Presentation by Dr. Thomas von Känel (Abteilungsleiter Medizinische Genetik, Spital Wallis (HVS) – Zentralinstitut der Spitäler (ZIS))
Title : Le Service de Génétique médicale de l’Institut Central des Hôpitaux
Abstract : tbd
Presentation by Andrey Malinin (University of Cambridge)
Title : Predictive Uncertainty Estimation via Prior Networks
Abstract : Estimating how uncertain an AI system is in its predictions is important to improve the safety of such systems. Uncertainty in predictions can result from uncertainty in model parameters, irreducible data uncertainty and uncertainty due to distributional mismatch between the test and training data distributions. Different actions might be taken depending on the source of the uncertainty so it is important to be able to distinguish between them. Recently, baseline tasks and metrics have been defined and several practical methods to estimate uncertainty developed. These methods, however, attempt to model uncertainty due to distributional mismatch either implicitly through model uncertainty or as data uncertainty. This work proposes a new framework for modeling predictive uncertainty called Prior Networks (PNs) which explicitly models distributional uncertainty. PNs do this by parameterizing a prior distribution over predictive distributions. This work focuses on uncertainty for classification and evaluates PNs on the tasks of identifying out-of-distribution (OOD) samples and detecting misclassification on the MNIST and CIFAR-10 datasets, where they are found to outperform previous methods.
Andrey is a 4th year Ph.D. student at the Cambridge University Engineering Department, supervised by Prof. Mark Gales. His work forms part of the ALTA Project to develop automated approaches to second language learning and assessment. Andrey’s primary research interest is the estimation of predictive uncertainty for Deep Learning, which is important for high risk applications, such as medicine, self driving cars, finance and high-stakes examinations. Additionally, Andrey investigates the application of Deep Learning to Natural Language Processing for automatic assessment of second language learning and proficiency. Specifically, he works on off-topic response detection, relevance assessment and grading.
Registration : https://doodle.com/poll/cis7d28nkvc36e32
Presentation by Dr. Richiardi (CHUV)
Title : Predictive radiology for precision medicine: medical imaging meets -omics
Abstract :High-resolution medical imaging data, together with large-scale genotype data and post-mortem gene expression data, are being generated and shared openly at an increasing pace. This opens tremendous opportunities to uncover new relationship between imaging, which can be performed non-invasively in-vivo and has very good spatial resolution, and abnormal molecular mechanisms potentially underlying abnormal findings such as brain atrophy.
For some diseases, such as multiple sclerosis or dementia, proteomic data is also collected in clinical practice, and can lead to much improved forecasts of clinical outcomes by combining imaging markers and liquid biomarkers.
Even without -omic data, predictive methods are increasingly gaining ground in medical radiology. In all cases, the key is the development of novel computational techniques that can link biological levels and enable individualised prediction and treatment.
In this talk, we will give an overview of how machine learning techniques are used in radiology, and present some of our ongoing work that seeks to combine proteomic data or genomic data with imaging in order to improve differential diagnosis, personal prognosis, and treatment planning.
Presentation by Sébastien Mabillard (Fondation The Ark)
Title : De l’idée aux marchés ou comment valoriser la recherche au travers des réseaux d’innovation
La fondation the Ark est active sur le terrain de l’innovation depuis près de 15 ans. Elle œuvre à valoriser les innovations dans les domaines technologiques au travers d’outils spécifiques de soutien et de promotion. Avec son initiative Swiss Digital Health, elle met à disposition une plateforme unique dédiée à la santé digitale afin de promouvoir les innovations dans ce domaine et mettre en synergie les différents acteurs : institutions médicales, recherche académique et industrie.
Responsable de l’Incubateur the Ark et directeur exécutif du Swiss Digital Health, Sébastien est actif depuis plus de 12 ans dans le domaine de l’innovation et du développement de start-up technologiques. Il est notamment à l’origine des Journée eHealth et des hackathons santé en Suisse, chapter du réseau international Hacking Health. Il intervient régulièrement comme expert sur des projets innovants de santé ou pour esquisser les défis de la santé digitale de demain.
Presentation by Kayhan Batmanghelich (University of Pittsburgh)
Title : Interpretability versus Prediction: A Gentle Balance in Medical Image Analysis
Developing an effective predictive model to characterize the severity of diseases is one of the main aims of the medical vision. However, in the medical domain, the stakes are high, and clinicians are interested in opening the black-box to learn what the decision of the machine is based on. Hence, interpretability is essential. In this talk, we show a case-study on how to maintain a balance between interpretability and predictive power in a context of Chronic Obstructive Pulmonary Disease (COPD). We present a generative probabilistic approach for discovery of disease subtypes determined by the genetic variants. In many Kayhan Batmanghelich, multiple types of pathology may present simultaneously in a patient, making quantification of the disease a challenging task. While generative models are highly explainable, they usually fall behind the discriminative model, such as deep learning, for a prediction task. We show how generative and discriminative models can interact with each other. We show the evaluation of the method on a large-scale cohort of patients with COPD.
