Biological domain shift and statistical nesting concerns in generative AI-based spatial tumor growth prediction for pediatric diffuse midline glioma

Abstract

Background

Laslo et al. recently reported a guided denoising diffusion implicit model for spatial tumor growth prediction on magnetic resonance imaging (MRI) in pediatric diffuse midline glioma. Their proof-of-principle study demonstrates the feasibility of generative artificial intelligence (AI) for producing patient-specific tumor growth maps as an early step toward informing personalized radiotherapy planning in data-limited pediatric neuro-oncology settings. However, the external validation cohort comprised only 13 patients, and growth-region prediction performance, measured using the continuous Dice coefficient (cDICE; median ≈ 0.22; range 0.071–0.376), was considerably weaker than full-tumor performance (cDICE median ≈ 0.81; range 0.439–0.877). In this Matters Arising, we provide a focused methodological commentary on several issues that should be considered when interpreting the translational implications of this work.

Main body

First, training both the diffusion model and the tumor-size regressor on pooled adult glioblastoma and pediatric high-grade glioma data may introduce biological domain shift, given differences in molecular drivers, anatomical distribution, growth kinetics, and treatment response. Second, although patient-level data splitting was appropriately performed, slice-level performance estimates may overstate precision because multiple correlated two-dimensional slices are nested within a small number of patients. Third, clinical utility should be evaluated against radiotherapy-relevant benchmarks, including target-volume delineation, isotropic expansion margins, geographic miss, normal-tissue exposure, and growth-region-specific performance.

Conclusions

Addressing these points would strengthen the evidentiary basis for future clinical translation of generative tumor growth modeling.

Laslo et al. [1] are to be commended for exploring guided denoising diffusion implicit models (DDIM) for spatial tumor growth prediction in pediatric diffuse midline glioma (DMG). By training both the diffusion model and the tumor-size regressor on pooled adult glioblastoma (GBM; n = 1,251) and pediatric high-grade glioma data from the pediatric Brain Tumor Segmentation dataset (BraTS-PED), including DMG and astrocytoma cases (n = 144), the authors demonstrate the potential of generative artificial intelligence (AI) to augment limited pediatric datasets. The model produces images that five pediatric neuroradiology domain experts could not reliably distinguish from real scans (accuracy 0.53 ± 0.03), and predicted growth maps achieve a mean continuous Dice coefficient (cDICE) of 0.79 ± 0.13 (95% confidence interval [CI], 0.767–0.805) against longitudinal follow-up in an external cohort of 13 DMG patients. In the longer term, with rigorous validation, such models could inform personalized radiotherapy planning by predicting likely invasion patterns and enabling adaptive target volume definition—an important clinical need in DMG where gross tumor volume increase during treatment is well documented. While the expert classification performance confirms the visual realism of the generated images, visual fidelity does not by itself demonstrate that the predicted direction and extent of tumor growth meet the accuracy required for clinical decision-making. Several methodological considerations merit discussion, centered on whether adult GBM data enhance or bias pediatric DMG predictions and the strength of the evidence supporting clinical translation. We note that this article is a focused methodological commentary on the published study design and interpretation, not an independent replication, and our recommendations are intended to guide future validation efforts.

A primary concern relates to the biological domain shift inherent in training both the diffusion model and the tumor-size regressor on pooled adult GBM and pediatric high-grade glioma data, followed by evaluation in pediatric DMG. These entities differ fundamentally in their molecular drivers—adult GBM is frequently characterized by isocitrate dehydrogenase (IDH)-wildtype status and alterations such as epidermal growth factor receptor (EGFR) amplification, whereas DMG is defined by H3 K27 alteration—as well as in anatomical predilection, growth kinetics, and treatment response [2]. The key question is whether adult GBM data improve generalization or introduce systematic bias: adult cases comprise approximately 90% of the training data, and a model learning predominantly from supratentorial adult tumors and their peritumoral environment may internalize growth patterns that do not faithfully reflect infratentorial pediatric DMG progression. The authors’ own data lend support to this concern: the pediatric subset yielded a higher Fréchet Inception Distance (FID, 16.0 vs. 13.6 for adult) and greater variability in structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) metrics (Fig. 3B of the original article), which may be consistent with less faithful modeling of pediatric-specific features. Domain shift between training and target populations is a recognized failure mode in medical deep learning [3]. Although the authors justify this pooled adult–pediatric training strategy by shared anatomical structures between adult and pediatric brains, the extent to which adult GBM-derived priors influence pediatric DMG predictions remains unquantified. Ablation analyses comparing adult-only, pediatric-only, and domain-adapted model variants would help resolve this uncertainty.

