An automatic annotation method based on federated learning

Abstract

With the explosive growth of neoconiosis worldwide, medical institutions worldwide generate a large number of chest radiograph images to be annotated for neoconiosis diagnosis every day. Moreover, due to the privacy of patient information, it is not possible to pool medical data from multiple medical institutions to jointly build a fast and accurate medical image annotation system for the task of new coronary pneumonia annotation. To this end, we propose a federal learning-based automatic annotation method that enables multiple medical institutions to design and develop an accurate and robust annotation system for neocoronary pneumonia without sharing data. Experimental results show that the proposed federated learning-based automatic annotation method (93.37% accuracy) is able to protect the privacy of medical image data across hospitals and accomplish a higher accuracy rate of medical image annotation compared to the automatic annotation method constructed by aggregating data (95.29% accuracy).