Leiden-Delft-Erasmus collaboration brings self-learning healthcare system a step closer
More effective diagnosis and prognosis than ever, with less intrusive medical screening? Scientists from Leiden, Delft and Rotterdam are well on the way to achieving just that. Imaging professors Serge Rombouts and Wiro Niessen are working on an extremely rigorous, self-learning adviser for radiologists. Their work is based on models similar to those developed by Marcel Reinders. MRI scans have become so commonplace that it is easy to forget what a miracle of technology they are. The recognisable images they produce, such as those of the brain, are the result of calculations based on radio waves. That may be extremely smart, but, according to Serge Rombouts, it’s not artificial intelligence. ‘To be defined as AI, the computer itself has to advise or decide something.’ And this is exactly what the Professor of Methods in Cognitive Neuroimaging aims to achieve. He works in Leiden, at the departments of Psychology and Radiology.
AI learns from brain scans and helps to diagnose
Rombouts is a physicist and an expert in functional MRI. This so-called fMRI produces images of activity in parts of the brain. Much of Rombout’s work focuses on the early identification of brain disorders, such as dementia. ‘We’re working to train computers to identify the disease based on advanced scans of brain networks. The computer combines all the scans of structure, function and activity from one patient and compares them to other people’s scans. Based on a decision model it develops by learning from all that data, the computer may ultimately be able to contribute to the diagnosis.’
Future vision: learning from all data
Imagine that this kind of intelligent system has even more data at its disposal from which to learn. Not only from scans, but also genetic data, blood tests or, for example, data from wearables that monitor your heart function for a day. Imagine that a system of this kind also has the values from other patients at its disposal, with which to make comparisons. This is the vision of the future that Wiro Niessen, like Rombouts also a qualified physicist, is endeavouring to work towards.
The Professor of Biomedical Image Analysis and Machine Learning at TU Delft and Erasmus MC can picture it all: ‘We’re creating Diagnostic Competence Centres that will play a key role in disease prevention, detection, diagnosis and prognosis. In them, we will bring together different types of data and apply AI techniques to learn how to use all that information to improve diagnosis and prognosis.
‘For example, we can develop an AI algorithm using data from prostate cancer patients in different hospitals. Based on those data, we can then learn the relationship between the diagnostic data and the outcome for the patient; this will enable us to develop a system able to look at each new patient and estimate what type of tumour they have and what the prognosis and expected response to therapy will be.’
“Directorate-general for data management”
According to Niessen, there are large amounts of data that have been hardly used at all until now. ‘You do a scan, decide on a conclusion for this patient and often that’s as far as it goes. The challenge will be to develop an infrastructure that enables us to reuse data to develop algorithms. This will not require all of the data from all hospitals to be included in a single system, but if all hospitals store their data in a standardised way, algorithms will be able to make sense of it. What we really need is a “directorate-general for data management” like the one we have for water, we sometimes joke. By way of example, we’re currently working on a nationwide data portal as part of our response to Covid-19.’
Leiden <=> Delft <=> Erasmus
Niessen forms the link between scientists from TU Delft and those at Erasmus MC. ‘Many of my Master’s students from Delft, a lot of them engineers or AI specialists, do internships at the Erasmus MC. Students from various faculties at Delft − Applied Sciences, EEMCS, Biomedical Engineering − are very interested in medical issues.’
Rombouts, from Leiden, is certainly no stranger to Niessen. For his Vici research , Rombouts did analyses using data from the Rotterdam study. The Rotterdam study, a population study involving around 15,000 participants aged 45 and above, is a great source of data for the systems that Rombouts and Niessen are developing. Scientists have been studying this group of people from the Ommoord neighbourhood since 1990. Niessen: ‘That study is a wonderful source of information for us, because, in addition to MRI brain images, it has also collected genetic data and tracks how participants are doing.’ Rombouts: ‘We’re using this data to investigate whether MRI could play a valuable role in predicting the emergence of dementia.’
Spin-off: from knowledge to patient
In order to ensure that all this knowledge and potential can reach the market, and therefore also patients, collaboration with businesses is essential. Niessen himself set up the spin-off company Quantib, which now supplies software worldwide to analyse MRI brain scans. ‘Without this software, a radiologist needs to estimate, based on knowledge and experience, whether, for example, a brain volume measurement or an inconsistency in white matter is normal for someone age 60. Our software can determine this statistically, and it also knows what the average is for someone of that age based on the Rotterdam study. This is making radiology more quantitative and more objective.’
The basis: models, made in Leiden-Delft-Erasmus
The basis for what Wiro Niessen and Serge Rombouts aim to achieve will be models made by selflearning algorithms. Marcel Reinders is working on this challenge at TU Delft and at LUMC. He illustrates what this kind of model is capable of: ‘Right now, we’re starting a project in Rotterdam, in which we aim to estimate whether an incoming Covid-19 patient is likely to end up in intensive care and for how long. For doctors, a sudden decline in these patients can often come as a surprise. This is because they can only assess a few parameters, such as temperature, age and weight. An algorithm is easily capable of a hundred, or many more.’
And he means many more. Reinders on a recent success: ‘You can use human cells to make RNA profiles with which you identify 20,000 genes per cell. In other words, 20,000 parameters. And you then want to do that for 10,000 blood cells, from which you can then identify T cells, B cells and so on. This helps you to tell if someone is ill or not, because there are more T cells and B cells in the case of illness. In Delft, we’ve developed an algorithm to quickly learn to gain an insight into all of these parameters, in very large numbers of cells. With that algorithm, my colleagues in Leiden discovered that there are not twenty, but at least 115 cell types in blood.’
It is good to see the institutes in Leiden, Delft and Erasmus increasingly working together. Reinders: ‘In Delft, they are experts at thinking about fundamental questions, which helps you to make a better model. In Rotterdam and Leiden, huge amounts of medical data are produced that the models can learn from and scientists there are putting their minds to what they can do with that data. Having said that, I’m also noticing that this difference is becoming less marked.’ Well that’s what comes from effective collaboration, of course.