Best Oral and Poster Presentations Women in Science with Mirabela Rusu A publication by DAILY October 8-12
Elan’s picks of the day (Tuesday): Elan Naideck is a software engineer with Claronav, a leader in optical navigation systems based in Toronto, Canada. He’s a member of the Micron Tracker development team at Claronav which develops the core computer vision hardware and software for a number of guidance and robotic systems ranging from reconstructive dentistry to spinal implants. The Micron Tracker boasts industry leading accuracy and flexibility. (Oral 5) How Reliable are the Metrics Used for Assessing Reliability in Medical … ? (T-03-059) From Mesh Completion to AI Designed Crown (T-03-088) Physics-based Decoding Improves Magnetic Resonance Fingerprinting (T-03-089) Point Cloud Diffusion Models for Automatic Implant Generation (T-03-101) Right for the Wrong Reason: Can Interpretable ML Techniques Detect Spurious Correlations? “I love being given a challenge that seems impossible, only to discover or invent a new technique to pull through in the end. Pushing the limits of the technology so consistently requires keeping an open mind to new ideas and techniques, even ones that seemingly have nothing to do with theproblem’s you’re trying to solve. The algorithm the Micron Tracker uses to track our new ergonomic, omnidirectional tools borrows techniques from database and search algorithms. It’s something we only could have discovered by exposing ourselves to a lot of different ideas. That’s one of the reasons I’m so excited to be here at MICCAI. This is actually my first time coming to MICCAI with Claronav, and I feel like a kid in a candy shop. I get to soak in so many new ideas, and talk to the wonderful minds behind them. The inspiration for our next feature could be on the poster next door.” Meet Elan at the Claronav booth in the Poster Hall. Oral: Posters: For today, Tuesday 10 MICCAI 2 DAILY MICCAI Tuesday Elan’s Picks
Parkinson’s disease, a neurodegenerative disorder affecting millions worldwide, has long been a focus of research to improve diagnostics and treatment. In this paper, Favour presents a method capable of deriving neuroimaging biomarkers associated with disturbances in gait –a common symptom in individuals with Parkinson’s disease. However, the significance of this work extends beyond the laboratory. Favour is determined to make a tangible impact on clinical practice. “A big piece of the work is an explainability framework, meaning that it’s not only computational for how other medical physicists can 3 DAILY MICCAI Tuesday An Explainable Geometric-Weighted Graph Attention Network for Identifying Functional Networks Associated with Gait Impairment Favour Nerrise is a PhD candidate in the Department of Electrical Engineering at Stanford University under the supervision of Ehsan Adeli and Kilian Pohl. Her work proposes a new method to derive neuroimaging biomarkers associated with disturbances in gait for people withParkinson’s disease and has been accepted for an oral and poster. She speaks to us ahead of her presentations this morning. Favour Nerrise
use the work, but more importantly, we created visualizations that you can turn on and off through our pipeline that allow people with relevant clinical or neurological knowledge to interpret what our computational model is doing on the back end,” she explains. “We hope that helps computational scientists, neuroscientists, or clinicians who want to adopt and expand this work translate it into clinical understanding.” One of the most significant challenges was the scarcity of clinical data, a common issue in research involving rare diseases like Parkinson’s. Clinical datasets are typically small and imbalanced, with more healthy cases than diseased ones. Favour developed a novel statistical method for learningbased sample selection to address this. This method identifies the most valuable samples in any given class for training, oversampling them to achieve a balanced representation across all classes. “There were some existing methods that did sample selection either randomly or through synthetic oversampling, but we thought it’d be better to address this directly in a stratified way,” she tells us. “Making sure there’s equal representation of that strong sample bias across every class before applying an external method like random oversampling.” The method performed very well, surpassing existing approaches, including those that inspired Favour to take up the work in the first place, with up to29% improvement for area under the curve (AUC). The success can be attributed to a solution comprising various techniques rather than a linear approach of only visualization explainability or sample selection. Favour plans to expand the research by evaluating it on a larger dataset, the Parkinson’s Progression Markers Initiative (PPMI) database, and exploring new methods involving self-attention and semisupervised techniques. 4 DAILY MICCAI Tuesday Oral Presentation
