Two-stage deep learning pipeline for endotracheal tube detection on pediatric chest radiographs

A two-stage deep learning model detected endotracheal tubes on pediatric chest X-rays with a 0.994 area under the curve (95% CI 0.986-1.000) and localized tube tips with 6.59 mm mean error; inter-reader error was 2.01 mm; single-center pilot.

• Retrospective single-center study of 1,000 pediatric chest radiographs (476 with endotracheal tubes, 524 without).
• Adding a ResNet classification stage before U-Net segmentation cut false-positive segmentations on tube-negative films from 17 to 3.
• Tip localization error was 6.59 mm (95% CI 5.19-8.21), compared with 2.01 mm inter-observer variability.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.