Deep Learning in Healthcare Summit schedule

Towards the Development of Clinically Relevant Applications of Deep learning in Healthcare

Medicine, by definition, is an information science that requires the capacity to actively acquire individualized and context-specific data, and to then iteratively evaluate, assimilate and refine this information against a vast database of medical knowledge in order to arrive at a small solution space and a corresponding set of implementable policies. Deep Learning, as a transformational tool, is thus extremely well suited to medical application; unfortunately, fundamental understanding of the domain and where and how Deep Learning can be applied in a clinically relevant manner still lag. In this talk I will share my group’s experience over the past decade with trying to develop more advanced and clinically relevant computational approaches in cancer that can integrate large and diverse multi-scale biological data sets including medical imaging, tissue, genomics, and clinical data, in order to predict an individual patient’s cancer genomics and likelihood of response to a particular therapy using only their medical imaging data. I will discuss how we are incorporating Deep Learning in our approaches and highlight other areas of future growth and opportunity in healthcare where Deep Learning can potentially have great impact.

Dr. Kuo received his Medical Degree from Baylor College of Medicine and did his clinical training in Diagnostic Radiology at Stanford University, where he also completed a clinical fellowship in Cardiovascular and Interventional Radiology. He served as Assistant Professor in the Department of Radiology at the University of California-San Diego from 2003-2009. In 2009 he moved to the University of California-Los Angeles where he is an Associate Professor in the Departments of Radiology, Pathology and Bioengineering and served as the Directors of both the Radiogenomics and Radiology-Pathology Programs. Dr. Kuo is an international leader in the field of Radiogenomics where he has published seminal foundational papers. His principle area of research focus is in the field of radiogenomics where his group applies integrative computational and biological approaches in order to derive actionable clinical insights and tools centered around patient stra tification and therapeutic response prediction by leveraging large multi-scale relational data sets including clinical outcomes, clinical imaging, tissue, cellular and subcellular biological data.

Machine Learning systems Design

Machine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision-making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.

Neil Lawrence is a Professor of Machine Learning at the University of Sheffield. His main technical research interest is machine learning through probabilistic models. He focuses on both the algorithmic side of these models and their application. He has a particular interest on applications in personalized health and applications in the developing world. Neil is well known for his work with Gaussian processes, and has proposed Gaussian process variants of many of the successful deep learning architectures. He is also an advocate of of the ideas behind “Open Data Science” and active in public awareness (see https://www.theguardian.com/profile/neil-lawrence) and community organization. He has been both program chair and general chair of the NIPS Conference.

Deep Learning-based Diagnostic Inferencing and Clinical Paraphrasing

Deep learning has emerged as the preferred approach for machine learning with limited or no annotated corpora and hand-crafted features. In addition to the success of convolutional and recurrent neural networks in addressing typical NLP tasks e.g. syntactic parsing or sentiment analysis, integrating “memory” and “attention” in neural networks while leveraging relevant knowledge sources yields results that compare to or outperform the state-of-the-art for analysis of semantically-rich narratives in the clinical domain. My presentation highlights some of our work at the AI Lab in Philips Research NA in which we implement attention-based and memory networks to perform clinical paraphrasing and diagnostic reasoning.

Oladimeji (Dimeji) Farri received his PhD in Health Informatics from the University of Minnesota, and MBBS (Medicine and Surgery) from the University of Ibadan, Nigeria, in 2012 and 2005 respectively. He is currently a Senior Research Scientist at Philips Research – North America (PRNA) in Cambridge, Massachusetts, where he leads the Artificial Intelligence (AI) Lab. His interests are in clinical NLP, text analysis, question answering and dialog systems to address medical dilemmas experienced by patients/consumers and healthcare providers. His recent work includes the use of deep learning in offering solutions for clinical decision support and patient engagement.

Deep Learning for drug discovery: applying deep adversarial networks for new molecule development

Neural networks and other machine learning models have recently been applied to many biological problems, including drug discovery. Applications of deep neural networks combined with domain expertise can help design de novo druglike compounds and generate large virtual chemical libraries, which can be more efficiently screened for in silico drug discovery purposes. This presentation will describe aspects of applications of deep adversarial networks and reinforcement learning for molecular de novo design. It will also briefly cover the Insilico Medicine drug discovery pipeline and how machine learning can be applied on each step.

Polina Mamoshina is a senior research scientist at Insilico Medicine, Inc, a Baltimore-based bioinformatics and deep learning company focused on reinventing drug discovery and biomarker development and a part of the computational biology team of Oxford University Computer Science Department. Polina graduated from the Department of Genetics of the Moscow State University. She was one of the winners of GeneHack a Russian nationwide 48-hour hackathon on bioinformatics at the Moscow Institute of Physics and Technology attended by hundreds of young bioinformaticians. Polina is involved in multiple deep learning projects at the Pharmaceutical Artificial Intelligence division of Insilico Medicine working on the drug discovery engine and developing biochemistry, transcriptome, and cell-free nucleic acid-based biomarkers of aging and disease. She recently co-authored seven academic papers in peer-reviewed journals.

