Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
The role of AI in personalized therapy
Prof. Arvid Lundervold BSc, MD, PhD
Department of BiomedicineNeuroinformatics and Image Analysis Laboratory
University of Bergen&
Mohn Medical Imaging and Visualization CentreHaukeland University Hospital
with Assoc. prof. Alexander S. LundervoldDepartment of Computing, Mathematics and Physics, Western Norway University of Applied Sciences
Biomedical Network meeting on a Medical Innovation Incubator in Bergen, March 8th 2018
https://mmiv.no
17:45-17:55
W. Ertel, 2017
Bruk av AI til utvikling av persontilpasset terapi
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
The role of AI in personalized therapy
Prof. Arvid Lundervold BSc, MD, PhD
Department of BiomedicineNeuroinformatics and Image Analysis Laboratory
University of Bergen&
Mohn Medical Imaging and Visualization CentreHaukeland University Hospital
with Assoc. prof. Alexander S. LundervoldDepartment of Computing, Mathematics and Physics, Western Norway University of Applied Sciences
Biomedical Network meeting on a Medical Innovation Incubator in Bergen, March 8th 2018
https://mmiv.no
17:45-17:55
W. Ertel, 2017
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
What is AI ?
Stuart J. Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 3rd ed., 2016https://developers.google.com/machine-learning/glossary
Machine Learning
• A program or system that builds (trains) a predictive model from input data.
• The system uses the learned model to make useful predictions from new (never-before-seen) data drawn from the same distribution as the one used to train the model.
• Machine learning also refers to the field of study concerned with these
programs or systems.
“ The art of creating machines that per-form functions that require intelligencewhen performed by people”(Kurzweil, 1990)
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
https://www.gartner.com/smarterwithgartner/top-trends-in-the-gartner-hype-cycle-for-emerging-technologies-2017
Upcoming
- Neuromorphic hardware- Human augmentation- Brain-Computer Interface- Conversational UI- Edge computing / sensors- Smart robots- Virtual Assistants
Peak
- Deep learning- Machine learning
Maturation
- Cognitive Expert Advisors- Augmented Reality- Virtual Reality
MEDICINE
FEAR HYPE HOPE
InnovationTrigger
Peak ofInflated
Expectations
Through ofDisillusionment
Slope of Enlightenment Slope ofProductivity
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
MEDICINE AND THE “NEW”
COMPUTATIONAL FIELDS
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Data Science `producing insights’ e.g. explorative and longitudinal data analysis
Artificial Intelligence `producing actions’ e.g. imaging-guided robot surgery
Machine Learning `producing predictions’ e.g. biomarkers treatment response
Computational science `producing governing equations’ e.g. tumor growth
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Artificial Intelligencein medicine
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Artificial Intelligencein medicine
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIVhttps://pct.mdanderson.org
Personalized therapy
in cancer
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIVhttps://pct.mdanderson.org
Personalized therapy
in cancer
… and then
Artificial intelligence-drivenbiopharmaceutical companies
e.g.http://www.twoxar.com
Virtualscreening
https://www.profacgen.com
Computer-AIded Drug Design
subject-specificas a service
on demand
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
MACHINE LEARNING
( an educational example )
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Predicting academic achievement from inattention (SNAP)
A.J. Lundervold, T. Bøe, A. Lundervold. Inattention in primary school is not good for your future school achievement - a pattern classification study. PLoS ONE 2017;12(11): e0188310
Feature importanceRandom Forest
10000 trees (“weak learners”)
Top 3
k-fold cross validation
Prediction
Accuracy
Precision = tp / (tp + fp)Recall = tp / (tp + fn)
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
DEEP LEARNING
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
http://fortune.com/ai-artificial-intelligence-deep-machine-learning
Human vision vs. computer vision
http://neuro.cs.ut.ee/lab
Along the ventral stream the humanbrain represents increasingly morecomplex visual features. The verysame phenomenon emerges in deepartificial neural networks designed toclassify visual images: eachconsecutive layer of a deep neuralnetwork codes for more complexvisual features than the previous layer.
