Fig. 1: Overview of the study.
From: From pretraining to privacy: federated ultrasound foundation model with self-supervised learning
a Medical data from 16 institutions and 9 countries are collected to pre-train and evaluate UltraFedFM, encompassing 1 million ultrasound images with extensive diversity. b The pre-training framework of UltraFedFM, where each client uses its private data to pre-train a local model through pixel-level reconstruction. During pre-training, only the local model parameters are uploaded for learning the global model, thus eliminating the risk of privacy breaches. Icons used are free to download from www.iconfont.cnand do not involve commercial use. c Clinical applications of UltraFedFM. UltraFedFM is a versatile ultrasound foundation model capable of handling multiple ultrasound scenarios, supporting multi-disease, multi-modal, and multi-task applications, and demonstrating superior performance compared with ultrasonographers in real clinical scenarios. Icons used are free to download from www.iconfont.cnand do not involve commercial use.
