioblastoma","diffuse-interface model","proper orthogonal decomposition","neural network surrogate","patient-specific parameter identification","tumor growth prediction","sensitivity analysis","neuroimaging"],"falsifier":"Run the trained pipeline on real longitudinal MRI and DTI from a glioblastoma patient not used in training: segment the tumor at two time points 20 days apart, estimate the six parameters, simulate the tumor distribution at a third time point, and compare with the actual segmentation. The central claim predicts volume accuracy near 96% and a morphology that follows the white-matter structure; a substantially larger volume error or a clearly wrong invasion pattern would falsify the transfer claim. A purely synthetic falsifier is to evaluate the surrogate on held-out parameter sets near the boundary of the Table 1 ranges: if the reduced basis learned from the 750 training sets cannot reproduce the full-order solution within the reported error, the surrogate claim fails.","tokens_in":16847,"feed_emoji":"🧠","tokens_out":10604,"duration_ms":94221,"temperature":0.7,"pith_summary":"The paper tries to establish that a patient-specific forecast of glioblastoma growth can be made fast enough for clinical use by combining a diffuse-interface PDE model of tumor invasion with reduced-order surrogates. Its central claim is that two tumor segmentations taken 20 days apart are enough to recover the six biophysical parameters $\\nu$, $M_0$, $\\kappa$, $\\delta$, $\\delta_n$, $S_n$ that drive the model, and that a POD-neural-network surrogate then reproduces the full-order simulation about 150 times faster while tracking tumor volume with roughly 96% accuracy. The proof-of-concept is carried out on a realistic brain-shaped mesh reconstructed from one patient's MRI and diffusion tensor imaging, with synthetic tumor evolutions used to train and test the networks. The authors present this as the methodological foundation for a glioblastoma digital twin, not yet as a clinically validated tool.","feed_headline":"Neural network predicts glioblastoma growth 150x faster from two MRIs","feed_subtitle":"Two scans 20 days apart recover growth parameters and forecast tumor volume to about 96 percent accuracy.","key_machinery":"The load-bearing mechanism is the POD-NN surrogate pipeline built on the diffuse-interface model. Proper orthogonal decomposition (POD) compresses snapshots of the full-order finite element solutions for $\\phi$, $\\mu$, and $\\hat{n}$ into a basis of 20 modes retaining about 95% of the solution energy; two neural networks then replace the expensive reduced-order projections. The forward network $NN_\\phi$ maps $[\\nu, M_0, \\kappa, \\delta, \\delta_n, S_n, t]$ to the reduced coefficients of the tumor field, and the inverse network $NN_\\mathrm{inv}$ maps the reduced coefficients of two tumor distributions separated by 20 days to the six parameters. The mechanism works because the reduced basis makes both directions low-dimensional: the inverse problem becomes a regression on $2N_\\mathrm{POD}$ inputs rather than an optimization over the full PDE, and a forward evaluation takes seconds instead of the roughly 780 s cost of the full-order model. Global Morris and local Monte Carlo sensitivity analyses identify which parameters the data can actually constrain.","core_discovery":"The paper's central claim is that the inverse problem of patient-specific parameter identification for glioblastoma growth is tractable in real time once the forward model is compressed by proper orthogonal decomposition and learned by neural networks. Starting from the diffuse-interface system in Eq. (9) for the tumor phase field $\\phi$, the oxygen concentration $\\hat{n}$, and the chemical potential $\\mu$, the authors build a reduced basis of $N_\\mathrm{POD}=20$ modes that retains about 95% of the solution energy, train a forward network mapping the six parameters and time to reduced coefficients, and train an inverse network mapping the reduced coefficients of two tumor distributions at $t_0$ and $t_0+20$ days back to the parameters. On a synthetic test case built on a patient's anatomy, the recovered parameters are close to the ground truth (for example $\\nu$ goes from 0.356 to 0.366 d$^{-1}$), and the forecast matches the full-order tumor volume to about 96% accuracy. The surrogate runs about 150 times faster than the full-order model, cutting the simulation time from roughly 780 seconds to about 5 seconds, and the parameter estimation itself takes seconds.","pith_inferences":["A natural extension the authors do not pursue is to use the inverse network's output as a proposal generator inside a Bayesian or ensemble framework, converting the reported mean parameter error of about 15% into calibrated uncertainty intervals for surgical planning.","The 20-day input spacing is tied to the clinical follow-up protocol; a testable extension would train the same architecture on variable time gaps (10, 30, 60 days) to see how identifiability of $\\nu$ and $\\delta_n$ degrades as the gap shrinks.","Because all validation is synthetic, the 96% volume accuracy measures consistency with the model that generated the data; the clinically meaningful test is against real recurrences, which the authors state is the object of an ongoing