“We also want to do something that I’mreally passionate about, which is trying to identify cross-modal relationships between our various features,” she reveals. “In this work, we focus on the neuroimaging side, taking the rs-fMRI connectivity matrices and then optimizing that using Riemannian geometry and leveraging some features. Now, I’m interested in combining some of the patient attributes and seeing how that could better inform linkages that could be learned during training amongst other types of techniques we’ll try.” Favour has several follow-up projects in the pipeline that promise to push the boundaries further, leveraging attention-based methods and geometric deep learning techniques. Beyond neuroimaging, she aims to incorporate video data from the same patients. This multimodal approach is a significant next step. She intends to derive specific motion biomarkers that can 5 DAILY MICCAI Tuesday Favour Nerrise
be associated with the existing features. This expansion aims to optimize the learning process and further enhance the understanding of those linkages. The ultimate goal is to combine all these modalities into a comprehensive framework that can be generalized to a broader population. She envisions creating a foundation model that can serve as a valuable resource for researchers and clinicians in various downstream tasks. Research journeys are rarely smooth sailing, and Favour tells us a remarkable aspect of this one was the need for constant course correction in her coding efforts. As the deadline for MICCAI submission approached, the results were not exactly where she needed them to be. “It was weighing on my heart so heavily,” she admits. “At first, we were even doing an approach of comparative binary classification and multi-class classification, and things just weren’t making sense. Then, I just focused on the multiclass classification. Once I did that and started to look into how I could directly optimize my metrics, ensuring everything was weighted in my loss functions, sampling techniques, and all those things, we started to see consistent results that could be repeated over trials. I was so concerned about that because I’d get good results here and there, but I couldn’t repeat them. I was so happy once it got stable enough to have reproducible results. That literally happened a few weeks before we were supposed to submit! I kept updating my paper every day until the night of submission.” Favour remains dedicated to her academic journey. With a couple of years to go until she completes her PhD, she is committed to ongoing research in the field and leadership roles both on and off campus. As we wrap up our time together, she acknowledges the importance of the support she has received along the way. This work has been a significant milestone as her first selfowned project and paper to be released during her graduate school career. It now has the honor of being accepted as an oral at a prestigious conference like MICCAI. 6 DAILY MICCAI Tuesday Oral Presentation
7 DAILY MICCAI Tuesday MICCAI Daily Publisher: RSIP Vision Copyright: RSIP Vision Editor: Ralph Anzarouth All rights reserved Unauthorized reproduction is strictly forbidden. Our editorial choices are fully independent from MICCAI, the MICCAI Society and MICCAI 2023 organizers. Favour Nerrise Don’t miss Yann LeCun’s keynote today! Why not read his interview before?☺ “I literally can’t believe it!” she smiles. “To any other PhD students thinking, I don’t know what I’m doing, does this even make sense? I don’t understand what I’mwriting. Just have faith in your work because when I read my paper, I’mlike, did I write this?! Just have faith in the work you’re doing, and somebody will love it, and you’re absolutely brilliant, and it’s going to be worth it!” To learn more about Favour’s work, visit Oral 5 at 08:00-09:30 in Exhibit Hall A–Main Hall and Poster 3 this morning at 09:30-11:00 in the Poster Hall.
In this paper, Joshua proposes a new deep learning model to advance causal inference analysis of clinical trials. Driven by the pursuit of better treatment options, it represents a culmination of his own research and the work of organizations like the International Progressive MS Alliance to leverage the power of technology to improve patient outcomes. One of the primary challenges of using causal and counterfactual inference in this scenario is missing data from one branch of the trial. For example, if a patient has received only one type of treatment throughout their observation period, it is difficult to predict how they might respond to alternative treatments. The goal is to construct deep models capable of identifying when sufficient data is available for making these predictions. But why is this work so important? The answer lies in the sheer complexity of clinical data. Researchers often encounter patients with unique characteristics not seen in previous trials. These 8 DAILY MICCAI Tuesday Poster Presentation Improving Image-Based Precision Medicine with Uncertainty-Aware Causal Models Joshua Durso-Finley is a PhD student in the Probabilistic Vision Group at McGill University under the supervision of Tal Arbel. His work, exploring the convergence of deep learning and causal inference, reshapes how we analyze clinical trials. He speaks to us ahead of his poster this afternoon.
