Rebecca Wright, University of Maryland School of Medicine; Nabile M. Safdar, MD; Zhimin Huo, PhD; Khan M. Siddiqui, MD; Naomi J. Saenz, MD; Eliot L. Siegel, MD, FSIIM

The purpose of this study is to evaluate the effectiveness of a visual enhancement algorithm as an aid to radiologists in reviewing portable ICU chest radiographs for evaluation of ET tube placement.

A common complaint by radiologists making the transition from film to computed/direct radiography is a loss in apparent conspicuity of life support lines such as the endotracheal tube (ET). An improperly placed endotracheal tube can result in life-threatening consequences such as complete atelectasis of one or more lobes of the lung or alternatively inadequate ventilation. Surprisingly little attention has been paid to this important topic in imaging literature, especially given the high percentage of radiographs obtained to ascertain the position of the ET tube and other life support lines specifically in intensive care units. The purpose of this study is to evaluate the effectiveness of a visual enhancement algorithm as an aid to radiologists and clinicians in reviewing portable ICU chest radiographs for evaluation of proper ET tube placement. This computerized detection and enhanced visualization method will automatically detect the presence of an endotracheal tube and determine the location of its tip.

We reviewed 487 chest images from 100 consecutive ICU patients. Of the 487 images, 145 were identified by an expert reviewer as having an ET tube. A computerized method was developed to detect ET tubes. The method was trained and tested on the 145 ET tube-positive images and 200 ET tube-negative images. The method includes: (1) generating gradient-Haar templates to detect edge-center-edge patterns as initial ET tube candidates, (2) tube growing and merging, and (3) feature extraction to remove false positive (FP) detections. The templates are created with their directions closely aligned with detected spine direction. The tube growing is guided by the fitted curve from the initial candidates. The features extracted are used to characterize the width, length, and curvature of a detected tube and its position relative to the spine and lungs. A quadratic discriminant analysis method was employed to differentiate true-positive detections from FP detections using 7 features. The performance of the algorithm was evaluated using the leave-one-out method.

Based on the detected ET tube tip position, several regions-of-interest or ROIs (700x1200 pixels) centered near the tip were selected for further local visualization enhancement. Three image enhancement techniques were developed to optimize visualization of the ET tube tip and the trachea (or carina) and to better determine proper placement of the ET tube tip, relative to the carina. A pilot reader study independently evaluated the effectiveness of the ROI enhancement techniques. After IRB approval, we randomly selected 70 ICU digital chest radiographs positive for ET tubes. Four radiologists participated in the reader study, and each reviewed 35 images. Each pair of readers reviewed the same 35 cases independently to assess reader variability. Readers were asked 5 questions before and after an image was processed to enhance visualization of the ET tube regarding the presence (yes/no) of the ET tube, visibility of the ET tube tip and carina on a 4-point scale (1-invisible to 4-highly visible), and their confidence in the determination of the ET tube tip and carina position on a 100-point scale. The enhanced image was presented in the format of a locally enhanced ROI overlaid on globally enhanced chest images.

Quadratic discriminant analysis yielded a sensitivity of 92% at 0.16 FPs/image when trained on the 345 images. It yielded a sensitivity of 92% at 0.33 FPs/image when tested on the 345 images using the leave-one-out method.

When reading without the enhanced image, 3 of the 140 total interpretations were misclassified as not having an ET tube. After review with the image processing, two (67%) of the 3 missed ET tubes from two readers were correctly detected. Computer-aided visualization improved the radiologists’ ability to detect the presence of an ET tube by 1.4%. Overall, confidence in determining ET tip position is increased by an average of 25% on 43% of the total reads after reviewing the enhanced image. The confidence in determining the carina position was increased by an average of 22% (ranging from 10% to 80%) on 33% of the total interpretations.

This algorithm successfully identified and located an endotracheal tube in a strong majority of cases. Preliminary results showed that our method could detect ET tubes at a high sensitivity with a reasonable low FP rate. In addition, the independent reader study showed that enhanced visualization increased the readers’ ability to detect ET tubes by 1.4%. It also increased the readers’ confidence to accurately determine the tip position by an average of 25% on 43% of the cases. In light of these initial results, computer-aided detection, in conjunction with enhanced visualization, is promising as an important adjunct in the detection of endotracheal tubes and should also be tested for use by clinicians less experienced in interpretation of chest radiographs. It could have the potential to aid clinicians in improving their ability to detect ET tubes and to determine the placement of ET tube tips. This software could serve as a clinical decision support and safety system for routine clinical care.

The algorithm achieved a high sensitivity (92%) with a low number (0.16) of false positives per study suggesting that it could be successful as a decision support tool for identification of the presence and tip of an endotracheal tube. The use of this type of algorithm has the potential to improve radiologist speed and accuracy in the detection of endotracheal tubes and to have an even greater impact on the improvement and reliability of detection of endotracheal tubes for less experienced clinicians. This study demonstrates both the limitations and potential benefits of digital radiography in the detection of life support lines with an emphasis on the detection of the termination of the endotracheal tube. An automated version of this software could have a major positive impact on patient safety, especially as a decision support tool for non-radiologists.

1. Hall JB, White SR, Karrison T. Efficacy of Daily Routine Chest Radiographs in Intubated, Mechanically Ventilated Patients. Crit Care Med. 1991;19:689-693.

2. Brunel W, Coleman DL, Schwartz DE, et al. Assessment of Routine Chest Roentgenograms and the Physical Examination to Confirm Endotracheal Tube Position. Chest. 1989;96:1043-1045.

3. Marik PE, Janower ML. The Impact of Routine Chest Radiography on ICU Management Decisions: an Observational Study. Am J Crit Care. 1997;6:95-98.