Abstract
Purpose :
The aim of this study is to estimate the Cup/Disc ratio (C/D ratio) from fundus images taken by a non-mydriatic fundus camera using an Artificial Intelligence model developed through Federated Learning (FL). The study addresses two main challenges. Firstly, many medical facilities resist sharing medical images externally due to privacy concerns. FL overcomes this by using locally learned AI model parameters without needing data sharing between facilities. Secondly, the study questions the necessity of universal AI models. FL allows for creating local models using data from individual facilities, optimized models with combined data, and universal models using data equally from all facilities, facilitating performance comparisons.
Methods :
The non-mydriatic fundus camera used in this study was the Retina Station (Nikon), and the OCT measurements were taken with the RS-3000 Advance (Nidek). The study was conducted at Tsukazaki Hospital and Ikuno Eye Clinic, involving 100 subjects (200 eyes) at Ikuno and 118 subjects (236 eyes) at Tsukazaki. A Convolutional Neural Network (CNN) model was developed to learn from two ground truth labels obtained from OCT: the C/D segmentation map and the C/D ratio. The CNN model had a segmentation part using existing optic nerve head detection models and a regression part, which input optic nerve head images and the C/D map estimated by U^2NetP into VoVNet27 for C/D ratio estimation. The study connected these two models for end-to-end learning and compared the performance of Local and Universal models using the Mean Absolute Error (MAE) between OCT-measured C/D ratios and the estimated values by each model.
Results :
42 eyes from Tsukazaki Hospital were excluded from the training set and used as a test dataset to evaluate the performance of each model. The Universal model achieved a Mean Absolute Error (MAE) of 0.060, while the Local model recorded an MAE of 0.066.
Conclusions :
This study demonstrated that Artificial Intelligence models created using Federated Learning (FL) are effective in estimating the Cup/Disc ratio (C/D ratio) from fundus images. The Universal model showed slightly better performance than the Local model, though the difference was minimal, with both models exhibiting excellent performance. Additionally, the use of FL is an effective method for constructing AI models while avoiding the need for data sharing between medical facilities.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.