Track B – AI for Medicine

Track B – AI for Medicine

Tue, July 20

8:30-9:00 CET – Opening (joint with Track A)
9:00-10:00 CET – Keynote 1 (joint with Track A)
Mihaela van der Schaar (University of Cambridge)
Mihaela van der Schaar is a Professor of Machine Learning, AI and Medicine at the University of Cambridge, a Fellow at The Alan Turing Institute, and a Chancellor’s Professor at UCLA. She has received numerous awards, including the Oon Prize on Preventative Medicine, a National Science Foundation CAREER Award, the IBM Exploratory Stream Analytics Innovation Award, and the Philips Make a Difference Award. Her work has led to 35 USA patents and 45+ contributions to international standards for which she received 3 ISO Awards. In 2019, she was identified by National Endowment for Science, Technology and the Arts as the most-cited female AI researcher in the UK. She was also elected as a 2019 “Star in Computer Networking and Communications” by N²Women. Mihaela’s research focus is on machine learning, AI and operations research for healthcare and medicine. She is founder and director of the Cambridge Centre for AI in Medicine.

Why medicine is creating exciting new frontiers for machine learning and AI

Medicine stands apart from other areas where machine learning and AI can be applied. While we have seen advances in other fields with lots of data, it is not the volume of data that makes medicine so hard, it is the challenges arising from extracting actionable information from the complexity of the data. It is these challenges that make medicine the most exciting area for anyone who is really interested in the frontiers of machine learning – giving us real-world problems where the solutions are ones that are societally important and which potentially impact on us all. Think Covid 19! In this talk I will show how machine learning is transforming medicine and how medicine is driving new advances in machine learning, including new methodologies in automated machine learning, interpretable and explainable machine learning, causal inference, reinforcement and inverse reinforcement learning.
10:00-13:00 CET – Course 1
Marco Lorenzi (Inria)
Marco Lorenzi is a tenured research scientist at Inria Sophia Antipolis. His research interest is in the development of statistical and machine learning methods for the analysis of large-scale and heterogeneous biomedical data. Current research topics include Bayesian modeling and uncertainty quantification, time-series analysis, latent variable models, and federated learning. He’s an editorial board member of Nature Scientific Reports (Neurology Panel) and he was awarded in 2015 with the second ex-aequo prize for the ERCIM Cor Baayen Award.

Federated learning methods and frameworks for collaborative data analysis

Federated Learning (FL) is an attractive solution for the secure analysis of sensitive data hosted remotely. FL is based on a decentralized learning paradigm which avoids sharing raw data across clients: while models are trained locally on the available data, only model parameters are subsequently shared to define an aggregated global model. The application of FL to challenging scenarios, such as in healthcare, is currently limited by several factors. First, model aggregation is often not possible when datasets are heterogeneous, for example when classes and views are not uniformly represented across clients (i.e. non-iid distributed). Second, although FL avoids data sharing across clients, sharing model parameters may still open up the possibility of information leakage and privacy breaking in presence of malicious clients. Finally, there is currently a limited availability of production ready FL schemes that can be readily used in real-life multi-centric data analysis applications. This course aims at providing a comprehensive illustration of FL in sensitive applications. After introducing the state-of-the-art and principles of FL, I will present current efforts in defining FL frameworks schemes in challenging application scenarios. In particular, I will present novel approaches to FL to better account for variability across clients, and will illustrate the problem of malicious attacks in FL and related defence strategies. The course will be based on the software Fed-BioMed, an Inria open-source frontend framework for federated learning. I will present the basic paradigms for software components for clients and server, and illustrate the workflow for deploying learning models in typical FL scenarios, such as in healthcare applications.
14:30-17:00 CET – Course 2
Gaël Varoquaux (Inria)
Gaël Varoquaux is a tenured research director at Inria. His research focuses on statistical-learning tools for data science and scientific inference. Since 2008, he has been exploring data-intensive approaches for social and health sciences. More generally, he develops tools to make machine learning easier, with models suited for real-life, uncurated data. He co-funded scikit-learn, one of the reference machine-learning toolboxes, and helped build various central tools for data analysis in Python. He has a PhD in quantum physics and is a graduate from Ecole Normale Superieure, Paris.

