The semantic segmentation approach for normal and pathologic tympanic membrane using deep learning

Objectives The purpose of this study was to create a deep learning model for the detection and segmentation of major structures of the tympanic membrane. Methods Total 920 tympanic endoscopic images had been stored were obtained, retrospectively. We constructed a detection and segmentation model using Mask R-CNN with ResNet-50 backbone targeting three clinically meaningful structures: (1) tympanic membrane (TM); (2) malleus with side of tympanic membrane; and (3) suspected perforation area. The images were randomly divided into three sets – taining set, validation set, and test set – at a ratio of 0.6:0.2:0.2, resulting in 548, 187, and 185 images, respectively. After assignment, 548 tympanic membrane images were augmented 50 times each, reaching 27,400 images. Results At the most optimized point of the model, it achieved a mean average precision of 92.9% on test set. When an intersection over Union (IoU) score of greater than 0.5 was used as the reference point, the tympanic membrane was 100% detectable, the accuracy of side of the tympanic membrane based on the malleus segmentation was 88.6% and detection accuracy of suspicious perforation was 91.4%. Conclusions Anatomical segmentation may allow the inclusion of an explanation provided by deep learning as part of the results. This method is applicable not only to tympanic endoscope, but also to sinus endoscope, laryngoscope, and stroboscope. Finally, it will be the starting point for the development of automated medical records descriptor of endoscope images.