Futurists sometimes claim that artificial intelligence (AI) will make radiologists obsolete. Their argument has been that compared to humans, algorithms are better and faster at analyzing medical images such as X-rays. So why has this hype failed to become reality? In this opinion piece, Ulysses Isidro and Saurabh Jha write, “For radiology AI to be widely adopted, it needs to overcome several barriers. Most of all, it needs to make the radiologist’s work simpler.” Isidro is a MD/MBA student at the University of Pennsylvania Perelman School of Medicine and The Wharton School. He has a diverse range of experiences across radiology, health care management, health policy, and global health and plans to become a diagnostic radiologist. Jha is an associate professor of radiology at the Perelman School of Medicine. He writes extensively about artificial intelligence.
Last year, the Radiology Society of North America held its annual meeting at McCormick Place in Chicago. At the event, which includes a radiology trade show and is one of the world’s largest medical meetings, a special place was reserved in the sprawling conference center for vendors in artificial intelligence (AI) – the basement.
The basement is both symbolic and ironic. Radiology, a field where radiation once abounded, began its evolutionary journey in the hospital basement. The radiologist was a mythical character found only in the catacombs of the hospital. With the advent of digitization, images could be viewed remotely on a computer screen, anywhere, through an interface known as the Picture Archiving and Communication System (PACS). After that the volume of imaging exploded, and the radiologist could be found anywhere in the hospital, or the world, including the high-rise buildings of Bangalore, India.
A few years ago, as computing became more powerful, using deep neural networks, computers were able to recognize patterns in medical imaging. Futurists predicted the end of radiologists. The computer scientist, Geoffrey Hinton, one of the pioneers of deep neural networks, cautioned against training more radiologists because it was evident that AI would soon replace them. Radiologists, some prophesied, as skilled as they were, would face what economist John Maynard Keynes called “technological unemployment.”
It’s 2020. Where are the robot radiologists?
Beyond the Original Hype
The radiology AI industry astutely dismissed the original hype of AI. Instead of threatening to replace radiologists, the industry touted its AI products as tools that would help radiologists perform better, as a “rad helper,” so to speak. Despite this strategic shift in role from master to assistant, AI has not penetrated the radiology marketplace to any reasonable degree. It is unclear if the industry is stuck in the primordial soup phase of evolution – promissory but not ready for prime time or if it has reached its asymptote, a goal it keeps approaching without being able to reach it.
The suppliers have been unable to convince most radiologists that they need AI.
Fundamentally, AI products aim to help radiologists read imaging faster and more accurately. The problem facing the industry is the lack of evidence-based clinically meaningful radiology AI models that are integrated into a radiologist’s workflow with an enticing financial return on investment. The suppliers have been unable to convince most radiologists that they need AI. Without demonstrating a need, it is hard to derive payment models that pay for that need.
To better understand the radiology AI industry, let us divide it into producers, distributors, and customers. From a producer’s perspective, the three categories of need are diagnosis, reading efficiency, and scanner cost-effectiveness. Most radiology AI products, or algorithms, specialize in diagnosis, such as detecting life-threatening brain bleeds, or other flagging urgent diagnoses such as pneumothorax, pneumonia, or fracture. Some companies have developed AI models that diagnose a group of common diseases or findings for a specific imaging modality such as chest radiograph, mammography, or head computed tomography (CT).
Some algorithms increase the radiologist’s efficiency by moving patients with urgent findings to the top of the worklist, or by saving time doing repetitive tasks such as measuring lung nodules, or by pre-populating radiology reports with patient information and imaging findings. Finally, a few of the radiology AI algorithms increase the cost-effectiveness of CT and magnetic resonance imaging (MRI) scanners by minimizing radiation and decreasing acquisition time without compromising image quality.
The producer sector comprises mostly private startups, some of which have spun out of academia, which have raised more than $500 million in angel, seed and venture capital funding over the past few years. Though academic centers have done much research on AI, their algorithms have not been widely commercialized. There are also more established companies producing algorithms as part of a larger portfolio of AI solutions, such as merging radiology with pathology, or integrating imaging with data from a patient’s electronic medical record. Some producers are pharmaceutical and biotechnology companies that have created radiology AI models for clinical trials which involve imaging substrates.
