Heart failure affects over 40 million people globally and accounts for 1-2% of the total healthcare spend. Yet, some treatments, such as Cardiac Resynchronization Therapy, have non-responder rates of between 30-50%. At the Deep Learning in Healthcare Summit in London this September, Peter Mountney, Program Manager and Senior Key Expert Scientist at Siemens Healthineers addressed some of the challenges of using deep learning with limited data and when ground truth data is not available and show how we are using imitation learning to help doctors perform challenging tasks in stressful situations. At the event, we caught up with Peter to find out some more about his work in the field.
During my undergraduate studies, I was really interested in computer vision and how it could be used to detect and track object or people. I worked on tracking footballers movements from video footage. It was a lot of fun but after I graduated I wanted to experience industry and went to work for a dotcom for a couple of years. I then returned to academia and started my PhD, I wanted to used what I had learned in computer vision and apply it in medical imaging and surgical robotic vision. It was during my PhD that I really started to see the power of using machine learning in computer vision through techniques like decision trees and randomized forests which were popular at the time.
I am a Key Expert Scientist on AI and Quantum Technologies and I run our special projects team in the UK. I lead programmes of research across the technology reediness spectrum from highly innovative early research to deploying clinical prototypes. My focus is on translating research so that it is practical and usable and can have an impact on patients and doctors. I have a strong interest in intervention imaging.
Everything in medical imaging. Siemens Healthineers has been very strong on machine learning for a long time. We have multiple products on the market that use deep learning. You will find Siemens products across the clinical spectrum and through hospitals from diagnostic imaging to interventional X-ray, molecular testing to Imaging IT. We are using ML to improve imaging workflows, detect/measure/ quantify disease, predict/plan/prescribe and population health management.
Two of the main challenges are data and clinical workflows. There are many data related problems, multiple modalities, incomplete datasets, lack of ground truth data, small datasets. This creates a research environment that is very different from working on public datasets like MNIST or CIFAR.
When deploying deep learning into hospitals it is important to have the clinical workflow at the forefront of your mind. Our products and solutions must be easy to use and clinically adoptable if they are going to impact patients.
We expect AI algorithms to help speed up clinical workflows, prevent diagnostic errors and lead to sustained productivity increases. Artificial intelligence could lead to more precise results and more meaningful prognostic risk scores. More broadly in healthcare we will see AI being used for drug development, population healthcare management and being used to deliver new digital health services that enable patients faster and broader access to clinical services and increased assistive living at home.
I have a keen interest in quantum machine learning or machine learning algorithms that use quantum computers. Quantum computer do not use bit, they have a fundamentally different building block called a qubit that exploit the quantum mechanical properties of superposition and entanglement. This is a nascent research topic but could facilitate the next leap forward in machine learning. One of the potential early applications will be chemistry.
It is great to see what the healthcare community using deep learning for and the creative ways they are using it to positively impact our society. I am always keen to hear about new applications and people’s experiences translating research into clinically usable systems.