Presenter Guibo Luo, PhD, of Massachusetts General Hospital (MGH) and colleagues sought to explore the relationship between distribution metrics and segmentation performance, as well as to develop a reliable federated deep-learning algorithm for liver and tumor segmentation.
First, they gathered 692 contrast-enhanced abdominal scans from three different sources: a publicly available dataset, China, and MGH. The datasets included images from different clinical sites and were acquired using different protocols on a variety of scanners.
The researchers then developed a number of liver and tumor segmentation models, including three trained only on each source's local data; a centralized model that was trained on all data at the same time; and an algorithm trained using a federated-learning approach that involves alternately training a global, shared model on private local data.
In testing, "our proposed federated deep-learning with local [batch normalization] approach provided a comparable performance with centralized deep-learning for liver and tumor segmentation in multi-center datasets," the authors wrote.
These results show the potential for federated learning to yield high-performing algorithms while retaining data privacy, according to the researchers.
What else did they find? Attend this presentation on Monday morning to get all of the details.
