Author: Gregory T. Armstrong, James E. Bates, Christine V. Chung, Lei Dong, Ralph Ermoian, Jie Fu, Christine Hill-Kayser, Rebecca M. Howell, Meena S. Khan, Sharareh Koufigar, John T. Lucas, Thomas E. Merchant, Taylor Meyers, Tucker J. Netherton, Constance A. Owens, Arnold C. Paulino, Sogand Sadeghi 👨🔬
Affiliation: Department of Radiation Oncology, University of Washington and Fred Hutchinson Cancer Center, Department of Epidemiology and Cancer Control, St. Jude Children’s Research Hospital, St. Jude Children's Research Hospital, Department of Radiation Oncology, Fred Hutchinson Cancer Center, University of Washington, Department of Radiation Oncology, St. Jude Children’s Research Hospital, Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Department of Radiation Oncology, University of Washington/ Fred Hutchinson Cancer Center, Department of Radiation Oncology, University of Pennsylvania, University of Pennsylvania, Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Department of Radiation Oncology and Winship Cancer Institute, Emory University, The University of Texas MD Anderson Cancer Center UTHealth Houston Graduate School of Biomedical Sciences 🌍
Purpose: Clinical workflows often rely on auto-segmentation tools trained on adult data, which may exhibit suboptimal performance in pediatric imaging due to inherent anatomical variations and smaller structures. This study comprehensively evaluates the performance of pediatric trained deep learning-based segmentation models (nnUNet, SwinUNeTr, and 3DSegResNet) relative to adult trained commercial platforms (MIM and RayStation) trained on adult-cohorts.
Methods: A total of 138 non-contrast CT scans (35-abdomen/pelvis, 46-thorax, 57-brain) from pediatric patients (aged 1–20 years, 69-male and 69-female) (collected for a Childhood Cancer Survivor Study pilot assessing cohort expansion feasibility) were included. Training involved 96 patients; 42 were reserved for validation. Preprocessing included resampling and intensity normalization, followed by data augmentation (random cropping, rotations, axis flips).. A 5-fold cross-validation strategy was utilized, averaging performance metrics (Dice similarity coefficient (DSC) and average Hausdorff distance (AHD)) across all iterations.
Results: Across all body regions, pediatric trained deep learning models (nnUNet DSC: 0.91±0.03, AHD: <2.05 mm; SwinUNeTr DSC: 0.88±0.02, AHD: <1.9 mm; 3DSegResNet DSC: 0.87±0.04, AHD: <2.78 mm) outperformed all adult-trained models (MIM DSC: 0.7–0.79, AHD: 11.46–28.21 mm; RayStation DSC: 0.65–0.75, AHD: 10.56–27.07 mm). Across pediatric trained deep learning models, performance (DHC & AHD) was consistent across each age group (nnUNet DSCs: >0.91, AHDs <2.01 mm; SwinUNeTr DSCs: 0.87±0.03, AHDs: 2.1±0.54 mm; 3DSegResNet DSCs: 0.88±0.05, AHDs: 2.81±0.39 mm) and improved relative to commercial platforms (MIM DSCs: 0.71±0.22 - 0.77±0.21; RayStation DSCs: 0.67±0.22 - 0.72±0.21).
Conclusion: Deep learning-based models trained on pediatric cohorts demonstrated superior performance in auto-segmentation compared to adult-trained clinical tools across all body regions and age groups. These findings highlight the potential to improve pediatric-centric radiotherapy workflows with dedicated AI models. Furthermore, the consistent performance of pediatric-trained models across all age groups suggests that age-specific models may not always be necessary, highlighting their adaptability and robustness across diverse pediatric cohort.