Deep learning model performance
The AUC of classifying the normal pars flaccida and attic retraction pocket was 0.89, and the accuracy, sensitivity and specificity were 89%, 0.93 and 0.62 respectively (Fig 1). The DCNN model achieved an AUC of 0.87 in classifying the normal pars tensa and atelectasis, and the accuracy, sensitivity and specificity were 79%, 0.71 and 0.84 respectively (Fig 2).