When the full extent and global impact of COVID-19 became clear in early 2020, researchers all over the world had the same natural reaction: “How can we help?”
Dr Mike Roberts (Senior Research Associate of Applied Mathematics) and Professor Carola-Bibiane Schönlieb (Professor of Applied Mathematics) of the University of Cambridge, were no different. They wanted to step up and do their part.
Their desire to help was driven by their expertise in machine learning and data science. They wondered how they could use automated AI technologies, imaging, and clinical data to predict patient outcomes. Or, to put it another way, how can AI technology help fight one of the most significant pandemics in recent memory?
Securing Funding from a Trusted Partner
Around the same time, Intel launched a $50 million fund called the Pandemic Response Technology Initiative (PRTI). Set up to support companies and organizations who were looking to use technology to combat the coronavirus through accelerating access to technology at the point of patient care, speeding scientific research, and ensuring access to online learning for students. Its long-term view is to recognize and fast track the technological and scientific discoveries needed to prepare us for future health crises.
When the newly set-up team became aware of the fund, they approached Intel with an exciting project proposal that would bring their vision to life. Before long, the proposal was approved, and funding was secured. The AIX-COVNET research project was born.
Awarding the PRTI funding to the team was an easy decision for Intel. Intel and the University of Cambridge have enjoyed a long and rewarding relationship over many years, having worked together in many areas, including high-performance computing. In particular, they collaborated on the opening of the Cambridge Zettascale Lab, dedicated to making exascale computing practical. The AIX-COVNET project marks another important milestone in their relationship.
Building the AIX-COVNET Team
With the PRTI funding secured, Dr. Roberts and Professor Schönlieb set to work building their wider team to give them the firepower they needed. They knew they would need extra support to gain a deeper understanding of the disease itself. They also had to define existing best practices for how AI is already used in healthcare. Plus, they felt the urgency to prioritize software development, as this would be the driver for their work.
By reaching out through their vast network of professional contacts, they created a team made up of people from various disciplines. They included clinicians, medical researchers, mathematicians, image scientists, computer scientists, and engineers. They found their request extended beyond Cambridge, and beyond the UK. Indeed, the response was so overwhelming that they brought on experts from places such as Italy, Austria, The Gambia, Ghana, South Africa, and China, all of them working together to develop an AI toolkit that might one day help manage and treat COVID-19 patients. Everything was in place to begin their initial research.
“It was beautiful to see a multidisciplinary team come together, made up of clinicians, medical researchers, mathematicians, computer scientists, and engineers.” —Professor Carola-Bibiane Schönlieb, Professor of Applied Mathematics
AIX-COVNET: How It Works
As mentioned, the team’s work, and the software involved, are currently in their research phases. However, it’s hoped that it may be ready for real-life patient diagnosis by late 2023.1 But we do know enough to take a deeper dive into how the AIX-COVNET program might work.
A server, set up in the hospital, is connected to the imaging system. When a patient with suspected COVID-19 comes into the hospital, they would receive a CT scan of their chest. 3D images are formulated, which will be sent to the algorithm devised by AIX-COVNET.
The images are overlaid by the AI-predicted disease region. The algorithm is trained using a database of 20,000 chest images, taken from numerous hospitals across England and Wales. Based on initial measurements taken from the patient, and by pinpointing the different patterns when compared to the database, the clinician would receive quantification of the disease, helping them to predict how it might develop.
By making an informed decision about how severe or progressive the disease is, choices can be made faster on the best care pathway for the patient. Should they be sent to ICU or to the ward? Should particular medications be prescribed? Or does the patient need non-invasive ventilation?
The team is working on a way of filling in any gaps in the datasets by using a form of informed probability, called imputation. But this was just the start of their database challenges.
Although a huge database of past chest scans is vital for predicting future patient outcomes, the team also discovered that often some x-ray images were found to be scratched, others rotated inconsistently, or even inverted, and therefore unable to be used for training. The team is looking at a way of putting the images into different buckets to make them viable. And having these images in usable forms can save so much AI training time in the event of another pandemic.
The team is not just focused on looking at the lungs, but cardiac features too. They know that if the disease was putting the lungs under strain, then the heart might also be at risk. So they needed to factor in heart health alongside lung health to create a clearer picture of patient wellbeing.
During their research work, the team also identified another area of interest for them to investigate. Dr. Mike Roberts and a team of students who are part of the BloodCounts! Consortium are looking into whether COVID-19 can be detected from a routine and inexpensive blood test. The aim is to use comprehensive data from blood tests, the most common clinical test in the world, to predict outcomes for patients.
How Intel Has Helped
As a technology provider, Intel has been offering support for hardware from the beginning. Using the Intel® Movidius™ Neural Compute Stick, for example, has helped the team speed-up the whole imaging process considerably. A type of accelerator stick, the Intel Movidius Neural Compute Stick allows clinicians to run the necessary algorithms in the hospital without the need for a costly GPU (a highly specialized Graphic Processing Unit, designed to render graphic imagery quickly). Not only is the hardware low-cost, but it’s mobile and scalable enough to work across the UK and the world. Intel also provided the 3rd Gen Intel® Xeon® Scalable Processor-based server that was invaluable in the early testing of the algorithm.
“The conversations with Intel have been invaluable for understanding the technology that is available, and how we integrate it. As well as the advice for how we use different technologies around the world.” —Dr. Mike Roberts, Senior Research Associate of Applied Mathematics
What the Team Has Found So Far
At the moment, the earliest prototype algorithms designed to identify, segment, and quantify diseases are in the clinic, awaiting regulatory approval. One of the next stages for the team is to integrate the next level of disease prediction.
Yet the work has already heralded various learnings that could inform subsequent AI healthcare tools and processes going forward. It has highlighted the need for data access, for example. AI is fueled by high quality training data, allowing it to find patterns and generally determine the quality of the algorithm. The better the data, the better the outcomes. Access to reliable, consistent data is crucial.
It has also flagged the importance of creating a new robust software framework to accommodate the reproducibility and reusability of the algorithms. This way they can be developed further to become a scalable, robust, and useful tool for NHS clinicians to use in a real environment.
Perhaps most importantly, the work carried out by the team so far has underlined the importance for researchers, clinicians, and mathematicians to engage in multidisciplinary dialogue. The act of learning from other experts, and sharing the challenges across the various disciplines, will be essential as the trend of AI in healthcare continues to grow.
“The AIX-COVNET project will become a role model for how AI methodology should be developed for healthcare. AI technology can’t be developed in isolation, so bringing together a multi-disciplined team who work closely with other is crucial.” —Professor Carola-Bibiane Schönlieb (Professor of Applied Mathematics)
Looking to the Future
What makes the work done so far by AIX-COVNET particularly exciting is that its learnings and successes will stretch far beyond COVID-19.
Specifically, the opportunity to create a tried-and-tested blueprint for future health emergencies, including new potential global pandemics. It’s a blueprint the team believes can also be scaled across the whole of the NHS in the UK. And because the common patterns of COVID-19 look similar to other existing respiratory epidemics, it’s hoped the segmentation tools can be used to tackle, say, influenza. It could establish improvements in best practice for development and communication, as well as informing the development of other AI tools on how to function in a real-life environment.
“It’s exciting to see how this can be rolled out, making AI significantly more accessible and gaining outcomes and answers faster and easier. That’s when you get to a point where you change the world with technology.” —Phillippa Chick, Global Account Director, Health & Life Sciences at Intel
With work on the AIX-COVNET program continuing, we are looking forward to more exciting and world-changing collaborations with the University of Cambridge in the future.