The AI Radiologist

Sure. Like all doctors, I started out at medical school. I went to Imperial College in London, which was a five-year degree with one year intercalated BSc. I chose to do that in medical imaging and informatics, which actually laid some of the foundations for what I ended up doing later in life — but I had no idea at the time.

I qualified in 2000 and went on to become a junior doctor. I did my F1 and F2 years down on the south coast in the KSS deanery, then applied for run-through training in radiology and got a post at Brighton. That was quite good, because it was a five-year place at one hospital, so unlike many other specialties you didn't have to move house every six months or a year. I managed to buy a small house on the seafront in Brighton and did my radiology training down there.

While I was a registrar, the field of 3D printing started to become a thing — the first early home 3D printers were coming out. So I set up a very small 3D-printing enterprise where I'd take CT scans, very carefully tease out the underlying anatomical structures, and create 3D models of various parts of the body. My office to this day is littered with people's skulls and bones that I've 3D-printed.

So I learnt a little bit about setting up a company through that. After I finished my radiology training, I applied to do a research degree at the Institute of Cancer Research in Surrey, just south of London. I did two years there studying morphological and functional imaging of prostate cancer — a lot of advanced imaging techniques, data analysis and statistics. One of the projects I was working on was the start of the new deep learning paradigm, using segmentation networks called U-Nets to pull out different areas of prostate tissue. So I got into deep learning completely by accident, through my research degree.

When I finished, I decided I didn't really want to be a consultant radiologist. I was always having side hustles and other things distracting me, so I took a completely different step and applied for a job outside the NHS — a very small health-tech company no one had heard of, called Babylon Health. I was one of the first few team members, and we all know where that ended up. That was my baptism of fire into health tech.

I was there for a year, working with the data science team and the clinical leads, but I also built out the regulatory department, because we quickly realised what we were building was a medical device that needed validation and regulatory oversight. We worked very closely with a government agency called the MHRA to get it regulated, and set up a regulatory team.

After a year, I thought, well, maybe I should go back to consultancy. So I took a six-month locum consultant post at Guy's and Tommy's — but after three months I realised I did not want to sit in a dark room reporting scans anymore. I'd been speaking to some friends about setting up another startup, this time focusing on breast cancer screening using deep learning. So I joined that and worked there for two years, and we got CE marking as a class 2a device for breast cancer screening software, which is now being trialled in the East Midlands. That was great — two years at the forefront of deep learning and health technology.

Over that time I'd been blogging on Medium and tweeting. I never set out to pick up a following, but some of my stuff got noticed. I got asked to co-chair the Topol Review — a year-long review looking at the future of health technology for the NHS, chaired by Professor Eric Topol, a prominent American professor invited by Jeremy Hunt to come over and conduct it. He picked me to co-chair the AI stream. That was really good fun — I got to meet everybody who's anybody in AI and tech across the whole UK and US.

I continued academic publishing on the side, and became an associate editor at the Nature Digital Medicine journal — I did that for a couple of years, and now I've been promoted onto the board. And since leaving the last startup, I've set up my own consultancy. Taking in all the experience I've got as a radiologist, an academic and in industry, I now advise startups on the proper routes for clinical validation, regulation, IP and patenting, go-to-market strategy, health economics and so on. And I continue to tweet controversial things about this space. So that's me in a nutshell. There's no overarching theme — it was a bit of a long, winding road. If you'd asked me when I was a 15-year-old applying for medicine where I'd end up, I'd never have guessed it would be here.

I have no idea how directly or indirectly it helped — I think some of my academic papers helped too. But when there's a new field coming into being, those writing about it the most will obviously be affiliated with that sector. I was writing on Medium predominantly before I was tweeting, and the year that really took off was 2017, about two years after the invention of deep learning. I was blogging about it a lot as I transitioned from Babylon into the second startup, spending a lot of my time thinking about it.

I think it was probably through my blogs. Eric is an avid reader of everything out there, and he picked up that I was doing some serious deep thinking on this stuff, ahead of what was coming out in the academic publications. Academic publications take time — there's a six-to-twelve-month lag before anything gets published, so it's much quicker to blog about it. Then I guess he just verified who I was on Twitter. I've asked him, "Why did you pick me?" and he said, "You came across as someone thinking deeply about this, who understands both the industrial and the academic insights." It was great to make a friend in Eric through that, and we're still friends — this very moment I have an essay open on my desktop that we're writing for The Lancet together.

