Gap-latched infiltration: modeling hidden glioma reservoirs that wake up late

Glioblastoma is the most common aggressive brain tumor. It recurs almost every time because its cells spread silently into healthy brain beyond what MRI can see. Radiation plans today add a fixed 15 mm margin around the visible tumor. The best patient-specific alternatives fit a reaction-diffusion equation, a model in which tumor density spreads like heat and grows logistically, to the patient's scans and irradiate where the model predicts hidden cells. Those models assume smooth spreading: wherever cells diffuse, they keep proliferating at the same rate. Biology suggests otherwise. Invading glioma cells switch between a motile, non-dividing state and a stationary, networked, proliferative state (Tweety-Homolog 1 Drives Brain Colonization of Gliomas). Infiltrating cells can also sit dormant in vascular niches (3D Brain Vascular Niche Model Captures Glioblastoma Infiltration, Dormancy, and Gene Signatures). A population that spreads first and multiplies later leaves a flat, low-density reservoir far from the visible edge. A smooth model would instead predict a thin tail that fades quickly with distance.

We propose to model this reservoir with a mechanism borrowed from neutron-star physics, and to fit it with an observation model borrowed from underwater sensor networks. In old neutron stars, Cooper pairing (the pairing of nucleons into a superfluid) opens an energy gap that blocks the particle reactions which would relax the star's chemical imbalance. The imbalance therefore builds for millions of years and only releases as heat once it exceeds the gap (Contrasting neutron star heating mechanisms with Hubble Space Telescope observations). We carry this "store, then release past a threshold" structure into glioma growth. The second ingredient addresses how the model is compared with scans. MRI segmentations are essentially one-bit readings of an unseen cell density, so we fit the model through an explicit likelihood for such coarse measurements. This replaces the hand-picked density thresholds that current methods hard-code, a sensitivity the field itself flags as unresolved. The concrete test is whether recurrence coverage on public glioblastoma cohorts improves when the planned volume is held fixed.

Core Claim

A two-population growth model with a gap-type latch predicts recurrence locations that a tuned Fisher-Kolmogorov model cannot. Here migratory cells convert to proliferating cells only after a locally accumulated drive exceeds a threshold. The difference matters most far from the visible edge, where the latched model places a seeded but quiet reservoir. That claim is only testable if the link between density and image is not fixed by hand, because a free threshold can mimic a reservoir. So the second half of the claim is methodological. MRI labels should be treated as quantized measurements with a fitted probabilistic link. Underwater networks handle few-bit sensor readings in a similar way (Optimal Quantization Scheme for Data-Efficient Target Tracking via UWSNs Using Quantized Measurements), where the estimator is built around how coarsely the data were recorded rather than treating the readings as exact.

Model

Let $u(x,t)$ be the proliferating cell density and $m(x,t)$ the migratory, non-dividing density. Let $a(x,t)$ be a local "drive" that accumulates while migratory cells are present, analogous to the chemical imbalance in the star. The model is

$$\partial_t m = \nabla\cdot(D\nabla m) + \gamma\, u(1-u) - \kappa\, H(a-\Delta)\, m, \qquad \partial_t u = \rho\, u(1-u-m) + \kappa\, H(a-\Delta)\, m, \qquad \partial_t a = m - \lambda a,$$

where $D$ is the diffusion tensor built from an atlas of white-matter fibers measured by diffusion tensor imaging, and $\gamma$ is the rate at which crowded proliferating cells shed migratory cells. The conversion rate $\kappa$ applies only after the drive passes the gap $\Delta$. The step function $H$ is the latch, which we smooth slightly so the model can be optimized, and $\lambda$ is a slow decay of the drive. Setting $\Delta = 0$ and $\gamma \to \infty$ recovers ordinary Fisher-Kolmogorov dynamics. The model therefore contains the standard one and can be compared with it directly.

The observation model treats each voxel label $y_k$ (enhancing on contrast-enhanced T1, or abnormal on FLAIR) as a one-bit quantization of total cellularity $c = u + m$. The probability of a positive label is $P(y_k = 1 \mid c) = \Phi((c-\tau_k)/\sigma_k)$, with a per-patient threshold $\tau_k$ and softness $\sigma_k$ under weak priors. FET-PET, a positron-emission scan that tracks amino-acid uptake by tumor cells, enters as a graded measurement when available.

Research Program

We build the solver on the discrete-loss formulation of Individualizing glioma radiotherapy planning by optimization of a data and physics-informed discrete loss (GliODIL). That work optimizes cell density directly on a multi-resolution grid and treats the equation as a soft penalty. It is the method we reuse, extended to three fields ($u$, $m$, $a$) and to the quantized likelihood above in place of its data terms.

Two other papers constrain the design. A Lightweight Optimization Framework for Estimating 3D Brain Tumor Infiltration shows that unconstrained fits collapse to trivial solutions without a physical prior and that results are sensitive to threshold choice. We take this as evidence that our extra degrees of freedom need strong structure and priors. From Fiber Tracts to Tumor Spread: Biophysical Modeling of Butterfly Glioma Growth Using Diffusion Tensor Imaging shows that atlas-based anisotropic diffusion improves Dice overlap, by about 5% for butterfly gliomas. We adopt that anisotropy unchanged so it is not confounded with the latch.

