내시경 영상을 활용한 중이 질환 진단에 있어서 딥러닝 모델과 이비인후과 의사의 비교 분석

Abstract

Deep learning based on machine learning has rapidly advanced over decades to bridge the gap between expert image interpretation and automated analysis. We compared a custom EfficientNet-B4 deep learning model's diagnostic performance with otolaryngologists in 100 endoscopic images, and compared the change in the diagnosis conclusion of the otolaryngologist after knowing the deep learning results. The model predicted primary (otitis media with effusion, chronic otitis media, 'None') and secondary classes (attic cholesteatoma, myringitis, otomycosis, ventilating tube). Three otology professors, five senior residents, and five junior residents performed selection of the primary class and the secondary class from the same endoscopic images. After knowing deep learning results, they performed selection again. In the prediction of the primary class, the accuracy of deep learning model were 95.0%. Before knowing the deep learning result, the accuracy of professors, senior residents, and junior residents was 78.7%, 65.0%, and 54.4%. There was a significant difference between deep learning models and three groups. (p<0.001, respectively). After knowing the deep learning results, the accuracy of professors, senior residents, and junior residents were 89.7%, 93.8%, and 86.6%, respectively. The accuracy was statistically increased in all group compared to before knowing the deep learning results. (p<0.001, respectively). The diagnostic performance of all groups, especially residents groups, improved after knowing the deep learning result. This suggests that deep learning can be helpful not only for doctors with little experience in diagnosing middle ear diseases, but also for resident in terms of education.