By ALEX LOGSDON, MD
Leave your bias aside and take a look into the healthcare future with me. No, artificial intelligence, augmented intelligence and machine learning will not replace the radiologist. It will allow clinicians to.
The year is 2035 (plus or minus 5 years), the world is waking up after a few years of economic hardship and maybe even some dreaded stagflation. This is an important accelerant to where we are going, economic hardship, because it will destroy most radiology AI startups that have thrived on quantitative easing polices and excessive liquidity of the last decade creating a bubble in this space. When the bubble pops, few small to midsize AI companies will survive but the ones who remain will consolidate and reap the rewards. This will almost certainly be big tech who can purchase assets/algorithms across a wide breadth of radiology and integrate/standardize them better than anyone. When the burst happens some of the best algorithms for pulmonary embolism, stroke, knee MRI, intracranial hemorrhage etc. etc. will become available to consolidate, on the “cheap”.
Hospitals can now purchase AI equipment that is highly effective both in cost and function, and its only getting better for them. It doesn’t make sense to do so now but soon it will. Consolidation in healthcare has led to greater purchasing power from groups and hospitals. The “roads and bridges” that would be needed to connect such systems are being built and deals will soon be struck with GE, Google, IBM etc., powerhouse hundred-billion-dollar companies, that will provide AI cloud-based services. RadPartners is already starting to provide natural language processing and imaging data to partners; that’s right, you speak into the Dictaphone and it is recorded, synced with the image you dictated, processed with everyone else to find all the commonalities in descriptors to eventually replace you. It is like the transcriptionists ghost of the past has come back to haunt us and no one cried for them. Prices will be competitive, and adoption will be fast, much faster than most believe.
Now we have some patients who arrive for imaging, as outpatients, ER visits, inpatients; it does not matter the premise is the same. Ms. Jones has chest pain, elevated d-dimer, history of Lupus anti-coagulant and left femoral DVT. Likely her chart has already been analyzed by a cloud-based AI (merlonintelligence.com/intelligent-screening/) and the probability of her having a PE is high, this is relayed to the clinician (PA, NP, MD, DO) and the study is ordered. She’s sent for a CT angiogram PE protocol imaging study. This is important to understand because there will be no role for the radiologist at this level. The recommendation for imaging will be a machine learning algorithm based off more data and papers than any one radiologist could ever read; and it will be instantaneous and fluid. Correct studies will be recommended and “incorrectly” ordered studies will need justifications without radiologist validation.
The patient receives her scan and while she is being transported back her results are already available. An alert is sent directly to the ordering provider via an app on their phone that Ms. Jones does in fact have multiple PE’s. Treatment options are arranged and the patient is treated. Maybe the ER doc takes a quick look at the chest CT and while scrolling through she can see where the AI has pointed out the pulmonary emboli (www.aidoc.com). Part of the algorithm will also analyze osseous structures, the heart, lung parenchyma, etc. and point out abnormalities and provide clinical and radiologic recommendations, instantly. Anything unrecognizable to the system will be flagged and sent to the 1–2 hospital employed radiologists for analysis. This will be many at first but the numbers will forever decrease as the years progress because machine learning systems are inherently initially hypersensitive. Heck, studies from all over the country and/or world may even be sent to select academic centers that employee the largest amount of radiologists to further analyze the images and hospitals will not have to employee any radiologists except to reside over “radiology extenders” performing fluoroscopy studies; but by then they’ll likely have autonomy like NPs.
Orthopedist are pretty good at looking at the images they order already and would never perform surgery without reviewing their own films, this applies to all surgical specialties. Having appropriate access to cloud-based algorithms, that a group or hospital can purchase for them is highly incentivizing for the orthopod and they can collect the professional fee. This scenario goes like this; an orthopedist has a patient they’ve been treating for osteoarthritis of the knee and the time has now come to consider replacement. Imaging is ordered, the patient is scanned and the results and recommendations are sent to the orthopedist. The image is compared to 100,000 other individuals who have had knee replacements, their demographics and outcomes. The orthopod reads the report, scrolls through the imaging and thinks “yepp its time for a new knee”. The algorithm agrees and the surgery is scheduled. The most common soft tissue malignancies, osseous abnormalities etc. are included in the algorithm and again any extreme variations will be sent out for validation to the drastically reduced number of radiologists.
The next patient is status post fall with head strike and is sent for a head CT. Again, before the patient is back to the ED the doctor is aware the patient has a subdural hematoma. Neurosurgery is consulted, they look at the films and agree. The patient is admitted and observed. A repeat study performed hours later shows no change in the size or density; per the algorithm, the neurosurgeon agrees and the patient will follow up post-discharge. The incidental osteoma was seen and the white matter changes seen and characterized to a much deeper extent than possible with the human eye. Speaking of the ED, ultrasound use will become ubiquitous, acute cholecystitis can be diagnosed with a hand-held Butterfly IO ultrasound (love these they’re so convenient and easy to share images) the images sent to the surgeon, the clinical correlation made and the patient goes to the OR. Maybe they’ll even be able to find the appendix and pancreas (pushing it) when they are inflamed too. Plain films x-rays are the easiest target many radiologists in community hospitals do not even read plain films overnight. Algorithms to point out fractures and dislocations already exist. Layering on top will be algos to diagnosis other bony abnormalities. Stroke? Forget it! Neurologists are already battling for this turf and they don’t have ML to assist (http://www.i-rapid.com). Some neurointerventionalist already say “we don’t look at the radiologist read just use rapid”, wrong or right its happening.
