Sunday 2019 October 13-17 Daily

GE Healthcare

A Word from the President Kerstin Kläser Zongwei Zhou Women in Science

Workshop Posters MICCAI 2019 2 Editorial DAILY

Good morning MICCAI! Welcome to MICCAI 2019 and the first issue of your new MICCAI Daily. MICCAI and RSIP Vision (the publisher of Computer Vision News) have partnered for the fourth consecutive year to bring this magazine to our community throughout the conference in . Follow us today, Tuesday and Wednesday for daily news, highlights and previews of all the action from what we can promise will be a unique and memorable event! This is already a record-breaking year for MICCAI. We have seen an all-time high number of submissions at almost 1,800. That’s a whopping 75% increase on last year. With 538 accepted papers – 165 more than 2018 – that’s an acceptance rate of around 30%.

To accommodate more orals, one of the innovations we have introduced this year is to limit talks to 8 minutes, with questions delivered to the panel of speakers at the end of the session via the Slido platform. This should streamline both the flow of speakers and the asking of questions. In addition to themed oral sessions, we also have a session devoted to scientific developments coming from within industry, a session devoted to the significance of MICCAI challenges, and the Doctoral Symposium, which highlights the work of MICCAI’s up-and-coming generation.

For the second year running we have a conference app that allows you to personalise your schedule so that you won’t miss a thing. It gives you speedy access to all aspects of the conference, including the posters.

Before we leave you to savour this magazine and the day ahead, I (Terry) have to say a big thank you to my fellow Program Executive members – Larry, Caroline, Sean, Pew-Thian, and Ali. Also, MICCAI’s Paper Submissions Coordinator, Kitty, and Li and Gang from the Publication Committee. We couldn’t have got to the end of the enormous task of assembling this fantastic conference program without you all! On my side (Ralph), I want to thank Rotem Sahar for taking the charge of graphic design on this MICCAI Daily.

Terry Peters Ralph Anzarouth Program Chair, MICCAI 2019 Editor, Computer Vision News Scientist, Robarts Research Institute Marketing Manager, RSIP Vision MICCAI 2019 DAILY Summary 3

A Word from Leo Joskowicz Kerstin Kläser 4 12

GE Healthcare Women in Science 8 14

Zongwei Zhou Workshop Posters 10 18

MICCAI Daily Unauthorized reproduction Publisher: RSIP Vision is strictly forbidden. Copyright: RSIP Vision Our editorial choices are fully All rights reserved independent from MICCAI and the MICCAI Society. MICCAI 2019 4 Welcome DAILY

manage the growth and interest in AI and radiology, and in AI and surgery. There are specific actions that we are taking to grow and capitalise on this.

Of course, we all salute Wiro for all the great work that he has done over the years. Getting back to this year’s event, what is particularly new?

What you can see from the programme Leo, the MICCAI conference has is the growing strength of AI and deep started out very positively. learning, but also diversity into different clinical niches and technical niches. Yes, I’m very impressed by the city You see a record number of workshops of Shenzhen and by how well the and more papers. You see a very organisation team have put the healthy field, with a growing interest conference together in this wonderful from different parts of the scientific and venue. I’m looking forward to a great commercial community. MICCAI. I want to extend a very warm welcome to all participants. We What are your main expectations for have about half researchers and half MICCAI this year? students from all over the world. We have an important Chinese presence I think it will be very useful for students and are very happy that many have and researchers to establish contacts joined us for the first time. Since we are with Chinese researchers. We have an international organisation, we have more than 20 companies. Most of people from all around the world here. them are Chinese, some start-ups, I’m sure that participants will learn a lot so it is going to be great exposure for and enjoy their stay in Shenzhen. people around the world to come and see what’s happening in China. Also, As the new president of MICCAI, can Chinese researchers and students that you tell us what you will bring to the come to MICCAI for the first time will be organisation? exposed to a series of very interesting topics. This exchange will be particularly I’m fortunate to stand on very good valuable due to the large size of the grounds. My predecessor, Wiro Niessen, conference this year – we have about did a wonderful job. The organisation 2,300 registrations. is very healthy and participative. The MICCAI board helps in that. What we want to do in the coming years is to MICCAI 2019 DAILY Leo Joskowicz 5