With so many startups entering this sector in such a short period of time, it is expected that many will not survive in the coming years. The effect of the coronavirus pandemic on imaging’s fortunes also affects the financial health of these companies. Consolidation among AI producers and acquisitions by larger health care companies are likely soon.
App Store for Algorithms
The distributors, or vectors, of AI can be divided into three categories: AI marketplaces, which are essentially like an App store for algorithms; original equipment manufacturers (OEMs) in radiology; and PACS vendors. A few distributors are vertically integrated with producers.
Several AI producers have attempted to bypass distributors and go directly to customers, such as hospitals administrators and radiologists. However, the overall trend is that AI marketplaces — by connecting producers to other distributors and/or customers — are gaining salience.
AI marketplaces give radiologists access to multiple algorithms developed by different producers, eliminating the need to install each algorithm separately. These marketplaces are critical to the dissemination of AI; they function as intermediate agents and must extract some of the payment reserved for AI.
AI vendors face a choice between internally building their own marketplace or partnering with existing AI marketplaces. That is like Nike deciding between leasing space in a mall or building its own mall.
OEMs and PACS vendors are distribution channels for radiology AI producers and even AI marketplaces. On behalf of radiology AI producers, OEMs have marketed AI products directly to their customers or bundled them when selling diagnostic imaging equipment. AI vendors face a choice between internally building their own marketplace or partnering with existing AI marketplaces. That is like Nike deciding between leasing space in a mall or building its own mall.
The major customers of radiology AI are, of course, radiologists, across various practice settings such as hospitals, academia, private practice, and teleradiology. Radiology AI companies have targeted non-radiologists, too, such as emergency physicians, orthopedic surgeons, and neurologists. The aim is to make these clinicians, who are not formally trained in imaging, more confident about interpreting medical imaging.
There are financial, regulatory, clinical, and logistical barriers to adoption. The financial devastation that the coronavirus pandemic has caused has left hospitals and radiology groups with less capital for investment. Radiology AI vendors have not demonstrated much of a financial return on investment for AI. There is little financial incentive to invest in radiology AI as Medicare and private insurance companies do not reimburse for its use, at least for now. Because there is no robust evidence showing that AI improves quality or decreases costs, it is unlikely that physicians will be reimbursed for using AI in the foreseeable future.
Barriers to Adoption
Radiologists will not use AI if they fear that the technology will eventually replace them. But fear of replacement is not the biggest barrier to adopting AI; it is the fear that AI will slow them down. Radiologists are already working at near maximum efficiency. Though more radiologists are beginning to acknowledge that the best patient care will be provided by radiologists who use AI, many are concerned about the legal consequences of using AI. One thorny question, whose answer is unclear, is whether the radiologist or the algorithm is liable for missing a finding.
For radiology AI to be widely adopted, it needs to overcome several barriers. Most of all, it needs to make the radiologist’s work simpler. Radiologists would be unwilling to use AI that is not integrated in their existing workflows, which are rooted in their PACS. Radiologists may resist adjusting their workflows, particularly if the new workflow reduces their efficiency.
The radiology AI industry has grown significantly in less than a decade, most notably within the producer sector and more recently in the distributor sector. Before the coronavirus pandemic, there was a lot of excitement and investment fueling the rise and, arguably, saturation of AI startups. However, due to barriers to adoption combined with the impact of coronavirus, there remains limited demand from customers to buy and use radiology AI. The radiology AI market is at risk of extinction or irrelevance.
With several research-grade AI models in the market, AI producers and distributors need to strategically co-invest the capital and resources to gain regulatory clearance for AI models for which customers are willing to pay. For future AI models, because the workflow integration process of radiology AI begins with the customer, AI producers should gather and incorporate as much feedback from customers early in the development process. Until radiology AI gains significant traction with radiologists, it may be worth the risk for AI producers and distributors to temporarily pivot and instead adapt existing radiology AI models for non-radiologist customers.
In the end, it will be up to the radiologists to save the radiology AI market. This will be a worthwhile endeavor, as these algorithms will, one day, do wonders for patient care without stealing anyone’s job.