Purpose or Case Report:

Deep learning models have become instrumental in medical imaging, with a wide array of applications such as automated segmentation and diagnosis. However, the majority of these models are trained on adult imaging datasets, leading to underperformance when applied to the pediatric population. This discrepancy arises from anatomical differences between these populations, posing a significant challenge in fields requiring high precision like spine segmentation in Computed Tomography (CT) scans.
Methods & Materials:

In this study, we address this challenge by creating a training dataset for whole body CT segmentation in children. This dataset includes 358 CT scans of patients between 0-16 years of age.
Results: Segmentations of 117 anatomic structures were created through manual refinement of outputs provided by the TotalSegmentator CT segmentation model. Additional classes, specific to the pediatric population were added (e.g. thymus). Based on this training data set we trained an nnUNet-based deep learning model for automated pediatric CT segmentation that outperformed TotalSegmentator specifically for the pediatric population.
Conclusions:
Our study demonstrates the importance of developing specialized datasets tailored to the pediatric population. In future work, we aim to expand the dataset and explore additional anatomical regions, further enhancing the robustness and applicability of our approach in pediatric and adult medical imaging.