When choosing a location for the It has to do with alignment of the strategic conference, is there a danger that directions that the different companies you exclude some researchers who choose. It doesn’t necessarily have can’t make the trip? to be that all the large European and American companies are here. Their This is the way an international presence is not in the booths, but in the organisation works. You switch between participants. continents on purpose to give exposure to different parts of the world. It happens You spoke about the rotation, and I when you do the conference in Europe know that next year’s conference is or in the Americas too, where you would very close to your heart. Can you tell have a larger presence from that region. us more? This is done on purpose and is the nature of an international organisation. Yes, I’m very happy to organise MICCAI Being in Shenzhen this year gives a lot 2020 with Daniel Racoceanu in , of people the opportunity to more easily Peru. I’m originally from Mexico, I grew attend, both from China and the rest up there, so the vision we had was to of Asia. It is the nature of the rotation bring MICCAI to Latin America for the which tries to give different regions an first time. We put the bid in four years opportunity to take part. ago and are so happy about it. We want to bring a similar effect to Latin America "Being in Shenzhen this as we have to China this year, so that year gives a lot of people the people from that region who would opportunity to more easily otherwise not necessarily attend come to the meeting and get some exposure. attend, both from China and the rest of Asia."

This also extends to the industry. As you said, most of the companies in attendance are Chinese. Would it have been better to have some other companies from other regions attend and join the talks?

I think there is a presence, if not necessarily in the exhibit, of researchers and people from industry from other parts of the world. It is only natural that the locals have a larger presence. We do have GE as a large sponsor. MICCAI 2019 6 Welcome DAILY

Yesterday evening, I had dinner with a is as important as the formal one. My participant who asked me: “Will people advice would be to enjoy yourself and travel to Peru, and are there enough talk to as many people as you can. laboratories in South America to justify that?” One participant asked me: “I work for a company on medical imaging Latin America is relatively large and research and I am allowed to go although there might be fewer in to one conference every year. Did numbers, it is important to give them a I choose well coming to MICCAI?” chance to have some exposure. It’s not What would you say? just about the numbers of researchers, but also to expose the international I would definitely say yes. There is participants to local health problems a showcase of different research that could be of interest, and we’re directions here. As I said, because of planning sessions on that. It’s not just the size and diversity of the workshop a matter of how many do you have, but events and sessions in the main also what is the impact that bringing conference, I’m sure that everyone something to a region for the first time can find something that is of interest can have, and how will that inspire to them. different researchers. That’s the idea. What can you tell us about What advice do you have for the satellite events (tutorials, researchers who are coming to workshops, and challenges) this MICCAI for the first time? How can year? they take maximum advantage of the conference? These are always an integral part of the conference. This is a chance for I understand that for first-time different groups to gather researchers attendees this can be overwhelming. in their specific areas of interest, and The two important points are to listen this year we have a record number of to the talks and to the invited speakers those – almost 60. They even extend as much as you can, and also to talk to into the evenings to give people people and exchange ideas. This is the maximum opportunity to take part. main benefit of a face-to-face meeting You can choose to follow some of that you don’t get in a virtual meeting. them or all of them according to your Listen to other people’s experiences interest. They are an integral and and interests. People are usually interesting part of the event. very open and friendly and more than happy to discuss their research. The informal aspect of the meeting MICCAI 2019 DAILY Leo Joskowicz 7

Dinggang Shen and Tianming Liu "Let’s not forget that in the are the General Chairs this year end we want to improve and have done so much work to healthcare and help organise the event. Can you tell us something about how it has been people. That is the ultimate to work with them? goal of our research."