Because to say that data science or deep learning or any form of technology is going to entirely replace a profession is a really naive thing to say.

And it was said by Geoffrey Hinton, who is an incredibly brainy person — literally known as the godfather of deep learning. But it's incredibly naive, because unless you've been a radiologist, which of course I have, you don't understand what radiologists do and the nuances of their work. I would not presuppose to say that about any other profession, that they could be replaced, without knowing exactly what it is they do.

There's a deep misunderstanding here. The majority of people think radiologists just sit in a dark room all day reporting scans. If that were the case there'd be no radiologists, because it would be immensely boring. There's a lot more to it. We have direct contact with patients, we do ultrasound lists, we're all trained in inserting biopsies and drains and doing interventional procedures, we're integral to multidisciplinary team meetings and oncology cancer boards, and there's teaching and training and all sorts of ancillary activities. It is not just about the reporting. In fact, I know of no radiologist who spends their entire week just reporting scans.

The whole essence of being a radiologist is the conduct and delivery of the radiological service, not just the reporting, which is a tiny aspect. In that blog, I tried to break down the pathway from a patient's point of view. Once you divide it into eight or nine parts, you'll see that only one of those parts is looking at the images — and that's the bit deep learning can potentially be very good at. But that's like saying we're going to replace drivers because we can change from first to second gear. There's so much more. To automate and replace an entire profession, you need to be able to do everything that profession can do. That's why I remain bullish on radiology as a career.

There's a cheeky version I like to say to get people excited — there's no reason it couldn't look like Tom Cruise in Minority Report, dealing with some kind of virtual dashboard, VR goggles on, swiping images left and right and pulling out data insights. That's the sexy, idealised version.

In reality, you'll still be dealing with images, but you'll be dealing with a lot of the data around the images. You'll be like the conductor of an orchestra of algorithms. You'll have the deep training and knowledge to understand what each of these algorithms is doing, what they're looking for, what their limitations are, what their accuracy is, and you'll be able to assimilate that and make the conclusions, the diagnosis and the recommendations to your referring physician colleagues. So it's about the orchestration of more data.

It's going to turn from an anatomical, biological and physics speciality — we know about radiation physics and MR physics — and lean more on data science and statistics. The whole field of radiomics and AI and deep learning is going to change the profession almost to make it unrecognisable. And it's all driven by increasing volumes of imaging: year on year, CT, MRI, X-ray and ultrasound usage just increases across the NHS, and globally, with no significant increase in workforce. It's about getting more accurate, faster results for patients.

I don't think we'll get to the point of that Star Trek tricorder, where someone untrained can wave something magic over a patient to get an answer — at least not in the foreseeable future. So again, that's why I remain bullish on radiology as a career.

Yes. In the Topol Review I pushed hard for a recommendation — which was included — that everyone should have basic statistical training in their undergraduate degree, and then go deeper in postgraduate training depending on the statistics useful in their field. For deep learning algorithms, which generally rely on classification or segmentation, radiologists need an understanding of the statistics involved — confusion matrices, area under the curve, sensitivity, specificity, trade-offs and so on. I didn't learn any of these things until I went to do my research degree at the Institute of Cancer Research, where in the first few months we did a statistical course on all of it. These aren't part of medical undergraduate training.

You learn chi-squared tests and t-tests, but I can tell you, no one does chi-squared or t-tests in the academic literature, especially not around this technology. So people need to up their game on their underlying statistical knowledge. I'd never claim to be a statistics expert — I'm pretty good at the things I've mentioned, but outside of that I've got no clue; there's a whole speciality involved. But it's absolutely pertinent for radiologists, and indeed pathologists, who are undergoing the same dynamic shift, to understand these basic concepts.

As for coding, I don't think anyone needs to know how to code. The future is no-code. To be honest, I've got where I am being the worst Python coder in the world. Teaching doctors to code is like trying to teach coders medicine — it's never going to be the same without that dedicated training. So have a go if you want to feel more comfortable around the jargon and the language, but don't for one minute think you're going to be building software that's production-ready and valuable to anybody.