Deep Learning in Medical Imaging - Successes and Challenges

Machines capable of analysing and interpreting medical scans with super-human performance are within reach. Deep learning, in particular, has emerged as a promising tool in our work on automatically detecting brain damage. But getting from the lab into clinical practice comes with great challenges. How do we know when the machine gets it wrong? Can we predict failure, and can we make the machine robust to changes in the clinical data? We will discuss some of our most recent work that aims to address these critical issues and demonstrate our latest results on deep learning for analysing medical scans.

Ben Glocker is Senior Lecturer in Medical Image Computing at the Department of Computing at Imperial College London, and one of three academics leading the Biomedical Image Analysis Group. He also leads the HeartFlow-Imperial Research Team and is scientific advisor for London-based start-up Kheiron Medical Technologies. His research is at the intersection of medical image analysis and artificial intelligence aiming to build computational tools for improving diagnosis, therapy and intervention. He has received several awards including a Philips Impact Award and the Francois Erbsmann Prize. He is a member of the Young Scientists Community of the World Economic Forum. His ERC Starting Grant MIRA is devoted to developing the next generation machine intelligence for medical image representation and analysis.

Quantitative MRI-Driven Deep Learning

Deep Learning (DL) has recently garnered great attention because of its superior performance in image recognition and classification. One of the main promises of DL is to replace handcrafted imaging features with efficient algorithms for hierarchical feature extraction. Many studies have shown DL is a powerful engine for producing “actionable results” in unstructured big data. We present deep learning methods to effectively distinguish between indolent and clinically significant prostatic carcinoma using multi-parametric MRI (mp-­MRI). The main contributions include i) constructing DL frameworks to avoid massive learning requirements through pre-trained convolutional neural network (CNN) models and ii) applying the proposed DL framework to the computerized analysis of prostate multi-parametric MRI from improved cancer classification.

Dr. Sung received the M.S and Ph.D. degrees in Electrical Engineering from University of Southern California, Los Angeles, in 2005 and 2008, respectively. From 2008 to 2012, he finished his postdoctoral training at Stanford in the Departments of Radiology and joined the University of California, Los Angeles (UCLA) Department of Radiological Sciences in 2012 as an Assistant Professor. His research interest is to develop fast and reliable MRI methods that can provide improved diagnostic contrast and useful information. In particular, his group (http://mrrl.ucla.edu/meet-our-team/sung-lab/) is currently focused on developing advanced quantitative MRI techniques for early diagnosis, treatment guidance, and therapeutic response assessment for oncologic and cardiac applications

Deep Learning for Analyzing Perception of Human Appearance in Healthcare and Beauty

Deep learning techniques can be used to extract facial imaging biomarkers of human health status and to track the effects of cosmetic interventions. Here we present a set of tools for analysis of perception of human age and health status. We also demonstrate that when certain population groups are under-represented in the training sets, these populations are left out or may be subject to higher error rates. This is why Youth Laboratories launched Diversity.AI, a think tank for anti-discrimination by the deep-learned systems. The presentation describes the strategies for evaluating human appearance for machine-human interaction and reveals the risks and dangers of deep-learned biomarkers.

Anastasia Georgievskaya is the co-founder and research scientist at Youth Laboratories, a company developing tools to study aging and discover effective anti-aging interventions using advances in machine vision and artificial intelligence. She helped organize the first beauty competition judged by the robot jury, Beauty.AI and develop an app for tracking age-related facial changes and testing the effectiveness of various treatments called RYNKL. Anastasia has a degree in bioengineering and bioinformatics from the Moscow State University. She won numerous math and bioinformatics competitions and successfully volunteered for some of the most prestigious companies in aging research including Insilico Medicine.

The Challenges of Developing an Active Deep-Learning Platform

Deep learning is usually linked to Big Data. However, there are several scientific, engineering and medical imaging problems with limited data available (e.g. rare medical conditions). Can Deep Learning prove a useful tool in such cases? COSMONiO is designing NOUS, an active deep learning platform that aims to make high-accuracy predictions using significantly smaller training datasets. NOUS aims to allow experts from any field to train neural networks without any prior experience. We will discuss the main challenges and how we address them.

