PREDICT-GBM: A multi-center platform to advance personalized glioblastoma radiotherapy planning

1 Introduction

• MRI misses diffuse infiltration, and uniform margins risk both undertreating tumor and overtreating surrounding tissue.The approach does not capture patient-specific tumor biology, invasion patterns, or anatomy.Source: §1 · p. 1
• Prior studies lacked sufficiently large, heterogeneous cohorts and standardized evaluation protocols for meaningful comparison of modeling approaches.Longitudinal imaging curation and privacy concerns contribute to limited suitable datasets.Source: §1 · p. 2
• Unstandardized preprocessing and inconsistent recurrence-coverage metrics hinder objective comparison across studies.Accurate peritumoral tissue classification is challenging in the presence of mass effect, edema, and postoperative changes.Source: §1 · p. 2
• PREDICT-GBM provides a curated longitudinal dataset and unified pipeline for comparing computational models with standard clinical care.The open-source framework supports processing additional datasets and reproducible evaluation.Source: §1 · p. 2

Recurrence coverage

• U-Net achieved the highest mean coverage for both recurrence definitions, closely followed by GliODIL, with all tested models except LMI and nnU-Net performing at least as well as SOC.LMI and nnU-Net underperformed SOC by approximately 10% and 20%, respectively, across the aggregated cohort.Source: §2 · p. 3
• Including necrosis in the recurrence definition produced only modest changes in mean coverage.Source: §2 · p. 3

Statistical analysis

• U-Net and GliODIL significantly improved coverage over SOC for both enhancing-only and enhancing-plus-necrosis recurrence definitions.Reported p-values were p < 10^-5 and p < 10^-3, respectively.Source: §2 · p. 3
• For U-Net, GliODIL, and LOTI, improvement over SOC had a weak positive correlation with recurrence distance at 15–30 mm.The reported Spearman correlation was ρ between 0.2 and 0.3 in that distance range.Source: §2 · p. 3

Dataset SOC LOTI GliODIL U-Net PINN-GBM LMI nnU-Net GlioMap

• Table 2 reports coverage and standard error for enhancing recurrence and enhancing-plus-necrotic recurrence across datasets and models.Asterisks mark paired Wilcoxon signed-rank comparisons against SOC; bold indicates the best value per dataset.Source: Table 2 · p. 4
• Across registration algorithms, absolute coverage varied slightly, but model rankings and paired-test conclusions remained stable.The comparison included rigid, affine, and deformable registration approaches.Source: Dataset SOC LOTI GliODIL U-Net… · p. 4
• U-Net assigns higher recurrence probability to white matter and predicts cross-hemisphere migration mainly through the corpus callosum.The authors describe these spatial patterns as biologically plausible and not based only on proximity.Source: Dataset SOC LOTI GliODIL U-Net… · p. 4
• Distant recurrences were not adequately captured, and small initial tumors constrained iso-volumetric redistribution toward recurrence.For multifocal tumors, SOC outperformed model-based approaches; physics-informed models assume a single initial tumor position, while multifocal cases are underrepresented in data-driven training.Source: Dataset SOC LOTI GliODIL U-Net… · p. 4
• The example CTVs are iso-volumetric; model-based contours redistribute volume toward recurrence while sparing normal brain tissue missed by SOC.Source: Figure 4 · p. 5

