Sunday, November 29 | 1:00 p.m.–2:00 p.m. | S4-SSPD01-8 | Room E352
A deep-learning model can segment and localize multiple lines, drains, and airway devices on pediatric chest x-rays, research in this session suggests.
In a study of 1,000 chest x-rays from 553 neonatal and pediatric patients (median age, 118 days), presenter Hailong Li, PhD, of Cincinnati Children’s Hospital, and colleagues trained a DeepLabV3+ architecture with Atrous Spatial Pyramid Pooling to detect four targeted devices: endotracheal tubes (ETT), tracheostomy tubes, and left and right peripherally inserted central catheters (PICC). Two image analysts manually annotated each device under radiologist supervision, with a second radiologist independently annotating distal tips to establish ground truth. The dataset was split 80:20 for training (799 radiographs) and independent testing (201 radiographs).
The model achieved an overall mean absolute error (MAE) of 7.58 mm and a mean Dice similarity coefficient (DSC) of 0.70 across all four devices. ETT showed the highest tip localization accuracy (MAE, 5.79 mm; DSC, 0.78), followed by left PICC (MAE, 6.26 mm; DSC, 0.63), tracheostomy tubes (MAE, 8.69 mm; DSC, 0.77), and right PICC (MAE, 9.56 mm; DSC, 0.61).
Accurate device localization is especially critical in high-volume neonatal and pediatric settings, the authors noted. Manual review of device positions is time-intensive, and an automated approach could meaningfully reduce radiologist workload in environments such as the neonatal intensive care unit, they added.
“Although performance currently remains below that of human experts, the model offers a computationally efficient approach for automated LDA assessment in neonatal and pediatric populations,” Li and colleagues wrote.
Drop by this Sunday afternoon session on pediatric imaging for more details.