Dr. Laurens Hogeweg completed his PhD on medical image processing using machine learning in 2013 at the Radboud University Nijmegen in the Netherlands after having acquired MSc degrees in both medicine and biomedical technology from the University of Groningen. After his PhD he went to industry and developed cloud-based solutions for processing of large image datasets. In 2016 he joined COSMONiO as a research scientist on the topic of deep learning for image processing. His research interest is in the area of learning from small data.

Jorge has a BSc in Biomedical Engineering (2006) and an MSc in Medical Electronics and Signal Processing for Biomedical Engineering (2008) from the Universidade do Minho, Portugal, followed by a PhD (2008-2012) and PostDoc (2012-2015) in medical image analysis and biomarker development between CMIC and the Dementia Research Centre at UCL. In June 2015 he was appointed Lecturer in Quantitative Neuroradiology at the Translational Imaging Group, part of CMIC, in collaboration with the National Hospital for Neurology and Neurosurgery, working on translating and integrating quantitative biomarkers and automated image analysis techniques within the clinical environment. His research explores novel highly accurate and robust machine learning techniques to segment, parcellate and localize different types of tissues using anatomical, microstructural and functional images.

Reza is currently the Chief Scientist at AIG (leading the company's AI research and InsurTech innovations, globally), as well as one of the co-leads of Deep Medicine program at the University of Oxford's Martin School (focused on healthcare innovation, and the use of AI in digital health ecosystems). He obtained his DPhil (i.e., PhD) in computational neuroscience and machine learning from the University of Oxford in 2010, and since then has been leading teams, researches and disruptive innovation projects in both academia and industry.

Machine Learning in Healthcare: Why We Are Not Quite There Yet?

Machine learning promises to revolutionise medical applications such as diagnostics and clinimetrics. Recent progress in algorithms such as deep learning have pushed performance to human-level competence in some applications. However, these algorithms can give meaningless predictions for some kinds of data where humans would not. These confounded predictions could be perilous in mission-critical applications such as healthcare. I will argue that we will have to address difficult issues such as the nature of sampling and data collection from an imperfect world, the accountability of complex predictors, and the need for explanatory rather than just predictive power.

Prof. Max Little is an applied mathematician and statistician. He is a leading expert on clinical signal processing and machine learning algorithms for the use of consumer technologies such as telephones and smartphones to detect the symptoms of Parkinson's remotely. Along with being a Associate Professor of Mathematics at the University of Ashton, he is also a Senior Research Fellow at the University of Oxford and a Visiting Associate Professor at MIT.

Artificial Intelligence and Optical Coherence Tomography - Reinventing the Eye Exam?

Ophthalmology is among the most technology-driven of the all the medical specialties, with treatments utilizing high-spec medical lasers and advanced microsurgical techniques, and diagnostics involving ultra-high resolution imaging. Ophthalmology is also at the forefront of many trailblazing research areas in healthcare, such as stem cell and gene therapies.

Moorfields Eye Hospital in London is the oldest eye hospital in the world. Every year, >600,000 patients attend Moorfields - more than double the number of the largest eye hospitals in North America. Together with the adjacent UCL Institute of Ophthalmology, Moorfields is among the largest centres for vision science research in the world. In July 2016, Moorfields announced a formal collaboration and data sharing agreement with DeepMind Health. This collaboration involves the sharing of >1,000,000 anonymised retinal scans with DeepMind to allow for the automated diagnosis of diseases such as age-related macular degeneration (AMD) and diabetic retinopathy (DR).

In my presentation, I will describe the motivation - and urgent need - to apply deep learning to ophthalmology, the processes required to establish a research collaboration between the NHS and a company like DeepMind, the goals of our research, and finally, why I believe that ophthalmology could be first branch of medicine to be fundamentally reinvented through the application of deep learning.

Pearse A. Keane, MD, FRCOphth, is a consultant ophthalmologist at Moorfields Eye Hospital, London and an NIHR Clinician Scientist, based at the Institute of Ophthalmology, University College London (UCL). Dr Keane specialises in applied ophthalmic research, with a particular interest in ocular imaging. He joined Moorfields in 2010; prior to this, he carried out retinal imaging research at the Doheny Eye Institute in Los Angeles. He is originally from Ireland and received his medical degree from University College Dublin (UCD).

In January 2015, he was awarded a prestigious "Clinician Scientist” award from the National Institute of Health Research (NIHR) in the United Kingdom (UK) - the first ophthalmologist in the UK to receive such an award. His remit from this award is to explore the potential of new medical technologies and innovation in ophthalmology, ranging from advanced imaging to artificial intelligence to virtual and augmented realities. With this remit in mind, he recently established a collaboration between Moorfields Eye Hospital and Google DeepMind to apply deep learning algorithms to ocular imaging. This collaboration will initially involve the application of deep learning to approximately 1 million retinal optical coherence tomography (OCT) images and fundus photographs.