9 DAILY MICCAI Tuesday newcomers can slow down the decision-making process. “An example of this is you have a testing set, which is all nice and neat, and then a new patient comes into the clinic, and you want to know, how is this patient different?” Joshua poses. “What am I certain about? What am I uncertain about? Sometimes, it doesn’t matter if you’re uncertain. If they’re going to do very well on one drug, you can give them that one. But sometimes it does matter, and you’d like to knowthat.” What sets Joshua’s work apart is its applicability across various fields of medicine. It offers a versatile framework that can be adapted to different medical domains. While he has demonstrated its capabilities in the context of multiple sclerosis research – an area marked by its heterogeneity, availability of drugs, and absence of a known cure – it also holds promise in fields like depression. Even for general triage, it could streamline the initial assessment of patients, helping healthcare professionals quickly determine their risk levels. “The most challenging part of the work was the validation,” he tells us. “Making counterfactual predictions is very well studied, but making them with uncertainty, and then validating the uncertainty when you don’t have ground truth about what’s going to happen, is difficult. On top of that, we wanted a clinical focus. We looked at some causal inference metrics with uncertainty, and we said, when the metrics improve, that’s great, but what does that mean when an individual comes in andyou’re trying to treat them? We have a better area under the uplift curve, which means we’re correctly identifying the right responders, but that’s not important to the individual. They want to know, how does this help me?” Joshua used computer vision techniques to build and analyze deep learning models for MRI inputs. While other aspects of clinical information were considered, such as patient interviews and physical examinations, the real breakthrough came from integrating MRI data. In a Joshua Durso-Finley
10 DAILY MICCAI Tuesday Poster Presentation separate paper, he demonstrated that incorporating MRI information leads to more accurate outcome predictions. Therefore, using MRI should lead to better counterfactual outcomes. Now, by harnessing the power of deep learning on MRI data, he could significantly enhance his model for causal inference. Imagine a scenario where a patient walks into a doctor’s office, experiencing many symptoms that defy easy categorization. The doctor collects information about the patient, possibly conducting a series of MRIs to gain deeper insights. Traditionally, making a diagnosis and determining the appropriate treatment course would be a formidable challenge due to the uniqueness and complexity of the patient’s condition. However, this process becomes more systematic and informed thanks to Joshua’s work. “Because this patient is unique and has these complex features, the doctor can see which drugs would work well for them,” he explains. “These work, these might work, these don’t work. They can discuss the risks of each drug, balance the costs, and make an informed decision about how to treat this patient best.” Looking to the future, while the current focus is on static treatment decisions made at a single point in time, Joshua sees a clear path toward dynamic treatment decisions involving assessing a patient’s response to a particular drug over time. If it does not yield the expected results or thepatient’s condition changes, the model could help doctors decide whether to switch or adjust treatments. This dynamic approach aims to devise a comprehensive profile of each patient’s health journey, ensuring that treatment decisions evolve as new information becomes available. “I am hoping that this paves the pathway for deep learning and
machine learning to be available clinically,” he says. “There are other ways to do this. Some people look at health records at the text data of the patient. Some people look at the genomics data of the patient to build these models. I haven’t seen a lot of deep predictive models using this kind of data unique to the individual. I hope this is a guideline for people.” Joshua’s regular work revolves around building and refining these kinds of models. He is currently working with colleagues at MILA to construct time series models and investigate how they can be adapted for causal inference. He speaks highly of his supervisor Tal Arbel. “Tal has a very good vision of what to do,” he smiles. “She has the foresight to know where the field is going. She’s seen it all and knows what’s important, what needs to be done, and how to do it. It’s been excellent working with her.” We asked Tal what is special about this work. “This work is particularly exciting to us because causal models for image-based based personalized medicine represent a new area within medical imaging, with the potential for not only improving individual patient outcomes and but also drug development in clinical trials,” she said. “In this work we show how the integration of uncertainty-aware causal models for personalized medicine sets the stage for safer and more reliable image-based personalized medicine.” After MICCAI, Joshua intends to take an immediate vacation in Vancouver. Despite studying in Canada, he hails from Rhode Island in the US, where he tells us residents possess a unique privilege. “Every Rhode Islander has an automatic license to dig up clams!” he laughs. “I don’t think people know that about Rhode Island. You can go to any lake you want, find clams, and dig them up. It’s very popular!” To learn more about Joshua’s work, visit Poster 4 this afternoon at 13:00-14:30 in the Poster Hall. 11 DAILY MICCAI Tuesday In this Zürich shot: Tal, Joshua, Chelsea Myers-Colet (a graduated master's student from Tal's lab) and Jean-Pierre Falet (graduated masters student at McGill who is a co-author on this paper) Joshua Durso-Finley
Breast cancer screening is an important aspect of healthcare for women aged 40-75. While it is an invaluable tool, the prevalence of cancer in this age group is relatively low, with only about five out of every 1,000 women being diagnosed. However, one of the challenges of breast cancer screening lies in the occurrence of false positives, which can lead to anxiety and unnecessary medical procedures for patients. The key to improving the screening process is addressing these false positives, and researchers Nhi and Dan are turning to computer-aided detection (CAD) software to assist radiologists in this mission. “When a woman goes into screening, they will take four images, two images of each breast, and usually only one breast will be cancerous,” Nhi explains. “If CAD software produced one false positive per image, radiologists must dismiss 400 false positive marks for every meaningful cancer detected. That’s a huge workload. It’s a lot of resources wasted. Women have to go through a lot of extra procedures, anxiety, and in some cases, even biopsy, which is a very serious procedure.” The primary objective is to develop high-sensitivity CAD software while maintaining a low false positive rate. Achieving this goal is far from straightforward, as medical imaging 12 DAILY MICCAI Tuesday Poster Presentation M&M: Tackling False Positives in Mammography with a Multi-view and Multi-instance Learning Sparse Detector Yen Nhi Truong Vu (left) and Dan Guo (right) are Research Scientists at Whiterabbit.ai working under the Director of Research, Thomas Paul Matthews. Their paper proposes a system to reduce the number of false positives for breast cancer screening. They speak to us ahead of their poster this afternoon.
