Deep Learning-Assisted Prediction of Air-Bone Gap Using Tympanic Membrane Perforation Image Features.

Area of Science:

• Otolaryngology
• Medical Imaging
• Artificial Intelligence

Background:

• Audiometry is crucial for assessing conductive hearing loss.
• Access to audiometry can be limited in certain populations and settings.
• Tympanic membrane (TM) perforations can impact hearing, necessitating accurate assessment.

Purpose of the Study:

• To evaluate a deep learning (DL) approach for predicting air-bone gap (ABG) from tympanic membrane (TM) perforation images.
• To automate segmentation and feature extraction from TM images for ABG prediction.
• To address limitations in audiometry availability.

Main Methods:

• A Mask region-based convolutional neural network (Mask R-CNN) was trained on 1014 intact and 150 perforated TM images.
• Segmentation performance was evaluated using class pixel accuracy (CPA), intersection over union (IoU), and Dice coefficient.
• Quantitative features were extracted to predict ABG using regression models, with performance assessed by R² and root mean square error (RMSE).

Main Results:

• TM and perforation segmentation achieved high scores (CPA, IoU, Dice).
• Deep learning models predicted ABG with R² values of 0.433 (theoretical) and 0.516 (quadratic).
• DL-assisted models achieved 83% and 86% accuracy, comparable to manual annotation.

Conclusions:

• Deep learning analysis of TM images enables accurate ABG prediction.
• This DL approach may offer a scalable tool to support conductive hearing loss assessment.
• It is particularly useful in environments lacking access to traditional audiometry.