A Federated Learning-Based Precision Prediction Model for External Elastic Membrane and Lumen Boundary Segmentation in Intravascular Ultrasound Images

Abstract

Deep learning technologies have been widely used in intravascular ultrasound (IVUS) image analysis. An efficient and real-time tool for IVUS image segmentation is essential during percutaneous coronary intervention (PCI). However, collecting a significant volume of data is infeasible lying to the following reasons: (1) patients’ privacy concerns and (2) manually annotating data is time-consuming and laborious. Therefore, training a deep model with generalization capacity using distributed data remains a challenge. In this study, a federated learning-based deep learning model is proposed. It is a framework that enables model training on multi-source data without sharing each local dataset. A lightweight U-Net model is also adopted in the federated learning framework due to the limited volume of data. The results indicate that the lightweight model’s performance can almost achieve the performance of a deeper model but significantly spends less time consumption. The proposed lightweight U-Net model parameters are less than 30 times that of the baseline model. The external elastic membrane (EEM) and the lumen boundary segmentation achieved 0.8567 and 0.8457, respectively. The efficiency of implementing such a federated learning framework on a distributed IVUS image segmentation is estimated to be increased by more than four times.