10,000 steps; so what? Are wearable technologies the future of clinical trials?

Wearable technologies such as activity trackers have the potential to speed up the evaluation of medical treatments and reduce the costs associated with their development. Supporters of the use of wearable technologies in clinical trial monitoring argue, that the access to continuous, objective data may allow for a faster, more detailed understanding of the impact of a treatment even in situations that are currently difficult to assess (rigidity in Parkinson’s patients). There is also however an appreciation of the challenges associated with the use of unregulated devices including, amongst others, reliability, interchangeability and data security. Even if the latter challenges were overcome, there is a need to understand how to best utilise data, such as step counts or energy expenditure, in a meaningful manner. This session aims to assess the current use of general wellness tools within clinical trials, potential benefits and limitations of their use as well as the role that ‘deep medicine’ can play to overcome these limitations.

Johanna Ernst is a DPhil student at the University of Oxford working in affilitation with the Institute for Biomedical Engineering and the George Institute for Global Health, where she is involved with the center’s Program on Deep Medicine. As part of her research, Johanna explores the use of wearable technologies for heart failure risk-stratification. She previously worked as a visiting researcher at Misfit Inc., a world-leading wearable technology developer, where she investigated the use of commercially available physical activity monitors for clinical trial monitoring.

Micro EMG: Imaging the Inner Structure of the Human Muscle Guided by a Deep Learning Approach to Muscle Fiber Localization

The presentation will discuss our latest development of a multi-channel Electromyography needle. Using flexible electrodes technology, 64 electrodes are placed in a custom design pattern to maximize the information available for the localization of muscle fibres in human. After the motor units are isolated, an unsupervised stacked denoising auto-encoder is employed to further decompose the motor unit into its constituent muscle fibres leading to localization of fibres with 100 micro meter accuracy of over 50 fibres simultaneously. This will potentially revolutionize neurophysiology and the diagnosis of neuromuscular disease.

Dr. Awwad Shiekh Hasan is a senior research associate in computational neuroscience at Newcastle University. His research is focused on the use of computational modelling to expand our understanding of the fundamental neural mechanisms of cognition and perception, and how that understanding can be translated into action. He worked in several interdisciplinary areas including Brain-Computer Interfaces, neural imaging, and most recently the development of medical devices. He has a British patent and has published extensively in leading scientific outlets in neuroscience, and machine learning.

Collaborative Artificial Intelligence for Healthcare Data

Everyday, new deep learning algorithms are trained to solve specific tasks, such as medical images classification. What if we could share and connect those algorithms and create the conditions for a cross-fertilization between these powerful artificial intelligence systems ? In this talk, we will discuss the fundamental role of transfer learning to foster the emergence of collaborative artificial intelligence and show how it can bring the power of big-data-trained deep learning algorithms into the world of medical not-so-big data. As an illustration, we will present a new platform for medical image recognition based on deep transfer learning and collaborative AI.

Gilles Wainrib is Chief Scientific Officer and co-founder at Owkin, where he leads the data science team. He holds a PhD in applied mathematics from Ecole Polytechnique and was a former researcher at Stanford University and Ecole Normale Supérieure in Paris, working on machine learning algorithms and their applications in biology and medicine. He is the author of 30+ scientific publications in mathematics, physics, biology, medicine and computer science.

Clinical Relevance of Deep Learning to Facilitate the Diagnosis of Cancer Tissue Biomarkers

Tissue biomarker scoring by pathologists is central to defining the appropriate therapy for patients with cancer. However, inter-pathologist variability in the interpretation of ambiguous cases can affect diagnostic accuracy. Modern artificial intelligence methods such as deep learning have the potential to supplement pathologist expertise to ensure constant diagnostic accuracy. We developed a computational approach based on convolutional neural networks that automatically scores HER2, an immunohistochemistry biomarker that defines patient eligibility for anti-HER2 targeted therapies in breast cancer. Our results show that convolutional neural networks substantially agree with pathologist based diagnosis. Furthermore, we found that convolutional neural networks highlighted cases at risk of misdiagnosis providing preliminary evidence for the clinical utility of deep learning aided diagnosis. More studies are needed to show not only the validity of deep learning, but also its utility i n clinical practice to improve diagnostic accuracy. Beyond correlations of artificial intelligence and human made diagnosis, new study designs should be investigated to enable the demonstration that deep learning can improve clinical decision making.

Michel Vandenberghe is working at AstraZeneca, developing deep learning algorithms to analyse immunohistochemistry biomarkers and evaluating the potential uses of deep learning to support biomarker development and clinical decision making. Prior to that, he gained a PhD in Computer Science at University Pierre and Marie Curie, and a Doctorate in Pharmacy, at the University Paris Sud XI.