Ivana Išgum is the University Professor of AI and Medical Imaging at the Amsterdam University Medical Center. Graduating in Mathematics from the University of Zagreb in 1999, Išgum became a PhD student in the Netherlands in 2001 at the Image Sciences Institute. Obtaining her PhD degree in 2007, titled ‘Computer-aided detection and quantification of arterial calcifications with CT,’ Ivana then worked at both Leiden University Medical Center and UMC Utrecht, before embarking upon her position at the AMC.
Why do you do what you do?
Growing up in Croatia, both my mother and father were engineers. I remember being taken to my father’s work as a little girl and finding the whole experience and atmosphere so exciting. I knew I wanted to follow in my parents’ footsteps. I came to the Netherlands 20 years ago as part of my journey to pursue that goal and never left.
Studying mathematics was quite abstract for me. I think that math is everywhere around us but we’re often not aware of it. It is fulfilling for me to connect knowledge with something that is applicable and visible in everyday life. So, I wanted to do something medical. I wanted to strive for an academic career and surround myself with inspiring people. Of course getting funding for the research we want to do can be a struggle, but ultimately my job comes with the capacity to work on what I wish to work on, there is such creative freedom.
How has AI proved so transformative to medical imaging?
It has been the appearance of deep learning and the ability of algorithms to now be able to learn from the data themselves that has boosted the field enormously. Whilst AI is everywhere and indeed everyone is excited about it, for some this may appear as ‘hype’, but I really try to avoid using the word ‘hype’. It implies that after a few years AI will somehow go away. That’s not the case, I am convinced that AI is here to stay, it will develop further, and it will radically change the way we work.
Unlike in other industries, in the field of medicine we seem to not be as far as we could be. The main reason for this is probably accessibility of data. AI requires large sets of data that represents different populations. Furthermore, responsible application of AI in clinic while adhering to privacy legislation and ethical standards is far from trivial. We are actively researching these issues so that the medical field can benefit from current AI much more than it’s already doing.
In a broad sense, how can AI currently benefit healthcare systems?
Current healthcare systems are not sustainable as they are. The number of medical experts is limited and AI can help relieve their burden. AI might save costs, but it will more likely allow for a more sustainable system and it will certainly enable expert healthcare to be brought to parts of the world where there is a shortage of expertise.
Where does reservation around AI stem from, and how can it best be addressed?
One of the fears of AI comes from the caution of un-transparent decision-making. There are big, important and needed conversations currently going on about this. As experts, we do have to look at causality and make sure we validate thoroughly. I think the general public have a fear of AI being depersonalised but I don’t think it necessarily is, or should be.
Are mindsets becoming more confident in AI powered tools in clinical practice? How does that feel for you at the forefront of research?
More people are seeing the potential benefits of AI to the future of healthcare, and resistance is decreasing. But people generally resist change, if you look back at history it’s always been the way. Five years ago, people were panicking that radiologists would be replaced, and therefore not trained and that we’d see the takeover of robots. Now that hasn’t happened and undoubtedly never will. The medical world moves slowly and AI brings added value to routine work that radiologists can partner with and check. Ultimately AI makes it faster, more reproducible and cheaper than clinicians doing it all themselves, and that creates impact.
Physicians understand AI in medical care is here to stay and people want to ride the wave. It opens the door for so many opportunities and for us who have been in the field for a long time, it’s rewarding.
Where are we currently in terms of the potential of AI being utilised in clinical practice and what are the key challenges when it comes to implementation?
From a research perspective, you see that a lot is possible. I would say the technology to answer many questions is available. But if you look at the products, the diversity doesn’t match up. For example the number of AI powered radiology tools used on a large clinical scale, is not very many, if at all. So the translation from research to clinical practice is really important. Regulatory bodies such as the FDA for example, are changing their strategy quite considerably in order to keep up with regulations that enable these tools to come into use.
Careful implementation is crucial. We cannot make mistakes. It’s important to consider all aspects of AI application including associated ethical and legal issues. We are working with an extremely powerful tool and sensitive data; there is a lot to be discovered. Whilst we want to utilise AI to find information which is important to be found, we do not want to generate investigations which are not needed. These are all issues that are fundamental as we explore the possibilities of translating research into clinical practice.
What do you think are the main technical challenges related to applying AI to medical imaging in comparison to using AI in other fields?
Medical image data is large and high dimensional, often 4D and 5D. At the same time, there are small sets of labelled representative data to learn from and the labels are often noisy. This can be challenging for analysis. Furthermore, data is quite diverse regarding image acquisition protocols, variability in anatomy and pathology. Designing methods that are robust to all these is not straightforward.
After the introduction of deep learning, there was a boom of papers applying similar techniques to different domains. Has the rate of progress slowed down or are there new interesting challenges that (if solved) could represent a major milestone in the field of medical imaging processing?
I am not sure whether the rate has been slowing down. We see that methodological developments from one research domain are being adjusted for application in others, which is good. However, we also see enormous competition in publishing where outperforming previous work, even if very slightly, is expected. While improving on the performance is possibly important and relevant, I think it is even more important that these are reproducible and that we try to understand why some approaches work better than others.
What’s next in terms of key medical areas to benefit from AI?
It’s such an exciting time, we’re at the tip of what’s possible. There are very many possible applications of AI in medicine.Not only analysis of medical images but the merging of different types of information can be very impactful and could help deeper understanding greatly.
As a successful woman in the field of AI, you must get asked about gender dynamics a lot, do you feel this question is still useful?
Absolutely, it’s an incredibly important but sensitive issue. I dislike when people say it’s not a problem or that it’s not a problem for me any more. That is just not true. Ultimately we do not have enough women in science. As a society we are educating women but missing out on benefiting from their talent. Why? It’s such a complicated myriad of things at play (such as PhD and PostDoc timing coinciding with starting families, judgement in the workplace, at home, in the school playground etc) that ultimately work together to prevent society not utilising the full capacity of experts trained in science.
Diversity is key for benefiting science. Some people say we need to encourage women, I think it would be a start if we didn’t discourage them.
Nico.lab improves clinical practice with innovative and trusted AI. Our product StrokeViewer, is an AI-powered clinical decision support system, offering a complete assessment of relevant imaging biomarkers within 3 minutes.