Dirty data science: machine learning on non-curated data

According to industry surveys, the number one hassle of data scientists is cleaning the data to analyze it. We will survey what “dirtyness” forces time-consuming cleaning. We will then cover two specific aspects of dirty data: non-normalized entries and missing values. We will show how, for these two problems, machine-learning practice can be adapted to work directly on a data table without curation. The normalization problem can be tackled by adapting methods from natural language processing. The missing-values problem will lead us to revisit classic statistical results in the setting of supervised learning.
17:30-19:00 CET – Participant posters and demos (joint with Track A)

Wed, July 21

9:00-10:00 CET – Keynote 2
Lena Maier-Hein (German Cancer Research Center - DKFZ)
Lena Maier-Hein received a Diploma (2005) and doctoral degree (Dr.- Ing. 2009) with distinction from Karlsruhe Institute of Technology and conducted her postdoctoral at the German Cancer Research Center (DKFZ) and at the Hamlyn Centre for Robotic Surgery at Imperial College London. During her time as junior group leader at the DKFZ (2011-2016), she finished her Habilitation (2013) at the University of Heidelberg. As DKFZ division head she is now working in the field of biomedical image analysis with a specific focus on surgical data science, computational biophotonics and validation of machine learning algorithms.

Why domain knowledge (still) matters in medical image analysis

The breakthrough successes of deep learning-based solutions in various fields of research and practice have attracted a growing number of researchers to work in the field of medical image analysis. However, due to the increasing number of publicly available data sets, domain knowledge and/or close collaboration with domain experts is no longer an essential prerequisite. To demonstrate the potentially severe consequences of this recent development, this talk will highlight the importance of domain knowledge for various steps within the development process: from the selection of training/test data in the presence of possible confounders, to the choice of appropriate validation metrics and the interpretation of algorithm results.
10:00-13:00 CET – Course 3
Thomas Moreau and Demian Wassermann (Inria)
Thomas Moreau is a researcher at Inria Saclay. His research interests are centered around unsupervised learning for time series, with application to neural activity recording such as MEG or fMRI. In particular, he has been working on Convolutional Dictionary Learning — studying both its computational aspects and its applications to pattern analysis — and on theoretical properties of learned optimization algorithms for inverse problems such as LISTA. He is also involved in Python open-source projects, mainly around parallel scientific computing. Before joining Inria, he obtained a PhD from ENS Cachan on convolutional representations for physiological signals.

Demian Wassermann is a researcher at Inria since 2014. His research work is divided between the development and application of principled mathematical models for diffusion MRI micro and macro structure and the formalisation of human discourse to analyze neuroimaging data. For both these tasks, he harnesses the power of machine learning and artificial intelligence in probabilistic settings. Demian received a prestigious Starting Grant from the European Research Council (ERC) in 2017 for his project “NeuroLang”. Before joining Inria, he worked at the Harvard Medical School, specifically to the Brigham and Women’s Hospital.

Introduction to neuroimaging with Python

There is growing interest in data-driven analysis, multivariate statistics and predictive modeling for neuroimaging. Datasets are also constantly growing in sample size, resolution, and complexity. Specifically, the tutorial will be based on Nilearn which is a Python package designed to address these new challenges in contemporary data analysis for imaging neuroscience. The tutorial will cover: (i) plotting and image manipulation with nilearn; (ii) decoding and predictive models; (iii) functional connectivity and resting-state data analysis. In all, students will gain knowledge in the processing and visualization of neuroimaging data through NiLearn. NiLearn provides state-of-the-art machine-learning methods for convenient pre-processing, analysis, and visualization of various types of neuroimaging results (i.e., experimental fMRI, VBM, and resting-state correlations).
14:30-17:00 CET – Course 4
Francesca Galassi (Inria) and Rutger Fick (Tribvn Healthcare)
Francesca Galassi is with the Empenn research team, Inria Rennes. Her research interests are in the fields of image processing, computer vision, and machine learning. She holds a Ph.D. in Biomedical Engineering from Imperial College London (UK), and completed a Post-doctoral fellowship in cardiovascular computational modeling at the Acute Vascular Imaging Unit in Oxford (UK). Her current research focuses on brain image processing, specifically segmentation and detection of white matter lesions from multi-modal MRI. Her teaching activity at the École supérieure d’ingénieurs de Rennes comprises Artificial Intelligence, Data Mining and Medical Imaging courses.