3 Discussion

• PREDICT-GBM unifies curated longitudinal data with an open, extensible pipeline and models, addressing barriers from demanding preprocessing and inconsistent comparative evaluation.Accurate registration and segmentation require domain expertise, and small misregistrations can distort voxel-level overlap metrics, especially for small recurrences.Source: §3 · pp. 5–7
• U-Net achieved the highest mean recurrence coverage, followed by GliODIL, LOTI, and PINN-GBM; each outperformed SOC, while GlioMap exceeded SOC on the diffusion subset.U-Net and GliODIL showed statistically significant improvement over SOC for both enhancing-only and enhancing-plus-necrotic recurrence in the pooled cohort.Source: §3 · pp. 6–7
• With iso-volumetric targets, increased recurrence coverage corresponds to increased specificity, rather than simply reflecting a larger irradiated volume.The geometric evaluation does not capture dose falloff or organ-at-risk constraints.Source: §3 · p. 7
• Coverage improvement had a weak positive correlation with center-of-mass distance within 15–30 mm, and method rankings and significance conclusions remained robust across registration strategies.Absolute coverage varied slightly across registration algorithms.Source: §3 · p. 7
• The benchmark remains modest for deep learning, uses standard MRI with diffusion available only for a subset, and evaluates geometric targets rather than delivered doses.Observed coverage gains cannot be directly interpreted as expected survival, toxicity, or quality-of-life improvements; retrospective evaluation under SOC is not prospective validation of model-based plans.Source: §3 · pp. 7–8
• Future work includes larger shared cohorts, spatially resolved biopsy data, richer modalities and tumor models, and dose-aware evaluation that respects organ-at-risk constraints.Proposed model extensions include mass effect, multi-compartment tumor representations, and explicit handling of multifocal disease.Source: §3 · pp. 6, 8

Cohort and inclusion criteria

• Follow-up exams at least 12 weeks after baseline were selected to identify early post-treatment recurrence in the curated cohort.For patients with multiple eligible follow-ups, the exam with the best resolution was selected, and experts confirmed visible recurrence.Source: §4 · pp. 8–9

Image preprocessing

• MRI exams were converted, skull-stripped, co-registered to SRI24 space, and longitudinally aligned by deformable registration.Transforms were applied to follow-up modalities and segmentations; ADC was derived from mean diffusivity for patients whose only diffusion modality was DTI.Source: §4 · p. 9

Segmentation

• BraTS methods segmented tumor subregions, while atlas tissue priors supplied gray matter, white matter, and CSF maps for models requiring tissue information.Tumor presence was verified by a medical expert; tissue priors were warped into patient space using deformable atlas registration.Source: §4 · pp. 9–10

Model approaches

• The benchmark compared Fisher-KPP growth models that estimate tumor-cell concentration with data-driven models that directly predict recurrence regions.The evaluated families included GliODIL, LOTI, PINN-GBM, and LMI for concentration mapping, and GlioMap, U-Net, and nnU-Net for recurrence prediction.Source: §4 · p. 10

Biophysical models

• GliODIL combines data and discretized reaction-diffusion physics losses, while LOTI fits MRI tumor segmentation and penalizes large spatial gradients.PINN-GBM infers patient-specific growth parameters and concentration using a Fisher-KPP solution and neural-network fine-tuning; LMI learns PDE growth parameters from simulated data.Source: §4 · p. 10

Recurrence prediction models

• GlioMap estimates peritumoral voxel-wise recurrence probabilities from MRI radiomic features, and nnU-Net served as a cross-validated baseline with dataset-stratified splits.GlioMap uses undersampling for class imbalance and distance-based correction; nnU-Net receives preoperative tumor and tissue maps as inputs.Source: §4 · p. 10

Custom U-Net

• The custom U-Net used 15 trained networks across stratified splits, with out-of-sample predictions and channel dropout, augmentation, and a tumor-size-adjusted progression coverage loss.Its extended training dataset contained 327 patients, including 84 additional cases from four datasets not included in the public data release.Source: §4 · pp. 10–11

Radiation plans and evaluation

• Model targets matched the SOC volume, defined by a 15 mm isotropic tumor-core expansion, and recurrence coverage was evaluated for enhancing and enhancing-plus-necrotic recurrence.Coverage was computed as sensitivity and compared per patient with paired Wilcoxon signed-rank tests; Table 3 summarizes inference time, training time, and inference memory requirements.Source: §4 · p. 11, Table 3 · p. 11

