Generative AI for spatial tumor growth on MRI: a proof-of-principle study in pediatric diffuse midline glioma

Area of Science:

• Artificial Intelligence in Medical Imaging
• Neuro-oncology Research
• Pediatric Cancer Imaging

Background:

• Magnetic resonance imaging (MRI) is crucial for neuro-oncology diagnosis and monitoring.
• Predicting spatial tumor progression based on patient anatomy is increasingly important.
• This study focuses on pediatric Diffuse Midline Glioma (DMG).

Purpose of the Study:

• To present a proof-of-principle for personalized spatial tumor progression prediction on MRI using generative AI.
• To model anatomical tumor growth in pediatric DMGs.

Main Methods:

• Employed guided Denoising Diffusion Implicit Models (DDIM) for tumor growth modeling.
• Trained a slice-based framework on multiparametric MRI scans from adult and pediatric patients.
• Generated probabilistic tumor growth maps conditioned on baseline scans and target tumor size.

Main Results:

• Generated anatomically coherent, patient-specific T2-FLAIR MRI slices.
• Expert evaluations confirmed high image quality, with radiologists unable to reliably distinguish generated from real scans.
• Tumor growth probability maps showed strong alignment with observed tumor growth (DICE score of 0.79).

Conclusions:

• Guided DDIMs show potential as a predictive tool for spatial tumor growth, particularly for DMGs.
• This technology could be integrated into personalized radiotherapy planning.
• Careful evaluation of synthetic data is essential for research and clinical integration.