presents unique challenges compared to traditional natural image processing. Unlike natural images, where multiple objects are present, mammograms often contain only a single focal point, even in cancerous cases. Consequently, applying models designed for natural images to medical imaging is not advisable. “If you take a model designed for natural images, every image has an object,” Nhi continues. “Every image can be used to train the model. In medical imaging, many images are negative, so there’s no object at all. How can you use these images to train the model?” Also, breast cancer screening typically involves capturing two images of each breast, providing complementary views of potential findings. A finding can look suspicious on one view but normal on the other. These views must be considered together to make accurate decisions. “Our director, the last author, is very interested in multi-view reasoning, so he came up with this idea of working with two different views,” Nhi recalls. “To make multi-view reasoning work, we must limit how many boxes or proposals we have on each view. That’s when I started to look into the sparsity. At this point, Dan joined and did a lot of work on multi-instance learning.” As each malignant image typically contains only one finding, M&M uses a sparse detector with a set of sparse proposals, limiting the number of false positives. Dense detectors, where you have many dense proposals and anchors, tend to generate many false positives. Furthermore, introducing a multiview attention module facilitates reasoning between the two views, enhancing the accuracy of detection, and multi-instance learning allows the team to harness seven times 13 DAILY MICCAI Tuesday M&M withNhi andDan
14 DAILY MICCAI Tuesday Poster Presentation more negative images for training, a substantial improvement in the model’s robustness. While the team was optimistic about finding solutions to these challenges, it was a journey that required time and dedication. Dan, who is from China, initially joined the project as an intern. She returned the following year, and together with Nhi, who hails from Ho Chi Minh City in Vietnam, they refined their methods. “I think for all of these solutions, intuitively, we expected those methods to work, but how well it works was a somewhat unexpected component,” Dan reveals. “Overall, the model’s performance was surprisingly good at the end.” Nhi and Dan are working to bring their innovative solution into the real world. They have submitted their product to the Food and Drug Administration (FDA) for approval, a crucial step in ensuring its safety and effectiveness. Their optimism stems from pilot studies and positive feedback from radiologists, who have long grappled with the issue of false positives from CAD systems, which slow down the screening process and introduce unnecessary complexity. The team’s model holds the potential to alleviate this burden significantly.
“We want to achieve something that, even at a very low false positive, like 0.1 or 0.2 false positives per image, gets a high recall,” Nhi tells us. “We’ve made progress in this part. If you look at the figure below, our blue curve is quite a bit better than the others. All the curves get close to each other as you go to three or four false positives per image, but that’s not a point you want your software to operate at. We care about the part between zero and one, highlighted in yellow, and you can see that our curve is better than our competitors. It must be around 5% higher in terms of recall at 0.5 or 0.2 false positives.” We have to ask about the origin of the model’s title: M&M. Where could the inspiration for that possibly have come from, we wonder? “Nhi came up with the title M&M, which makes me think of M&Ms, the chocolates!” Dan laughs. Whiterabbit was founded in 2017 with a unique approach to earlystage breast cancer detection through mammography image data. Unlike traditional AI companies solely focused on algorithms and data, Whiterabbit initially owned and operated nine radiology clinics. This direct engagement with the clinical aspect of breast cancer 15 DAILY MICCAI Tuesday M&M withNhi andDan
16 DAILY MICCAI Tuesday Poster Presentation detection allowed the founders to forge close relationships with radiologists, staff, and patients. It cemented their desire to make products that have a meaningful impact on clinical outcomes. “We want to come up with products that help detect breast cancer earlier,” Nhi says. “This doesn’t just come from something like our paper, which looks at the mammogram and detects the cancer, but we also try to improve the experience. It’s not just about accuracy. You can say your model has very high sensitivity, but what if it creates a lot of false positives? As a company, we’re very careful about the different aspects and consequences related to mammography. I feel lucky to be surrounded by people like Thomas and our CTO, Jason Su. They care about the problem and the clinical outcome as opposed to just making the biggest, most data-efficient model.” In closing, the message from Nhi and Dan’s research is clear. When addressing a meaningful problem like reducing false positives in breast cancer screening, it’s essential to identify and tackle the underlying challenges. “Essentially, the key idea is to find aspects of medical imaging that may not exist in the usual natural image computer vision sphere and try to use them to tackle these very special aspects of mammography,” Nhi adds. “It’s not like we can just bring something from detecting dogs and cats over and apply it to patients!” To learn more about Nhi and Dan’s work, visit Poster 4 this afternoon at 13:00-14:30 in the Poster Hall.
17 DAILY MICCAI Tuesday UKRAINE CORNER “Hi, I am Khrystyna Faryna - a PhD candidate at RadboudUMC, supervised by Geert Litjens and Jeroen van der Laak. My research focuses on bridging the clinical implementation gap in histopathology, particularly, robustness under domain shift. Visit our Dynamic AI (DIACOW2023) tutorial on October 12. Here I am wearing a colorful vyshyvanka, the traditional Ukrainian shirt.”
