Deep learning volumetrics reveal distinct clinical trajectories for pediatric low-grade gliomas under surveillance: A...

Pediatric low-grade gliomas (pLGGs) have heterogeneous clinical presentations, and given the morbidity of treatment, some patients receive observation with magnetic resonance (MR). The natural histories of untreated pLGGs remain understudied. We leveraged deep learning-based volumetrics to analyze longitudinal growth trajectories and progression risk factors for untreated pLGGs.

We conducted a pooled, retrospective study of radiographically diagnosed pLGG patients from two institutions diagnosed between 1992 and 2020 who were surveilled for at least 1 year post-diagnosis. Tumor segmentation was applied to longitudinal T2-weighted MR to calculate 3D tumor volumes. We assessed volume trajectories, disease progression, and associated risk factors using Cox-Hazards regression, survival analysis, and time-series forecasting with autoregressive integrated moving average (ARIMA). Patients were categorized based on volumetric changes into progression (≥25%), regression (≤-25%), or stability.

Of 99 patients (970 scans; median follow-up: 7.0 years; median diagnosis age: 12.0 years), 55 (55.5%) had tumors that volumetrically progressed, 28 (28.3%) remained stable, and 16 (16.2%) regressed. 42 (42.4%) patients initiated treatment. Risk factors associated with progression included infancy/preschool age, cortical location, and female sex (p ≤ 0.05 for each). Most progressions occurred within five years of diagnosis (80.0%), most commonly in school-aged children (7-13 years old). Time-series forecasting predicted future tumor volume with a mean absolute error of 2.04 cm³.

Deep learning enables systematic, longitudinal, pLGG growth tracking and characterization of patients on surveillance, yielding insights into untreated tumor trajectories and progression risk. This pipeline is useful at population-level to study growth trends and at patient-level to guide personalized management.