Connect with us

Interviews

Tobias Rijken, Co-Founder & CTO at Kheiron Medical Technologies – Interview Series

mm

Updated

 on

Tobias Rijken is the Co-Founder & CTO at Kheiron Medical Technologies, a medical imaging company that uses advanced machine learning technologies to develop and provide intelligent tools for radiologists, radiology departments, imaging centers and hospitals to improve efficiency, consistency and accuracy of radiology reporting.

Kheiron Medical Technologies was founded with the sole focus of helping radiologists detect breast cancer earlier with machine learning software.

What initially attracted you to machine learning?

My informal introduction to machine learning was when I first started learning programming when I was a young teenager. There was this games engine where you could program your own games. I wanted to play this game, but who do I play against? It could be another player, but it could also be a program. How do you build a program that learns how to play a game? It wasn’t a modern AI, it was rule-based AI. I beat it every time, so it was not a good AI.

When I was studying for my master’s degree at University College London, we entered a new era of modern AI. DeepMind had just been acquired, and I was studying machine learning at the place where DeepMind was actually founded. David Silver, one of the leading scientists of DeepMind, was teaching a course on reinforcement learning. As part of the course, I built an agent to learn how to play a game – a simplified form of blackjack. When I finished the AI and played against my own creation, I couldn’t beat it anymore. And that was a real revelation.

My formal introduction to machine learning came in my first year of university. I was studying math and computer science at Amsterdam University College. At the same time, massive new open online courses were established and I could join AI courses from leading figures in the field. The course that made the biggest impact was with Peter Norvig and Sebastian Thrun from Stanford, and I could study AI with 100,000 people at the same time.

Now with Kheiron, what I like so much about machine learning is the application it has on the real world. The opportunity it has to solve real world problems. That’s what really excites me.

Could you discuss the genesis story behind Kheiron Medical?

Peter and I met in 2016 at the UK accelerator Entrepreneur First. We connected straight away – probably because of our similar backgrounds. We both come from medical families – Peter’s mother is a radiologist, I grew up in a medical family, surrounded by doctors, and both of us chose to go down the STEM route.

Our backgrounds gave us a clear view of the struggles and challenges that healthcare workers face and an understanding of how AI can improve their lives as well as the lives of patients. We share a belief that AI together with clinicians will change healthcare for the better.

There’s a big problem to be solved here. There’s a global shortage of radiologists. A lot of people die from cancer – more than 40,000 American women die every year from breast cancer. Part of the reason so many people die from cancer is that doctors don’t have the right information to make the right decisions. AI can help with that.

First and foremost, Kheiron is here to help the radiologist so the radiologist can help the patients. I personally get very motivated by digging into the problem to understand its intricacies. We all work hard to really understand the radiologists’ work, the data they are working with on a daily basis, their workflow. Then we can figure out how we can deploy AI as part of this workflow.

We invest heavily in deeply understanding the users and the data. We have radiologists on our team who help us integrate the user needs into the design and development process. We have a patient involvement initiative that engages directly with women to understand their fears and needs for breast screening.

Many people ask me about the name of our company – Kheiron. Kheiron was a wise, gentle centaur –  a creature that is part man, part horse in Greek mythology – who trained the heroes in medicine. This idea of two halves making up something more powerful than the individual parts is the inspiration for Kheiron – AI and the clinician working together to change what’s possible in the standard of care.

Can you discuss the breast cancer screening solutions that are offered?

We’re working on several breast cancer solutions along the breast cancer screening pathway.

Our first product is called Mia™ – which stands for mammography intelligent assessment. Mia has been optimised to perform the same task as radiologist – namely, to determine whether the woman should be called back for further examination or not.

The level of performance we’re achieving allows us to rethink the workflow. For American women, this is incredibly meaningful. The U.S. has what is called a “single-reader workflow”. This means that each mammogram is read by one radiologist. However, the U.K. and Europe have a “double-reader workflow”, meaning that every mammogram is read by two radiologists. Many American radiologists consider double reading to be the “gold standard” because it is more likely than a single reader to detect early, small breast cancers.

Our clinical studies prove that Mia is a clinically safe and cost-effective option as an independent or concurrent reader. This means that in the U.S., Mia can be used together with the radiologist to achieve that gold standard of a double-reading workflow – with a single human reader. Like I said earlier, this is truly how we see the power of AI and the human coming together to solve big problems.

The first product we’re bringing to the U.S. market is called Mia IQ™, which sits one step earlier in the breast cancer screening pathway. Mia IQ helps the radiography technicians analyze the quality of the image. If we can improve the quality of the image, then the quality of the reading will also go up. And this means that more cancers will be detected sooner, when the outcomes can be better for the woman.

This is very important for MQSA quality assurance audits and for continuous training, which is where we expect to see Mia IQ’s first applications. Mia IQ provides screening programs with reports detailing positioning problems. This will allow the program directors to provide spot training for techs and address any broader image quality problems that may affect their quality assurance audits.

How many images has the neural network been trained on?

For us, the quantity of data is not the point. The good thing about breast cancer screening space is there’s a lot of data available. There’s a lot of historical data because screening programs have been around for quite a while. And because you screen an entire population, there’s a lot of new data being generated.

The challenge is that the data is heavily skewed. 99% of the screening population doesn’t have cancer, luckily, which means there’s a skew toward normal images. There are also population differences, with a lack of diversity in the data.

So for Kheiron, the challenge isn’t about getting the quantity of data. Instead, where a large part of our success comes from is understanding which data to collect to get the best results, and to be very selective and targeted in where and how we get our data. Our collaboration with Emory University is a great example of how we’re doing this.

Are false positives currently an issue?

False positives are only part of the issue. The science of screening relies on tradeoffs between false positives and false negatives. For instance, we can achieve a very low false positive rate, but we will miss some cancers. On the other hand, we can have very few missed cancers if we have a high false positive rate. It’s important to understand those trade-offs and to balance your screening service based on priorities and resources.

Cultural and policy differences also exist between countries, which influence the way these tradeoffs are made. In some countries, missing a cancer is considered very problematic. In other countries, it’s seen to be more important to minimize unnecessary recalls because of the healthcare cost associated with each recall.

Ultimately, it’s about understanding how this tradeoff impacts patients, in terms of the number of unnecessary recalls and further tests experienced by the ‘false positive’ group versus the ‘false negative’ group whose cancers go undetected. It’s a difficult balancing act, but AI promises to raise the bar for both groups.

Kheiron Medical Technologies and Emory University recently announced a collaboration to evaluate data from prior mammograms on over 50,000 African American women who have been screened at Emory Healthcare. Could you share more details about this project?

Our partnership with Emory University broadens the diversity of data we are using to ensure Mia, our AI breast cancer screening solution, works to the same standard for any woman, regardless of ethnicity.

The first project in the collaboration will support the validation of Mia’s performance against Emory’s diverse screening population. This is a multi-site study spanning four screening sites and hospitals. A cohort of more than 200,000 screening mammograms will be included in this project – approximately 45% of women in that cohort are of African American descent.

We announced this collaboration in late November and have made great progress since then, even considering the challenges of operating under a global pandemic and remote working conditions. I think that is a huge credit to the team at Emory, who are also clinicians providing excellent care to patients in the community.

Right now our collaboration is investing the upfront energy to establish a strong foundation for our future work. We’re at a stage where we’re working closely with the Emory team to understand clinical workflows. This serves as the necessary foundation for deep analysis and interrogation of potential differences between subpopulations, like ethnic groups.

This work is not easy or fast. Not many things in healthcare are. However, it is very rewarding, and we’re lucky to benefit from a set of global partners who are extremely collaborative and aligned with us on our shared mission: to help any woman anywhere have a fighting chance against breast cancer.

How large of an issue is the under-representation of non-whites for image data for cancer diagnosis?

Historically, the diversity of patient populations has not been mirrored in medical research to the extent that it should be. Those of us developing AI have a responsibility to include ethnic diversity in our studies.

Under-represented populations are an issue for breast cancer screening. Outcomes for African American women are an example of this. African American women tend to have higher mortality rates (39% higher than non-Hispanic whites), and this may be attributed to access to health care, delays in diagnosis, and delays in treatment initiation, among other factors.

Generalizability of AI across various populations is a key step toward addressing racial disparities in breast screening. I’m excited that our Kheiron team is committed to building a solution that performs equally well on mammograms from racially diverse populations – and this is a big motivating factor for our work. Our collaboration with Emory and UCSF will help us achieve this vision of generalizability – making sure that Mia gives every woman, everywhere a better fighting chance against breast cancer.

Thank you for the great interview and for working on such important life saving technologies. Readers who wish to learn more should visit Kheiron Medical Technologies.

Antoine Tardif is a Futurist who is passionate about the future of AI and robotics. He is the CEO of BlockVentures.com, and has invested in over 50 AI & blockchain projects. He is the Co-Founder of Securities.io a news website focusing on digital assets, digital securities and investing. He is a founding partner of unite.AI & a member of the Forbes Technology Council.