Figure 1
Network architectures of deep learning algorithms using both sides of mastoid anterior–posterior (AP) view with symmetry evaluation function and using single side. (a) A network architecture that predicts mastoiditis from both sides using symmetry evaluation. This architecture consists of a main network and an auxiliary path. The main network receives images of both ears as input, and three categories of mastoiditis are predicted through a convolutional neural network (CNN), feature map, and mastoiditis classifier for each ear. Each mastoiditis classifier predicts three mastoid categories as a probability vector. The symmetry evaluation layer in the auxiliary path receives both feature maps corresponding to the three mastoiditis categories and calculates the absolute value per pixel between both feature maps. The symmetry classifier predicts the difference between mastoiditis categories on both sides as three-valued probability vector. The symmetry loss is calculated from the symmetry classifier, and this loss is added to each mastoiditis loss calculated from the mastoiditis classifiers to obtain the final losses on both sides. (b) A network architecture predicting mastoiditis on only one side. This architecture receives only the image of one ear as input and predicts mastoiditis with similar structure to the main network in (a).