The program has three stages. First, a synthetic study checks whether $\Delta$ and $\kappa$ are identifiable at all from one pre-operative scan, versus pre- plus post-operative scans. Second, retrospective fitting on two public cohorts uses a matched-volume protocol, described below. Third, the latch is checked qualitatively against intravital imaging results. In those data, removing the motile phenotype slows invasion fourfold but leaves a radioresistant networked core (Tweety-Homolog 1 Drives Brain Colonization of Gliomas). The pattern is consistent with a latch: suppress the migratory pool and invasion speed falls, while proliferation in the latched core is untouched.

Baselines

The naive reference is the clinical uniform 15 mm margin around the enhancing tumor. On the GliODIL cohort of 152 patients it achieves 70.04% recurrence coverage, as reported by Individualizing glioma radiotherapy planning by optimization of a data and physics-informed discrete loss. Beating it is necessary but not interesting.

The strongest existing method is GliODIL itself, which reports 72.94% recurrence coverage on the subset with FET-PET in the same paper. It is the strongest competitor because it already fits a physics prior softly to multi-modal imaging on exactly this cohort. Any gain from our method must therefore come from the latch or the likelihood, not from using more data.

We also run the lightweight Dirichlet-plus-wave-loss framework of A Lightweight Optimization Framework for Estimating 3D Brain Tumor Infiltration. Its paper reports improvements of up to 8% in predicting any recurrence over standard plans. We found no absolute coverage figure for it on Lacuna, so we rerun it. Finally, we run the anisotropic Fisher-Kolmogorov model fitted by evolution strategy (CMA-ES, a gradient-free optimizer), following From Fiber Tracts to Tumor Spread: Biophysical Modeling of Butterfly Glioma Growth Using Diffusion Tensor Imaging.

Fairness rests on the matched-volume rule. Every method receives the same inputs and the same atlas fiber tensor. Every method's predicted density is turned into a plan with exactly the same irradiated volume as that patient's standard 15 mm plan, so coverage gains cannot come from simply treating more tissue. Every method also gets the same per-patient optimization budget in GPU-minutes, and hyperparameters are tuned only on a held-out tuning split.

Two ablations isolate the ingredients. The first keeps two populations but replaces the latch with linear conversion ($\Delta = 0$). The second keeps the latch but restores fixed segmentation thresholds in place of the quantized likelihood. Beating GliODIL means a matched-volume coverage gain over 72.94% on the PET subset, plus a significant per-patient paired improvement across all 152 patients. The gain must also persist when we transfer to the RHUH cohort of 40 patients without retuning.

Why not just…?

Why not just tune the imaging threshold in an ordinary Fisher-Kolmogorov model? This is the most serious alternative. A lower visibility threshold can push predicted density outward and partly mimic a reservoir. The lightweight framework's need for a wave loss to resist threshold choice (A Lightweight Optimization Framework for Estimating 3D Brain Tumor Infiltration) shows thresholds matter a great deal. Our answer is the second ablation, latch plus fitted likelihood versus no latch plus fitted likelihood. If ordinary Fisher-Kolmogorov dynamics with a fitted quantized likelihood match the full model, then the latch is unnecessary, and we will report the likelihood alone as the contribution. The shapes differ in a testable way. A threshold shift rescales a decaying tail, whereas a latch produces a flat plateau. We expect the difference to show up as recurrences beyond roughly 2 cm from the enhancing edge.

Why not just use a two-population "go-or-grow" model, in which cells either move or divide, without a threshold? That is precisely the first ablation, linear conversion. It isolates whether the storage-then-release behavior, rather than having two cell states, explains distant recurrences. If it matches the latched model, the neutron-star analogy adds nothing, and we will say so.

Why not just train a 3D network end to end on scans? Continuous-time attention models had to fall back to 2D slices for memory reasons (Continuous-Time Deep Glioma Growth Models). Two hundred or so patients with recurrence labels is also very little supervision for a voxel-wise 3D predictor, and GliODIL's authors report that unregularized fits overfit noisy images. We include a probabilistic U-Net in the style of Deep Probabilistic Modeling of Glioma Growth, trained with cross-validation on the same splits, as a reference under the matched-volume rule.

Proposed Experiments

The decisive check is a three-way comparison on matched-volume recurrence coverage. The GliODIL cohort supplies the main sample (152 patients, 58 with FET-PET and follow-up), and RHUH supplies external transfer. The comparison is the full model against GliODIL and against the two ablations, reported per patient with paired tests. It is further stratified by distance of the recurrence from the enhancing edge, because the claim predicts gains concentrated at large distances.

RHUH includes pre-operative, post-operative and recurrence scans. This allows a stronger test: fit the model on the first two time points, then predict where the tumor recurs. That is the setting in which an accumulated drive should be identifiable. Before touching patient data, we run the synthetic identifiability study. We simulate latched tumors, generate one-bit labels, and check whether $\Delta$ is recovered from one versus two time points. If it is not identifiable from a single pre-operative scan, we restrict the clinical claim to the longitudinal setting.

On the butterfly-glioma cohort we test whether the latch harms the 5% Dice gain from anisotropy. It should not, and if it does, the extra flexibility is fitting noise. A negative result on the first ablation, with no gain over the no-latch two-population model, would falsify the core claim. A gain that appears only in the ablation without the latch would redirect the contribution to the quantized-likelihood observation model.