18 DAILY MICCAI Tuesday Tutorial Preview by Camila González Have you ever worked with data from five, or even ten years ago? You probably noticed how different it is from more recent cases. If you didnot, your model definitely did. Many state-of-the-art methods for medical imaging rely on deep learning models that are susceptible to distribution shifts. Several factors cause changes in data acquisition, including ever-evolving scanning technologies and the presence of image artefacts. Likewise, naturally occurring shifts in disease expression and spread can cause the annotated training base to become outdated. As a result, deep learning models deteriorate over time until they are no longer helpful to the clinician. To maintain the expected performance, models must adapt to incorporate new data patterns while preserving their proficiency in the original evaluation set. Continual learning allows us to acquire new information without losing previous knowledge. This opens up attractive possibilities, such as extending the lifespan of medical software solutions and leveraging large amounts of multi-institutional data. I am Camila González, a postdoctoral researcher working at the Computational Neuroscience Laboratory at Stanford University, School of Medicine. Since my undergrad days, I have been passionate about developing deep learning approaches that translate well to dynamic clinical settings. I am excited to be organizing the first MICCAI tutorial on Dynamic AI in the Clinical Open World (DAICOW) together with a wonderful team of colleagues, to be held in conjunction with MICCAI 2023 on the morning of October 12th (starting at 8 am, but don’t shy away if you can’t make the earlycall ). DAICOW @ MICCAI2023
19 DAILY MICCAI Tuesday Yet actually building, approving and deploying medical lifelong learning solutions faces several practical challenges. Our aim with this tutorial is to give participants hands-on insights into how various domain shifts affect the performance of deep learning models in dynamic environments and help them develop strategies to address these issues and correctly monitor performance. We hereby seek to breach the gap in the MICCAI community between technical research on continual learning and the reality of deploying lifelong learning software in clinics. Join our tutorial to learn the technical, clinical and regulatory aspects of developing continual learning solutions. Let us take you through the process of building and deploying medical AI products that learn continuously over their lifetime in our interactive half-day event! We will address the following topics: ❖ Data drift in medical imaging: Common sources of domain shift and their effect on model performance, with a keynote from the fantastic Prof. Jayashree Kalpathy-Cramer. ❖ Continual learning strategies and evaluation: State-of-the-art methods and how to select the appropriate strategy considering performance, flexibility and resource use, with a keynote from Dr. Martin Mundt, a ContinualAI board member. ❖ Current regulations for updating models in different global regions. The event is aimed at a broad audience within our community. Registration is not required, but it does help us assess the number of participants, so please let us know you’rejoining here. Follow us on X at @ContinualMedAI for more updates Dynamic AI in the Clinical Open World
20 DAILY MICCAI Tuesday Wolfgang Wein is the Founder and CEO of ImFusion, a German technology company established over a decade ago that blends software development and licensing with consulting and research to address the unique needs of its customers. AI in the Real World
21 DAILY MICCAI Tuesday ImFusion has crafted a software framework encompassing accelerated platform-independent libraries, front-end labeling tools, and domain-specific plugins to help medical device companies transform cutting-edge research into innovative, minimally invasive surgical solutions that rely heavily on medical image computing. Its commitment to not reinventing the wheel sets it apart from its competitors. “We don’t have to advertise much,” Wolfgang begins. “There are more and more requests for what we do, which means the need is there. We’re very advanced in medical imaging and guided surgery.” The team was initially composed primarily of engineers and PhD experts in the field; now, it has product managers and a back-office operations team. A unique blend of academic innovation, advanced software engineering, highperformance computing, and efficient numerics is at its core. Its software development kit (SDK) reflects this, offering customers a product-grade experience and streamlining the development process. All this innovation is not without its challenges, but Wolfgang tells us the company’s adaptability and foresight have allowed it to navigate any hurdles successfully. While its core framework remains robust, ImFusion constantly monitors trends within different technology groups. “There’s global competition, and the environment is very fast-paced,” he points out. “First and foremost, in machine learning, we must be very selective about what we implement ourselves and where we rely on large frameworks that the global community has adopted. Many years ago, we implemented our own random forest framework that performed better than the one available through OpenCV and others. Now, we’ve removed it from our build because of superior technology.” Another shift he has observed is the development community’s growing preference for Python over C++. ImFusion
