DL-Enhanced Otoscope: A Breakthrough in Detecting Middle Ear Abnormalities

Improving Middle Ear Abnormality Detection with DL-Assisted Otoscopy

Introduction

Accurately diagnosing middle ear conditions can reduce hearing impairment and antimicrobial resistance, especially in low- and middle-income countries. However, otoscopy remains challenging for many medical practitioners and may be unavailable in remote areas.

Deep Learning (DL) in Otoscopy

DL algorithms can automatically extract predictive features directly from raw images, enabling even non-experts to diagnose middle-ear abnormalities. Several DL-assisted systems have been developed for this purpose, but none offers a user-friendly, highly accurate, end-to-end smartphone-based solution.

This Study

This study aimed to develop and validate such a system, combining a smartphone-attached otoscope, a DL model, and a smartphone app.

Methods

1. Data Acquisition: 98,137 digital otoscopic images were collected and divided into training, validation, and test sets.
2. Model Development: An Inception-v2 CNN was trained on the training set.
3. Model Validation: The model was evaluated on the validation set (3,962 images) and a held-out test set (326 images).
4. Analysis: Accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) were calculated for each diagnostic class.

Results

Validation Set:

• For the binary classification of normal vs. abnormal images, the model achieved:
• Sensitivity: 99.1%
• Specificity: 100%
• AUROC: 1.00
• High accuracy for most specific diagnostic classes, with AUROCs ranging from 0.98 (for eardrum perforation) to 1.00 (for wax plug).

Test Set:

• For normal vs. abnormal images:
• Sensitivity: 99.0%
• Specificity: 95.2%
• AUROC: 1.00
• High accuracy for wax plugs:
• Sensitivity: 100%
• Specificity: 100%
• Misclassification of foreign bodies as normal in 4 cases.

Discussion

The DL-enabled system achieved high accuracy on the validation set and for specific diagnostic classes on the test set. It has potential applications in remote areas, primary care, and telemedicine for triage or as an add-on diagnostic tool.

Limitations

• The system relies on single still images, while video-otoscopy may provide better assessment.
• The test set evaluation pointed to high accuracy only for distinguishing between normal and abnormal images, and for detecting wax plugs.
• Further external validation studies are needed, especially for less common diagnostic categories.

Conclusion

A user-friendly, end-to-end smartphone-based DL system for detecting middle ear abnormalities has been developed and validated. It shows promise for improving diagnostic accuracy and accessibility, but prospective external validation is required before widespread clinical deployment.