Where to invest in Radiology AI

As we reach the peak of the hype curve surrounding AI and its impact on the field of radiology, it’s more important than ever for savvy investors to be aware of the perils and pitfalls of this advancing space. For all the dreams and start-up bluster, there’s a mountain of hard work, regulatory bureaucracy, scientific validation and institutional bias to overcome. None of this is insurmountable, but to stand the best chance of succeeding we need to take a step back and survey our surroundings before deciding on the best investment strategy.

In this article I will discuss the promise, hope, and harsh realities of applying deep learning to radiology, and the exciting aspects to look for when choosing where to place your bets.

The Hype

We should stop training radiologists right now – Geoffrey Hinton, Google.

I’m a huge evangelist of AI in radiology, but also a pragmatist and a realist. I do not subscribe to the ideology that radiologists will be replaced in a mere 5 years time, but I do believe that radiologists will be incredibly well served and augmented by AI within the next decade. My respect for Geoffrey Hinton is immense; he is quite literally the godfather of image perception, after all. However, his famous quote over-eggs the pudding quite considerably (and I’m sure that if pressed he would clarify and cushion this statement!). For starters, his implication is that the only thing a radiologist does is interpret images — a huge misrepresentation of an entire profession. He also assumes hospitals will accept new technology unquestioningly. I only need point to the abysmal uptake of CADx software over the past decade to demonstrate how difficult it is to infuse new tech into the clinical frontline…

AI promises huge amounts of future reward, but total replacement of radiologists is not happening in the foreseeable future. Yes, radiology is at a turning point with demand exceeding capacity by an order of magnitude, but that’s the very reason the hype is so pronounced in this sector – the problem is so absolutely huge and the solution so tantalisingly close! Of course we need AI! Of course it can help! Of course we desperately need it! But none of these statements actually mean that it is going to happen any time soon. It just helps inflate the hype bubble.

Some hype is good; it helps drive research, bring investment, raise awareness, creates competition. But hype can also be detrimental; it can lead to over-promising, lack of investment in improving current practice, and rushed unscientific approaches to problems.

Navigating the trough

As we start to drop over the hype apex into the trough of disillusionment, we will start to see excitement wear off rapidly as reality sets in. This is where smart investors can make rich pickings.

There will be two main types of investment opportunities here: Companies looking for seed funding in this climate will have to prove themselves significantly more than their predecessors, and the strong that have survived so far will be looking for Series A or B funding to go the next step. Diligent analysis of their results at this stage should be reviewed against their initial promises to be sure they are solid.

What technological advances should we be looking to invest in at this point? What challenges will have arisen that weren’t foreseen during the hype phase? What type of infrastructure is needed to support scaling of these companies? These are the key questions investors should be asking, and I’ll try and answer them now…

Big dreams, small realities, smart money

Ambition and big dreams are great in overhyped seed funding rounds, but not so useful during a trough of disillusionment. The company that promises the world in this phase needs to be avoided. Focus on the companies that have nailed down their vision to key problems, and the smaller those problems, the better. In radiology AI this means several things:

