Developing artificial intelligence, AI, for applications in healthcare, for example in medical image analysis, can save substantial time in clinical workflows. But there are several challenges. Data-driven AI, i.e. large-scale machine learning models, needs a large amount of medical data to train on. For example, to train models to automatically analyze magnetic resonance (MR) images of brain tumor patients, a large number of images are needed. But sharing medical images is complicated due to ethics, anonymization and data protection regulations.
Combining medical information from multiple hospitals results in a larger medical dataset. Because each hospital individually trains an AI model locally, the sensitive information does not need to be sent outside the hospital. In a European research project, researchers combine locally trained models in real-time into a joint so-called global model on a server in another city. That way, they don't have to send the medical data each local model is trained on. The technique is called federated learning. The researchers have now performed federated learning between Linköping University and the radiation therapy clinic at Skåne University Hospital in Lund.
First in Sweden
“We believe that we are among the first in Sweden to do federated learning with medical images. We will soon add several more cities, Umeå is next in line,” says Anders Eklund, senior associate professor at the Department of Biomedical Engineering (IMT) and the Department of Computer and Information Science (IDA).
One challenge that remains to be solved is that the images can differ between hospitals due to different equipment and MR scanners. The hospitals may also have different clinical guidelines for how oncologists, i.e. medical doctors who have special training in diagnosing and treating cancer, should mark the tumors and handle surrounding organs, so-called risk organs. Anders Eklund is hopeful that they will come up with a solution that compensates for that. To study how differences between hospitals affect the federation, they have made simulations where they used an open dataset, meaning an open collection of data, called BraTS. BraTS includes images and marked tumors for brain tumor patients from more than twenty different hospitals, and has been created specifically to develop similar models.
“If, for example, a hospital has marked the tumors with a smaller margin, which affects the federation negatively, we can hopefully develop methods so that it still works.”
More areas of application
The project also involves companies and universities in the Netherlands, Belgium and Türkiye. In the Netherlands, for example, researchers are developing similar models for images of the liver.
“With these images, the same methods are used but with a different application. That's the point, to learn from each other and use the same methods in different ways.”
Anders Eklund also mentions that other possible areas of application with federated learning could be genetics or text in medical journals, i.e. language models in healthcare. Companies that have a lot of data in different cities and don't want to put everything in the same place could also use the technology. Similarly, companies that want to take part in research projects and that do not want to send their data.
“Federated learning can be used for all types of information that is sensitive in various ways. It is becoming more and more popular, otherwise the companies would not be involved and invest in the project.”