Dinggang and Tianming really did an Is there anything else you would outstanding job with their team. It has like to say to participants? been a long time in preparation. They decided to include as many groups, Yes, I very much wish that people workshops and satellite events as go back home highly stimulated and feasible, both for the venue and from excited about all the possibilities for the technical content, so I just want to what technology can do in healthcare. congratulate them on an outstanding Let’s not forget that in the end we job which I’m sure people will want to improve healthcare and help remember for many years to come… people. That is the ultimate goal of our research. "An outstanding job which I’m sure people will The MICCAI Daily magazine is not remember for many years only distributed to participants to come…" here, but also to those in our community all over the world who for many reasons were not able to join us. Do you have a word for them?

I would say the impact is beyond the physical meeting. It goes to proceedings of the conference and of the different workshops. This establishes a record for many years to come in terms of the latest ideas and techniques and is of great value to people who did not attend the physical event. MICCAI 2019 8 EXPO DAILY

Kristin McLeod is the AI and data science team leader for cardiovascular ultrasound at GE Healthcare. She is joined at MICCAI by colleagues from Women’s Health Ultrasound, Acute Care, Magnetic Resonance, Cardiovascular Ultrasound and their Global Research Center. GE Healthcare are a gold sponsor and ran a workshop yesterday called The AI / Data Science team of Cardiovascular Ultrasound Smart UltraSound Imaging. Kristin speaks to us about their work. technologies is crucial. Everything in Throughout GE Healthcare, they are AI is cutting-edge and it’s important building up their AI offering, as they to stay at the frontier of that. see the potential for AI to make a significant improvement to clinical She started working at GE two years workflows. It will make measurements ago, when they were working on their more robust and make it easier for first AI-based feature recognising users to use their systems. It will standard image views from ultrasound also make things more comfortable exams. Although it was only a couple for patients, by making examinations of years ago, this feature is already faster, and enabling physicians to do on the market and in the hands of more measurements, thus providing doctors in hospitals around the world more information for them to make improving their workflows. Kristin diagnoses. found that turnaround time particularly motivating and says they are highly Ultrasound is known to be a bit driven to work fast and provide for subjective and user-dependent, so customer needs as soon as they they want to use AI to remove this possibly can. subjectivity and enable more robust measurements by standardising the In terms of what’s coming up next, acquisition and analysis of ultrasound one step is to automate analysis of images. They have a lot of data readily images. They also want to work on available. the acquisition of images themselves – which is a tough challenge in Kristin’s team counts around 10 ultrasound – and going forward people, including many students. into automating measurements She was a researcher herself before, because these measurements drive so being familiar with the latest diagnosis. Once you automate all MICCAI 2019 DAILY GE Healthcare 9

of the measurements, you can as customer. We’re providing access a consequence create a pipeline to our API which is enabling us to that leads to diagnosis and clinical provide plugins that we can deploy on decision-support systems. our install base throughout the world. This gives start-ups and academics So, why did they choose ultrasound an opening into the market that over other modalities? Kristin says they wouldn’t have had access to she is highly driven by the clinical otherwise. For us, we want to provide application, and ultrasound is the first more solutions for our customers point of care when it comes to clinical and support third-party developers workflows. MRI and CT are only and academics at the same time. prescribed to very severe patients That’s what I’m proudest of, enabling – they’re expensive and there is more research, because I know how radiation, so there are drawbacks. As challenging it is to get access to demoed at the workshop yesterday, data, access to readers, and access they have portable handheld and to customers. We’re providing that laptop-size devices which are easy avenue for others to have interesting to get into the hands of doctors and projects.” much less expensive than other devices. Not only that, but they do 3D An example of this is a Polish imaging too. company called MedApp, who are streaming data from GE’s scanners and doing a HoloLens application. You can stream the image from the scanner to a computer where you can do whatever analysis you want.