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Non-invasive estimation of Glomerular Filtration Rate (GFR)
will need fast image segmentationof the kidney
Image-derived biomarkers
https://www.mayoclinic.org
Chronic kidney disease ↑Diabetes, hypertension, …
Functional renal imaging √
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
A. S. Lundervold, J. Rørvik, A. Lundervold
Fast semi-supervised segmentation of the
kidneys in DCE-MRI using convolutional
neural networks (CNN) and transfer learning
~ 50 hr
~ 5 hr
~ 5 sec
Alexander S. Lundervold et al.
Functional Renal Imaging: Where Physiology, Nephrology, Radiology and Physics Meet, Berlin 2017
(hippocampus)
Transfer learning
Manual3D labelling
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
• Promote cross-disciplinary research activities related to state-of-the-art imaging equipment (preclinical and clinical high field MRI, CT and hybrid PET/CT/MR)
• Aim: new methods in quantitative imaging and interactive visualization to predict changes in health and disease across spatial and temporal scales.
• Applications in basic research and preclinical validation
https://mmiv.no
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
The Mohn Medical Imaging and Visualization Centre
https://mmiv.no/machinelearning
Computational medical imaging and machine learning – methods, infrastructure and applications– A collaboration between the Department of Biomedicine, UiB, and the Department of Computing, Mathematics and Physics, HVL
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Radiology: Volume 278: Number 2—February 2016
Radiomics:
Computational imaging, machine learning, biomarkers, visual data science
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Radiology: Volume 278: Number 2—February 2016
Computational imaging, machine learning, biomarkers, visual data science
Radiomics:
… but what if no image reconstruction
necessary
BMED360
?!
(measurements)
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
WE NEED EDUCATION& TRAINING
IN THE NEW FIELDS OF AI
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Ultimately, machine learning in medicine will be a team sport, like medicine itself. But the team will need some new players: clinicians trained in statistics and computer science, who can contribute meaningfully to algorithm development and evaluation. Today’s medical education system is ill prepared to meet these needs.
… Undergraduate premedical requirements are absurdly outdated.
Medical education does little to train doctors in the data science,
statistics, or behavioral science required to develop, evaluate,
and apply algorithms in clinical practice.
Z. Obermeyer & T.H. Lee, Harvard Medical School
link
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
A new elective course at the Faculty of Medicine (Spring 2019, 6 ETCS)
ELMED219
• The computational mindset, machine learning and AI in future medicine - pros et cons
• A guided tour of some mathematical and statistical modelling techniques in biomedical and clinical
applications. Examples and demonstrations will be related to in vivo imaging and integrated quantitative
physiology, imaging-derived biomarkers, omics data, and sensor data.
• Operational principles of selected sensors and measurement devices in biomedical research and clinical
practise - from smartphones to MRI scanners.
• The concepts of "big data", "data analytics", "machine learning", and "deep convolutional neural networks"
with examples from personalized and predictive medicine.
• Throughout the course, the students will use principles and tools from numerical programming, data
analysis, and scientific computing for medical applications. This will provide an introduction to e.g. R, Python,
and Jupyter notebooks, and "the cloud" for data storage and computations.
• The concepts and importance of "open science", "data sharing", and "reproducible research".
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Atom10-12 m
Protein10-9 m
Cell10-6 m
Tissue10-3 m
Organ100 m
Anatomy
Organ system& organism
Physiology
Gene-networks
Pathway models Stochastic models Ordinarydifferential-equations
Continuum models(Partial differential-
equations)
System-models
10-6 smolecular events
(ion channel gating)
10-3 sdiffusion
cell signaling
100 smotility
103 smitosis
106 sprotein
turnover
109 shumanlifetime
Brain
Spinal-cord
Peripher.nerves
TIME:
SPACE:
• -OMICS, IMAGING, PRECISION MEDICINE, DECISION-MAKING, PERSONALIZED MEDICINE and THERAPY
Fra: C. Dollery and R. Kitney, Systems biology: A vision for engineering and medicine, Tech. report, The Academy of Medical Sciences and The Royal Academy of Engineering, London, UK, Feb. 2007.
