Prediction of the Location of the Glottis in Laryngeal Images by Using a Novel Deep-Learning Algorithm

초록

A novel deep-learning algorithm for artificial neural networks (ANNs) was developed and presented in this paper, which is intuitively understandable, simple, efficient, and completely different from the back-propagation method, i.e., randomly selecting weight factors and bias values of an ANN and adjusting their values by small random amounts during the training session where it does not need to calculate the gradients of the training error to adjust weight factors as does the back-propagation method. The algorithm was applied to predict the location of the glottis in airway images obtained using a video airway device. The glottic locations were marked in 1,200 airway images captured using GlideScope (R) and fiberoptic laryngoscopy. With the randomly selected 1,000 training set data, 84 ANN models were trained using the above algorithm. We sought anANNmodel that minimized the average training error for all training set data by reducing the input image resolution. As the resolution was reduced, the average training error decreased to its lowest level at 30x30 pixels. Eventually, the 900-98-49 ANN model was selected as the prediction model for the location of the glottis; it was the model with the lowest training error, i. e., the highest learning rate. The selected prediction model was applied to the remaining 200 test set data to obtain the test accuracy, and we obtained that the accurate prediction and the adjacent prediction rates were 74.5% and 21.5%, respectively. Reducing the input image resolution to an appropriate level could yield better prediction of the glottic location in airway images. This ANN model can help clinicians perform intubation by presenting the predicted location of the glottis.

키워드