Dataset
This retrospective study was approved by the Institutional Ethics Committee of Children’s Hospital of Nanjing Medical University and was conducted using the data obtained from Picture Archiving and Communication System (PACS) database and Operation Anesthesia Information System (OAIS) database. Informed patient consent was waived by our IEC. Clinical information and CT images were collected from infants with PRS who underwent intubation anesthesia in 2018 at Children’s Hospital of Nanjing Medical University.
Seven clinical risk factors [18] that may have an impact on tracheal intubation difficulty were provided by experienced clinicians, including gender, height, weight, body surface area (BSA), throat area, age, and pneumonia (Table 1). The calculation of the throat area was elaborated below, and the remaining indicators could be directly obtained or simply calculated. Tracheal intubation difficulty is divided into three levels based on whether glottis can be completely observed under visual laryngoscope, where level Ⅰ refers to complete observation, level Ⅱ refers to partial observation, and level Ⅲ refers to the case when the only epiglottis can be observed.
Labeling criteria
To assess the impact of the throat area on tracheal intubation difficulty, the collected CT images (Figure 1a) were labeled according to the irregularity of the area being labeled using Labelme, an annotation tool which is based on the Python language and allows for irregular area annotation [19]. A radiologist with 20 years of clinical experience, who was blinded to the infants’ difficulty level, was responsible for labeling. Through a three-dimensional reconstruction technique, the median sagittal image of the upper airway of the infants was obtained, after which then the area of the oropharyngeal cavity (ie, the pharyngeal area between the plane of the tongue and the glottis) was labeled.
Annotation file processing and area calculation
The overall workflow is shown in Figure 2. The annotation file generated by Labelme is in the format of .json (Figure 1b) [20]. To calculate the throat area, the annotation file was first converted to a single-channel image in .png format (Figure 1c).
OpenCV performed subsequent processing in the Python environment. First, the single-channel image that was obtained during the previous step underwent color space conversion using the cvtColor function of OpenCV and was converted into a grayscale image (Figure 1d) [21][23]. The grayscale image was then thresholded (the threshold was set to 1) using the threshold function and becoming a binary image (Figure 1e) [22][23]. The throat contour information of the marker was then obtained by the findContours function, with pixel position difference between two adjacent points in all contour points no larger than 1 [23][24]. Finally, the contour information obtained in the previous step in the form of a point set was input into the contourArea function of OpenCV to calculate the area [23][25].
Correlation analysis
Correlation coefficients were used to assess the impact of each risk factor on tracheal intubation difficulty. Clinical risk factors highly correlated with difficulty level had better predicative effects in the clinic.
Statistical analysis
Since clinical risk factors include numerical and categorical variables and tracheal intubation difficulty is categorical, the correlation was measured by the Spearman rank correlation coefficient. Besides, to analyze whether there is a significant difference in each clinical risk factor under tracheal intubation difficulty, the Kruskal-Wallis test was used for numerical factors, and Pearson's Chi-squared test was used for categorical factors.