GE have also started an initiative called Edison AI Workbench, where they are building up infrastructure Oliver Zettinig from ImFusion scanning Qing Cao from GE Healthcare Cardiovascular Ultrasound, demonstrating the so that on all of their Edison-enabled ease of using VScan Extend probe. products, third parties are able to The aspect of their work that Kristin develop algorithms that can be is most proud of is not a feature, but deployed on their scanners. a collaborative environment they are facilitating between GE and third- Finally, Kristin tells us that GE is party companies: always hiring! So, make sure that you pay them a visit. You will find GE “We’re working together with others Healthcare in Exhibition Booth 15 in to provide more functionality for the the foyer outside Espana Ballroom. MICCAI 2019 10 Oral Presentation DAILY Models Genesis: Generic Autodidactic Models for 3D Medical Image Analysis

Zongwei Zhou is a third-year PhD student in Biomedical Informatics at Arizona State University, where he is working on deep learning for medical imaging. His advisor is Jianming Liang. He speaks to us ahead of his oral and poster session today.

Zongwei tells us that this work provides a pre-trained 3D model for medical image analysis. A pre-trained to medical domain, does it model gives a better starting point. directly from medical domain to It’s easier to optimise, eventually medical domain instead. It’s more leading to better or more stable straightforward if the models are performance. built in 3D directly because most medical imaging, such as CT and Nobody has done this before in 3D. MRI, is in 3D. Common practice uses Professor Fei-Fei Li’s ImageNet, which has Another advantage of this method is it millions of hand-annotated images. doesn’t need any human annotation. Models are pre-trained on that and Zongwei thinks ImageNet is great, use transfer learning to the medical but they had to ask people to label domain. On the other hand, this work millions of images, whereas this employs a method that, rather than method doesn’t need that. The transfer learning from natural image approach is self-supervised learning. It downloads 3D CT or MRI raw data from the Internet and lets the model learn directly from it.

"We want a machine, image by image, to really quickly and really accurately detect a disease." MICCAI 2019 DAILY Zongwei Zhou )周纵苇( 11

Zongwei explains the practical Thinking about next steps, Zongwei consequences of this work: “Our says they need some collected effort work is for disease detection. We from other industries. This work is don’t focus like doctors do on helping self-supervised learning without the a patient to be healthy again. You need for human effort. Everything can have a scan and we tell you that you be downloaded from the internet and have a disease, but you still need to the model can learn it. However, their go to a hospital. Disease detection is lab is just a research lab, and they usually done by radiologists. We want have limited GPU and human mind a machine, image by image, to really power. They have already published quickly and really accurately detect a their work so far and invite people to disease.” contribute.

This will certainly be faster, but will it also be more accurate? He says yes, If you want to learn more for some diseases. Lung nodules, for about this work, come along to example. There are some cases where Zongwei’s oral today at 11:26 and machine learning is significantly poster [T-5-B-013] tomorrow at better than radiology at detecting disease, some where the two are 13:00. The paper is nominated similar, and others where radiology for MICCAI 2019 Young Scientist is better. It’s a sliding scale. Award. MICCAI 2019 12 Workshop Poster DAILY Improved MR to CT Synthesis for PET/MR Attenuation Correction Using Imitation Learning Kerstin Kläser is a third-year PhD student at UCL in under the supervision of Sébastien Ourselin and Jorge Cardoso. She speaks to us following her poster session yesterday at the SASHIMI workshop.

The work she presented explores MR to CT synthesis for PET/ MR attenuation correction. It hypothesises that the main aim of PET for CT synthesis when used for PET attenuation correction is that we shouldn’t minimise the error in the CT image, but we should actually make sure that the error in the resulting PET image is as low as possible. "we shouldn’t minimise the error in the CT image, but we should actually make sure that the error in the resulting PET image is as low as possible." attenuation correction for the PET. She explains that they developed a Basically, the second network can network that generates pseudo CTs be described as a function that can in three stages. The first stage is a estimate the residual in the resulting feedforward network that minimises PET. In the third stage, they train a an L2-loss. This is a well-known combined network of the two. method. The second stage has a second network that predicts the PET residuals that would result when this "MR to CT synthesis for PET/ initial pseudo CT would be used for MR attenuation correction" MICCAI 2019 DAILY Kerstin Kläser 13

