Construction of three image recognition models of manikin′s glottis using visual laryngoscopy based on deep-learning algorithm / 中华麻醉学杂志

ABSTRACT

Objective:To construct three image recognition models of manikin′s glottis using visual laryngoscopy based on deep-learning algorithm.Methods:The tracheal intubation manikin′s epiglottis was visualized using a videolaryngoscope, and then epiglottis was elevated to expose the glottis and acquire glottic images. A total of 149 images were obtained from various angles and orientations and randomly divided into training set and test set, and the annotation of image data was completed. Three glottal image recognition models of CenterNet, YOLOv3 and YOLOv4 were developed. The training set was used to complete the training of the models, and finally the test set was used to evaluate the model performance.Results:CenterNet, YOLOv3 and YOLOv4 three models were successfully constructed, the mean average precision of CenterNet, YOLOv3 and YOLOv4 was 92.33%, 89.52% and 89.02% respectively, the recall rates were 87.50%, 90.00% and 90.00% respectively, the precision rates reached 97.22%, 94.74% and 94.74% respectively, and the accuracy rates were 90.91%, 85.11% and 88.89% respectively. All three algorithms demonstrated an identical F1 score of 91.00%.Conclusions:The CenterNet, YOLOv3 and YOLOv4 models are successfully constructed, and three recognition models can accurately identify the glottis in the image, with the CenterNet model demonstrating the highest recognition precision.