Rutger Fick is a senior AI-specialized R&D engineer in medical imaging. His passion is to further the digitization of the hospital and healthcare industry, allowing medical specialists to focus on making high-level decisions and not waste time on repetitive, automatable tasks. His background encompasses three medical imaging domains: first a PhD on dMRI-based brain microstructure imaging at Inria Sophia Antipolis, followed by leading the development of CT-based radiotherapy treatment optimization algorithms at the award-winning start-up TheraPanacea. His current challenge is to improve the clinical workflow in digital pathology imaging by leading the design and development of interpretable-AI based algorithms at Tribvn Healthcare.

Towards domain-shift robustness in medical imaging analysis: A case study in 2D digital pathology and 3D brain MRI

In recent years, Convolutional Neural Networks (CNNs) have shown better performance in image segmentation tasks than the traditional methods. Yet, their clinical use remains limited due to a reproducibility issue across different sites or image domains – meaning the CNN’s performance is sensitive to unexpected shifts in measured signal properties. In this tutorial, we will explore what “unexpected shifts” are relevant in a case study of two medical imaging modalities, namely 2D digital pathology and 3D brain MRI. While seemingly unrelated, we will show that within the scope of “medical imaging” their properties can be examined in similar ways. The participants will get hands-on experience on how to explore the domain shifts that are present in the data, and how to train CNNs in a way that is robust to these shifts. This will include implementing techniques like data augmentation, image normalization and beyond. We will apply these techniques together to cancer segmentation in 2D digital pathology, and multiple sclerosis lesion segmentation in 3D brain MRI.
17:30-19:00 CET – Open discussion with industry
Boris Dimitrov (Check Point Cardio) and Nicklas Linz (ki elements)
Boris Dimitrov is the co-founder and CEO of Check Point Cardio. He is the founder of telemedicine in Bulgaria with 25 years of entrepreneurial experience. Check Point Cardio specializes in a wide range of medical services to provide excellent patient care. It developed one of the first systems for real-time online patient tele-monitoring with a fully operating cardiological medical center for continuous 24/7 observation, diagnostics, patient health management and emergency reaction services. Incepted by cardiologists, this disruptive technology offers quick and precise online diagnostics for preventive and emergency care.

Nicklas Linz is the managing partner of ki elements, based in Saarbrücken, Germany. After completing his computer science studies at Saarland University, Nicklas worked as a scientist at DFKI in Prof. Wahlster’s research group Cognitive Assistance Systems. Here, he dealt with applied machine learning methods and natural language processing for the early detection of dementia. Nicklas is the author of numerous conference and journal publications, which have been awarded several Best Paper Awards. Today, Nicklas is the full-time manager of ki elements, which he founded in 2017 together with other DFKI scientists. The start-up pioneers the use of speech and language as clinical biomarkers for several neurological and psychiatric indications.

Industry use cases for AI in Medicine

Thu, July 22

9:00-10:00 CET – Keynote 3 (joint with Track A)
Joanna Bryson (Hertie School)
Joanna Bryson is recognised for broad expertise on intelligence and its impacts, advising governments, transnational agencies, and NGOs globally. She holds two degrees each in psychology and AI (BA Chicago, MSc & MPhil Edinburgh, PhD MIT). From 2002-19 she was Computer Science faculty at Bath; she has also been affiliated with Harvard Psychology, Oxford Anthropology, Mannheim Social Science Research, and the Princeton Center for Information Technology Policy. During her PhD she observed confusion generated by anthropomorphised AI, leading to her first AI ethics publication “Just Another Artifact” in 1998. In 2010 she coauthored the first national-level AI ethics policy, the UK’s Principles of Robotics. She presently researches the impact of technology on human cooperation, and AI/ICT governance.

AI ethics

10:00-13:00 CET – Course 5
Tim Dahmen (DFKI)
Tim Dahmen holds a PhD in computer science (Saarland University) and is senior researcher at the DFKI, where he leads the team Computational 3D Imaging. He conducts interdisciplinary research in the combination of computer graphics and artificial intelligence, mainly in the biomedical and microscopy domain. His current work focuses on using parametric generative models to generate synthetic training data for deep learning systems. He has contributed to the field of sparse and adaptive data acquisition in electron microscopy and to Deep Learning enabled mechanical simulation.

