Longitudinal risk prediction for pediatric glioma with temporal deep learning

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Oncology

Background:

• Pediatric glioma recurrence is difficult to predict, leading to frequent, non-personalized surveillance imaging.
• Current methods lack precision in identifying individual recurrence risk, necessitating extensive monitoring for all patients.

Purpose of the Study:

• To develop and validate a novel deep learning approach for predicting pediatric glioma recurrence using longitudinal magnetic resonance (MR) imaging.
• To improve the accuracy and personalization of surveillance strategies for pediatric brain tumors.

Main Methods:

• A self-supervised deep learning model, termed temporal learning, was developed to analyze spatiotemporal information from sequential brain MR scans.
• The model was applied to a dataset of 715 pediatric glioma patients with 3,994 MR scans from four clinical settings.

Main Results:

• Temporal learning significantly improved recurrence prediction performance by up to 41% compared to traditional methods.
• Prediction accuracy increased with the number of historical MR scans available for each patient.
• Performance gains were observed for both low- and high-grade pediatric gliomas.

Conclusions:

• Longitudinal MR imaging analysis using temporal deep learning enhances the prediction of pediatric glioma recurrence.
• This approach may enable point-of-care decision support for pediatric brain tumors and similar surveillance needs in other cancers.