Welcome back to Beyond the Patterns. So today we have another invited talk in our series and it's a great pleasure to welcome Professor Dr. Mauricio Reis from the University of Bern to our lab.

He will be giving a presentation entitled Medical Image Analysis in the Era of Deep Learning from Performance to Challenging the Alchemist Within.

Professor Reis conducted graduate studies in electrical engineering at the University of Santiago, Chile, where he was awarded the best electrical engineer thesis by the Chilean Institute of Engineering School.

He conducted postgraduate studies to obtain a PhD in computer sciences with a specialization in medical image analysis from Inria, France in 2006.

He is an associate professor at the medical faculty of the University of Bern and is currently leading the medical image analysis group at the Art Org Center for Biomedical Engineering Research at the University of Bern.

His research focuses on basic and applied machine learning methodologies as well as biomedical engineering solutions to improve healthcare throughout medical image computation technologies.

A particular strength of his research has been the emphasis on developing solutions that are designed to be integrated into the clinical workflow.

Dr. Reis has participated in several SWIFT National Science Foundation projects, Commission of Technology and Innovation projects, EUFP7 projects on computational oncology, computational anatomy and further projects supported by Swiss foundations.

From 2006 he has secured over 7.6 million euro research funds, he has an age index of more than 30 and authored over 230 articles with more than 6000 citations.

His entrepreneurial work has also led to the creation of one consolidated company and the second one is in its first steps.

So it's a great pleasure to welcome Maurizio here and Maurizio after the short introduction, the stage is yours.

Thank you very much, Andreas, for the kind invitation. Thank you everyone for being here.

I'm from home. This is my virtual background. This is in case you haven't been to Bern, this is the main building of the University of Bern, it's a nice campus.

So when you arrive with the train on the upper floor, there is this nice campus. So in case you visit Bern, it's a small town, it's a very nice place to come so you are most unwelcome to our laboratory.

And yes, I wanted to show you some of the experience we have on working with deep learning from medical images and the challenging, the alchemist within comes from some of those challenges that we have many times every day while working with these systems.

So as an intro I want to first show you some of the systems that had received first FDA approval. So what you see here already in February 2018, April 2018, first systems or among the first systems receiving FDA approval.

And this is just to sort of motivate the power or the importance, the impact of AI technologies in our society, in our technologies, even to the point that FDA itself has stated that they want to change their way of proceeding with FDA approval to facilitate

and to profit from these technologies.

So now, related to this, many times we read many articles about deep learning methods, results. And there's one article in this sense that I want to show you that is this study, you see Eric Topol, Hugh Harvey and other known researchers, we have a strong clinical emphasis.

And this article, AI versus Clinicians, is a very interesting paper in case you are interested in the translational aspects that talks about this review they performed in the design reporting standards and claims of deep learning studies.

So many times we focus on the methods here, and one of the main messages is we need to also focus on the way we evaluate the way we actually look at the results and we prepare the experimental design of our system.

And one of our main areas is in brain oncology. So here we have seen from the standard machine learning to now deep learning, how this has contributed to our work.

So what you see on the top left is basically a system that performs brain tumor segmentation.

There is available called Bratumia. It uses machine, very classical machine learning using decision trees, that we've been using for many years. Now we are better testing to release the deep learning version of Bratumia, we call it deep Bratumia,

that takes four sequences, normally you need four sequences, MRI sequences, to perform the segmentation of the brain tumor.

And on the right you see a segmentation of organ citrus. This is very important in radiation therapy planning, and this designation needs to be done for every patient.

And this is part of the planning, and this is all organs that need to be spared from radiation. So you can imagine the time consuming and tedious task that hopefully a computer can replace.

What you see on the bottom left is an overlay of these segmentations on a industrial system, commercial system that's used for brain tumor biopsy.

And so here is the idea of enhancing the information by overlaying the brain tumor segmentation on top.

And what you see on the right is longitudinal analysis for patients, typically you have preop and immediate post-operative, and then you have time-consumption, that you can see on the right, in this curve.

The colors correspond to the different compartments of the tumor, and the enhancing part, so you see the active part, this is one of the main areas of interest for clinicians,

and clinical protocols used to assess the disease are also based on the volume or the size of this active part of the tumor.

So you can see the complexity of the task, you have four sequences, volumetric sequences, and then you have time points.

Neuroradiologists need to go back in time, define a starting point, we call it the native point or the baseline, and then start deciding, is the tumor progressing, is the tumor responding to therapy, is the patient stable?

So it's a rather complex task for them. So this is what our hope is to help them with these techniques.

Now, in our research community, one of the initiatives that have grown over time is the Grads Competition, and so this is the brain tumor segmentation.

I wanted to show you first the standard machine learning results we used to see in the past.

So this is a box plot, display of dice, a low dice, it's bad, you want as high as possible, maximum of one, you see on the left, these are the inter-rater variability, this is greater to consensus, and this is inter-rater to rater variability,

and this is the worst case scenario. And then in colors, you see the computer results, and you see that there is a difference between humans and the systems.

You see a lot of variance in the results, so the robustness of the systems is not really similar to the humans.

And you would like computer results to sort of always be in that green area, right, where this is the rate of what humans normally perform, and you would like to avoid any computer result that goes in that direction.

So now that was for the first year, so Grads, when we were using decision trees, FDM, and other classical standard machine learning.

When deep learning arrived around the year 2014, we saw the first approaches, we saw that improvement.

And the next plot I'm going to show you is a summary, a very nice summary from my colleague, Bjorn Menze, who summarized the years 2012 through 2017, and its rank.

So you see again on the left, the best, the top rank, and then on the right, the bottom part of the performance.

Now, you can see already that there is an increase, we are getting better. The variance is also getting better, so it's reducing.

When we zoom into the top, we start seeing very interesting things. So let's zoom into the top.

And we see here, the humans, the human performance, and this one over here corresponds to the fusion of experts.

So when we take all experts, and we fuse them, we call that the reference.

It is anyways a gold, it's not a gold standard, it's sort of a silver standard, because they have errors.