Federated Learning in Cgan: An Application of Oct Images

Abstract

Age-related macular degeneration (AMD) is one of the most common causes of elder vision loss, the early screening of which can be accomplished through AI-based Optical coherence tomography (OCT). However, the shortage of labeled OCT images greatly hinder the implementation of this method. Therefore, we propose an improved federated learning conditional generate adversarial network (FL-CGAN) for OCT AMD-labeled data enhancement scheme. Three federated learning sub-models are involved on 6,500 OCT images collected from Kaggle, and 1,500 fake images are generated as well. The experimental results suggest that, measured by indicators of MMD, Frechet Inception Distance score (FID) and Inception Score (IS), the proposed scheme represents a high efficiency of data enhancements, which can also contribute some reference value for the medical image enhancement. Besides, compared with the nomal GAN architecture, a distributed deep learning framework is constructed with less resource consuming.

Ähnliche Arbeiten