22 DAILY MICCAI Tuesday ImFusion is acutely aware of this trend and is responding by enhancing its SDK with powerful Python bindings, which enables users unfamiliar with C++ to harness the capabilities of its C++ program libraries, ensuring it remains at the forefront of high-performance computing. ImFusion’s operations are distributed across seven departments, with computer vision being one of six technical divisions alongside machine learning, ultrasound, computed tomography, robotics, and SDK. “You could say that almost everything relates to computer vision,” Wolfgang adds. “In the vision group that Alexander Ladikos leads, we focus on RGBD, real-time point cloud and geometry processing, endoscopic image processing, and industrial vision in some projects. Computer vision is broader than that, and you also have projects where this is combined with robotics. Some of the endoscopic image processing is associated to our computer vision group. We want to cover all interventional modalities. We also need to use ultrasound and X-ray and be at the absolute state of the art in processing all of those, which requires heavy computer vision.” ImFusion is actively involved in transforming patient care, with 90% of its business centered in the medical sector. A range of customers have been able to commercialize fully with its help, and devices running on its framework are being used on patients in surgical settings everyday. “That makes us very happy,” Wolfgang smiles. “It also creates some of the most interesting and exciting research problems. There are certain interventional images in edge cases where we barely see anything in the images, so from all these regressions of the real-world use of our software, we can improve it even further.” This dynamic environment allows ImFusion to enhance its solutions continually, a unique experience that many in academia seldom encounter. The company also dedicates time each year to developing new technologies and methodologies, publishing its findings at leading conferences, including MICCAI and MIDL, and sharing discoveries with the global research community. As well as sponsoring the event, it has two papers on the MICCAI program this year and its is staffing a booth. AI in the Real World
23 DAILY MICCAI Tuesday With over 40 full-time employees and organic growth spanning 11 years, ImFusion has created a strong foundation, but it remains vigilant about potential challenges. The loss of key customers and critical personnel is recognized as a possible risk. However, Wolfgang considers this unlikely, citing happy clients and an ambition to be one of the best places to work in Germany, aided by a new HR colleague focused on improving employee benefits and overall long-term satisfaction. Also, with a rising interest in what it offers, the company continues to focus on hiring new people. “You could say there’s a risk in people licensing software for this because of the fast-moving pace of all the libraries and fundamentals changing,” he ponders. “Some large players, the big tech companies, NVIDIA, and others keep developing powerful libraries. We interface well with them, and we recommend them. It might be that, at some point, the focus will have to be less on our own software. We want to focus more on that, but if it doesn’t work, we’d retreat back to contract work. Therefore, overall, we’re actually not a very risky business. I guess we’renot much different from RSIP Visionanyway!” ImFusion is bootstrapped, meaning it operates without external investors, providing the freedom to move in the direction it feels would most benefit existing and future clients. “We want to be sustainable and are not geared toward any quick exit,” Wolfgang continues. “We’re serving multiple customers in a very responsible way with long-term relationships.” While the company’s role as a subcontractor for larger corporations can be volatile, he points out that its flexible approach, experience, and size allow it to navigate unexpected changes in project priorities and funding with relative ease. What might be a setback one year could be an opportunity the next. “Nowadays, what’s important is that you have people who are the absolute “… it has two papers on the MICCAI program this year and its is staffing a booth …” ImFusion
24 DAILY MICCAI Tuesday AI in the Real World absolute cutting edge in medical imaging, numerics, math, linear algebra, programming, C++, and object-oriented programming, and you combine engineers with data scientists,” he asserts. “You always need both. You cannot do a product in the medical space with one.” Understanding algorithms and a mix of smart engineering and numerical programming is essential in the highly regulated medical space, and this is baked into ImFusion’s framework. A dedication to collaboration with academia, medical device companies, and clinicians helps it to stay innovative. “We cherish being exposed to academia and want to continue doing that,” he points out. “We’ve found some of our best talent through academic outreach. That’s why we’re very happy to be at MICCAI again.” Wolfgang acknowledges the profound impact of his own academic journey, telling us he owes a lot to his PhD advisor, Nassir Navab, and the community of professors and research groups around the world who raised him in an international environment. He remembers how exciting it was to be a founding member of Navab’s group in Munich. “A piece of advice he gave me very early on because I worked at Siemens after my PhD was to keep my academic profile,” he recalls. “Keep publishing. Keep giving invited talks somewhere so that people know you. We now live this in our company as well. We try to keep publishing. That means my team’s market value and visibility go up, so they could go elsewhere, but that’s fine. They have a public profile and are happy and highly respected. That gives us the possibility for growth. If you compare us to Google or Apple, we’re more like Google – less secretive and restrictive.” If you would like to meet Wolfgang and the team and find out more about ImFusion, visit their booth at MICCAI 2023! “… you combine engineers with data scientists. You always need both. …”
25 DAILY MICCAI Tuesday Nahal Mirzaie is a PhD student at Sharif University of Technology with major in AI. “Our paper Weakly-supervised Drug Efficiency Estimation with Confidence Score Application to COVID-19 Drug Discovery has been accepted at MICCAI 2023 and marks the culmination of my work during my M.Sc. journey. Our research addresses a crucial issue related to the reliability of existing drug discovery based on cellular morphological features methods when there are out-of-distribution phenotypes. I extend my heartfelt gratitude to the MICCAI RISE committee, whose generous support made my in-person participation possible.” Many thanks to awesome Esther Puyol for the intro My First MICCAI
26 DAILY MICCAI Tuesday Women in Science Read 100 FASCINATING interviews with Women in Science! Mirabela Rusu is an Assistant Professor in the Radiology Department at Stanford University. She also has courtesy appointments in the Department of Urology and the Department of Biomedical Data Science. … I've been very lucky to be allowed to make my choices!
