Multiparametric MRI radiomics for noninvasive risk stratification of pediatric neuroblastoma: a pilot study.
This retrospective single-center pilot study of 30 children found that pretreatment multiparametric MRI radiomics classified low/intermediate- versus high-risk neuroblastoma with a best cross-validated AUC of 0.88, with T2-weighted and diffusion-derived features contributing most.
Open original publication →What the AI sees
This retrospective single-center pilot study of 30 children found that pretreatment multiparametric MRI radiomics classified low/intermediate- versus high-risk neuroblastoma with a best cross-validated AUC of 0.88, with T2-weighted and diffusion-derived features contributing most.
Research significance
Evidence: MRI radiomic models showed preliminary ability to reproduce clinical neuroblastoma risk groups noninvasively. Inference: after external prospective validation, such models might complement—not replace—tissue and molecular testing by supporting earlier treatment stratification or biopsy planning, but no improvement in treatment selection or outcomes was demonstrated.
Source abstract
Neuroblastoma is the most common extracranial solid tumor in children, with risk stratification guiding therapy and prognosis. Although current risk stratification incorporates imaging-based staging, definitive risk assignment still relies on tissue and molecular characterization, highlighting the need for complementary noninvasive imaging biomarkers. The objectives were to evaluate the classification performance of multiparametric magnetic resonance imaging (MRI) radiomics for risk stratification of pediatric neuroblastoma and to determine how MRI sequences, feature-selection methods, and machine learning classifiers influence classification performance. This retrospective single-center feasibility study included 30 children with histologically confirmed neuroblastoma who underwent pre-treatment T1-, T2-, and diffusion-weighted MRI. From each sequence, 208 radiomic features were extracted from the whole-tumor volume and reduced using six feature-selection methods. Principal components of selected features trained six machine learning classifiers. Performance was assessed using a nested leave-one-out cross-validation framework, with predictions aggregated to one per patient before computing performance metrics, for binary classification of low/intermediate-risk versus high-risk neuroblastoma, using clinical risk classification as the reference standard, with pairwise differences evaluated by DeLong test and Benjamini-Hochberg correction. Among 30 children (mean age ± SD: 38 ± 40 months), the highest discrimination between risk groups was achieved using T2-weighted features and the combined T1w + T2w + ADC features, both with XGB (AUC = 0.88 ± 0.06 and 0.88 ± 0.07, respectively); however, the limited sample size prohibited the detection of significant differences between classifiers after correction for multiple comparisons. Features derived from T2-weighted and diffusion-weighted MRI contributed most to accurate classification. The chi-square feature selection method most frequently contributed to high-performing model configurations (30.8%). Multiparametric MRI radiomics based on whole-tumor volumes showed preliminary evidence of feasibility for noninvasive risk stratification of pediatric neuroblastoma, supporting its potential as a complementary imaging biomarker.