study.","Since the weakly identifiable parameters $\\delta$ and $\\kappa$ are also those with wide literature ranges, a pragmatic variant would freeze them at population medians and reduce the inverse problem to four parameters, likely improving stability."],"forward_implications":["If the framework transfers to real clinical scans, a treating team could estimate a patient's proliferation and oxygen parameters from two routine MRIs and simulate the likely tumor distribution at a future date within a single clinical visit.","Because the surrogate separates the fixed brain anatomy (tensors $T$ and $D$, mesh) from the six parameters, forecasts for a new set of parameters on the same anatomy are near-instant; the main remaining cost is retraining for a new patient's anatomy.","The sensitivity results imply that clinical calibration should focus on the proliferation rate $\\nu$, the oxygen consumption rate $\\delta_n$, and the oxygen supply rate $S_n$, while $\\delta$ and $\\kappa$ are weakly identifiable and can be pinned to literature values with little loss in volume prediction.","By learning reduced coefficients directly from simulation data, the pipeline avoids hyper-reduction techniques such as DEIM, so the reported speed-up does not require projecting the nonlinear terms of Eq. (9)."],"supporting_citations":[{"why":"Diffuse-interface mixture models of tumor growth that provide the phase-field formulation behind Eq. (9).","marker":"[25–27]"},{"why":"Supplies the personalized-treatment computational framework and the anisotropy parameterization used to build tensors $T$ and $D$ from diffusion imaging.","marker":"[30]"},{"why":"Provide the proper orthogonal decomposition method used to build the reduced basis from full-order snapshots.","marker":"[42, 43]"},{"why":"Prior work learning patient-specific parameters for a diffuse-interface glioblastoma model from neuroimaging data, the direct predecessor of the inverse network.","marker":"[48]"},{"why":"Introduces the POD-NN approach that maps parameters and time to reduced coefficients, the core of the forward surrogate.","marker":"[51]"},{"why":"Supply the Morris screening method used for the global sensitivity analysis that identifies the dominant parameters.","marker":"[53, 54]"}],"fun_headline_variants":["Neural net predicts glioblastoma growth 150x faster from two MRIs","Two scans, seconds: patient-specific tumor forecast via neural net","Reduced-order neural network cuts brain tumor simulation from 13 min to 5 sec","Neural network recovers glioblastoma growth parameters with 96% accuracy","Real-time brain tumor growth prediction using neural network surrogates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes that the diffuse-interface model in Eq. (9), with its literature parameter ranges and DTI-derived tensors, is an adequate description of real glioblastoma growth, and that two tumor snapshots 20 days apart carry enough information to identify the parameters that actually drive the forecast.","fun_headline_variants_meta":{"raw":{"variants":["Neural net predicts glioblastoma growth 150x faster from two MRIs","Two scans, seconds: patient-specific tumor forecast via neural net","Reduced-order neural network cuts brain tumor simulation from 13 min to 5 sec","Neural network recovers glioblastoma growth parameters with 96% accuracy","Real-time brain tumor growth prediction using neural network surrogates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001308,"raw_usage":{"total_tokens":5333,"prompt_tokens":950,"completion_tokens":4383,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":4287}},"tokens_in":566,"tokens_out":4383,"duration_ms":31894,"temperature":1.0,"reasoning_tokens":4287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:16:15.990836+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained pipeline on real longitudinal MRI and DTI from a glioblastoma patient not used in training: segment the tumor at two time points 20 days apart, estimate the six parameters, simulate the tumor distribution at a third time point, and compare with the actual segmentation. The central claim predicts volume accuracy near 96% and a morphology that follows the white-matter structure; a substantially larger volume error or a clearly wrong invasion pattern would falsify the transfer claim. A purely synthetic falsifier is to evaluate the surrogate on held-out parameter sets near the boundary of the Table 1 ranges: if the reduced basis learned from the 750 training sets cannot reproduce the full-order solution within the reported error, the surrogate claim fails.","supporting_citations":[{"cited_title":"Agosti et al","cited_arxiv_id":null,"evidence_quote":"Supplies the personalized-treatment computational framework and the anisotropy parameterization used to build tensors $T$ and $D$ from diffusion imaging."},{"cited_title":"Agosti et al","cited_arxiv_id":null,"evidence_quote":"Prior work learning patient-specific parameters for a diffuse-interface glioblastoma model from neuroimaging data, the direct predecessor of the inverse network."},{"cited_title":"Hesthaven and S","cited_arxiv_id":null,"evidence_quote":"Introduces the POD-NN approach that maps parameters and time to reduced coefficients, the core of the forward surrogate."}],"review_version":1}