AI-Assisted MRI Molecular Subtyping in Pediatric Brain Tumors

Pediatric brain tumors are the most common solid tumors in children and represent a highly heterogeneous group of diseases with marked variation in histology, molecular alterations, anatomic location, treatment response, and prognosis. Several molecular features, including H3K27M mutation, BRAF V600E mutation, ZFTA fusion, SMARCB1 loss, CTNNB1 mutation, and TP53 alteration, are clinically important for diagnostic classification, risk stratification, prognosis assessment, and treatment planning. However, most molecular characterization currently depends on postoperative tissue-based testing, and noninvasive preoperative prediction remains limited.

This study is designed to evaluate an AI-assisted MRI-based diagnostic system for pediatric brain tumors in a real-world multicenter observational setting. The study will include seven target tumor categories: glioma, medulloblastoma, ependymoma, atypical teratoid/rhabdoid tumor, intracranial germ cell tumors, craniopharyngioma, and choroid plexus tumors. The retrospective component will collect multimodal data, including clinical variables, preoperative MRI, pathology reports, molecular testing results, treatment information, and follow-up data, from eligible pediatric patients treated from January 1, 2020 through December 31, 2025. The prospective component will consecutively enroll eligible patients from July 15, 2026 onward for additional validation of model performance.

Preoperative MRI data will be preprocessed using standardized procedures, including bias field correction, skull stripping, isotropic resampling, and intensity normalization. The AI model will be developed to support classification of tumor type and prediction of key molecular subtypes/markers from presurgical MRI. Model performance will be evaluated using postoperative pathology and molecular testing as the reference standard. The primary endpoint is the area under the receiver operating characteristic curve (AUC) for prespecified molecular prediction tasks. Secondary analyses will evaluate agreement between AI output and integrated final diagnosis, comparative performance against blinded radiologists on independent test sets, multiclass tumor classification performance, biomarker-specific sensitivity and specificity, and progression-free survival stratified by AI-predicted subgroup.

This study is observational and is not intended for medical device registration. Biospecimen banking is not a registration objective of this study.