Bio-mechanical simulation for individualized implants

Fractures of bones in the lower extremities are serious injuries even today. Complications with severe consequences for the patient, including amputation, remain a real possibility. One way to improve treatment quality is the application of patient-specific implants that guarantee optimal healing conditions. However, finding the ideal shape of an implant that will not break even in the case of an accidental misstep, is rigid enough to hold bone fragments in place to allow healing, but at the same time flexible enough to allow to crucial micro-movements that stimulate bone-growth is a difficult optimization task that requires bio-mechanical simulation. While conventional Finite-Element-Method (FEM) simulations can in principle solve such simulation tasks, the approach can strongly benefit from the incorporation of modern machine learning techniques and opens an entire field of research. Using convolutional neural networks allows the automatic generation of simulation-models from computed tomography scans. More interestingly, FEM simulations rely on mathematical estimators for various tasks. By replacing these closed-form estimators with numerical fitted models, i.e. by using neural networks to guess intermediate results, drastic speedups can be realized. One example is to “learn” coarser representations of the model in a multigrid context. These performance improvements are crucial because (1) finding an optimal treatment solution requires running simulations for a large number of configurations, and (2) because in a clinical context, designing and manufacturing an implant is restricted to a rather narrow time window of typically 2-4 days for clinical reasons.
14:30-17:00 CET – Course 6
Denis Engemann (Inria)
Denis Engemann holds a PhD in experimental psychology (University of Cologne) and is a research scientist at Inria Saclay, Parietal team. He conducts interdisciplinary research targeting brain and mental health by combining machine learning with large-scale analysis of brain signals. During his postdoc (ICM Paris & Neurospin), he studied consciousness-revealing EEG dynamics in severely brain injured patients. His current applied work focuses on detecting aging-related cognitive dysfunction from MEG and EEG. He has been contributing to the MNE software since 2012 and is actively working on methods and tools facilitating the analysis of clinical EEG data.

Building EEG-based proxy measures of brain aging in Python with MNE and its ecosystem

Electroencephalography (EEG) and magnetoencephalography (MEG) are promising tools for studying brain health. Machine learning provides a framework for robust modeling of biomedical outcomes as a function of M/EEG, e.g. diagnosis or drug response. In this setting, it is common to predict one such outcome from an entire M/EEG recording, hence, the unit of observation is the patient. If other biomedical data is available, e.g. imagingl or demographic covariates, it is helpful to combine it with M/EEG data. This can be challenging as clinically important M/EEG features e.g. alpha band power extend over multiple frequencies (8-12Hz) and are measured on multiple sensors. This puts strong emphasis on non-linear regression, handling signal distortion due to field spread, building sub-models for different blocks of features and dealing with missing values. This workshop introduces the elements needed to build custom prediction models for biomedical problems from heterogenous EEG signals and other biomedical covariates using the MNE package and its ecosystem. This covers 1) scalable preprocessing of large-scale data and 2) clean APIs for building structured and complex prediction pipelines. This is exemplified by building non-linear regression models for brain age prediction on the Temple University Hospital EEG data.
17:30-19:00 CET – Participant posters and demos (joint with Track A)

Fri, July 23

9:00-10:00 CET – Keynote 4
Gerd Reis (DFKI)
Gerd Reis is a senior researcher and deputy director at DFKI which he joined in 2001. He received his PhD in computer science from the Technical University of Kaiserslautern on the topic of 4D Cardiac Ultrasound. Since 2008 he is lead reasearcher and developer of the ANNA/C-TRUS system, an medical product to aid urologists in the diagnosis of prostate cancer. He has also developed a continuous bladder irrigation system to prevent complications caused by blood clots after operations on the bladder, prostate, or kidneys.

