Deep learning models for segmenting pediatric tumors are typically trained on small, single-center cohorts, which limits their robustness. In contrast, adult kidney cancer datasets are far larger. A new preprint on arXiv investigates whether models trained on adult renal lesions can be applied to pediatric Wilms tumors on MRI.

The authors frame this as a cross-domain generalization problem, asking whether adult-trained models can transfer to the pediatric domain, where tumor appearance and imaging characteristics may differ. The preprint's abstract outlines the motivation and the question, though it does not include specific results in the available text.

If adult-trained models generalize well, they could provide a starting point for pediatric segmentation without requiring large pediatric datasets. If not, the study would highlight the need for domain adaptation or dedicated pediatric data collection. The preprint is available on arXiv.