Implementation of a Digital Otoscope with Deep Learning for an Automatic Tracking Function in Otitis Media Assisted System

• Main Authors: Yi-Syuan Sung, 宋怡萱
• Other Authors: Chen Lin
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/6czs9b

碩士 === 國立中央大學 === 生醫科學與工程學系 === 107 === Otitis media is defined as infection in the middle ear. Acute otitis media (AOM) is one of the most common infections in children under 15 years of age. According to epidemiological studies, children with otitis media have an infection rate of more than 60% before one year old. More than 80 percent of children have at least one episode of otitis media by the time they are 5 years of age and 46% of them have had more than three times of acute otitis media. Therefore, the diagnosis of otitis media in children is very challenging. However, many parents confuse otitis media with a common cold and only half of the patients with otitis media would have a fever. If children are not able to describe the symptoms related to otitis media, parents often ignore the symptoms and even for the physician other than otorhinolaryngologist can misjudge the symptoms, as a consequence, losing the golden time for treatment. At this time, the equipment with the otoscope-assisted diagnosis system in the home can timely observe whether the eardrum is abnormal or not. Therefore, we proposed a semi-automatic eardrum tracking function implemented to the device, which can guide the user to capture the complete eardrum based on the eardrum illustration diagram. We sketched the outline of the eardrum on the screen so that the user can know the shape of the eardrum. We also add a guide sign to allow the user to move the direction to find eardrum. Finally, it is decided to capture the eardrum according to the ratio of the area of the eardrum to the area of the total picture. Our results demonstrated that this semi-automatic eardrum tracking algorithm can capture the complete eardrum with 90.43% accuracy for total images. Among them, the accuracy of 95.66 % for normal images, the accuracy of 84.92 % for AOM images, the accuracy of 87.88 % for COM images and the accuracy of 84.11 % for OME images. In the back-end image recognition analysis, we also add the concept of deep learning using FCN-AlexNet and FCN-Vgg16 modules to optimize the eardrum image segmentation technology. The computer automatically can learn to get the best and complete eardrum image in order to feature extraction on the eardrum perform automatic classification. The smart otoscope will be combined with the mobile APP to design an eardrum shooting guide interface, in order to the user efficiently operate the otoscope to achieve high quality eardrum photographs. The smart otoscope can help parents to continuously detect and monitor the internal structure of the ear in time. Through the concept of machine learning, you can diagnose the symptoms of the eardrum and give appropriate treatment to reduce recurrence of the disease. This can avoid hearing loss and slow language development in children.