2399 - Imaging-Based Deep Learning for Brain Age Prediction following Radiation Therapy in Pediatric Medulloblastoma
Presenter(s)
A. Zapaishchykova1, S. Zhao2, D. P. I. Capaldi3, J. L. Nakamura4, S. Vajapeyam5, S. M. Perkins6, K. X. Liu1, H. Averts7, D. A. Haas-Kogan8, S. Meuller9, T. Y. Poussaint10, H. Elhalawani1, N. Rammohan11, and B. H. Kann12; 1Brigham and Women's Hospital, Boston, MA, 2Department of Radiation Oncology, WashU Medicine, St. Louis, MO, 3Imaging Research Laboratories, Robarts Research Institute, Western University, London, ON, Canada, 4USCF, Department of Radiation Oncology, San Francisco, CA, 5Department of Radiology Boston Children's Hospital, Boston, MA, 6WashU Medicine, St. Louis, MO, 7Brigham And Women's Hospital, Boston, MA, 8Department of Radiation Oncology, Mass General Brigham, Boston, MA, 9UCSF Benioff Children's Hospital, San Francisco, CA, 10Department of Radiology, Boston Children's Hospital, Boston, MA, 11Washington University School of Medicine in St. Louis, Department of Radiation Oncology, St. Louis, MO, 12Department of Radiation Oncology, Mass General Brigham/Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA
Purpose/Objective(s):
CSI improves survival in patients with pediatric medulloblastoma but causes long-term neurotoxicity. We aimed to characterize longitudinal brain age gap (BAG) and volumetric trajectories following CSI, comparing patients to matched healthy controls and evaluating CSI dose effects.Materials/Methods:
We applied AgeDiffuse (deep learning brain age prediction) to T1-weighted (T1w) MRIs and SynthSeg for substructure segmentation in patients with medulloblastoma with longitudinal MRI from two institutions. Patients were compared to matched controls from nine public datasets using Wilcoxon rank-sum tests.Results:
We analyzed 261 T1w MRIs from 46 patients (Institution 1: N=211 scans, 28 patients, median 7 scans/patient; 29% female; median age at diagnosis 8.4 years [IQR 6.4–12.2]; median follow-up 2.2 years [IQR 1.0–4.8]; Institution 2: N=50 scans, 18 patients, median 2 scans/patient; 61% female; median age at diagnosis 8.5 years [IQR 6.3–11.0]; median follow-up 6.8 years [IQR 4.9–10.1]). Pre-surgery BAG was elevated vs controls (median 3.52 [IQR 1.77, 4.25] vs 0.70 [-0.42, 1.54] years; p<0.001). BAG increased at 3 months (6.15 [3.73, 9.17]) and 6 months (6.78 [3.38, 8.74]), declined by 3 years (1.58 [0.55, 5.81]), and stabilized at 5 years (4.47 [0.64, 6.02]), remaining above controls at all timepoints (all p<0.01). BAG did not differ significantly by dose (RT High 36 Gy, n=12; RT Low 23.4 Gy, n=34) at baseline (p=0.77) or 3 years (p=0.45); the 5-year comparison was precluded by sample size (RT High n=1). At 1 year, CSI patients had lower white matter (-57.2 cm3; p<0.001) and greater ventricular volumes (+17.7 cm3; p<0.001) vs controls. High-dose CSI trended toward greater ventricular volumes vs low-dose (1 year p=0.05; 3 years p=0.09). Substructural volumes explained 9% of BAG variance (R² = 0.085), suggesting that BAG captures features beyond volumetrics.Conclusion:
Pediatric patients with medulloblastoma had elevated BAG before CSI, suggesting tumor and surgical effects on brain age estimation. Post-CSI, BAG increased before partially normalizing, possibly reflecting ongoing neurodevelopment, though remaining persistently elevated. Higher CSI dose trended toward greater ventricular enlargement but not differential BAG trajectories. Deep learning-derived brain age and substructural volumes capture distinct, persistent neuroanatomical changes following CSI and should be validated against longitudinal neurocognitive outcomes in survivors of pediatric brain tumor.