Longitudinal risk prediction for pediatric glioma with temporal deep learning

Abstract

Pediatric glioma recurrence following surgery causes morbidity and mortality, and thus, children undergo frequent longitudinal magnetic resonance (MR) surveillance postoperatively to inform management. However, the pattern and severity of pediatric glioma recurrences are highly variable and challenging to predict with current clinical and genomic stratifications. Quantitative imaging analyses have shown promise for cancer risk prediction, and longitudinal analysis of glioma MR may improve the ability to predict future recurrence. Here, we propose a novel self-supervised, deep learning approach to longitudinal brain MR analysis, temporal learning, that models the spatiotemporal information from a patients prior, longitudinal brain MRs to predict future recurrence. We apply temporal learning to pediatric glioma surveillance imaging for 715 patients (3,994 scans) from four distinct clinical settings. We find that longitudinal imaging analysis with temporal learning improves recurrence prediction performance by up to 41% compared to training from scratch, with improvements in performance in both low- and high-grade glioma. We find that recurrence prediction accuracy increases incrementally with the number of historical scans available per patient. Temporal deep learning may enable point-of-care decision-support for pediatric glioma to personalize surveillance and postoperative therapy.