Public data

• The 243-patient cohort is made publicly reproducible through patient and modality lists, with skull-stripped exams released for 142 TUM-GBM subjects.The authors state this makes all 243 patients publicly available while protecting patient anonymity.Source: §5 · p. 11

Released assets

• Processed tumor and tissue segmentations, along with growth-model tumor-cell maps, are provided as NIfTI files.The passage also identifies public access locations for the LUMIERE and RHUH datasets.Source: §5 · p. 11

Reproducibility

• The released patient and modality information is intended to let researchers with access to the public datasets reproduce the experiments.Source: §5 · p. 11

Repository contents

• The repository provides the complete end-to-end PREDICT-GBM pipeline and containerized implementations of evaluated growth and recurrence prediction models.Source: §6 · p. 11

Extensibility

• Documentation and integration guidelines support adding new models to the repository.Source: §6 · p. 11

Access

• The code is available at github.com/BrainLesion/PredictGBM.Source: §6 · p. 11

Imaging terms

• Imaging terms include ADC for apparent diffusion coefficient, DTI for diffusion tensor imaging, FLAIR for fluid-attenuated inversion recovery, and CSF for cerebrospinal fluid.Source: Abbreviations · p. 12

Clinical terms

• Treatment and disease abbreviations include CTV for clinical target volume, RT for radiotherapy, SOC for standard-of-care, GBM for glioblastoma, and OS for overall survival.Source: Abbreviations · p. 12

Modeling terms

• Modeling abbreviations include LMI for Learn-Morph-Infer, LOTI for Lightweight Optimization for Estimating Tumor Infiltration, ODIL for Optimizing a Discrete Loss, and PINN for physics-informed neural network.Source: Abbreviations · p. 12

• The GlioMap algorithm was formally disclosed through a Disclosure of Invention (DOFI).Source: §8 · p. 16
• The DOFI was filed with the University Hospital of Northern Norway (UNN).Source: §8 · p. 16
• The filing is identified as reference DOFI 2024/8833.Source: §8 · p. 16
• The authors report no financial interests or royalties associated with the disclosure.Source: §8 · p. 16

9.1 Additional metrics

• 10.67 ± 0.78 was the U-Net’s highest mean Dice score, compared with SOC at 10.34 ± 0.77; SOC had the lowest median 95 % Hausdorff distance at 32.02 mm, while PINN-GBM had the best mean at 36.44 ± 1.21 mm.Source: §9.1 · p. 17, Figure 7 · p. 17, Figure 8 · p. 17

9.2 Custom U-Net training details

• Three ensemble members each trained five U-Nets on distinct splits; predictions used each patient’s held-out test model, yielding out-of-sample results.Source: §9.2 · p. 17
• The extended dataset contained 327 patients, with 197 training, 65 validation, and 65 test cases per split; 84 came from four additional datasets.The additional datasets were CPTAC-GBM, IVYGAP, TCGA-GBM, and UPENN-GBM.Source: §9.2 · p. 17
• Inputs included tissue probability maps, tumor segmentation, T1c, FLAIR, and ADC scans; unavailable ADC scans were zero-filled, and registered recurrence segmentations supplied ground truth.Source: §9.2 · p. 17
• The 32-feature U-Net used instance normalization, Gaussian smoothing, and equally weighted BCE and tumor-size-adjusted PCC losses, with voxel weighting and a brain mask.The loss weighted edema voxels at 0.25 and enhancing-core and necrosis voxels at 0.75; outside-brain voxels had zero loss.Source: §9.2 · p. 17
• Grid search selected batch size 7, learning rate 1 × 10−4, Adam, and weight decay 1 × 10−4.Source: §9.2 · p. 17
• Intensity, geometry, and label-smoothing augmentations each improved recurrence coverage; transformations included intensity perturbations, flips and affine changes, and Gaussian label-edge softening.Source: §9.2 · p. 17
• Predictions blended tumor-core-center distance with model output at weights 0.8 and 0.2, respectively, increasing performance by approximately 0.5 %.Source: §9.2 · p. 17