  • Sharp focus on a specific medical problem. Avoid companies claiming to have solved entire modalities. Instead look for those that have deep expertise in just one issue and have shown results. You can scale these companies later to expand to other focused problems. An example would be a company claiming to have solved Chest CT. This is hype. Avoid. Instead, invest in a company working on solving just one problem within Chest CT imaging. A great example is Arterys who focused on one problem – measuring cardiac flow on MR.
  • Early engagement with regulatory bodies. If a company doesn’t know about or is unprepared for regulations, avoid them. Any product that interprets an image for a clinical purpose is a medical device, and that means stringent processes are needed in order to satisfy the powers that be. Invest only in companies that either have someone on the team who has already successfully navigated the regulatory space before, or have hired appropriately, or already made financial arrangements to outsource this work. Nothing kills a start-up faster than 4 years of unexpected clinical trials.
  • Avoid companies claiming to replace humans. Not one single company has ever got FDA approval for a clinical diagnostic device that is not overseen by a human. Instead, to reduce regulatory burden, look for companies producing software that works alongside and augments humans, known as Clinical Decision Support. These may be triage systems, quantitative analysis tools, registration or segmentation systems. If you absolutely must invest in a diagnostic service, be sure to have deep pockets – FDA fees for PMAs start at $250,000. Good luck to you!
  • Clinically valid use cases. This seems rather obvious doesn’t it? But, you would not believe the amount of bluster I’ve heard from start-up founders. “We are going to measure someone’s age by looking at their bone density.” Sounds like a fun science experiment – but where is the clinical benefit? Is anyone going to pay for this as a service? Aren’t there cheaper and easier ways of getting the same result? Invest only in companies that have a strong scientific board with proven clinical problems being addressed. Make sure that AI is the right way to solve that clinical problem, and that the problem actually needs to be solved. Avoid hammers looking for nails.
  • Look for companies tackling bias removal. AI has a huge bias problem – we can train neural nets on vast amounts of data but that means that the algorithms are always biased toward the population it was trained on. In some cases this means that algorithms are plain racist, ageist, sexist or worse. A company with clever solutions to reducing learning bias will outperform biased algorithms, be more likely to succeed in a regulatory environment, and be more useful in the real world. Medicine is about precision and requires precise tools.
  • Power to the people. Many radiology AI start-ups spin out from universities and academic institutions claiming up to “95% specificity” and other impressive performance metrics. This is hype. Avoid. Check their data set size. It will be nowhere near large enough (statistically powered) to prove anything. They may have trained on a couple of thousand images only, for instance. Instead – focus on companies who may report lower accuracy metrics but have access to much larger datasets (think millions). These companies are the realists! Of course, data access is a crucial roadblock in machine learning, but that brings me to the final point…
  • Invest in companies that will help grow radiology AI as a sector, not just the end products. If I had a multi-million fund to invest, I wouldn’t even look for companies involved in image interpretation. What is sorely needed in the field is not the algorithms (these are the fruit) – it’s the infrastructure behind it (the trees) that’s important. Invest in the orchard! Look for innovative solutions to zero knowledge transfer data storage systems, tools for anonymising medical data, APIs for access to EHR data, DICOM compliant blockchain services, and NLP services for structuring clinical free text. These will be the backbone of the radiology AI revolution, and finding the one company that the entire industry ends up relying upon to build their image interpretation algorithms could prove immensely rewarding.

How to find these companies

So, you know what to look for, but where to start? For seed funding investments, go to the source and network deeply with radiologists, scientists and researchers. Across the globe there are so many hospitals and research networks working on deep learning in imaging, each with a vast cohort of scientists just wishing they could bring their ideas to market. Investors should approach university spin-outs, institutional collaborators and even scientists themselves. Arxiv.org is a great way to find influencers in the scientific community, and also a really useful resource for staying up-to-date with new research in the field (here’s my saved search for radiology papers). Networking with researchers early is crucial to bring them on board if you want to avoid hefty IP issues related to funded university research spin-outs. Investors should be looking to form partnerships with these people credible in the fields of radiology and machine learning, and develop a strategy for finding novel technology using their partners’ expertise (a degree in banking isn’t going to cut it when trying to decide whether or not to invest in pixel cluster analysis of dynamic contrast enhanced MRI sequences).

For Series A funding, the ideal would be to find companies in stealth mode, rather than shiny hyped-up seed funded start-ups looking for growth. There are very few growth-phase radiology AI companies who actually have a fully regulatory-approved marketable product being purchased at scale, and those that do exist are already over-subscribed. Stealth mode companies are a safer bet, but harder task to find, but again, this is where networking with scientists pays off — they know who is building what, who is credible and who is leading in the field.

My final piece of advice is simple: be a tortoise, not a hare. You are in for the long haul. Do not expect significant return in under a 3 year timescale. Spread your investments and plan for a 5-10 (even 15) year period of scaling. Those who invest wisely now and choose companies that can scale smartly on focused problems can lead the market infrastructure. Those who rush and over-promise will only have to play catch-up later down the line.

If you enjoyed this article, it would really help if you hit recommend and shared it.

About the author:

Dr Harvey is a board certified radiologist and clinical academic, trained in the NHS and Europe’s leading cancer research institute, the ICR. He currently works at Babylon Health, heading up the regulatory affairs team, gaining world-first CE marking for an AI-supported triage service.