Author: Wesley E. Bolch, Emily L. Marshall, Dhanashree Rajderkar, Wyatt Smither 👨🔬
Affiliation: University of Florida 🌍
Purpose: To determine the accuracy of TotalSegmentator, an AI-based automatic segmentation toolkit, on pediatric CT scans as the original software was trained on adult image datasets with a mean patient age exceeding 60 years.
Methods: An IRB protocol was written and approved to obtain 46 CT image datasets from pediatric patients ranging from newborn to 9.25 years old from the UF Health Shands hospital; 2 per month from newborn to 1 year (24 total), 5 total between 2.75 and 3.25 years, 5 total between 4.75 and 5.25 years, 5 total between 6.75 and 7.25 years, and 5 total between 8.75 and 9.25 years. CT image datasets were from scans using a chest-abdomen-pelvis (CAP) protocol only. No specific disease state or medical diagnosis indication for the CT scan was targeted. Collected CAP scans were run through TotalSegmentator with the resulting organ segmentations being exported and developed into 3D mesh models (OBJ) as well as overlaying the produced segmentations on the original CT scans in each imaging plane. Segmentations of the pelvis, liver, kidneys, and adrenals were chosen as organs of interest to assess the effects of different tissue densities as a function of age and the pelvis was used as a stationary landmark for registration.
Results: A pediatric radiologist scored the accuracy of both a reconstructed 3D model and overlayed segmentations on the original CT scan on a scale from 1 (complete accuracy) and 5 (complete inaccuracy) to assess the accuracy of TotalSegmentator for each of the organs of interest across the patient population.
Conclusion: TotalSegmentator is a robust, AI-based automatic segmentation toolkit that can accurately segmentation over 100 individual organs from both CT and MRI scans. However, the software was trained on adult image datasets. This study investigated the accuracy of the AI model on pediatric patient CT image datasets.