Radiology AI -- Where Are We?

Radiology AI -- Where Are We?

We started Tau in 2019 with the fundamental premise that the inflection point of AI was near. Specifically our thesis was predicated on the quadrant below, where people refers to people qualified enough to apply AI.

No alt text provided for this image

When it comes to healthcare nowhere is the impact of AI perhaps more visible than radiology, and this article will zone in on four major aspects around it.


1) Approvals – The first FDA approval around AI was in 1995 and it took years for the pace to pick up. But now, in the month of July 2022 alone there were 19 approvals. In fact, the data from just about a year ago shows 521 approvals around AI, the vast majority under radiology. Perhaps most importantly, much of that innovation is starting to permeate in real life. Case in point the landscape below, even though from 2020, shows the race to the market:

No alt text provided for this image

2) Acquisitions – Another reflection of acceleration is acquisitions. IBM famously bought Merge Health for $1B in 2015 primarily to feed medical data to Watson. But overall we were still in the doldrums until things changed in 2022. That year started with Radnet, the largest provider of outpatient imaging services in the US, acquiring two startups . Sirona and Rad.ai , two of the most well-funded and visible disruptors, acquired another startup each – history shows that when startups buy startups it’s almost always a leading indicator for bigger deals. And lo and behold in 2023 large corporates made big plays, with Bayer , Tempus and Philips all acquiring radiology AI startups. At Tau we see the wheels of venture capitalism are now greased enough – the prospect of exits fuel more funding which fuels more startups.


3) Challenges – All that said, there is much to do:

  • Comprehensiveness aka how can you ensure that the AI has access to enough data, especially considering rare diseases / diagnoses and underserved populations.
  • Explainability aka why did the algorithm make a certain decision when you cannot explain it, which is especially tricky in healthcare considering AI is in many ways a black box.
  • Liability aka who is responsible for an AI’s mistake, is it the doctor applying the AI or is it the company that created the AI.
  • Integration aka how does it fit into existing workflows for providers, after all you can have the coolest tech in the world but if it’s not easy to adopt then it makes very little difference
  • Training aka how do we train and retrain humans fast enough to adopt the new ways.


4) Models > Data – Human data is incomplete, ambiguous and contradictory. Structuring it in a meaningful way for a machine is where the devil meets their mettle. But the progressive evolution of AI means we are increasingly developing models and methods such that the algorithm doesn’t need massive amounts of data. There is generative AI which allows us to create data, say images of a particular rare diagnosis. There are adversarial networks which allow an instance of AI to learn from another, which is how Google famously built an algorithm that beat the human champion of Go. And there are a number of improvements in the broader category of unsupervised learning i.e. a computer can learn without being necessarily told everything.


5) So What Does This All Mean – We subscribe to the view that AI overall will make radiology better, helping healthcare professionals do more accurate diagnostics and treatments, quicker and cheaper. Yes it will displace some jobs but it will also create many others that we cannot even imagine yet. The amount of imaging done worldwide has just exploded. One reason is doctors increasingly ask for an x-ray or MRI to make better decisions and patients will expect this ask. This been especially apparent in the US which is about half of the global imaging market, but is very much a trend worldwide as very large countries like India and China are becoming more prosperous. How do we navigate this brave new world? Instead of making it more robotics, AI actually help radiology and medicine generally more humane.

No alt text provided for this image


Originally published on “Data Driven Investor ”. Amit is Managing Partner and Cofounder of Tau with 20 years in Silicon Valley across corporates, own startup, and VC funds. These are purposely short articles focused on practical insights (we call it gl;dr — good length; did read). See here for other such articles. If this article had useful insights for you, comment away and/or give a like on the article and on the Tau Ventures’ LinkedIn page , with due thanks for supporting our work. All opinions expressed here are from the author(s).

Anne-Maree Cantwell

Venture Capital Partner | Advisory Board Member | Physician Leader | Accelerating Digital Healthcare Technology Adoption | Telemedicine | Mergers & Acquisitions | Kaiser Permanente & Wharton

1 年

Your comments are spot on Amit Garg, especially as it relates to the challenges to adoption by healthcare providers!

Steven Young

Engineer, Founder

1 年

Exciting to see AI's potential in enhancing radiology! The balance of challenges and progress outlined here gives a comprehensive view on the road ahead. Looking forward to the transformative impact of AI in healthcare and its potential to make diagnostics more efficient and accurate.

Richard Simpson

Insight Scientist | Intelligent Engagement | Life Science Investor and Consultant | Former Crime Solver

1 年

Just an FYI on Ventripoint Diagnostics Ltd. (GE Edison partner) and its AI cardiac imaging product that produces accurate metrics out of 2D/3D CVUS images (used for complex cases, kids, RV, poor images). Regulatory approvals in place, 10+ years of empirical testing, with latest version generating meaningful sales traction.

Andrew Johnston

Radiology @ Stanford, Product @ Clearstep

1 年

Great post! My opinion on the areas of highest impact for AI in radiology are to 1) help us quantify our findings better without extra work (like what HeartFlow does with coronary CTs) and 2) help us organize the increasing deluge of images we look at and how we report findings in a more streamlined way. With the massive increase in imaging exams and images per exam without the corresponding increase in radiologists to read them, I don't think we are too far away from radiology practices "dying" to adopt AI tools. A lot to iron out in terms of reimbursing the algorithms though.

要查看或添加评论,请登录

Amit Garg的更多文章

社区洞察

其他会员也浏览了