Glioblastoma, the most common malignant brain tumor, spreads beyond the part of the tumor that lights up on a standard contrast MRI scan. Radiotherapy therefore irradiates a uniform margin of one to two centimeters around it. That margin overtreats healthy brain in some directions and misses tumor growth in others. There are two existing ways to personalize the margin. One is deep networks trained on a few hundred patients or fewer, which tend to fit the quirks of one hospital's scanners. The other is fitting a small physics-style equation of tumor spread to each patient's scan. The equation-fitting approach copes well with small data, but its equations treat brain tissue as uniform. Blood-supply imaging says otherwise: tumors spread preferentially where they have recruited new vessels.
We propose to put that missing biology into the equation while keeping the parameter count small. Each patient's perfusion MRI, a scan that measures local blood volume and flow, would set how fast tumor cells multiply and move at each point in the brain. The strength of that link would be two numbers shared across all patients and learned jointly from the whole cohort. Because the vascular information adds only two population-level parameters, the model should stay stable at cohort sizes (tens to low hundreds) where deep networks overfit. The central empirical question is whether this coupling improves prediction of where the tumor recurs on data from a hospital the model never saw.
Core Claim
Current single-scan calibration methods fit a reaction-diffusion equation, a model in which tumor cell density spreads like a diffusing substance and grows logistically, with constant patient-specific rates. Image-Driven Biophysical Tumor Growth Model Calibration does this with a "moving atlas" registration scheme. That scheme cuts initial-condition reconstruction error to 7.1% and 6.3%, against 32.3% and 42.4% for the naive approach. Personalized Predictions of Glioblastoma Infiltration does the same fit with physics-informed neural networks, meaning networks whose training loss penalizes violations of the equation. It validates on only 8 and 16 patients. Both papers supply the method we reuse, and both ignore blood supply.
Two imaging studies motivate adding it. Advanced imaging reveals enhanced malignancy in glioblastomas involving the subventricular zone links invasive glioblastoma phenotypes to higher cerebral blood volume (CBV), with r = 0.26 for median CBV. Cerebral blood flow and histological analysis shows, against biopsy histology, that non-enhancing infiltrated tissue has more microvessels than pure edema. In that study, cerebral blood flow (CBF) ratio separates infiltrated tissue from edema with AUC 0.890, and mean transit time (MTT) ratio with AUC 0.946.
Our claim is that a perfusion-modulated equation fits these data better and generalizes better. We define it as
$$\frac{\partial c}{\partial t} = \nabla\cdot\big(D_0 e^{\alpha v(x)} \nabla c\big) + \rho_0 e^{\beta v(x)}\, c(1-c).$$
Here $c(x,t)$ is normalized tumor cell density. $D_0$ and $\rho_0$ are the patient-specific base rates of cell movement and multiplication. $v(x)$ is the patient's perfusion map, normalized voxel-wise against contralateral normal-appearing white matter. $\alpha$ and $\beta$ are cohort-level coupling strengths.
The normalization step borrows the within-patient reference trick from Hemodynamic Study of Plaque Progression and Regression. In a cohort of only 22 patients, that study obtained significant predictors by dividing each blood-flow measure by the value in an adjacent healthy segment. We expect the same normalization to remove scanner-level and patient-level offsets in perfusion.
The equation above differs from prior work in three specific ways. The prior PDE papers use no perfusion at all. The deep recurrence model in Novel radiotherapy target definition using AI-driven predictions uses diffusion and spectroscopic MRI but neither perfusion nor any physics. The anti-angiogenic model in Mechanistic modeling quantifies the influence of tumor growth kinetics is a spatially lumped model of ordinary differential equations with a fixed vascular fraction, fit to mouse xenografts only.
Research Program
We would build a two-level Bayesian calibration. At the patient level, $D_0$, $\rho_0$ and the tumor's origin are fit to the patient's enhancing-core and FLAIR segmentations. We use the moving-atlas inversion for this, with a forward finite-difference solver. We choose the solver over raw network outputs because the PINN paper found raw network evaluations noisy. At the cohort level, $\alpha$ and $\beta$ receive a shared prior and are estimated by pooling likelihoods across patients. Unlike both prior calibration papers, which report single best-fit estimates, this yields posterior uncertainty for every parameter.
The two-level structure is where the small-sample argument lives. The per-patient fit stays as constrained as the published Fisher-KPP inversions. The coupling is learned from all patients at once, so its uncertainty should shrink roughly as one over the square root of the number of patients, as for any pooled estimate of a fixed number of parameters.
We would analyze three things:
- Identifiability. Does a single scan identify $\alpha$ and $\beta$ separately from $D_0$ and $\rho_0$? We check posterior correlations on synthetic tumors grown on real perfusion maps.
- Stability. Do the cohort-level couplings stay stable when estimated from random halves of the cohort?
- Prediction. Does the coupled model predict the spatial pattern of recurrence better than the uncoupled one?
Perfusion is available in 79.6% of the 630 patients in The University of Pennsylvania glioblastoma (UPenn-GBM) cohort. That cohort also provides expert-refined segmentations for 232 subjects and diffusion tensor imaging for 88.2%, which makes it the development set.
The research file does not state that UPenn-GBM includes follow-up recurrence scans. Recurrence evaluation therefore needs a second-institution longitudinal cohort with pre-treatment perfusion MRI and a recurrence scan, of the kind used in the PINN and AI-recurrence papers. We will assemble or obtain one with a target of 60–100 patients. That cohort doubles as the true external test.
Baselines
The naive reference is the standard RTOG uniform clinical target volume, the radiotherapy region drawn by expanding the tumor outline by a fixed margin. The AI-recurrence paper reports its specificity (healthy brain spared) as 0.79 ± 0.11 and its sensitivity as 0.95 ± 0.08. We will recompute both on our external cohort with identical recurrence segmentations.