A separate but related issue concerns the statistical independence of slice-level performance estimates. The authors appropriately split the BraTS data at the patient level with no subject overlap between training, validation, and test sets. However, adjacent two-dimensional slices from the same three-dimensional magnetic resonance imaging (MRI) volume are highly correlated, and the 178 paired slices used for longitudinal evaluation are nested within only 13 patients. While the mean cDICE of 0.79 was computed across all 178 slices, the effective sample size for quantifying uncertainty is considerably smaller than 178 when within-patient clustering is unaccounted for. Assuming a moderate within-patient intraclass correlation, the effective sample size for patient-level inference could be substantially reduced, yielding wider CIs than those currently reported. Reporting patient-level bootstrap CIs or mixed-effects sensitivity analyses would better reflect the true precision of performance estimates and strengthen the interpretation of the external validation results.

On a clinical note, the evaluation relies primarily on image-quality metrics (SSIM, PSNR, FID) and a visual Turing test. While these assessments demonstrate the model’s capacity to generate realistic images, they do not directly address the question that clinicians face: would these growth predictions improve treatment decisions? The authors highlight personalized radiotherapy planning as a potential application, yet no comparison was made between model-predicted tumor boundaries and those delineated by radiation oncologists for target volume definition. Clinically meaningful benchmarks would include comparisons with radiation oncologist-delineated target volumes according to established International Commission on Radiation Units and Measurements (ICRU) reporting standards [4], conventional isotropic expansion margins, and treatment-planning endpoints such as geographic miss and normal-tissue exposure. Such evaluation should include not only quantitative overlap metrics but also blinded qualitative review by radiation oncologists regarding whether predicted growth maps would alter target-volume delineation or adaptive planning decisions. The clinical relevance of such benchmarks is underscored by recent evidence that pediatric-type diffuse high-grade gliomas of midline structures frequently require adaptive radiotherapy due to substantial gross tumor volume increase during treatment [5]. Additionally, the supplementary patient-level data show that growth-region cDICE values (median ≈ 0.22; range 0.071–0.376; Table S3) are considerably lower and more variable than full-tumor cDICE values (median ≈ 0.81; range 0.439–0.877). This distinction matters because radiotherapy adaptation depends more on the predicted direction and extent of newly involved tissue than on overlap with the pre-existing tumor bulk.

Lastly, the external validation cohort, while drawn from multiple institutions, comprises only 13 patients. The authors acknowledge this limitation but frame it as partially compensated by the multi-institutional nature of the data. That argument warrants closer examination: multi-institutional variability increases heterogeneity, which is valuable for assessing generalizability, but does not substitute for adequate statistical power. The reported 95% CI for cDICE (0.767–0.805) pertains to slice-level estimates and should be interpreted cautiously given the nesting structure noted above. The considerable between-patient variability in growth-region prediction across merely 13 patients underscores the need for larger, more representative validation samples before clinical translation can be considered.

These considerations do not diminish the value of Laslo et al.’s contribution—the application of DDIM-guided generation to tumor growth modeling is genuinely novel, and the publicly available code is commendable. The most pressing priority, in our view, is clinically oriented evaluation against radiation oncologist-delineated target volumes and growth-region-specific benchmarks, which directly addresses the translational gap between image realism and treatment-planning utility. A second, closely related concern is quantifying whether adult GBM training data introduce systematic bias in pediatric DMG predictions through the ablation analyses proposed above. Addressing these concerns, together with the statistical nesting considerations outlined above, would further strengthen the translational potential of this promising approach.

Abbreviations

AI

artificial intelligence

BraTS-PED

pediatric Brain Tumor Segmentation dataset

cDICE

continuous Dice coefficient

CI

confidence interval

DDIM

denoising diffusion implicit model

DMG

diffuse midline glioma

EGFR

epidermal growth factor receptor

FID

Fréchet Inception Distance

GBM

glioblastoma

ICRU

International Commission on Radiation Units and Measurements

IDH

isocitrate dehydrogenase

MRI

magnetic resonance imaging

PSNR

peak signal-to-noise ratio

SSIM

structural similarity index measure

Author contributions

WL and XS contributed equally to Conceptualization, Methodology, and Writing – original draft. WL, XS, and SP contributed to Writing – review & editing. SP contributed to Supervision. All authors read and approved the final manuscript.

Funding

No specific funding was received for this work.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests