“We are synergistic – what we’ve accomplished would not be possible without each other’s support”
While Mark may have been describing his relationship with co-founder Alex, we couldn’t help but think about the synergistic role their company, 16 Bit, has played in infusing machine learning with medical diagnosis. It goes without saying, when building such a complex AI system within healthcare, it takes an incredibly harmonious approach that is not for the fainthearted; yet Alex & Mark have managed to make it look all too easy. As winners of the 2017 RSNA Challenge, we wanted to learn why medical imaging has become a cornerstone in medicine and how 16 Bit enhances the field of radiology moving forward.
It was our esteemed pleasure to sit down with co-founders Alex Bilbily & Mark Cicero to learn about 16 Bit’s mission & the role of artificial intelligence in healthcare. Let’s jump right in!
16 Bit is a truly unique name! What was the inspiration behind that choice?
Alex: Our name is really a tribute to our main competitive advantage: possessing both the medical domain knowledge and the technical computer science knowledge. So traditional images you’d search for on google search would normally be saved in what’s called 8-bit rgb. 8 bits of memory are used to store each color channel of the pixel values. But medical images are fundamentally different – they are stored in 16-bit grayscale format, which means they’re much richer in terms of the amount of detail that is inherent in each pixel, and that’s important from our perspective – we as radiologists need special equipment to look at medical images. Therefore, our name is a tribute to how we understand the various nuances to consider when applying AI to our field of medical imaging.
At what stage in your career did you feel there was a need for inserting AI into the domain?
Mark: We started applying modern machine learning algorithms to medical images in late 2014. At the time we were starting to look at chest X-Rays to see if this technology could be applied to medical imaging – we realized that algorithms could learn clinically relevant features with off-the-shelf technology. This gave us a strong sign that AI was poised to make a big difference in medical imaging and enhance the work that we do as radiologists. We believed if we understood the technology deeper and had a better understanding, we’d be able to customly design solutions to better suit our needs. That’s when we seriously started considering building a company to further advance the field and build dedicated AIsystems that were guided by medical expertise. But I know Alex had a bit more of an ‘ah-ha’ moment he’d probably like to share.
Alex: Yeah, I really had a very distinct moment where I thought ‘Oh my God, we really have to do this!’: I was in third year of medical school – I’m a night owl – I was browsing the internet, reading a few articles in the middle of the night, and I came across Alex Krizhevsky’s paper, which in retrospect is a landmark paper in machine learning, and it really was the first time where modern AI has proved to be dominant in image classification. In fact, its so dominant, it’s virtually displaced all proceeding technology when it comes to image classification and image analysis tasks. When I read this paper at 3am it became immediately obvious that this technology is going to dramatically change medicine and medical imaging. I was always considering pursuing a career in Radiology but my choice was solidified when I discovered the field of AI and realized the synergy that could exist. When I interviewed for my radiology residency – I explained how I wanted to combine my computer science knowledge within healthcare and here we are today, trying to do just that.
Given the nature of your startup, it’s safe to say both of you have a strong interest in computer science & machine learning, not just medicine.
M: Definitely. Alex and I are sort of a similar breed. I did my engineering before medicine and I knew full well that I wanted to bring my technical expertise into this field. As co-founders, we both have a keen interest in technology and computers; pair that with our natural inclination towards medicine and it means we’ve always been keen on finding that balance: searching for a way to use both to better the field of medicine.
Balance is definitely important! But there must have been challenges that came with building 16 Bit.
M: Yes. Certainly the struggle within this industry (being an industry that’s still in its infancy) means you’re creating a new category, so undoubtedly it’s going to come with its own unique hurdles. The biggest barrier for us has been getting access to high quality data, while respecting patient identities, going through the right channels, REB submissions and making hospital agreements – all of that takes a tremendous amount of time, and we’re navigating that as best as we can. In addition to that, it’s managing our time as we’re still going through our training and fulfilling our clinical duties.
A: I think one of the other challenges is competing with the culture. Nowadays, the culture around AI, especially in our field of medical imaging, is really fuelled by fear. Everybody is scared they’re going to lose their job, because they’re concerned it’s going to be automated. But people don’t really understand the technology beyond news headlines. So you might read: ‘this AI system diagnoses pneumonia better than radiologists’ but these are extremely misleading statements. The technology solves a very specific problem in a very specific situation, and humans for the foreseeable future will remain a key component because they will direct how these intelligent tools will be used and in what context. Radiologists need to evolve with AI – they need to embrace it and see it as something which will revolutionize the field. It’s going to make us more valuable as a speciality because, not only will it improve our accuracy and efficiency, but it will enable us to be more specific, better integrate non-imaging data, and really act as a new form of research to further expand our fields reach.
At what stage does your system come into effect during a diagnosis? How do radiologists play a part in enhancing this work?
M: Radiologists need to continue to provide value: often, that value will not just be simply in the form of image interpretation. It’s in the way radiologists are consultants – how they help guide the critical decision-making process and diagnostic work up of a patient, and how they’re perceived by non-radiologists colleagues. If radiologists narrow their role down to image interpretation, that’s when they become liable to being displaced. We have to go above and beyond. We should continue to do those intangible things. There are so many nuances to our job – every day we’re faced with cases that are so subtle and complex, so there’s no way a machine can replace every facet of our job; this is coming from two trainees near the end of their training who are more excited than ever to begin their career as radiologists – we know our jobs will only increase in demand and we’ll play a pivotal role in patient care.
So you believe radiologists will also play a large part as communicators of how AI will enhance the field?
M: One of our new mandates should be how to guide AI into the field and how to leverage it to become better diagnosticians – Alex and I are both on the AI working group of the CAR (Canadian Association of Radiology), and in order to guide AI properly for it to have the greatest impact, we have to make sure it’s used in the right way – radiologists need to be involved, and that’s why Alex and I built 16 Bit. The technology needs to be used properly with the right specialists guiding the process.
In its early stages, what were some of the encouraging results you experienced from 16 Bit? Indicators that pushed you to continue your work!
M: In terms of bone age, we won the RSNA competition. Within three months of winning the competition, this algorithm is being used for research purposes in 10 of the top US hospitals – so for us, it was pretty astonishing how fast we were able to build something for a competition, and then have it being used in a hospital for research purposes.
A: In mammography, we have 2 studies approved; one at University Health Network and one at Sunnybrook. In terms of gaining preliminary results for our system, we went to a local radiology community clinic and they had 8,000 mammograms which we used. Even though it was a smaller set of data, it was enough to show that our system could learn to classify breast density (this is always the first step to diagnose breast cancer). As physicians, we comment on the breast density, because we know it’s important. So you’ll see there are entire companies built around measuring breast density, but for us, with a week of time and small data set 8,000 – we built a system which is 93% accurate, and we know that’s an important first step. It means our system is capable of learning from real mammograms and learning something of critical value – this was one of the early wins we had.
16 Bit’s Bone Age Demo
The three areas 16 Bit is currently specialising in are: Breast cancer screening, CT scanning, & paediatric bone age – how much do you envisage your AI system enhancing the interpretation process of those cases?
M: At this time in breast cancer screening, we see AI playing the most valuable role in the screening of normal cases. We feel it will save a tremendous amount of time which can be used to alter the whole paradigm of breast cancer screening; once you drive the cost down of interpretation, many new things become immediately possible. Existing CAD (Computer Aided Detection) software in mammography functions as a decision support tool or a second reader, where physicians use it as an internal check to see if they’ve missed anything, we want to change that paradigm! So for us, the computer will screen every mammogram before the radiologists – and it will diagnose the ones that it thinks are normal with a high degree of confidence. Over time, as we build evidence, we hope that radiologists will only need to interpret cases which are flagged as potentially abnormal.
