Validation of a Deep Learning–based Automatic Detection Algorithm for Measurement of Endotracheal Tube–to–Carina Distance on Chest Radiographs

Abstract

Background Improper endotracheal tube (ETT) positioning is frequently observed and potentially hazardous in the intensive care unit. The authors developed a deep learning–based automatic detection algorithm detecting the ETT tip and carina on portable supine chest radiographs to measure the ETT–carina distance. This study investigated the hypothesis that the algorithm might be more accurate than frontline critical care clinicians in ETT tip detection, carina detection, and ETT–carina distance measurement. Methods A deep learning–based automatic detection algorithm was developed using 1,842 portable supine chest radiographs of 1,842 adult intubated patients, where two board-certified intensivists worked together to annotate the distal ETT end and tracheal bifurcation. The performance of the deep learning–based algorithm was assessed in 4-fold cross-validation (1,842 radiographs), external validation (216 radiographs), and an observer performance test (462 radiographs) involving 11 critical care clinicians. The performance metrics included the errors from the ground truth in ETT tip detection, carina detection, and ETT–carina distance measurement. Results During 4-fold cross-validation and external validation, the median errors (interquartile range) of the algorithm in ETT–carina distance measurement were 3.9 (1.8 to 7.1) mm and 4.2 (1.7 to 7.8) mm, respectively. During the observer performance test, the median errors (interquartile range) of the algorithm were 2.6 (1.6 to 4.8) mm, 3.6 (2.1 to 5.9) mm, and 4.0 (1.7 to 7.2) mm in ETT tip detection, carina detection, and ETT–carina distance measurement, significantly superior to that of 6, 10, and 7 clinicians (all P < 0.05), respectively. The algorithm outperformed 7, 3, and 0, 9, 6, and 4, and 5, 5, and 3 clinicians (all P < 0.005) regarding the proportions of chest radiographs within 5 mm, 10 mm, and 15 mm error in ETT tip detection, carina detection, and ETT–carina distance measurement, respectively. No clinician was significantly more accurate than the algorithm in any comparison. Conclusions A deep learning–based algorithm can match or even outperform frontline critical care clinicians in ETT tip detection, carina detection, and ETT–carina distance measurement. Editor’s Perspective What We Already Know about This Topic What This Article Tells Us That Is New