27 DAILY MICCAI Tuesday Mirabela, how did you get so many appointments? Well, because my work is very interdisciplinary. The goal of my team is to improve the interpretation of radiology images. We work a lot with prostate cancer, and that places us from radiology to urology. But we are data scientists, so why not also have a foot in the data science space? What are the things that you're doing, and why could they be useful? Very good question! We are working on helping radiologists and clinicians with the interpretation of radiology images. We do this because we really believe that AI approaches can work hand in hand with these clinicians and address some of the shortcomings of their own interpretation of these images, which is what's known as the standard of care. That's pretty much where we go. But we have collaborations that are very active in the space of urology because we're working on prostate cancer. We think of this as being a lowhanging fruit. And it has something to do with the fact that when I started in science, the things that I work on a lot have been registering multimodal data. I actually started with biomedical data. So, it's not really medical, but more biological structural biology. I come from a background of all these techniques that are what we now consider preprocessing techniques. They really are essential for labeling data for AI approaches. So, we build on that expertise that I have accumulated over maybe 15 years of doing science. Nowit’s like, okay, we can develop AI approaches. We can create these labels that nobody has or has taken advantage of. As a team, we focus a lot on not just having yet another study that develops one of these preprocessing techniques but actually taking them to the next step. So, if I say yes, we are going to use registration to align radiology and pathology data, my science doesn't stop at the paper that shows how we do this. My focus is to take this and do the next step, which is labels. Now, let's start training. Let's see what are the advantages and disadvantages of this. A lot of your speech is plural, and you talked about a team. There are also other people. Tell us about that. We have a very diverse team. I run my lab in collaboration with a urologist, Geoff Sonn, an associate professor in the urology department. 50% of the people are in my lab, and 50% are in his lab. But practically speaking, he's a clinician. He doesn't really have time for basic research or AI research. We want to make sure that by working together, we take advantage of the complementary expertise that we have. So that's my other clinical half. But then, together, we are leading this team. Mirabela Rusu
They are the people that really do the science. I talk about the science. I talk about the direction. I guide the science. But ultimately, they are the heroes. They are the ones that are doing the studies. They are the ones that spend the time to actually get them working. So that's why I talk in the plural. We have a big team. How heroic are they? [laughs] I think they're very heroic. They work very hard to achieve their goals. They put in extra hours without me even asking them. They also work very hard to achieve some of these goals that I set ultimately. But they are the ones trying to achieve them. Would I be very far from the truth if I said that they share your passion? I would believe they do. Definitely, they do share the passion. What is your passion? [laughs] I don't think about it as a passion, but I have a goal in life that drives my passion. So, I was always trained as an engineer, which means I have a background in mathematics. A little dry, if you want. And as I was navigating this world, I figured biology and medicine are very fascinating, right? But it's very far away from math and engineering, which is very systematic. You have a mold, you have a pattern, you follow it. You're good at it. So, coming into this world as a scientist, I figured the more I started to dive into science, the more I realized that my skill could be useful in the medical space. As I started figuring out what I wanted to do as a scientist, one of the things that I realized I wanted to do was save a life. So, my goal in life is to use my science to really save a life. Now, of course, that's a big lofty goal. We don't really, even if we do AI for cancer detection, we don't really save a life. We extend life. So it's a little bit more complicated there. But I think that's what drives my life. I really want to have a clinical application with the skills and knowledge that I'm very good at and then actually apply them to a space where I hope to be able to actually help patient care. All lives can only be extended. Right, exactly. I give you 24 additional hours of free time daily. What will you do? If I don't have to work at all? Free to do what you want! You probably would use 23 to work more. [both laugh] No, actually, probably not. I definitely enjoy stepping a little bit 28 DAILY MICCAI Tuesday Women in Science
29 DAILY MICCAI away from science. I enjoy stepping away from computers and AI and whatnot. I actually really enjoy going to nature. I would absolutely go on a hike, go to the mountains, or go skiing. I would spend that time doing something that takes me away from computer screens. You’renot originally from California. No, I'm not. You're right. I am from Romania. Tell us about Romania. [hesitates for a moment] Romania in my life has been very important until I finished high school. I was in Romania until I finished high school. I left for college and since then, I've been away from Romania. I have actually been away from Romania more than I've lived in Romania. But I've lived now in California for six years, in the United States for 17, and away from Romania for about 24years. Wow. So, do you recognize yourself as mixed identities, or do you still have only one identity? What are you? [laughs] Good question! I think about