This combined network then it’s important that we minimise for minimises the pseudo CT error, but the right thing. The right thing in this also the resulting PET residual. It case is that we in the end generate does that by a combined loss. The a PET that is as close as possible to first loss is the classical L2-loss the real PET.” and the second loss is the whole second network described above. Current training on brain images is It’s minimised as a whole. So, they working well, and looking at next minimise for the CT residual, but steps, Kerstin says they would also for the resulting PET residual. like to try it on different and more difficult areas in the body. The lung Why is this important? Kerstin in particular has huge problems explains: “In order to do PET with attenuation correction because attenuation correction on a PET/MR, there is no signal in the lung from the the gold standard is still to acquire MR. She hopes that their work can an additional CT, which is not really achieve some improvement for this. healthy, and nobody really wants that. This is why CT synthesis in the first place is important. But for the sake of PET attenuation correction,

Little Ingrid is attending her first MICCAI with mom Aasa Feragen, who is supporting her student Rune Kok Nielsen, who has an oral on non-diffeomorphic image registration on Tuesday afternoon. MICCAI 2019 14 Women in Science DAILY

Carole Sudre is a Research Fellow in the Department of Biomedical Engineering at the School of Biomedical Engineering and Imaging Sciences at King’s College London.

Can you tell us what you are doing here at MICCAI?

At MICCAI, there are two things… the For instance, you may have work that I’ve been doing on vascular lesions that appear because of lesions. Also, there is a workshop problems in the blood vessels. that we have been preparing for quite Those problems may be related a long time now. It’s an UNSURE to hypertension, related to some workshop on the Uncertainty for genetic background, related to the Safe Utilization of machine leaRning lifestyle. You can also have other in mEdical imaging. That will happen pathology going on. You have the next Thursday. That’s a workshop with vascular part that can modify the people all over different universities. brain, but you have other issues in the vessels that are not related that Can you tell us about your work? much to cardiovascular risk factors. If you focus on what happens to the So I’m looking at brain images and brain vessels when you age, even trying to discover a bit more about that may be due to different things. vascular lesions when people age. It can be related to cardiovascular What are these vascular lesions? risk factors, problems with activity, What are the different types? How hypertension, or diabetes. You may can we better examine brain images? have also other pathologies. For How does all that relate to cognition instance, in some situations, you and brain function? find in the vessel wall what makes the vessel a bit more stiff and In case one day we will age, what difficult to adapt to the pressure should we expect? and function like they should. They make the vessels stiffer The problem is when we age, there are and less amenable to changes. so many things and so many aspects When you require some supplies, that can modify the brain. Actually, the blood vessels adapt. When it’s really difficult to distinguish any you age, basically, there is some of the pathologies that we discover. fatigue, and they don’t have the same ability to adapt, to expand MICCAI 2019 DAILY Carole Sudre 15

or reduce. That makes the full supply What is a lacunar space? and they don’t have the same ability to adapt, to expand to the brain A lacunar space is the kind of thing more difficult. In the end, it causes you observe in the brain, especially these different markers in the brain. in the white matter, that is due to a small ischemia. In that area, the Is your work to know what is brain lacked oxygen completely happening in the brain or to at some point so the tissue is prevent what happens? destroyed. What we see when we say lacunar space is this absence Well, both. If we want to prevent, of tissue, the scar that results from we need to understand why it ischemia. happens. To understand why it happens, we need to also detect What does it prevent people from where they are. All those things are doing? very tightly related. The first thing is trying to detect them and then It depends. These things happen characterize them. Then actually in a lot of situations. They are using this characterization, you actually very difficult to detect. can better distinguish them from They are not like other subtypes of one another and understand what vascular lesions or changes that we drives the thing to happen. Further see in the brain. For radiologists, down the line, you have what new it is something that is extremely things will do to the blood vessels, difficult to assess and to say that how will they evolve, and what will it’s really a lacunar, in fact, instead be the consequence in terms of of other subtypes of changes that cognition. especially in the lesions happen in the brain that are not that I’m looking into. Since they problematic. What is very important can be anywhere in the brain, you is that you have some of them that have also this relationship between where they will appear and what they will produce in terms of the clinical symptoms and cognitive symptoms in particular.