The key ablation is the same model with $\alpha=\beta=0$. This is the moving-atlas Fisher-KPP calibration: identical solver, identical data, identical compute. It is the only baseline that isolates the vascular coupling, and for our question it is the strongest competitor, because it already captures most of what physics buys in small samples. For its recurrence coverage, no reported result was found on Lacuna; the PINN paper reports only qualitatively that coverage is comparable to or better than RTOG. We therefore run it ourselves.
The strongest existing data-driven method is the 3D U-Net (an image-to-image segmentation network) with the tumor-volume-scaled loss from the AI-recurrence paper. That paper reports specificity 0.89 ± 0.05 and sensitivity 0.92 ± 0.11 against RTOG on its own 101-patient single-center cohort. We retrain it on UPenn-GBM inputs available at both sites, with a hyperparameter budget matched in wall-clock time to our calibration sweep, and test it on the same external cohort.
We count a result as beating the baselines only if two conditions hold on the external cohort:
- The coupled model shows higher specificity than RTOG at non-inferior sensitivity, meaning sensitivity within 0.03 on paired per-patient comparison.
- It significantly beats the $\alpha=\beta=0$ ablation in recurrence coverage at matched target volume.
Matching specificity to the U-Net's in-domain 0.89 while degrading less from internal to external data would count as a secondary win.
Why not just…?
Why not just add perfusion as an extra input channel to a U-Net? This is the most serious threat. It uses the same biology with no hand-built equation. The concern is cohort size. The AI-recurrence model needed a custom loss merely to stay stable at 101 patients. Temporal radiomics for non-invasive preoperative prediction shows AUC falling from 0.89 in training to 0.78 on external data, even with aggressive feature selection. Still, Predicting microvascular invasion in solitary hepatocellular carcinoma held AUC at 0.94–0.96 across three centers with an attention-based model on 319 patients, so data-driven transfer is not hopeless. We will train the perfusion-augmented U-Net as an explicit arm. We will also run a learning curve on 25, 50, 100 and 200 training patients. If the U-Net matches the coupled model externally at every size, the equation is not earning its keep and we will say so.
Why not just threshold the CBF or MTT ratio map and treat the high-risk voxels as the target? The MTT threshold's AUC of 0.946 makes this cheap and credible. A threshold, however, marks tissue that is infiltrated today; it does not extrapolate to where cells will travel between the scan and recurrence. We run it as a baseline: a perfusion-thresholded region expanded to the same volume as our target. Recurrence growing outside the currently abnormal perfusion region is exactly where the two methods should disagree, and we analyze that subset separately.
Why not just use the published Fisher-KPP calibration and let direction-dependent spread along white-matter fiber tracts carry the extra information? This is a plausible confound. Perfusion could look helpful only because it correlates with white-matter geometry. We will add a variant whose cell-movement term follows fiber directions from diffusion tensor imaging, with and without the perfusion coupling. The perfusion term must add predictive value on top of the fiber-direction model.
Proposed Experiments
Synthetic identifiability test. On synthetic tumors seeded with known $\alpha$ and $\beta$ on real UPenn-GBM perfusion maps, we measure how well the parameters are recovered at 20, 50 and 150 patients. The coupling is only worth testing clinically if recovery at 50 patients has posterior intervals excluding zero.
Internal consistency on UPenn-GBM. Here we look for evidence independent of recurrence. We fit the model on the enhancing core only, then test whether its predicted cell density inside the FLAIR region correlates with a diffusion-derived cellularity proxy better than the uncoupled model's prediction does. The subventricular-zone study used this kind of proxy, a tissue-volume measure derived from diffusion MRI. We also check stability by estimating $\alpha$ and $\beta$ on disjoint halves of the cohort.
Decisive external test. Parameters are learned only on UPenn-GBM. We then report per-patient recurrence sensitivity and specificity at matched target volume on the second-institution cohort for all baselines. Each method is compared with the $\alpha=\beta=0$ ablation using paired Wilcoxon tests, which suit paired per-patient scores that need not be normally distributed.
Pre-registered failure condition. The hypothesis is falsified if either of two things happens. The external coupling estimate may be indistinguishable from zero. Or the ablation may match the full model in coverage at matched volume while the perfusion-channel U-Net matches it externally. In that case the vascular-coupling hypothesis, not merely this implementation, would be in doubt.
References
- Image-Driven Biophysical Tumor Growth Model Calibration
- Personalized Predictions of Glioblastoma Infiltration: Mathematical Models, Physics-Informed Neural Networks and Multimodal Scans
- Advanced imaging reveals enhanced malignancy in glioblastomas involving the subventricular zone: evidence of increased infiltrative growth and perfusion
- Cerebral blood flow and histological analysis for the accurate differentiation of infiltrating tumor and vasogenic edema in glioblastoma
- Hemodynamic Study of Plaque Progression and Regression Based on Coronary CTA Imaging using Computational Fluid Dynamics Method: Preliminary Results
- Novel radiotherapy target definition using AI-driven predictions of glioblastoma recurrence from metabolic and diffusion MRI
- Mechanistic modeling quantifies the influence of tumor growth kinetics on the response to anti-angiogenic treatment
- The University of Pennsylvania glioblastoma (UPenn-GBM) cohort: advanced MRI, clinical, genomics, & radiomics
- Temporal radiomics for non-invasive preoperative prediction of pathologic complete response to neoadjuvant chemoimmunotherapy in non-small cell lung cancer
- Predicting microvascular invasion in solitary hepatocellular carcinoma: a multi-center study integrating clinical, MRI assessments, and radiomics indicators