The recurring question here is where and what is the radiologists roll in the future? Reports will be generated automatically faster than by the time the patient is back in their room, read through by the ordering provider and agreed with by the referrer or specialist and the clinician will be paid for the read. Now this is where it gets ugly. If I am a hospital CEO I would rip up the radiology employee contracts, hire 2–3 radiologists (medium sized hospital) or contract them with a local academic center and purchase cloud-based AI systems. There would be tremendous cost savings and I would still collect a technical fee on all the scans performed; maybe even a professional fee. For contracted groups, again those contracts terminated, add a couple employed radiologists and purchase cloud-based software while I roll around in my technical and professional fee collection with minimal relative overhead. A precedence would be set that, “any scan you order, you look over”, which in my experience most physicians already do and they will receive part of that professional fee into their salary as an incentive.
This idea works even better under a socialized medicine system, good luck suing the federal government (VA, military). Under the current system or a free-enterprise system a collective responsibility of tech giants (subsidiaries), hospitals and clinicians and governemtn would have to exist, tort reform would need passing and there would need to be years of radiologist watching AI and AI watching radiologists. The latter will likely decrease the overall need for imaging study related litigation.
Like a pilot utilizing autopilot the radiologist must be aware of AI short-comings, know their plane inside and out and be ready to take over at any moment. As a Air Force flight surgeon rumblings are the F-35 or the next version will be the last fighter jet to have a sitting pilot within it. No need to worry about G-force (except mechanical stresses), sacrificial sorties and extremely expensive training of fighter pilots. What is different than the airline industry though is the number of destinations. If one plane could carry 5,000 passengers to one destination and do it faster than before, the need for pilots would drastically decrease. I believe that is where we are approaching and quickly. Nvidia’s GPU’s in 2015 took 25 days to train ResNet-50 an extremely complex autonomous driving algo (arguably more complex than most all imaging algorithms), and 1 month ago it took 2 minutes; should that pace hold in 4 years it will take 0.7 seconds. These services are already also provided over the cloud through Google hosting and are fractions of the cost they use to be. This exists and is currently being used for more complex tasks such as autonomous driving, its not a far-off maybe.
Current training: I would personally rather be in scarcity than in over-supply because scarcity, as we know, is what makes something economically valuable. Our training should include structured and focused machine learning, natural language processing and artificial intelligence teachings. We will know no better than the IT guy (our hospital requires high school graduation for that position) what is malfunctioning, where and its affects. I also believe we should start to diversify our skillsets into genetics, radiogenomics and population health. Be ahead of the game and not scrambling to play catch-up. The fact is radiologists and radiology residents are some of the most intelligent well-rounded people on the planet. A way will be found but why make it harder?
Talking to patients about studies may be a popular idea now but it will be irrelevant in the future as clinicians will have many more tools to review studies with patients, provide instant results and feedback and patients will have better ways of accessing their images and results. Patients want convenience not another doctor telling them another opinion on another day (why is Amazon so popular? Convenience and options). Our worth lies in our ability to synthesize multitudes of medical knowledge, imaging and clinical data, and precisely articulate to others what they imply. That is what we have been doing and inorder to keep doing so we need much more capability than interpreting images at the highest RVU rate. Will there be a continued need for 30,000 radiologists? I’m no math genius but NO there will not be the need for 30,000.
So many tech companies feed us the line “we wont replace radiologists we will supplement them”. I’m calling BS. There is so much more impact to be made and money to be had by replacing and they would never stop at supplementing. Not that they are bad but because this is common sense.
This is why I believe the American College of Radiology needs to consider how many radiologists are made each year, “allowing” for the use of radiology extenders, our current educational objectives and residency curriculum and our future roll not solely as imaging interpreters. As it stands now IR/DR and ESIR residents complete about 3 years of DR and the rest is IR. They are eligible for both boards. I believe this means we have room to add curriculum on the above topics for diagnostic residents into the current 4 years of radiology training. The arbitrary and small number of needed studies to graduate also serves no purpose, other than maybe pleasing the ACGME, and should be reconsidered as well. In a year those numbers if focused on can be met. As current radiologists in training we need to be prepared for our mid to late careers and currently we are totally underprepared for that.
Radiology is not alone. Mid-levels, new technology and millennials’ perception of health care will decrease the need for many physicians. The physician shortage is highly exaggerated and mostly in rural areas many physicians will still not go no matter how many you produce. There are many advantages to government, insurers, hospitals etc. to over-produce physicians and NP’s but that is another discussion.
Summary: Radiologists today needed guided training on machine learning to understand what can go wrong and identify when it does go wrong. Each countries respective college needs to make sure that we have this ability and thus should be the only specialty capable of consistently interpreting these images.
Alex Logsdon is an early interventional radiology resident with over 10 years of experience in the military and healthcare. This article originally appeared on Medium here.