Multimodal deep learning improves recurrence risk prediction in pediatric low-grade gliomas

Area of Science:

• Neuro-oncology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Pediatric low-grade gliomas (pLGGs) pose challenges in predicting postoperative recurrence using conventional factors.
• Novel methods are needed to improve risk stratification for pLGGs post-surgery.

Purpose of the Study:

• To investigate the utility of deep learning (DL) applied to magnetic resonance imaging (MRI) tumor features for enhancing postoperative risk stratification in pLGGs.
• To compare the predictive performance of DL-derived imaging features against clinical and multimodal models.

Main Methods:

• A pretrained DL tool was used to extract imaging features from preoperative T2-weighted MRI scans of pLGG patients from two institutions.
• Three DL logistic hazard models were trained: clinical features only, DL-MRI features only, and multimodal (clinical + DL-MRI features).
• Model performance was evaluated using time-dependent Concordance Index (Ctd) and Kaplan-Meier analysis for risk group stratification.

Main Results:

• The multimodal DL model demonstrated superior event-free survival (EFS) prediction (Ctd: 0.85) compared to DL-MRI (0.79) and clinical models (0.72).
• The multimodal model significantly improved risk stratification, differentiating between high-risk (3-year EFS: 31%) and low-risk (92%) groups (P < .0001).
• A total of 396 patients were analyzed, with a median follow-up of 85 months; 28% experienced recurrence.

Conclusions:

• Deep learning effectively extracts imaging features from MRI that are informative for predicting pLGG recurrence.
• Multimodal DL models integrating imaging and clinical data significantly enhance postoperative risk stratification for pLGGs.
• Further validation with larger, multicenter datasets may be necessary to improve the generalizability of these predictive models.