Pediatric Personalized Deep Learning Models for Segmentation of Hepatoblastoma at CT and MRI

Purpose To evaluate the generalizability of adult-trained models for hepatoblastoma segmentation to pediatric patients and to develop two deep learning (DL) models, $M_P^{CT}$ and $M_p^{MRI}$ , specifically trained on pediatric contrast-enhanced CT and T2-weighted MRI scans, respectively. Materials and Methods Imaging data from the multicenter Children's Oncology Group AHEP0731 trial (NCT00980460; May 2008-July 2018) were analyzed. DL models employing the three-dimensional U-Net architecture were trained using D_CT-Train and D_MRI-Train. These models were evaluated on D_CT-Val and D_MRI-Val using the Dice similarity coefficient (DSC), and model segmentations were compared with manual segmentations from three annotators (R₁, R₂, and R₃), their consensus (R_c), and adult-trained model ( $M_A^{CT}$ ) segmentations. Volume percentage error analysis was performed to evaluate segmentation precision. Results A total of 104 participants (mean age ± SD, 28.2 months ± 30.5; 64 male; D_CT-Train = 56, D_CT-Val = 48) were included in the CT dataset and 123 (31.5 months ± 38.4; 87 male; D_MRI-Train = 50, D_MRI-Val = 73) in the MRI dataset. $M_P^{CT}$ achieved good agreement with consensus segmentation (DSC = 0.86 [95% CI: 0.80, 0.91]) and exhibited higher agreement than $M_A^{CT}$ with R₁ (0.83 vs 0.55), R₂ (0.85 vs 0.55), R₃ (0.84 vs 0.54), and R_c (0.86 vs 0.55) segmentations. Volume percentage error analysis revealed that $M_P^{CT}$ achieved segmentation results on par with or better than those of a novice annotator (R₃) in high-precision scenarios. $M_P^{MRI}$ also achieved a DSC of 0.86, demonstrating good agreement with R_c. Conclusion The pediatric-trained DL-based models outperformed adult-trained models for accurate segmentation of pediatric hepatoblastoma. Keywords: Pediatrics, Deep Learning, Liver, MR-Imaging, Abdomen/GI, Algorithm Development ClinicalTrials.gov NCT00980460 Supplemental material is available for this article. © The Author(s) 2026. Published by the Radiological Society of North America under a CC BY 4.0 license.