Automatic Prediction of Conductive Hearing Loss Using Video Pneumatic Otoscopy and Deep Learning Algorithm
Hayoung Byun1,2, Chae Jung Park3,2, Seong Je Oh4
1Department of Otorhinolaryngology-Head and Neck Surgery, Hanyang University College of Medicine, Hanyang University Medical Center, Seoul, South Korea.
Ear and Hearing
|March 28, 2022
Summary
A deep learning model analyzing video pneumatic otoscopy (VPO) images accurately detects conductive hearing loss. This AI tool shows promise in diagnosing middle ear conditions, outperforming experienced otologists.
Area of Science:
- Otolaryngology and Artificial Intelligence
- Medical Imaging and Diagnostics
- Hearing Loss Research
Background:
- Middle ear diseases can cause conductive hearing loss by disrupting sound transmission.
- Video pneumatic otoscopy (VPO) provides dynamic imaging of the tympanic membrane and ossicles, offering insights into middle ear function.
- Predicting middle ear transmission problems necessitates analyzing VPO findings.
Purpose of the Study:
- To develop a deep learning model using convolutional neural networks (CNNs) for analyzing VPO images.
- To detect the presence of an air-bone gap, indicative of conductive hearing loss, via VPO image analysis.
- To evaluate the diagnostic performance of the developed AI model against experienced otologists.
Main Methods:
- Retrospective review of adult patients' VPO tests and pure-tone audiometry (PTA) data.
- Definition of conductive hearing loss as an average air-bone gap >10 dB at 0.5, 1, 2, and 4 kHz.
- Utilized multi-column CNN architectures (Inception-v3, VGG-16, ResNet-50) with pre-trained backbones for image analysis.
Main Results:
- The best-performing model, an Inception-v3-based three-column architecture, achieved a mean area under the curve (mAUC) of 0.972.
- The AI model demonstrated high diagnostic accuracy (94.1%) and sensitivity (91.6%) in predicting conductive hearing loss.
- The algorithm's performance surpassed that of experienced otologists (mAUC 0.773, accuracy 79.0%) and identified specific middle ear pathologies.
Conclusions:
- A deep learning algorithm analyzing VPO images successfully identified conductive hearing loss from various middle ear pathologies.
- The AI tool shows potential for differentiating conductive from sensorineural hearing loss, particularly in non-cooperative patients.
- VPO interpretation using AI presents a promising diagnostic approach for middle ear transmission problems.
