(A820) Federated Learning for Privacy-Preserving AI: A Multi-Institutional Model for Automated Tympanic Membrane Perforation Size Estimation

Introduction: Visual estimation of tympanic membrane perforation (TMP) size is often inaccurate. This study describes developing a multi-institutional computer vision (CV) model for automated TMP sizing using federated learning (FL), a privacy-preserving AI technique that decentralizes data across institutions.

Methods: Three CV models were developed using distinct datasets: a global model trained on publicly available endoscopic images of TMPs, and two independent institutional models trained on site-specific datasets. A third institutional dataset served as an external validation set for model testing. TMP and remnant TM (RTM) boundaries were manually annotated in all images, and datasets were divided into training and test sets. The global model was initially trained on the public dataset in each FL round. Its resulting weights were the starting point for training each institutional model on local data. Trained institutional model weights were then averaged to update the global model’s logic for the next round. This iterative process shares patterns learned from each dataset, without sharing raw data, to refine the global model. Five FL rounds were conducted. Model performance was evaluated using mean average precision-50 (mAP50), a measurement of overlap between manual annotations and predictions, and TMP size accuracy. After each FL round, global and institutional models were tested on the external validation dataset.

Results: All models showed improved performance with successive FL rounds. The global model achieved the highest mAP50 over 5 total FL rounds, though 95% confidence intervals overlapped between models. The final global model predicted the TMP size with 92.5% accuracy.

Conclusions: FL enabled the creation of a global, multi-institutional CV model for TMP size estimation with use of small institutional datasets. In this approach, collaborative, privacy-preserving AI models can be developed by sharing only model weights rather than patient data.