AI in Medicine - An engineering perspective

Since the beginning of the 21st century, Artificial Intelligence has made enormous progress. Many previously unsatisfactorily solvable problems have become manageable due to the availability of modern AI methods, especially Deep Learning. Naturally, this has led to a desire to exploit the advantages offered by AI in as many areas as possible – including medicine. However, it is becoming apparent that there is a huge gap between academic research and practical applicability of a new and promising method, especially in medicine. Issues such as reliability, reproducibility, safety and explainability must be addressed in the course of certification as a medical product. In this keynote, these aspects will be addressed from an engineering point of view by means of concrete, practical examples. The talk is not intended as a guide to a medical certification process, but as a means to fine-tune one’s thinking during research and development of medical applications and devices.
10:00-13:00 CET – Course 7
Pierre Zweigenbaum (CNRS - LIMSI)
Pierre Zweigenbaum is a Senior Researcher at LISN (Orsay, France), a laboratory of CNRS and Université Paris-Saclay, where he led the ILES Natural Language Processing team for seven years. Before CNRS he was a researcher at Paris Public Hospitals and a part-time professor at the National Institute for Oriental Languages and Civilizations. His research focus is Natural Language Processing, with medicine as a main application domain. He is the author or co-author of methods and tools to detect various types of medical entities, expand abbreviations, resolve co-references, detect relations, and link text to biomedical terminologies and ontologies. He graduated from École Polytechnique and Télécom Paris, holds a PhD in Computer Science from Télécom Paris and an habilitation in Computer Science from Université Paris-Nord.

Natural Language Processing for medical applications

In many sources of data, relevant information is conveyed by free text: this is the case when analyzing the contents of patient records, scientific publications, social media, etc. Because of the non-formal nature of human language, contrary for instance to programming languages, computer-based processing of natural language text is challenged by the high variation in expression and the importance of context for correct interpretation. This tutorial will review examples of medical applications of natural language processing, with a focus on information extraction: entity recognition, entity linking and relation detection. The participants will get hands-on experience on supervised sequence tagging for entity recognition from biomedical abstracts.
14:30-17:00 CET – Course 8
Philipp Klumpp, Tomás Arias-Vergara, and Paula Andrea Pérez-Toro (Friedrich-Alexander-University Erlangen-Nuremberg)
Philipp Klumpp is currently pursuing his Ph.D. degree at Pattern Recognition Lab of Friedrich-Alexander-University Erlangen-Nuremberg (FAU), Germany. He received his B.Sc. and M.Sc. in Biomedical Engineering from FAU in 2014 and 2017, respectively. Working for the speech group, his research focus lies on the analysis of pathological speech signals of Parkinson’s Disease, Aphasia, Dysarthria and other neurological conditions. His general research interests lie in the fields of pattern recognition, pattern analysis and machine learning, particularly deep learning for sequential data analysis.

Tomás Arias-Vergara received the B.S. degree in Electronics Engineering from University of Antioquia (Medellin, Colombia) in 2014, and the M.Sc. degree at the same institution in 2017. Currently, he is a doctoral candidate at the GITA Lab from the University of Antioquia (Colombia) and at Pattern Recognition Lab from the Friedrich-Alexander-Universität Erlangen-Nürnberg (Germany) in a joint degree program. Additionally, he is a guest researcher in the Department of Otorhinolaryngology, Head and Neck Surgery from the Ludwig-Maximilians-Universität München (Germany). His research interests include speech processing, signal processing, pattern recognition, machine learning and their applications in pathological speech.

Paula Andrea Pérez-Toro received the B.S. degree in Electronics Engineering from University of Antioquia (Medellin, Colombia) in 2018 and the M.Sc. degree at the same institution in 2021. Currently, she is a PhD. candidate at Pattern recognition Lab from the Friedrich-Alexander-Universität Erlangen-Nürnberg (Germany) and at the GITA Lab from the University of Antioquia (Colombia). She has performed research activities related to signal processing, machine learning, and deep learning for health-care during the last three years, both in academic and industrial partners. Her research interests include speech processing, signal processing, pattern recognition, natural language processing, gait analysis, and their applications in health care and customer service.

Automatic analysis of pathologic speech – from diagnosis to therapy

In this lecture, we will give an overview over different disorders, which influence language and speech and discuss their underlying causes. Examples include speech of children with cleft lip and palate, dysarthria after stroke, Parkinson’s disease, Aphasia and Dementia. We will show how the different forms of speech disorders can automatically be classified in comparison with age/gender matched control speakers (diagnosis), the severity of the disease can be estimated (monitoring), and computer-aided therapy can be designed. The use of different machine learning methods will be discussed. Important design aspects like data privacy, the explainability of the analysis results to the patient and the medical caregiver and the user acceptance of the computer-aided therapy will be presented.
17:30-19:00 CET – Collaborative wrap-up (joint with Track A)

Comments are closed.