Learning How Glioma Margins Co-Vary: Covariance-Conditioned Radiotherapy Targets

Glioblastoma cells spread centimeters beyond the visible tumor, so radiotherapy treats a "safety margin" around it. The clinical default is a uniform 15 mm expansion. Patient-specific biophysical models do better: they fit a reaction-diffusion equation (a model in which cells both multiply and spread outward) to each patient's scans. Both approaches share a blind spot. Neither learns, across patients, how a model's errors in one direction relate to its errors in another. If a tumor turns out to have spread further than predicted toward the corpus callosum, is it also likely to have been under-predicted along the internal capsule? Or over-predicted there?

Robotics faced the same question about joints. Probabilistic movement primitives model a whole trajectory as one Gaussian with full covariance across joints. That lets a robot fix one joint's deviation and re-plan the others in closed form, which is the "automatic joint compensation" of Paraschos et al.. We propose to treat the tumor's infiltration extent along a set of anatomically defined directions as the "joints." We would learn, from a few hundred longitudinal patients, a low-rank covariance over the residuals of the best biophysical model in those directions. At planning time, that covariance is used to condition the margin on whatever direction-specific evidence each patient provides. The claim is narrow and testable. Errors of current infiltration models are correlated across directions in a way that a population covariance can capture. Exploiting that correlation should raise recurrence coverage at a fixed treated volume.

Core Claim

Current patient-specific infiltration models make errors that are not independent across directions around the tumor. A covariance over these errors can be estimated from PREDICT-GBM-scale data, and Gaussian conditioning on per-patient evidence can then correct them. The resulting margin should cover more of the eventual recurrence than either the biophysical model alone or an independent per-direction correction, at the same treated volume.

We borrow the statistical machinery of movement primitives, not the mechanics of limbs. Their key ingredients are three: a structured Gaussian over a compact parameter vector, full rather than diagonal covariance, and closed-form conditioning. This sidesteps the reviewer objection that tumor cells are not a kinematic chain. We make no claim that they are.

What Each Direction Contributes

From inter-joint coordination we take a well-established empirical lesson: modeling cross-dimension dependence matters. Grasemann et al. found that a single network driving three leg joints jointly cut tracking error by more than half relative to three independent networks. Li et al. showed that full covariance is needed to represent negative correlations, where one dimension moves opposite another. Both papers motivate the choice of full rather than diagonal structure.

The same literature also constrains the design. Full covariance explodes in parameter count, and ProMPs need many demonstrations to estimate it (Paraschos et al.). With roughly 255 patients, we must therefore use a low-rank-plus-diagonal (factor) covariance rather than a dense one.

We avoid autoregressive orderings over directions, as used in Zhang et al. and Benechehab et al.. Those orderings impose an arbitrary sequence, and directions around a tumor have no natural order.

From diffuse infiltration we take the problem, the data, and the base model. Isotropic spread misrepresents non-spherical invasion (Bach Cuadra et al.). Tensor-guided transport along white matter helps, most clearly in butterfly gliomas (Weidner et al.; Clatz et al.). Fitting the equation as a soft loss rather than solving it exactly improves coverage, as in GliODIL (Ezhov et al.). A fast optimizer that combines a data loss, a smoothness penalty and a wave constraint brings the fit under a minute (Balcerak et al.).

PREDICT-GBM (Zimmer et al.) supplies the evaluation pipeline, the 255-patient cohort and an iso-volumetric coverage protocol. All of these fit each patient in isolation. None pools information across patients about where and how the fitted model tends to be wrong. That pooled, cross-direction error structure is the gap we target.

The biology suggests why such structure might exist without proving it. Jung et al. show that invading cells extend long tumor microtubes and form connected multicellular networks. Human Glioma Migration and Infiltration Properties shows that some patient-derived lines migrate more after low-LET irradiation. Both effects are non-local, and a local smoothness prior cannot express them. We treat these as reasons to look for correlated residuals, not as mechanisms we model.

Research Program

All scans are registered to the SRI24 atlas through the PREDICT-GBM pipeline. Around each tumor's centroid we define $K$ directions, either 42 icosahedral rays or rays toward a fixed set of white-matter parcels. Along each direction $k$ we measure three distances:

  • $r_k$, the distance from the enhancing-core boundary to the outermost recurrence voxel;
  • $\hat r_k$, the same distance implied by the base biophysical model's thresholded density;
  • $\delta_k$, the pre-treatment FLAIR or FET-PET extent.

The object of interest is the residual vector $e = r - \hat r$. We model it as

$$e \sim \mathcal{N}\big(\mu(z),\; W W^\top + \mathrm{diag}(\psi)\big),$$

where $z$ holds coarse tumor descriptors (lobe, volume, distance to the corpus callosum) and $\mu(z)$ is a small linear mean. $W$ is a $K \times m$ loading matrix with $m \ll K$, and $\psi$ holds per-direction noise variances. This form keeps the parameter count near $K(m+1)$ rather than $K^2$.

For a new patient, some directions carry informative evidence. One example is a sector where FET-PET uptake reaches well beyond FLAIR. We treat that evidence as a noisy observation of the corresponding residuals. Standard Gaussian conditioning then gives a posterior mean and variance for every direction. This is exactly the ProMP via-point operation applied to infiltration extents instead of joint angles.

The margin is set by placing $\hat r_k + \text{posterior quantile}$ along each ray. A single quantile level is chosen so that the treated volume matches the clinical plan. That makes every comparison iso-volumetric.