A: On one level, we’re trying to semi-automate something that we can already do. The second level to this system (which I think is more interesting), is enabling us to do something which we can’t do. What this means is, we’re trying to make a software system that matches the level of performance of radiologists – so we’re really altering the economics of the situation, we’re trying to drive down the cost. This second level is where we see that the algorithm might be able to discover patterns that we didn’t know about as radiologists that are specific for cancer or non cancer, and this leads to a system that adds value – It allows the radiologist to become more specific with their interpretation and diagnosis, as Mark said. That’s why, both of us ultimately believe that AI will make us more valuable as a specialty with respect to patient care.
Apart from those three domains, where else do you believe AI could report?
M: It can be used for performing new research and discovering new features as Alex has mentioned. Another application we are investigating is AI’s role in point-of-care ultrasound. One problem is that non-radiologists are not formally trained as well as their radiology counter-parts in how to interpret these images, so there’s tremendous potential for providing them with support in the form of algorithms that help assist in image acquisition and classification. We picked these domains as they were the lowest hanging fruit; based on the resources we have and the data we have access to. But in order to make 16 Bit a lasting company, we hope to get into areas outside of these domains as well.
A: If I could add onto that – Catching any type of cancer early is always better for patients because it’s more often curable, and better in terms of cost (it saves us having to pay for the treatment later, when things are more expensive and more critical to deal with). In the U.S., they can afford to have a lung cancer screening program for people that smoke – this catches lung cancer early but it’s fairly expensive; that’s why we don’t have it in Canada. However, these AI systems will bring down the cost of screening image interpretation which will allow countries such as Canada to offer more cancer screening programs such as that for lung cancer offered in the U.S. , “Narrow AI” is very powerful, and for screening it’s perfect because it deals with a very specific problem in a defined population. We predict that AI technology will enable widespread screening which I think is going to be amazing for the population as a whole.
Your work is incredibly specific and specialised! In the early stages of building 16 Bit, how did you know you were collaborating with the right co-founder?
M: A lot of people say that starting a company is a commitment similar to a marriage . As physicians, we lead very demanding lives and with Alex, those demands and the level of commitment is understood by default. Both of us are technically oriented, which I believe is our key value proposition.
A: In our field, the pedigree of people that endure 20 years of education are usually known as perfectionists. They’re the ones who end up doing the bulk of the work in a school project because they want it completed to a certain standard. We both have a level of performance we expect, and with Mark, it’s gotten to a point where I trust him wholly. If I don’t have time to look something over I know I can trust Mark’s opinion and decision – we’re sort of an extension of each other, which means we trust each other to make those executive decisions.
M: Just to echo that: None of this would be possible without Alex. We are synergistic in that way – what we’ve accomplished would not be possible without each other’s support – we definitely need each other to lean on, and for that, we’re really thankful.
Before we let you go: If you could pick, what would be one key piece of advice you would give to someone building a startup?
M: Something I have been working on is to – learn how to say ‘No’ – if you’re successful, you will find yourself saying no more than you’re saying yes, because the demand for your time, service and ability will scale astronomically. This is something I’ve learned from Alex!
A: For me, before we started 16 Bit, I used to read or hear that having a co-founder is almost a necessity, but I never really understood it. It’s not just about just having someone in the trenches with you – forget the technical stuff, forget the time, and increased bandwidth that you have, or the networks that you both bring to the table: the emotional support alone is enough to make a co-founder a necessity. A startup is a roller coaster and building one seems to almost always be harder than expected. Both Mark and I have our times of frustration, and there are lots of decisions you’ll make where you may never know if it was the right choice, but our support for one another really guides us through that – A co-founder is basically critical.
To learn more about Mark, Alex, and 16 Bit’s mission, head on over to www.16bit.ai
Maryam Zaidi is a user experience/ digital strategist currently working in healthcare. As a graduate from the University of Toronto in the Master of Information Program, her interests lie at the intersection of human-computer interaction and empathetic design. In her spare time, she loves to read, run, and occasionally code! You can find her on Twitter as @MaryZai