Kayhan Batmanghelich is an Assistant Professor of Department of Biomedical Informatics and Intelligent Systems Program with secondary appointments in the Computer Science and Electrical Engineering Departments at the University of Pittsburgh and an adjunct faculty in the Machine Learning Department at the Carnegie Mellon University. His research is at the intersection of medical vision (medical image analysis), machine learning, and bioinformatics. He develops algorithms to analyze and understand medical image along with genetic data and other electrical health records such as the clinical report. He is interested in method development as well as translational clinical problems.
Presentation by Benoît Dubuis (Campus Biotech, EPFL)
Title : Idée-Innovation-Impact : Du processus créatif à l'économie de l'immatériel
Benoît Dubuis possède une expérience internationale de plus de 30 ans dans les Sciences de la vie, tant dans l’industrie que dans le monde académique. Après sa formation d’ingénieur, son doctorat de l’ETH Zurich (Médaille ETH 1995) et une activité académique à Cambridge et Cranfield en Angleterre, il a occupé différents postes de direction dans des entreprises pharmaceutiques (Chemap, Ciba-Geigy/Novartis, Lonza), avant de rejoindre l’École Polytechnique Fédérale de Lausanne (EPFL) où il a fondé la Faculté des Sciences de la vie, et en fut le premier doyen.
Presentation by Prof. Patrick Aebischer (EPFL)
Title : Med/Bio-tech from research to the market, a perspective and the needed ingredients for success
Patrick Aebischer is a well-known swiss professor with various academic and industrial activities, both very much linked to biotechnologies. Below is a short despcription of the 3 areas in which Patrick Aebischer is well-known.
Sir Aebischer is trained as a medical doctor (MD) and holds a PhD in neurosciences from the University of Geneva. He then moved to the US, Brown University to be more specific, where he held a professorship as the chairman of the section Artificial Organs, Biomaterials and Cellular Technology of the Division of Biology and Medicine until 1992. In 1992 he moved back to Switzerland as a professor, heading the Surgical research Division and Gene Therapy at CHUV. His areas of interest are varied, ranging from Biomaterials over artificial organs to neurodegenerative diseases, a field in which he is active today.
Technology transfer, entrepreneurship
Sir Aebischer is a well-know entrepreneur in the field of biotech. He has founded himself 3 startups, CytoTherapeutics inc., Modex Therapeutics inc. and Amazentis SA across 2 countries (USA and SUI). He sits on the board of the Lonza Group, Nestlé and presides the advisory board of the Novartis Venture fund, on of the largest corporate investment funds in Switzerland having over 40 companies in their portfolio.
Sir Aebischer was the president of EPFL for 16 years until the first of January. He instored an american model, rendering EPFL very competitive. Moreover he is the instigator of various major events that took place such as, the construction of the rolex learning center, press-conference establishing a framework for the Iranian nuclear negotiations, announcement of the double diploma (École Polytechnique Paris - EPFL), a talk given by Gordon Brown on the environment or creation of "under one roof" project to bridge culture and science. Last but certainly not least, a lot of sponsors have been attracted to EPFL under Aebischer's reign, e.g., Ernesto Bertarelli, Daniel Borel and Jean Claude Gandur invested in EPFL which gave further rise to thrilling R&D.
Presentation by Michela Bassolino (EPFL)
Title : Evaluating body representations in chronic stroke patients with motor deficits
Our body mediates any interaction with the environment, being at the same time the medium for perception and action. For instance, even to simply reach and manipulate an object, the brain needs to represent the size and the location of the involved body parts. Since no direct sensory signals exist to inform the brain about the metrics of different body parts, it has been proposed that body representations are generated during the life span from the integration of multisensory signals coming from the skin, joints and muscles and through visual information concerning the body. During daily life experience, body representations are constantly updated by a continuous flow of multisensory and motor cues received and sent by the brain from and to the body. This suggests that if the bi-directional flow of information is altered, body representations could be affected, as in patients with sensorimotor deficits after brain lesion (e.g. stroke).
However, besides impressive manifestations of body-related deficit observed in certain pathological conditions after stroke (e.g. somatoparaphrenia, personal neglect), systematic evaluations of body representations through quantitative tasks in stroke patients with motor deficits remain limited.
