Deep learning model improved cancer recurrence prediction for pediatric glioma using serial imaging
1. Tak and colleagues developed a temporal deep-learning model to predict one-year pediatric glioma recurrence using surveillance magnetic resonance imaging (MRI).
2. The deep learning model improved recurrence prediction for both low- and high-grade glioma compared to conventional approaches.
Evidence Rating Level: 2 (Good)
Study Rundown: Pediatric glioma recurrence can cause significant morbidity and mortality, but recurrence patterns and severity are challenging to predict with established clinical and genomic markers. Tak and colleagues developed a temporal learning model that analyzes serial MRI scans to predict recurrence within one year of the latest scan. The temporal learning model was compared with models using a single scan, contrastive pretraining, or conventional longitudinal training. The primary outcomes were the area under the receiver operating characteristic curve (AUROC), F1 score, specificity, and sensitivity. Across both the low- and high-grade glioma datasets, the temporal learning model achieved an AUROC of 0.75-0.89 and improved F1 scores by up to 58.5% over standard longitudinal models. Recurrence prediction performance increased incrementally with the number of historical scans available per patient, reaching plateaus between 3-6 scans. This study demonstrated that deep learning models have the potential to improve cancer surveillance, leading to possible earlier intervention.
Click here to read the study in NEJM AI
Relevant Reading: Harnessing artificial intelligence for predicting breast cancer recurrence: a systematic review of clinical and imaging data
In-Depth [retrospective cohort]: 3,994 post-operative MRI images from 715 pediatric patients were analyzed. The data came from four datasets: Dana-Farber Cancer Institute/Boston Children’s Hospital (DFCI/BCH) low-grade glioma patients, Children’s Brain Tumor Network (CBTN), RadART, and DFCI/BCH high-grade glioma patients. All patients had primary surgery and at least one year of clinical and radiographic follow-up. The primary performance metrics for the deep learning model were the area under the receiver operating characteristic curve (AUROC), F1 score, specificity, and sensitivity. The study found that the model achieved AUROCs of 0.77 (95% confidence interval [CI], 0.62-0.91) on the DFCI/BCH low-grade dataset, 0.53 (95% CI, 0.35-0.70) on the CBTN dataset, 0.72 (95% CI, 0.50-0.92) on the RadART data set, and 0.82 (95% CI, 0.68-0.94) on the DFCI/BCH high-grade dataset. Additionally, sensitivity for detecting glioma recurrence was increased, with inconsistent improvements in specificity across the datasets. Recurrence prediction performance increased incrementally with the number of historical scans available per patient, reaching plateaus between 3-6 scans. This study was limited by possible institutional differences in treatment and imaging schedules and lack of prospective evidence that improved prediction improves decisions or outcomes. Overall, this study demonstrated that temporal deep learning models may enable effective longitudinal medical imaging analysis and point-of-care decision support for pediatric brain tumors and possibly predict risk in patients with other cancers and chronic diseases undergoing surveillance imaging.
Image: PD
©2026 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.