Oh, for sure. Absolutely. There are several who've started their own companies and done so. But like I said, to create production-ready code that's robust, safe and gets through regulatory approval, you have to be pretty good. So by all means go for it — but would you trust the doctor's code over someone who studied it since they were a teenager and did it for a living for a decade?

The major mistake I see is people thinking they can build an entire product and then, at the last minute, go, "Right, let's get it regulated." It just doesn't work like that. You need to be doing regulation from day one, because you need to set out exactly what you're going to build and plan in all your quality management, robustness and assurances as you go along. You cannot turn up to the regulators with a finished product and say, "Can you assess this, please?" — because they'll have nothing to assess. They assess and audit your processes, not your product. If you haven't documented your processes, you cannot get regulatory approval.

The second biggest mistake is people not budgeting for it. If companies aren't aware of the burden of regulation, they spend a lot on the technical side — hiring data scientists and coders, an expensive chief medical officer to lead clinical validation, a business and marketing team to generate sales. What they won't do is hire a regulatory team, which to me seems absolutely insane, because you simply cannot sell your product in a healthcare system without regulatory approval. If your greatest barrier to market is regulatory approval, why on earth are you not putting a significant portion of your budget on getting that part right?

That's an incredibly good question. It depends on what you want to get out of it, but my advice would be to do two things. One: use your institutional access from your university to go to the latest journals. Most subspecialty journals now have an AI branch or special issue. Have a look at the papers — and not just the academic publications, but the opinion pieces and editorials, because there's been a lot of chatter about guidelines for reporting AI studies, best practices for ethics in AI, and so on.

The second thing is to go online and look up webinars and conference videos — there are millions out there. You can find the video series from pretty much any conference over the past few years and watch keynote talks from hundreds of people.

And if you want to get into the coding side, go sign up for MOOCs — massive open online courses. You can do a two-week course, a six-week course, a year-long course, whatever you want; some are free, some have a slight fee. If you're bored on a Sunday night, why not start practising and playing around? I've certainly done it — a few years ago I got some chest X-rays, did a little MOOC, and learned how to build an X-ray classifier. I built it in a few days. It's nowhere near production-ready or usable in a hospital, but it shows you can play around with the data. And the first time you do that, you'll quickly realise that deep learning is not going to replace anybody.

I'm very disciplined. I don't like chaos — I like organisation and structure. That's just my personality; I'm very organised, not to an OCD level, but planning ahead and organising things is very important.

Part of what's helped me get to where I am is that I compartmentalise my day, and I'm very strict and rigorous about it — into social time, family time and working time. As a student, family time is less on your radar than social time, but as you get older, get married and have kids, family time takes over. It's important to compartmentalise those, and to be realistic with your timeframes. Set yourself realistic goals, chop them into even smaller realistic goals, write them down and put them in your diary. My diary is absolutely chock full — even little half-hour blocks where I'm literally writing "have a think about X" or "make some notes." That way you get through your day and you always know where you are. I'd feel slightly lost now without that structure.

Make time for exercise, make time for socialising, and once you've got a wife and kids, make time for them above everything. Stop your work at 6pm, turn off your emails, go and live with them — because that's, after all, why you work. So: compartmentalise, and structure. That's the answer to your question.

Oh, blimey. A few words... I think I'd say something like: the grass isn't greener, but you might like the taste of the other fields. A lot of medical students get into medicine looking at a single train track, all the way up to consultancy. But the most interesting people I know — and I presume most of the people you're reaching out to on your podcast — have not stayed on that single train track. I can see you're nodding, so I think you'd agree. And the other billboard I'd put up is: follow what you find interesting.

Everything I've done, I've done because I found it interesting — that's the only way you can survive. I wake up now and I don't feel like I'm working a job, because I wake up and go, "Brilliant, I'm writing a paper on X and I'm really excited to get it published," or, "I'm helping a company validate this algorithm, which I think is really cool." Follow what you find interesting. And it may not be anything to do with AI or medical technology — it could be your own bag: fitness, wellbeing, whatever. I've got friends who've started nutritional YouTube channels, written books, or gone into stand-up comedy. Do what you find interesting. That's the way to succeed in life. Success is not about titles, positions or money — it's about, are you happy, and have you found the balance you wanted?

Just keep on going. All you medical students out there — I know it's a long route. If I could look back at my medical school years and name one regret, I'd say I probably didn't party enough. So go out there and party on.