Radiomics can extract quantitative imaging biomarkers with potential for tumor characterization and predictive modeling, but feature reproducibility remains a major barrier to clinical translation. Variations in MRI preprocessing and tumor segmentation can substantially alter radiomic measurements, making standardized workflows essential.

This international multicenter study evaluated radiomics reproducibility in 419 pediatric patients with neuroblastic tumors, analyzing 746 T2/T2*-weighted MRI sequences and 107 radiomic features. The impact of denoising, bias-field correction, normalization, resampling and segmentation-mask modifications was systematically assessed.

The results reveal a striking dependence on image preparation. Without z-score normalization, only 5% of radiomic features remained highly reproducible, making normalization the most influential processing step. By contrast, removing inhomogeneity correction had the smallest effect, with 84% of features remaining reproducible.

Importantly, shape features remained highly robust to small segmentation variations, whereas first- and second-order features were more sensitive.

As the first reproducibility study of radiomics in pediatric tumors, these findings provide practical evidence for standardized MRI preprocessing and more robust, generalizable radiomics-based biomarkers in pediatric oncology.