Kim, Sung Kyun1; Hong, Seok Min1; Park, Sung Min1; Choo, Jaegul2; Hong, Seok Jin1
1 Department of Otorhinolaryngology-Head & Neck Surgery, Hallym University Dongtan Sacred Heart Hospital, Hwaseong, South Korea;
2 Department of Computer Science and Engineering, Korea University College of Informatics, Seoul, South Korea.
Applicability of CNN-based methods in analysis of medical image increased. Herein, our results show that accuracy of automatic detection for finding of tympanic membrane (TM) and middle ear effusion (MEE) from endoscopic images was high in our proposed model.
– Study design: retrospective case review
– Setting: University hospital
– Materials: 2493 otoendoscopic TM images were selected
– Intervention: Diagnostic
– Main Outcome Measure: The images were augmented such as rotating and flipping of the originals to increase the number of training samples. Training was performed using various CNN architectures such as GoogleNet, VGGNet, ResNet, and DenseNet and the accuracy of each program was compared at the end point of analysis. we conducted 10 trials to alleviate the effects of random initial values of the network during our experiments and then averaged the results to confirm the robustness.
– Results: The accuracy of the 152 versions of the ResNet model was highest (94.8%) and the accuracy of the 161 version (95.4%) was higher than that of the 201 version (94.5%) for DenseNet. In the case of VGGNet, accuracy was slightly higher in batch normalization group (94.0-94.1%) and there was little difference in accuracy between models. The performance of the inception V3 model, which is the most popular version of GoogleNet, is similar to ResNet and VGGNet.
– Conclusion: Detection of TM abnormalities using CNN-based deep learning showed high accuracy. The performance of the CNN program for detecting endoscopic TM finding and middle ear effusion was highest for DenseNet 161 in our dataset.