Accurate tumor segmentation is essential for quantitative imaging analysis in neuroblastoma, but manual delineation is time-consuming, observer-dependent and difficult to scale. Automated deep learning could overcome these limitations, but robust external validation is critical before such tools can be translated into clinical workflows.

This study independently validated an nnU-Net deep learning model for automatic detection and segmentation of neuroblastic tumors on MRI using a large, heterogeneous dataset completely independent from model development. The cohort included 300 children and 535 T2/T2-weighted MRI sequences*, acquired across multiple institutions, scanner vendors, field strengths and imaging protocols.

The model successfully located and segmented the primary tumor in 94% of cases, achieving an exceptionally high median Dice Similarity Coefficient of 0.997. Performance was not significantly affected by tumor location, MRI field strength, sequence type or whether imaging was performed after chemotherapy.

Importantly, the approach also substantially reduced radiologist workload: visual review required only 7.8 seconds on average, while cases requiring manual correction took approximately two minutes, representing a 96.5% reduction in segmentation time compared with manual segmentation from scratch.

As the first international, multicenter and multivendor external validation of automatic neuroblastic tumor segmentation on body MRI, the study provides strong evidence of model generalizability and represents an important step toward scalable AI-assisted quantitative imaging in pediatric oncology.