E2691. Analysis of a Deep Learning System for Evaluation of Presence and Location of Endotracheal Tube on Chest Radiographs
  1. Jiyeong An; Seoul National University Hospital
  2. Eui Jin Hwang; Seoul National University Hospital
  3. Ye Ra Choi; SMG-SNU Boramae Medical Center
  4. Gunhee Nam; Lunit
  5. Sanghyup Lee; Lunit
  6. Chang Min Park; Seoul National University Hospital
  7. Jin Mo Goo; Seoul National University Hospital
Validation of a deep learning (DL) system for evaluation of presence and location of endotracheal tube (ET) on chest radiographs CR).

Materials and Methods:
The study population comprised consecutive CRs taken immediately after ET insertion between January and March 2020 (cohort A); and consecutive CRs obtained at intensive care units between January 1st and January 3rd, 2020 (cohort B) at a single institution. We used a DL system that can identify the presence vs. absence of ET on a single CR, identify the location of tip of ET (ETT) and tracheal carina, and measure the absolute distance between the ETT and carina (TCD). One thoracic radiologist and one trainee radiologist independently measured TCD on each CR, and the thoracic radiologist defined the malposition of ET as requiring reposition. Measurement accuracy for TCD was evaluated using absolute error and intraclass correlation coefficient (ICC). For evaluation of the accuracy for the presence of ET and its malposition, receiver-operating characteristic (ROC) analyses were conducted, and sensitivity and specificity were calculated at the thresholds defined by ROC analyses.

In cohort A (539 CRs; 318 men; mean age 64.4 years; ET malposition in 76), the DL system (mean absolute error, 6 mm [95% confidence interval (CI), 5.8–6.7 mm]; ICC, 0.94) exhibited similar accuracy with the trainee radiologist (mean absolute error, 5 mm [95% CI, 4.7–5.5 mm]; ICC, 0.94) for measuring TCD, against measurement by a thoracic radiologist. For deep malposition of ET, the DL system exhibited an area under the ROC curve (AUC) of 0.96, and the threshold TCD was 34 mm (sensitivity, 92.1%; specificity; 91.0%). For shallow malposition of ET, the DL exhibited an AUC of 0.97, and the threshold TCD was 69 mm (sensitivity, 100%; specificity; 89.2%). In cohort B (637 CRs; 49.8% men; mean age 68.0 years; ET presence in 101; ET malposition in 12), the DL system exhibited an AUC of 1.00 for discriminating the presence and absence of ET on CRs. Adopting thresholds for ET malposition defined in cohort A, the DL system exhibited sensitivity and specificity of 91.1% and 89.0%, respectively.

A DL system exhibited excellent performance in identifying the presence of ET and malposition of ET on CRs.