Before any planning experiment, we run an analysis step. We estimate the residual covariance on training folds and ask three questions. First, does the off-diagonal structure exceed what is expected from noise, judged by permutation of direction labels within patients? Second, are any direction pairs negatively correlated? Such pairs would be the "compensation" signature, where under-prediction along one route trades off against over-prediction along another. Third, do the leading factors align with known tracts such as the corpus callosum and internal capsule?

If the residuals are effectively uncorrelated, the project stops here. That would be a useful negative result: today's biophysical models leave no shared cross-direction error to exploit.

Baselines

The naive reference point is the standard plan: a uniform 15 mm expansion of the core, as implemented in PREDICT-GBM, evaluated on the same patients and at the same volume. The best cited numbers for it come from the GliODIL paper. On the FET-PET subset of 58 patients, recurrence coverage rose from 70.04% for the standard plan to 72.94% for GliODIL (Ezhov et al.).

The strongest existing methods are the biophysical models inside PREDICT-GBM. SBTC is reported there as achieving the highest mean recurrence coverage. However, no reported coverage value for SBTC was found on Lacuna. We will therefore re-run it ourselves in the released pipeline on identical folds. GliODIL and the lightweight optimizer of Balcerak et al. are run the same way. The latter reports improvements of up to 8% for "any recurrence" on the RHUH cohort.

Our method sits on top of whichever base model is strongest. Its added cost is fitting $W$ and $\psi$ offline, plus one $K$-dimensional conditioning at plan time, which takes milliseconds. Every base model therefore gets the same per-patient compute budget. Beating the strongest baseline means a statistically significant gain in iso-volumetric recurrence coverage over SBTC on held-out patients, assessed with a paired Wilcoxon test across patients. The gain must hold on the cross-center split, not only within one center.

Two ablations isolate the key ingredient. The first, a diagonal-covariance residual model, keeps the same mean $\mu(z)$ and the same evidence, but sets $W = 0$. Conditioning then corrects only the observed directions themselves. If this matches the full model, the covariance contributes nothing. The second is a dense empirical covariance with shrinkage, which tests whether the low-rank restriction costs accuracy. Finally, a data-driven nnU-Net recurrence predictor, as benchmarked in PREDICT-GBM, serves as the representative of purely learned approaches.

Why not just…?

Why not just use the biophysical model, since the PDE already couples directions? This is the real threat. A reaction-diffusion model with diffusion tensors (multi-voxel water-diffusion maps from MRI that encode fiber orientation) does propagate information across space through physics. It may already produce whatever correlation exists. Our analysis step tests exactly this. The residuals after SBTC or GliODIL are where we look for structure, so any covariance we find is by construction what the physics missed.

One reason to expect some leftover structure comes from Weidner et al.. They rely on a healthy-brain atlas of fiber tracts, which mass effect distorts. That distortion should produce errors along tracts that are systematic, not random. If the residual covariance is not significant, we report it and stop.

Why not just correct each direction independently from its own local evidence? That is the diagonal ablation above. It can win if the evidence is dense, meaning every direction has an informative FET-PET or FLAIR reading. The covariance should matter most where evidence is sparse. So we report results stratified by the fraction of directions with informative evidence. We expect the gap over the diagonal model to grow as evidence becomes sparser. The robotics precedent (Grasemann et al.) suggests joint modeling helps, but it is no proof for tumors.

Why not just train a deep network to predict recurrence directly, such as nnU-Net or a probabilistic U-Net? PREDICT-GBM reports that physics-informed models outperformed nnU-Net on the same 255 patients, with limited data being the likely cause (Zimmer et al.). The probabilistic U-Net of Deep Probabilistic Modeling of Glioma Growth was trained on only 38 patients and uses a diagonal latent space. Our model has on the order of a few hundred parameters, a size those data can support. We still include nnU-Net as a baseline on identical folds.

Why not just use an anisotropic, tract-shaped margin? Tract-guided spread is already part of the base models we build on. The question is whether the population covariance adds anything beyond them. The butterfly-glioma subgroup is the natural stress test. There, Weidner et al. report roughly 5% Dice improvement from anisotropy, against only 1.2% for single-hemisphere tumors.

Proposed Experiments

We use the full PREDICT-GBM cohort: TUM-GBM (152 patients), LUMIERE (63) and RHUH-GBM (40). The main split trains the residual model on TUM-GBM and LUMIERE and tests on RHUH-GBM. A five-fold within-cohort split is a secondary analysis. FET-PET evidence is used only for the 58-patient TUM-GBM subset where it exists; elsewhere the evidence is FLAIR extent alone.

The first decisive check is held-out log-likelihood of residual vectors under the full, low-rank, diagonal and no-residual models. The full or low-rank model must beat the diagonal one by more than its permutation null.

The second check is masked-direction prediction. We hide the recurrence extent in a random half of the directions, condition on the other half, and measure mean absolute error in millimeters on the hidden half. This is the direct analogue of joint compensation.

The third check is the clinical endpoint: iso-volumetric coverage of "any recurrence" and of enhancing recurrence. It is reported per cohort, for the butterfly subgroup, and for the FET-PET subset, where the comparison with the published GliODIL coverage of 72.94% applies.

We would also report whether negatively correlated direction pairs survive the cross-center split. The same check applies to the radiation-induced invasion question raised by Human Glioma Migration and Infiltration Properties. If recurrences far from the original site co-occur with under-predicted extent near the core, the residual factors should show it. We would present this only as a descriptive finding, not a causal one.

The project succeeds if the low-rank residual model beats both SBTC and the diagonal ablation on RHUH-GBM coverage at matched volume. It fails informatively if the residuals show no cross-direction structure.