With a particular focus on the upper limb, in this talk, I will present novel results based on new experimental tasks assessing body representation in stroke patients with upper limb motor deficits. Our findings revealed distortions in implicit and explicit representations of the body metric and of the space around the body. Further analyses aiming at understanding the link between the observed alterations in body representations and patients’ clinical characteristics such as brain lesion, sensory deficits and the severity of the motor impairment will be proposed.
Dr. Michela Bassolino received her Master degree at the Center for studies and researches in Cognitive Neuroscience at the University of Bologna, by working with Prof. Andrea Serino and Prof. Elisabetta Ladavas. She then moved to the Italian Institute of Techology (IIT) in Genova, where she obtained her PhD working in the Robotics, Brain and Cognitive Science unit working with Prof. Thierry Pozzo. During her PhD, in collaboration with the Prof. Marco Bove and with Prof. Luciano Fadiga, she conducted a project on sensorimotor plasticity, action observation and motor imagery during short-term upper limb immobilization in healthy volunteers. After 1.5 year of post-hoc at the IIT, where she mainly studied the effect of immobilization on body perception, she joined the Blanke lab at the EPFL Center for Neuroprosthetics (CNP) to work in a team having the mission of establishing a first CNP antenna at the Campus Suva in Sion, in the collaboration with the Clinique Romande de Réadaptation. Since 2016, she is leading a project funded by a SNSF Ambizione grant, aiming at studying the sensory-motor bases of embodiment and body representations by combing non-invasive brain stimulation (Transcranial Magnetic Stimulation) and virtual-reality in healthy participants and stroke patients.
Presentation by Ramon Roig (Assoc Director Innovation and Business Development - R&D Labs, Covance)
Title : Clinical Trials – From a central model to a patient centric model
Presentation by Navid Rekabsaz (IDIAP)
Title : Learning, Adapting, and Exploiting Word Representations for Text Analysis and Search
Word representation methods suggest a computational model to capture semantics of language by providing vectors as proxies to the meaning of terms, and have become the cornerstone of several text and language processing tasks. In this talk, I first briefly review the subtleties of various word representation models, followed by introducing the Generalized/Extended Translation Models, two recent methods to exploit the term-term similarities given by word representation models in "classical" Information Retrieval (IR) models. Motivated by the issues of directly using vanilla word representations in IR model (i.e. topic shifting), I then present studies on the exploration of the vector representation space, as well as learning novel word vectors, specifically tailored for document retrieval. Finally, I briefly discuss the interpretability aspect of the word representations, followed by presenting two applications of the introduced methods in financial sentiment analysis, and gender bias detection.
Navid Rekabsaz is a post-doctoral researcher at the Natural Language Understanding team of the Idiap Research Institute. His research focus is on the intersection of Deep/Representation Learning with Natural Language Processing and Information Retrieval. Before then, he was a research assistance at the IMP lab of the Vienna University of Technology (TU Wien), where he pursued his master and PhD.
Presentation by Mansoor Fatehi (Medical Imaging Informatics Research Center, Tehran, Iran)
Title : Musculoskeletal CAD & Microscopic image analysis automation
Part 1 : Deep Learning & Microscopic Image Analysis: Towards Automated Daily Laboratory Medical Services
A general medical diagnostic laboratory provides routine services in daily practice based on microscopic findings. This process bears the potential to be automated by using computer vision while deep learning methods can prepare the basis for object detection and classification to be inserted into the report. Peripheral blood spear, urinary sediment, spermatogram and Pop smear are four most commonly performed tests based on microscopy, each facing diverse challenges of image understanding and clinical validations. The purpose of this presentation is to overview methodology and achievements of our ongoing project aiming to develop a holistic solution to daily needs of medical labs.
Part 2 : Musculoskeletal Imaging Computer Aided Diagnosis: Trends and Challenges
Musculoskeletal tissue structures and abnormalities have been a common area of interest in development of computer aided diagnosis and detection, using images of actually all modalities.
In this presentation, after an overview of the trends in MSK CAD research in the past couple of decades, three projects will be introduced, (bone age determination, MR based FIFA grading system for athletic bone age determination and automated peripheral skeletal measurements), to see how deep learning and computer vision methods may apply to skeletal system and what are the specific challenges of assessing bone, joint, muscle and connective tissues.
Dr. Fatehi is chief of sports imaging department of FIFA medical center of excellence in Tehran and is currently working as technical director of Vista medical imaging center at Tehran where he is a co-founder.
