"Why didn't AI replace radiologists yet?" I get this question 3x per week. "Computer vision is solved," I heard Lucas Beyer say once. So... how hard can it be? Didn't Google show that it could detect lung cancer better than human more than a decade ago? Then I shadowed a radiologist for a few hours. Dark room, massive screens. Click. Scan. Diagnosis in 30 seconds: "Nothing to see here." Quick dictation to Microsoft Nuance. Done. Next patient. Then he showed me 'the AI': Right click → AI menu → Load tool → Wait 10 seconds → Select "lungs" → Wait 15 seconds → Colorful circles appear on screen. Out of an abundance of caution, it overflags. Leading to more work, not less. And potentially more false positives, instead of fewer. Radiologists are already impressively optimized. They obsess about every click. They've trained their internal models ("intuition") not just on images, but on all other highly detailed contextual signals, too: patient history, what the referring doctor told them on the phone, the particular quirks of this machine, the radtech that made a small remark when walking over just now... Oh, and, don't forget: what gets reimbursed. Incentives rule the world. That means that even if we get the AI right, the UI needs to be absolutely perfect. For instance: instead of firing up your local machine when you click "AI", preprocess all the patients from the last hour. Then first bring them a batch with the ones that the model confidently sees as "no cancer", like DeepHealth does. I'm also super excited about some of the new things I've seen come up, like Aidoc or what Wouter Veldhuis is cooking - stepping away from point solutions towards end-to-end workflows. But that still requires better AI models. And there are many hard problems that won't make this easy. For instance: → Medical language is often strategically fuzzy. "Cannot rule out" → There is a long tail of rare, unusual, and complex diagnoses → No reimbursement; limited incentive → Very, very high standards for medical devices That doesn't mean AI isn't already helping radiology today. For instance, at Ahead Health, we use it to determine various imaging-derived biomarkers. Moreover, we use it to help explain things in a way that you _and_ your non-radiology doctor understands it. And at our partner MRI clinics, we've squeezed every last bit of regulatory-approved-AI for a custom protocol. More data, but fewer minutes in a loud machine. Anyway, Out-Of-Pocket did an excellent write up I loved reading. Link in comments.
Interesting post Nick Lenten thanks 🙏
Interesting post Nick Lenten. Eleven years ago, when I worked at Quantib AI was still in its early stages. However big steps are made ever since. Its not just about speeding up the workflow. Its also about protecting radiologist and patients for small errors. For instance when doing 100 chest x rays during a day a small nodule or costal fracture can be missed, but its so essential. As clinical lead of Hevi AI Benelux I am so fortunate to test new software and glad to have two more eyes checking me while doing chest x rays or mri prostate in ACIBADEM International Medical Center . Off course, further optimalisation is needed and seamless integration in any PACS system is mandatory! In Acibadem and Scan Clinic we work with Sectra and PixelData and AI is seamless integrated. In Scan Clinic we use AI to support us in MRI Total Body Scans. Not only for cerebral aneurysmal detection or prostate lesions detection but also for composition scans and for CT bone reconstruction of MRI images (in order to see spondylolysis more easily and without radiation exposure!! ). AI is our 5th and 6th eye. Innovation is key in radiology ! Keep up the good work in Zwitserland Nick Lenten and keep us informed about new AI steps !
I always assumed that with a strong model, radiologists would simply work more efficiently. What i didn't realize is that the real challenge is how the model is integrated: if it slows down the process or adds extra steps, even the best model won't be adopted. Thanks!
Beautiful and insightful thank you !
Ultimately healthcare requires a human to be responsible, which is why radiologists have absolutely nothing to worry about. I find the current state of treating radiologist and pathologists (most of whom are frankly brilliant) as easily replaced technicians to be a complete delusion… there is just another layer of security that comes from a human with a limited amount of time and a lot to lose if they get it wrong.
"Radiologists are already impressively optimized. .. They've trained their internal models on images ...and on all other highly detailed contextual signals, too..." What can I say, as a radiologist... attention is all you need 🤓 More seriously: I totally agree. AI is already helping, somewhat, but to truly valorize the implementation you need custom protocols, like you have at Ahead Health, or like we're building with QA4AI @UMCU.
Founder Ahead Health - ex-CPO Coolblue + Google Flights - hiring!
4wLink: https://www.outofpocket.health/p/why-radiology-ai-didnt-work-and-what-comes-next