Benefits of Federated Learning: healthcare practitioners give their thoughts (radiology domain)
How can the immense power of artificial intelligence be used for the healthcare industry while addressing privacy concerns at the same time? In this interview with Joao Correia, from Biotronics 3D, Petros Papachristou and Christina Kotsiopoulou, from Hygeia, we explain how the MUSKETEER platform, a research project the two companies were part of, helped to tackle their challenges and the benefits it brings for their organizations.
—
What was the benefit of the project for you? Did it help to improve the accuracy your AI algorithms (i.e. using Federated learning)?
Joao Correia (Biotronics3D)
Yes, we were able to train several AI models using the tools provided by the partners. We were able to create models with good accuracy, around 90% for some of them. We hope now with more data, we can continue to improve these AI models. We are currently running final tests, hopefully improving them again.
Petros Papachristou (Hygeia)
On our side, the direct benefit we can see, and this is the case in general for AI systems, they relieve workers from repetitive tasks, radiologists in our case, especially for studies that are classified as “normal”. They increase safety, they add a safety net if radiologists miss something and detect things that were not detected in their routine since they have a lot of work already. When you use an AI system based on pre-classified data, it gives you the possible findings, stating for instance “this is normal but this area here, you have to see it again”. It is probably fine, but it creates this so-called safety net. It prevents mistakes. It is not replacing the radiologist but helping him through the image analysis process. And additionally, it increases patient safety.
In your opinion, will federated learning help to increase research in AI tools for medical imaging using distributed data repositories?
JC: There is definitely this positive outcome, that we will be able to continue to work in the prostate cancer specific domain and try to create new models and improving what we have already achieved. At the same time,we could start looking to other diseases, for instance, liver cancer and other cancers for which we can apply similar approaches by using medical imaging to identify and eventually also classify the lesions existing in other organs of the human body. The use of the MUSKETEER platform instantiation for healthcare enables us to keep working with Hygiea, and maybe also with the other partners to continue the research in the area of medical imaging. So, this is very important for us.
Does it improve also your commercial offer accordingly?
JC: Yes. At Biotronics3D,we offer these medical imaging solutions that already integrate with some third-party AI applications. These are external models that we integrate in 3Dnet, our platform, to enable automatic features identification in medical imaging to support medical diagnosis. And we have lots of customers that are interested in AI and growing their application of AI in their specific cases.We will be able not only to work with them and do some research but also start the certification of some AI modules(needed for health sector,i.e.complying with the medical device regulations and future AI regulations) transforming them into products. This opens new opportunities for B3D.
PP: And for industrial productivity, when you have an AI system saying that the image is 99.99% normal, it is indeed, in general, normal. Because the work of radiologists implies a lot of screening, where in most cases there are no findings, such tools are of great help in a lot of cancer detection routines, say breast cancer or pelvis cancer, where you save a lot of work. It eventually makes it important for both safety and productivity.
The project offered different Privacy Operation Modes (POMs), did it help to comply with your legal and confidentiality constraints?
JC: thanks to federated learning, we only share coefficients resulting from the training shared (instead of complete data), this is satisfactory regarding confidentiality restrictions.
PP: On Hygiea’s side, first of all, even before the project, we receive the consent from our patients, any patient entering the hospitals, to use their data anonymously for scientific reasons. We have the foundation, to work in the research area. And the good thing is that the project proved it can be done without exposing the data of our customers while improving algorithms quality through collaboration of different hospitals. So, even in the minds of the management, or in the minds of the radiologist, it is clear now, that we could fulfill our legal constraints and build useful tools for the daily work at the same time.It is good for the future projects or future products that we’re going to install.
The cornerstone of the MUSKETEER platform was to preserve privacy as much as possible, with participants not knowing each other. Reciprocally, does it create a lack of transparency?
Well, the way we think the implementation in health is that all partners that will be providing their updates to the training of AI models will be partners that know each other and that agree to participate in these projects. So, in health care, we don’t see this model of participating and providing data without knowing who the partners are and who is building the model and so on. This will be even more the case, considering that in the end, the AI models need to be certified, it is very important to know the provenance of that data.
Is the project improving citizens’ trust as privacy-aware transparency and control features are increasingly streamlined across data platforms and Big Data applications?
JC: Radiologists and medical staff are more confident with the processing of the data now (about the fact that the data of their patients are not shared or leave the hospital). From the patient perspective, it is important they learn about this project, and show that the industry is taking care of the privacy aspect and their data is used in a GDPR compliant way.
Christina Kotsiopoulou (HYGEIA Hospital –InteropEHRate)
It is important to ensure that from the patient’s point of view, their data is secured. They know beforehand that a protocol is in place to protect their data, treated accordingly to GDPR.
Did the project help with better value-creation from personal and proprietary/industrial data? On the long-term perspective?
PP: AI algorithms need a lot of data to be trained and MUSKETEER enables collaboration and access to these large troves of data (in order to improve and validate algorithms). It is already a great achievement. Another point, by-product of the project let’s say, is that usually the data is not formatted correctly. But now, knowing the power of the technology, radiologists are “incentivized” so to say, to report in a structured way (to benefit from large data sets, and eventually great new tools).
JC: In the UK there is a lack of radiologists,professionals able to interpret images and write reports.It is important to have new tools to enable quicker identification of points of interest and allow radiologists to report about more patients. AI’solutions are needed to improve the daily workflow supporting doctors to make quicker and better decisions. This is important for the patient outcome ensuring no one is left behind. True in the EU.
—
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 824988.
