Deep Learning for Endoscopic Classification of Adenoid Hypertrophy.

Objective: Endoscopy is a convenient and widely used method to evaluate adenoid size, but the subjectivity of its image diagnosis can result in over- or underestimation. To create a new assessment strategy for adenoid hypertrophy, we developed a reliable method for automated classification using a deep learning algorithm with nasal endoscopic images. Methods: Endoscopic images of adenoids were obtained from our institution and were divided into four grades based on choanal obstruction. The labeled images were randomly assigned to training and testing sets. The convolutional neural network (CNN) algorithm Xception was used and tuned on the training set. The classification performance was assessed on the test data using the optimal model. Results: Twenty-six thousand and sixty endoscopic images of adenoids were included and classified into four grades (5500 for Grade Ⅰ, 5452 for Grade Ⅱ, 7453 for Grade Ⅲ, and 7655 for Grade Ⅳ). The Xception model achieved an overall classification accuracy of 95.53%. When evaluated by receiver operating characteristic curves (ROCs), the area under the curves (AUCs) of Grades Ⅰ- Ⅳ were 0.93, 0.94, 0.97, and 0.91, respectively. Conclusion: Artificial intelligence is effective in the classification of endoscopic adenoidal hypertrophy. The deep learning approach may be beneficial in improving the efficiency, objectivity, and stability of diagnosis.