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

To evaluate a deep learning (DL)-based approach for predicting air-bone gap (ABG) from tympanic membrane (TM) perforation images using automated segmentation and feature extraction, addressing the limitations of audiometry availability in certain populations or settings.

Prospective, cross-sectional diagnostic study.

Tertiary academic medical center.

A total of 1239 otoscopic images were collected between January 2019 and May 2023. Of these, 1014 intact TM images and 150 perforated TM images were used for model development and validation, and 75 intraoperative perforated images were reserved for independent testing. A Mask region-based convolutional neural network (Mask R-CNN) model was trained to segment TM and perforation areas. Segmentation performance was evaluated using class pixel accuracy (CPA), intersection over union (IoU), and Dice coefficient. Quantitative features-including perforation ratio and spatial metrics-were extracted to predict ABG using theoretical and quadratic regression models. Model performance was assessed using R², root mean square error (RMSE), and the proportion of predictions within 10 dB of measured ABG.

TM segmentation achieved CPA, IoU, and Dice scores of 0.794, 0.702, and 0.875; perforation segmentation yielded scores of 0.824, 0.729, and 0.894. ABG prediction showed R² values of 0.433 (theoretical) and 0.516 (quadratic), with RMSEs of 6.15 and 5.68 dB. Deep learning (DL)-assisted models achieved accuracy of 83% and 86%, comparable to manual annotation.

DL-based analysis of TM images enables accurate ABG prediction and may provide a scalable tool to support assessment of conductive hearing loss in environments without access to audiometry.