• INTERDISCIPLINARITY and COMPUTATIONAL APPROACHES to better understand, predict and control the
interplay between molecules, cells, tissue, and organs - in health and disease
AI will be incorporated in ….
D mindset
D skillset
D toolset
- open science
- reproducible research
Challenges:(mechanisms)
Computational medicinebody engineers ?
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
THANKS !
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
ML & Artificial Neural Networks (have been around)
Lundervold, A., Godtliebsen, F. Tissue classification in MR images using contextual and artificial neural network classifiers. In: Proceedings from the NOBIM conference.
15–16 June, 1992: 263–275.
Data Prediction
Learning / training
Training
database
Classifier
synaptic
weights
synaptic
weights
g – nonlinear activation function
PATTERN
RECOGNITION:
ANN
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Mahjoubfar et al. Artificial Intelligence in Label-free Microscopy Biological Cell Classification by Time Stretch. Springer, 2017
http://nautil.us/issue/40/learning/is-artificial-intelligence-permanently-inscrutable
Explainable AI (XAI)
https://arxiv.org/abs/1712.09923
Andreas Holzinger, Chris Biemann, Constantinos S. Pattichis, Douglas B. Kell
What do we need to build explainable AI systems for
the medical domain?
The new European General Data Protection Regulation (GDPR and ISO/IEC 27001) entering into force on May 25th 2018, will make black-box approaches difficult to use in business …
https://www.darpa.mil/program/explainable-artificial-intelligence
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Hohman et al. Visual Analytics in Deep Learning: An Interrogative Survey for the Next Frontiers.https://arxiv.org/abs/1801.06889
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
UNDERSTANDING STROKE AND ALZHEIMER
Genetics
Epidemiology
Medical imagingOrgan-on-a-chip
Metabolomics
ECR Today, March 4, 2018
(the neurovascular unit)
• Imaging and -omics biomarkers• Early identification of persons at risk• Strategies for optimal prevention
http://www.costream.eu
Horizon 2020 #667375
Understanding and treating stroke and Alzheimer
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Prostate cancer & machine learning
• Multi-parametric MRI (mpMRI)
• From digital histopathology to
Computational pathologyAre Losnegård et al.
Computerized Medical Imaging and Graphics 63 (2018) 24–30
Histology MRI
Feature selectionFeature selectionMachine learning
Multimodal image registration
Zhou et al. Large scale digital prostate pathology image analysis combining feature extraction and deep neural network. https://arxiv.org/abs/1705.02678
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Predicting academic achievement from inattention (SNAP)
A.J. Lundervold, T. Bøe, A. Lundervold. Inattention in primary school is not good for your future school achievement - a pattern classification study. PLoS ONE 2017;12(11): e0188310
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Predicting academic achievement from inattention (SNAP)
A.J. Lundervold, T. Bøe, A. Lundervold. Inattention in primary school is not good for your future school achievement - a pattern classification study. PLoS ONE 2017;12(11): e0188310
CART
Tree classification“ SNAP2 = 0 ? ”
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Predicting academic achievement from inattention (SNAP)
A.J. Lundervold, T. Bøe, A. Lundervold. Inattention in primary school is not good for your future school achievement - a pattern classification study. PLoS ONE 2017;12(11): e0188310
CART
Feature importance
Random Forest10000 trees
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Predicting academic achievement from inattention (SNAP)
A.J. Lundervold, T. Bøe, A. Lundervold. Inattention in primary school is not good for your future school achievement - a pattern classification study. PLoS ONE 2017;12(11): e0188310
Feature importanceRandom Forest
10000 trees
Top 3
Prediction
Nettverksmøte om medisinsk inkubator 8/3/2018Arvid Lundervold, UiB / MMIV
Predicting academic achievement from inattention (SNAP)
A.J. Lundervold, T. Bøe, A. Lundervold. Inattention in primary school is not good for your future school achievement - a pattern classification study. PLoS ONE 2017;12(11): e0188310
Feature importanceRandom Forest
10000 trees
Top 3
k-fold cross validation results
Accuracy = fraction of correct classifications
Precision = tp / (tp + fp)
Recall = tp / (tp + fn)