Semantic Decomposition and Anomaly Detection of Tympanic Membrane Endoscopic Images

Abstract

With the recent development of deep learning, the supervised learning method has been widely applied in otolaryngology. However, its application in real-world clinical settings is difficult because of the inapplicability outside the learning area of the model and difficulty in data collection due to privacy concerns. To solve these limitations, we studied anomaly detection, the task of identifying sample data that do not match the overall data distribution with the Variational Autoencoder (VAE), an unsupervised learning model. However, the VAE makes it difficult to learn complex data, such as tympanic membrane endoscopic images. Accordingly, we preprocess tympanic membrane images using Adaptive Histogram Equalization (AHE) and Canny edge detection for effective anomaly detection. We then had the VAE learn preprocessed data for only normal tympanic membranes and VAE was used to calculate an abnormality score for those differences between the distribution of the normal and abnormal tympanic membrane images. The abnormality score was applied to the K-nearest Neighbor (K-NN) algorithm to classify normal and abnormal tympanic membranes. As a result, we were obtained a total of 1232 normal and abnormal eardrum images, classified with an accuracy of 94.5% using an algorithm that applied only normal tympanic membrane images. Consequently, we propose that unsupervised-learning-based anomaly detection of the tympanic membrane can solve the limitations of existing supervised learning methods.