This algorithm automatically detects and localizes endotracheal tubes and the central airways.  It is valuable to assess the position/malposition of these devices, and can be used to compute the position of the endotracheal tube tip relative to the carina, which is an important metric that guides whether the tubes need to be repositioned.

Value Proposition

Detection of endotracheal tube position relative to the carina and mainstem airway is necessary for confirmation of tube positioning.

Intended Use

Any PA/AP Chest Radiograph

Information on training data

Endotracheal/Tracheostomy tube segmentation models were trained on 771 from a public dataset of frontal chest radiographs and tested on 291 private radiographs. Two models for tube tip and carina localization were trained: (a) without and (b) with central airways and tube annotations. Annotations and model training were performed by a physician post-doctoral research fellow.

Model performance metrics

Inclusion of central airways and tube entirety improved tube detection AUC from 0.610 to 0.894.
Mean error in tube tip and carina localization were 19.7 and 13.3 mm, and improved to 10.2 and 6.4 mm with
concurrent training with the central airways and the entirety of the tube.

Additional Media

Additional Media


AiDA Lab at UC San Diego

AiDx™ is a mark of The Regents of the University of California.
The University California, San Diego is one of the world's leading public research universities, located in beautiful La Jolla, California

Brian Hurt

Brian Hurt