Can we treat that with to improve the symptoms?

That’s the ultimate goal. That we can only do when we understand where it comes from. MICCAI 2019 16 Women in Science DAILY

are completely silent. People call Why did you choose to specialize them the silent infarcts because in the brain? sometimes you don’t have any symptoms associated with it. People Because it’s such a complex may have them, and actually we system. I have the impression don’t see anything. That’s actually that there is always something what’s really interesting with all to understand better. It’s just a the vascular lesions in the brain. fascinating world in itself. I try to Historically, people thought that all understand what we can do to make those things, lacunar space, lacunar it continue to work. It’s related to infarcts, white matter, that they our whole life. The changes that were normal. They were happening, we observe, especially in older but they had no association with age, they affect so many people. cognition. When looking deeper It’s something that you want to into things, they realize that it may understand better to make sure have some implications. that people, when they age, that they continue being themselves. Did you have the curiosity to perform a CT scan of your own How many years have you spent brain to see what’s going on? on the brain already?

Actually, it’s not by curiosity. At Seven. some point, I needed to help in the preparation of some protocols. I sat Can you point out one in the scanner so I know what my achievement in the past seven brain looks like. years?

Were you curious to check it?... There are so many things that or scared? are worth it actually. I don’t think it’s related to the brain or just Scared? No, honestly, I find it so related to why I like working in amazing to picture the brain. What medical imaging or biomedical we can do is amazing. engineering in general. When you work with clinicians, and they discuss with you what you are working on, at the end, you have this relationship. You have the feeling that what you did is useful to them. For me, that’s really, really cool when that happens. MICCAI 2019 DAILY Carole Sudre 17

Did you hear any feedback from Make that happen in my lifetime... them? before brain degenerates!

To have this relationship between We are so many people working on what the clinician wants and that all it. There are actually good dynamics the people I work with can provide of people trying to work on it. Each an answer to continuously progress. of us is trying with other people to It’s not a loop that goes on by itself. put our own little brick to build a Each time, there is a new solution wall that, in the end, will be a full then there is a new question. Then house. We are all like that, bringing there is a new solution again. There our small, little bricks. is this constant dynamic that makes it really interesting for me. I love that metaphor! Why did you choose England? Is there any specific achievement that you hope to have during your I didn’t choose England. I chose the career? Maybe you can pinpoint people I wanted to work with. I chose something that you want to solve. the group after doing research with them. I wanted to continue working There is a difference between what I with the same people. I loved dream of achieving and the realistic what they were working on. At the ambitions. If we manage to really universities in England, it is fully understand what these things are international. You have so many doing to the brain, how the brain people other than British that being ages, and why do people get sick, a foreigner is almost a norm. why does the brain degenerate, why people end up having problems. If we UNSURE workshop is on manage to really understand what’s Thursday (12:30 to 16:30) in happening. If we understand how all Madrid 5 room. Read many more the pieces of the puzzle, from the interviews of Women Scientists. biology to the microscopic changes that we see to the way that people function. If we can understand the full pathway between those things, then this means we will be able to find ways to treat. For me, that would be amazing. MICCAI 2019 18 Workshop Posters DAILY

Dongyang Kuang (Oden Institute for Computational Engineering and Sciences) presents his poster about a new 3d volumetric registration network with the option to penalize negative Jacobian determinants.

Joost van der Putten (Technische Universiteit Eindhoven) presenting his poster about a model that outperforms 53 medical experts in the detection of dysplasia in Barrett`s Esophagus. MICCAI 2019 DAILY MLMI 19

Abdullah-Al-Zubaer Imran, a PhD student at UCLA, presents his work on semi-supervised learning for joint segmentation of lungs and classification of diseases from chest X-Ray images.

Noha Ghatwary (University of Lincoln) is working on the automatic detection of esophageal abnormalities from endoscopic images.