Few can claim to have done as much to prepare upcoming generations of radiologists for the technological future as Elliot K. Fishman, MD.
Dr Fishman is professor of radiology and oncology at The Johns Hopkins University School of Medicine and director of Diagnostic Imaging and Body CT at The Johns Hopkins Hospital.
In the last 20 years, Dr Fishman has coordinated more than 100 continuing medical education (CME) courses.
Each year, Dr Fishman brings to the RSNA Scientific Assembly and annual meeting at least 20 education exhibits, scientific posters, scientific papers, plenary sessions, and refresher courses.
He has won awards for his work encouraging radiologists to embrace innovations like CT, 3D imaging, and social media. AIMed Magazine spoke to Dr Fishman to find out how he is educating radiologists about AI:
AIMed: As an educator, what would you like to teach our readers about AI and medical imaging?
Elliot Fishman, M.D.: What AI can do is make you a much better radiologist than you are. AI presents the opportunity to increase our sensitivity and specificity for detecting disease.
Radiology has a 30% error rate – misdiagnosis up to 30%. At the same time, people are getting busier and the workload is getting harder. Something has to give.
AI will improve your ability to read studies, maybe improve the speed, surely it will improve the accuracy. It will be like having a helper, the world’s best helper, but it won’t replace the radiologist.
The programs must be strong and robust, interfaces have to be good, but at the end of the day it’s going to be the best of times, to quote Mark Twain.
I honestly believe there’s never been a better opportunity for radiologists to take radiology to the next level of expertise. I don’t see any downside.
AIMed: When did you begin to see the potential for AI in radiology?
EF: I’ve developed for 30 years in 3D imaging, since the Pixar days, so in a sense I’ve always known what computers are doing for us or could be doing for us.
Medicine is always behind, because it’s very conservative and slow, but when you see all the stuff that’s happening with deep learning in the rest of the world then you know it’s coming. If you don’t think that, then you’re a crazy person.
AIMed: Where do you see AI making the greatest impact?
EF: I think the two fields most apt for AI impacting are radiology and pathology.
Pathology is about looking at slides and is ideal because you’re looking for a set field and something that doesn’t belong, like a cancerous cell. There’s been articles published by Google already showing increased accuracy of AI compared to pathologists.
And look at the parts of radiology: tomography, breast imaging – it’s a limited field of view, one organ, you’re looking for the presence of calcifications or a mass and to compare things over time, tasks which computers are incredible at.
I think we will see a lot in the space of detection and patient management. People are working on being able to predict response of tumours to therapy based on radiomics and deep learning.
People who are smart are working on small problems, or finite problems. You can’t say, “I’m doing deep learning on CT” – that’s impossible, it’s too big.
AIMed: Have you seen any good examples?
EF: I saw recently the FDA approved a product detecting fractures of the wrist on plain x-rays . It’s no better than the radiologist, but it’s as good as the radiologist and it would help the radiologist.
Also, a research team developed a program which looks for a bleed in the head . It’s useful because teleradiology companies have a lot of slides from different cases come in which they read in order and there’s no way of separating which cases are more urgent.
The program looks at all the head CTs and if it detects blood then it moves the scan up the queue, so it’s read as the next case.
Let’s say the computer was wrong? The radiologist would read the scan correctly, which they have to read anyway, so the only thing that happened is they read it a bit earlier. That’s a very simple thing but it could be very helpful.
AIMed: How are you tackling the fear surrounding job losses?
EF: Every time there’s something new in radiology, some people say this is the end of radiologists. They need to take a deep breath and look at what AI is promising.
For example, at Johns Hopkins radiology department we have a project using AI which has been very successful for detecting pancreatic cancer . We’re getting great results and can detect pancreatic cancer with between 80-90% accuracy, automatically.
But you also have to be practical. Let’s say our program was 100% accurate at detecting pancreatic disease, that’s great – but you still need the radiologist, because after detecting cancer in the pancreas you need to look at the liver, kidneys, spleen, bowel, etc. Until it can look at everything, a computer can’t take over.
However, if you have this program running in the background and it picks up one thing you missed, then you’re gonna say to yourself boy this is damn good.
AIMed: Do you think radiologists should learn about data science and coding?
EF: That’s not going to happen. We’re busy enough trying to read the damn films. It’s naïve to think you’re going to take two courses and start programming. It’s just sheer madness.
I’m modestly an expert of body CT. Can I program a CT scanner? The answer is NO. Can I even run the CT scanner? I’m sure I could learn how, but we have technologists who do that.
The point to remember is: do only what only you can do. I read images really well, I know the history and what the referring doctors are looking for. My job is to make sure I do that.
However, I think a good thing for radiologists is to know where the opportunities are, and if they deal with vendors then encourage them to develop apps that are actually helpful and valuable for radiologists.
AIMed: Your earliest manuscripts focused on educating the practicing radiologist about body CT and the value of 3D imaging. Are you planning on doing similar work educating practicing radiologists about AI?
EF: I already am. On our website, www.CTisus.com, we are starting a section on Deep Learning. It’s about trying to show people where resources are.
Most radiologists know nothing about deep learning. They need practical lectures and someone who approaches it in the right way.
If you have someone who gets up there and says, “Oh my god, it’s coming! You guys are gonna be parking cars”, then people get very upset or stressed out.
But if someone stands up and says, “Guys, this is coming but this is the greatest thing that ever happened to us”, then people say, “Ok, let’s see what’s so good about it”.
This interview originally appeared in AIMed Magazine, a Deep Dive on AI in medical imaging, to download click here.