 Today I want to talk to you about brain and behavior. So about the relationship between how humans behave and the anatomical structure and function of the brain and how artificial intelligence has allowed us to go a step beyond this group level analysis and make individual inferences, which from a clinical perspective is critical. So historically the cognitive neuroscience domain has been dominated by the search for neuro correlates of all kinds of cognitive processes. And these group level analysis has shown that specific regions have been involved in working memory, response inhibition, cognitive flexibility and so on. These kind of studies have given us a great insight into the functioning of the brain. And one general conclusion that we can take from all these studies is that brain regions do not work in isolation, but they work in collaboration or in connection with other regions. Now, regions can be connected with other regions, both structurally and functionally. And structural connectivity refers to the anatomical or structural connections between two brain regions, which can be examined using diffusion imaging. Now functional connectivity relates to the temporal correlation and activation between different brain regions. So as you can see here, this yellow and orange lines here display the activity over time in two brain regions, posterior singlet cortex and immediate prefrontal cortex. And you see they are nicely correlated, which indicates that these two regions are functionally connected. Now functional connectivity allows to discern functional networks. And here we see two different functional networks, the default network in yellow reds and the attention network in blue. And as you can see here from this display, these two networks are anti-correlated. So the default network is mostly active during rest, rest and attention network is mainly active when performing a task. Now my main research interest is in the link between how humans behave and the functional and structural connectivity, both in healthy and clinical samples. And some recent studies that we did showed that indeed also in this brain tumor patients, we can show a link between structural connectivity, functional connectivity, and cognition. For instance, in this case here, we showed that the distance of a tumor to a specific white matter bundle, which is the frontal astral tract here in the right hemisphere is related to executive functioning. So brain tumor patients who have a tumor that is further away from this white matter tract, they perform better on executive functions. Furthermore, we have shown that also functional connectivity is related to executive functioning in this brain tumor patients. So if you look at changes in functional connectivity from rest to the performance of a task, patients who show larger changes in specific networks, in this case here, the frontal parietal network, these patients perform better on executive functioning. These studies all relied on group level analysis. And as I mentioned in the beginning, in a clinical context, it is crucial that one can make individual inferences. And that's why we turn to machine learning, the use of machine learning in some new studies and projects that we have started. And these I want to present next. So a first project uses the database that we have been generating the last decade or so at the Elizabeth Tristeen Hospital in Tilburg. And that contains clinical, neuropsychological, and imaging data for more than 1,000 brain tumor patients. And using this database, we try to optimize for each individual patient the balance between a maximal resection of the brain tumor, which should lead to better survival and minimizing the cognitive deficits resulting from this tumor resection so that people will have the best, the most optimal cognitive functioning and quality of life after the tumor resection. So to do so, we have to take or we are taking several approaches. And one of them is using machine learning on the data in this database and use clinical, neuropsychological, and imaging variables to try to predict the patient's functional outcome after tumor surgery with the aim to have a better informed shared decision-making. On top of that, we are developing new clinical methods to guide the neurosurgeon during brain tumor surgery. For instance, virtual neurosurgery. Now, virtual neurosurgery relies on the connection between functional and structural connectivity. Now, functional connectivity is largest or highest if the two brain regions are anatomically connected. But this correspondence is not perfect. So two brain regions can be functionally connected even without any anatomical connection between these brain regions. So how functional connectivity and structural connectivity are linked to each other remains a question. In a recent study that is shown here, Thawar and colleagues used deep learning to try to predict functional connectivity from structural connectivity and their accuracy of predicting such a functional connectivity matrix was quite high at the group level. So it's about 9% accuracy, which they did a great job here. And more importantly, also at an individual level, they were able to predict the functional connectivity from an individual's tractogram. There is some variability here between subjects. So there's still some room for improvement. But these results suggest that one might be able to examine the link between brain function and cognitive performance of a subject without actually measuring the functional connectivity in this patients or in this healthy subjects in this case here. So that's where virtual neurosurgery comes in. So we can virtually lesion this structural connectome and we can modulate the size of this lesion and see what size of a lesion shows the optimal balance between maximally removing the tumor and minimizing the cognitive deficits that follow from that. So these results can then be summarized in a risk map, which can be used by a neurosurgeon as an indication of the safe boundaries for his tumor research during the surgery. Now, in most cases and other cancer cases, you don't have this kind of huge database as we have for the brain tumor patients. So if you want to apply machine learning to this, you need to start combining datasets from different hospitals, which can imagine that these hospitals are not very eager to share data since this data is really privacy sensitive. So that's where this project aims at. We want to develop a technology that you can apply distributed deep learning on imaging data without the need for the data to move outside of the walls of the hospital. And for that, we use the personal hell train, which was developed before. And to explain how this personal hell train works, let's first look at a traditional approach. So in a traditional approach, data is centralized and then on that centralized data, the algorithms can be applied. So a lot of data needs to be transferred from within the hospital to this central location, which is very tricky since this is very privacy sensitive data. So what's the personal hell train does is sending the algorithm to the place where the data happens to be, so within the hospitals. So the data doesn't have to move. And the only data that is transferred back to the central part are actually the results of the algorithm. So in our case, this deep learning on the images. So this personal hell train has been used before, but for this project, it needs to be extended because it needs to allow distributed deep learning on the medical images. And these medical images are now in the medical archives of the hospitals. So they first need to be made fair. So this is actually the first step in this project to make this images fair. Then we can combine it with a personal hell train and start applying it to specific use cases. And we're going to use specific use cases to show the value of this technology. And in this case, we will focus on predicting the functional outcome of patients based on this imaging data that is distributed across several hospitals. And we're going to focus on the different phases within the cancer care. At this point, we already have a concept version of a similarity dashboard that shows patients the treatment that all the similar patients have had and what kind of outcome this similar patients had. For the treatment and the follow up phase, we will develop personalized prediction models to provide these patients information about which treatment to choose, which has the best outcome, and what kind of functional outcome to expect on the short and the long term. Here we focus on specific cancer use cases, but imaging is actually used in a very diverse set of disease management. So you can imagine that this technology can be applied to those areas as well. So this project is actually the first step to apply to this whole range of diseases. Thank you for your attention.