He got his medical degree from Iran University of medical sciences at Tehran continued to residency of radiology graduated in 1998. He became a fellow in University of Maryland for imaging informatics and musculoskeletal imaging between 2009-2010 and then got certified by American Board of Imaging Informatics to become a CIIP. He established medical imaging informatics research center in Tehran after returning to homeland to conduct research and training on computer applications in radiology.
Presentation by Dr. Mathew Magimai Doss (IDIAP)
Title : Automatic speech assessment: combining knowledge and data
Abstract : Speech assessment is crucial part of development of speech technologies such as speech transmission systems, text-to-speech systems, language learning systems, speech therapy systems to name a few. Assessment of aspects such as speech intelligibility, speech quality is typically carried out through subjective tests, which can be time consuming and expensive. Furthermore, it may not be always be reproducible. In this talk, I will present two different automatic speech assessment frameworks that we are pursuing: (a) linguistic prior knowledge-driven and (b) end-to-end data-driven that uses minimal prior knowledge. Through experimental studies on intelligibility assessment, assessment of accentedness of non-native speech and detection of presentation attacks, I will demonstrate the potential of these frameworks in combining knowledge and data for effective speech assessment.
Bio : Dr. Mathew Magimai Doss received the Bachelor of Engineering (B.E.) in Instrumentation and Control Engineering from the University of Madras, India in 1996; the Master of Science (M.S.) by Research in Computer Science and Engineering from the Indian Institute of Technology, Madras, India in 1999; the PreDoctoral diploma and the Ph.D. from the Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland in 2000 and 2005, respectively. He was a postdoctoral fellow at the International Computer Science Institute (ICSI), Berkeley, USA from April 2006 till March 2007. Since April 2007, he has been working as a Researcher at the Idiap Research Institute, Martigny, Switzerland. He is also a lecturer at EPFL, where he teaches courses on speech and audio processing. He is a Senior Area Editor of the IEEE Signal Processing Letters. His main research interest lies in signal processing, statistical pattern recognition, artificial neural networks and computational linguistics with applications to speech and audio processing and multimodal signal processing.
Presentation by Dr. Paul Matusz (HES-SO Valais-Wallis & UNIL)
Title : Understanding the role of attention in visual rehabilitation: Amblyopia as a model
Abstract : Pediatric amblyopia (PA; “lazy eye”) is the most common vision disorder in children, reflected by decreased vision acuity in an eye that is otherwise functions normally. Traditional PA treatments, such as patching, are slow, with the delayed patient rehabilitation creating a cascade of problems for the patients themselves (from low self-esteem to bullying), their families and the society. A substiantal number of children never recovers normal levels of visual acuity in the weaker eye. As such, there are several pertinent questions related to both the etiology and rehabilitation of amblyopia. First, it remains to be shown whether intervention programs shown to improve visual and/or attentional functions in adults can rehabilitate PA - when turned into child-friendly, engaging yet attention-demanding games on in-home devices. Second, we need to identify tools to assess “functional vision” skills, such as reading or face processing, and whether game-based interventions also improve those in PA. Lastly, we need to understand the role of low-level sensory and perceptual versus higher-level cognitive processes in driving basic and functional vision plasticity in PA. I will present a recent project that aims to fill in these lacks in knowledge by assessing the relative importance of sensory and attentional brain processes in vision recovery in PA as a function of treatments training sterovision (3D vision) skills in context of engaging games delivered on virtual-reality devices. The project builds on the synergies across eHealth data-analysis solutions, new rehabilitation approaches and neuroimaging expertise present between the Information Systems Institute (HES-SO Valais) and University of Lausanne (Dept. of Radiology & the University Ophthalmology Service). If successful, the new rehabilitation approach would have the potential to be adapted to other pediatric as well as adult vision/sensory, cognitive and/or motor disorders.
Bio : I completed my Ph.D. in 2013 at Birkbeck College London under the supervision of Martin Eimer. In my Ph.D. project, I employed event-related brain potentials (ERPs) to demonstrate how early in the adult brain the attentional object selection is controlled by salience-based and goal-driven types of multisensory processes. Since completing my Ph.D and together with Gaia Scerif at Oxford University, I have been studying how the dynamic interplay between multisensory processing, selective-attention skills and experience shape object recognition in school-aged children. In 2014, I started a 3-year-long postdoctoral training in employing state-of-the-art EEG signal analysis methods to understand brain and cognitive mechanisms orchestrating the perception of, selective attention to and learning of simple and complex multisensory objects, across the lifespan. In 2016 I received my first competitive grant as principal investigator and have since received several additional competitve grants as principal or co-investigator to study the role of multisensory attention in learning and object recognition in healthy and atypical populations.