myself as being a citizen of the world. I don't necessarily associate myself with being Romanian because I left young, and I also grew a lot. My trips and my life afterward have made me into the human that I currently am. I am very, very different than I used to be, of course, which I think it's normal. When I left, I was only 19, but I also changed my character. I mean, the world showed me life. Maybe for better or worse, I'm the human that I am right now. Your answer is not obvious, actually. I left Italy when I was 19, and I feel 100% Italian. [both laugh] You are the result of your own experience and choices. Did you have good choices and good experiences? I had good choices and good experiences. I've been very lucky to be allowed to make my choices! This is a big component of my development as a human right now. Also, taking responsibility for them. I think I made good choices. My current life is very busy for me, but it's also very fulfilling. Between the team that I lead and my personal life, kids included. I have two little kids; one is five, and one is two and a half. So, I'm at that level in my life where I have the career that I love. The position in which I am in, being a PI of a research team, is very exciting. It's all my ideas, my baby ideas. It’s like my baby but it’s my “idea” baby. And along with it, I can also kind of develop and ensure that I can take care of my own family. Tuesday Mirabela Rusu
Tell me, what is your next idea? [laughs] I have many ideas. That's actually the reason why, even after I come home from work, I don't stop thinking about science. I have to actually find ways to stop my head from talking about science. I have specific ideas, of course, for the projects that I work on. But then I also have things like, okay, this is kind of what I would want to be. Who do I want to be when I grow up? Right. I still don't feel like I'm an old person or that I've grown up to my full potential. So, I definitely want to make an impact. My goal is to get as far as I can with what I can do or what I know how to do in order to extend that life. Okay, I have a big ask for you now - a really big ask. Your next sentence will not be about science. Go ahead. Perfect! I love being in nature. I love walking. I love trees and forests. I really enjoy playing with my children, and more recently, we started going skating and going on scooters together. We have reached the level where we can all play together in a fashion that is enjoyable for all of us. How did you get to this balance? I think everybody will tell you there is really no work-life balance. [laughs] But I really try my very best to make sure that I'm available for my family. Because of all these deadlines, all these projects and all these papers always need to be submitted. There is no end to it. There is always another project, and there is always another deadline. There is always something else. Kids and families are growing very quickly. I prioritize them. And if my career cannot accommodate my choices in terms of family, then it's not a career I want. There are many options in terms of careers out there. What is the most wonderful thing that you found in California? [laughs] California is very pretty. The Redwood forests are absolutely amazing. I absolutely love them! Do you have a final message for the community? 30 DAILY MICCAI Tuesday Women in Science
31 DAILY MICCAI Tuesday Mirabela Rusu [hesitates for a moment] Hmm, my message to the world would be, first of all, if you're a young trainee, and you're trying to figure out whether science in academia or science in industry is your path, follow your gut. If you feel like that's the path you want to go, go with it. Don't get too distracted by other voices. Mentors are very useful in this. But sometimes I find mentors always providing advice that is useful from their perspective, not yours as a trainee. So, I have gone against the advice of many mentors many times over the career that I have, and I don't regret it one single time. One of them was a postdoc, and I decided to go into industry for three years. I was told, don't go to industry. I said I'm going to industry. Don't go to this company. I decided to go to that company. I got a lot of other ones, such as don't go into medical imaging, and I went into medical imaging. So I advise every trainee to make their decisions based on what they feel they want to do, even if other people are advising them in a different direction. Understand their perspective and understand their positive points, but move on with their own gut and feelings. One of the things that I wanted to do was to go to the industry because I thought I needed to go a little bit on the dark side. I don't think the industry is really the dark side, at least not anymore. But it's an interesting perspective. So I did that, too, and I understood the challenges of both worlds. I still came back as a professor at Stanford. Some people trusted that I could do this job, even though I come from an industry position, I'm feeling okay doing it. I feel like I am accomplishing something. Today, I just gave a keynote for one of the workshops. It's a big step. So follow your gut and go wherever you feel, whether that's industry or academia. Don't close your doors. Try to keep them open as you are doing it, and then let's see how we can do good with this AI approach. Ultimately, that's all that matters. They're very powerful. We have data and skills that are unique and can be used for the better, so why not use them?
32 DAILY MICCAI Tuesday BEST OF MICCAI 2023 Do you enjoy MICCAI 2023? Do you enjoy reading our dailies? MICCAI 2023 does not end this week! Like every year, Computer Vision News will publish the BEST OF MICCAI in the issue of November. Yes, in just 3 weeks! GET THE BEST OF MICCAI! Subscribe for free and get the BEST OF MICCAI in your mailbox. Computer vision News. Meet the scientist.
RkJQdWJsaXNoZXIy NTc3NzU=