A research team led by Professor Park Sang-hyun of the Department of Robotics and Mechatronics Engineering at the Daegu Gyeongbuk Institute of Science and Technology (DGIST) said on the 2nd that, in collaboration with a Stanford University team in the United States, it developed a new one-shot federated learning artificial intelligence (AI) technique that can efficiently train large-scale models without sharing personal information.
Medical imaging data contain patients' sensitive personal information, so sharing among hospitals is limited, which has made it difficult to develop AI models that use large-scale data.
Federated learning, proposed to overcome this, conducts joint training by sharing only trained models instead of patient data, but it had the limitation of large time and expense due to repeated transmissions. As an alternative, one-shot federated learning has been studied, but existing methods still had high computational costs and overfitting problems.
To address these limitations, the researchers proposed adding structural noise to synthetic images and using the mixup technique to generate virtual intermediate samples. This secured diversity in the training data to reduce overfitting and significantly improved computational efficiency by reusing synthetic images to cut unnecessary computation.
When the technique was applied to various medical imaging datasets such as radiology images, pathology images, and fundus images, it achieved higher accuracy with less computation than existing one-shot federated learning methods.
Park Sang-hyun said, "It is meaningful in that we can train a model that can be used broadly in the medical imaging field even under practical conditions of privacy protection and communication constraints," adding, "We will continue to advance this technique to ensure privacy protection while developing AI models that cover diverse patient groups, contributing to the establishment of accurate and highly reliable diagnostic support systems."
The research findings were published in the international journal in the field of medical image analysis, "Medical Image Analysis," on Jul. 7.
References
Medical Image Analysis (2025), DOI: https://doi.org/10.1016/j.media.2025.103714