If the past few years of speculation, debate and research & development initiatives are any indication, we can deduce that artificial intelligence (AI) is the wave of the future in healthcare, and in the field of radiology specifically.
It makes sense – radiology is a branch of healthcare inundated with huge amounts of data and patient throughput. Most of us at some point in our lives will need to receive a radiological examination of some sort, and the radiologist is tasked with not only identifying pathology but also drawing actionable conclusions from the results. This job function is combined with an increasing pressure on radiology departments to demonstrate value-based care, cost-saving practices, and optimized efficiency, putting tremendous strain on the day-to-day workflow of the typically overloaded radiologist.
This is where the incorporation of AI starts to look very attractive for radiology departments, as a tool that makes the radiologist's life easier. For instance, wouldn't technology that integrates machine learning capabilities to detect lesions in a CT scan free up the radiologist to spend more time interpreting that lesion, collaborating with other physicians and determining the best path forward? Of course, this begs the question of why haven't a majority of healthcare providers incorporated AI into their day-to-day workflow? The simple answer is that while talk of AI is continuing to evolve as a common thread among thought leaders, and we are beginning to see some real implementation of machine learning capabilities in the imaging suite, there are significant barriers to widespread incorporation of AI in radiology.
The first challenge, one that I have commented on before, is the dilemma of coming to a consensus around the precise capacity for AI to be integrated into our workflows. Perhaps most optimistically, and certainly most immediately, AI will be an instrumental tool that aids the radiologist much in the way suggested above – by handling data segmenting and rudimentary tasks, so the radiologist is freed up to engage in the more critical and complex aspects of his/her function.
However, a fear among those in the industry is the possibility of AI becoming so independently capable of reading and interpreting images that the role of the radiologist is called into question. But how legitimately pressing is that fear? If a 2017 survey that found that 35% of hospital leadership has no plans to move forward with AI1 is accurate, we are clearly nowhere near there yet. But the idea does beg the question – what are the long-term implications?
Another challenge in AI implementation is the logistical questions that arise when we consider how we would reasonably leave elements of healthcare decision-making in the hands of machine learning algorithms. The more 'independently-minded' a piece of technology is, the more critical the medical community will be in its review (rightly so), ultimately heeding the path to FDA clearance. How would we go about board certifying an output of AI, after all? We must keep in mind the lack of precedence that has been established around medical equipment designed to run entirely autonomously.
Last but certainly not least is the issue that most community hospital systems and private practices are simply too overloaded trying to manage patient throughput and day-to-day responsibilities to consider integration of AI. As it happens, these are also the locations where most people get their images taken. Yes, Johns Hopkins and Dana Farber may be prioritizing the integration of top-of-the-line data capabilities, but the smaller practices are just trying to stay afloat in a value-based care environment. They are not (as of yet) focused on AI because it has not yet been presented to them in a digestible format.
But that is starting to change. Innovators are seeing these challenges to widespread implementation of AI and are finding new and creative ways to move past them. Rather than viewing AI as a fear-inducing replacement of medical experts, technology companies are instead looking for ways to integrate machine learning into PACS and other digital equipment in more seamless ways that anticipate user preferences and inform data management.
Companies like Philips are even changing the way we talk about AI, coining the more nuanced term 'adaptive intelligence' as a marriage between intelligent data-oriented solutions and human clinical expertise. Through this approach, innovators are telling those who may be more skeptical of AI that they anticipate and understand their concerns – and are meeting them where they are in their journey to deliver advanced care.
Trends we saw at RSNA this year also reflect a sort of shift in approach to AI. Whereas 2017 talks and showcases featured grandiose interpretations of AI's place in radiology, 2018 saw a more restrained, yet evolved outlook. Rather than presentations on the future of AI as a potential replacement for radiologists, this year's meeting focused on specific applications of AI that for the first time are ready to be used clinically. Even so, it is likely that the further along we are in the development of AI-infused technology, the more we will realize how long the road ahead really will be with AI. Thus, the industry may be settling into its current limitations as an assisting tool for radiologists, rather than a total solution.
Of course, where opportunity lies will be in innovators' abilities to develop these tools further, and in a way that makes them as seamlessly incorporated into the day-to-day functions of the imaging suite as possible. I envision innovations that expand past individual point solutions to streamline overall workflow across multiple modalities. Larger companies like Philips are already starting to do this, through features like PerformanceBridge and IntelliSpace Portal that draw actionable conclusions from large amounts of data sources and at multiple points along the care journey.
In the future, we will no doubt see this more and more. As AI continues to evolve in radiology and the larger healthcare space, one thing is certain – it is here to stay.
Footnotes
- 2017 HealthLeaders Media Analytics in Healthcare Survey as covered in HealthLeadersmedia.com. September, 2017.
