Web Exclusive

The Road to Radiology's AI Adoption Runs Through Workflow Integration

By Chris McIntyre

The convergence of robust artificial intelligence (AI) algorithms and the growing masses of patient image data have created a perfect storm of hype around the future impact of AI on helping to solve radiology's clinical problems. This hype, however, often focuses primarily on technology, zeroing in on the accuracy and rigor of algorithms and their potential to disrupt radiology practice. In their day-to-day work, radiologists are actually much more concerned about how to demonstrate value and deliver higher-quality patient-centered care. AI has the power to augment radiologists' work and achieve these goals, but access to that power requires new approaches to radiology workflow design.

While it is a fact that AI imaging is still very much in the infancy of its technical development and practical applications, and there are regulatory issues and business model barriers to AI adoption once these hurdles are cleared, broad provider adoption will require integrating AI into clinical workflows; it's not just about dropping the technology and results into existing interfaces and calling it a day. Current workflows and user interfaces need to be rethought and adapted to the ways AI interacts with patient image data and delivers insight. Unobtrusively bringing AI insights to clinicians at the right time, in the right way, requires consideration of clinical context and the diverse preferences and mental models of clinicians.

Getting Looped In

Imaging AI algorithms are created to target specific aspects of the body, such as the brain, heart, or lungs, as well as specific modalities. While cardiac CT specialists need access to insights from one specific set of algorithms, generalists will need tools that offer access to all available algorithms. The different needs of generalists and specialists put AI into the familiar territory of health IT interoperability. The results of AI insights need to be available to readers through all radiology tools with an understanding of the needs and requirements of the individual image reader.

Another element critical to clinical use of AI is the machine learning "feedback loop." Like radiologists, AI algorithms become more knowledgeable and accurate as they are exposed to different patient studies and expert feedback over time. Support for this machine learning process requires seamless access to features for accepting, rejecting, or even correcting diagnostic insights, and thus validating and training algorithms. Not only will the ability to review and revise AI's diagnostic capabilities improve accuracy, it will also give clinicians input and control over the process, which in turn will build their trust in and acceptance of AI as a diagnostic tool.

For maximum adoption and value, action needs to be taken within the context of a normal workflow instead of via some specialized interface at some other time. The ideal workflow scenario allows radiologists to interact with the feedback loop naturally as they read a study, both because it is more efficient and because the feedback loop is more effective when the details of an exam and the associated AI insights are top of mind.

In addition, for AI to be another tool in a radiologist's toolbox that augments clinical decisions, this workflow flexibility is critical. In some cases, AI will provide great insight, while in others it may not be considered helpful. In either case, to use or put aside an AI insight should not require a specialized process.

Once AI insights and diagnoses have been viewed and reviewed, the radiologist also needs the option to include the results—or not—within the patient record for downstream consumption via diagnostic reports, rounds, conferences, and face-to-face collaboration with referring physicians. Within radiology tools and workflows, AI results must be presented in a manner that can be integrated naturally into these established communications channels.

Valuable Lessons

Looking back at how other new technologies have been adopted offers lessons for smoothing the path to AI's adoption. The adoption and use of speech recognition, which today has a critical role in efficient and effective report generation, parallels AI in many ways. Like AI for imaging diagnostics, speech recognition technology is built on algorithms that are refined with use. In addition to becoming more accurate over time through feedback loops, speech recognition tools progressively became more deeply and seamlessly integrated into PACS reading workflows, which is when the real productivity gains began to emerge. Workflow integration was critical to the adoption of speech recognition and the realization of its clinical and performance benefits.

In addition, the history of computer-aided diagnosis (CAD) highlights the critical nature of clinical workflow integration in health IT adoption. CAD has existed as a concept in multiple forms for almost 20 years, but it has yet to deliver fully on its initial promise. While AI is more technically advanced, its basic premise is similar to that of CAD and, in the same way, unlocking its value requires workflow integration that makes it easy to access and builds trust with clinicians.

Lessons learned from clinicians' previous adoptions of health IT show that acceptance and use of new technologies is just as dependent on proven results as it is on workflow integration. While the health care industry focuses on making AI technically and clinically reliable, it is equally important to address how it will be made useful to clinicians in their day-to-day work. Health care innovation mixes technology and humans, and the best solutions ensure smooth integration between technical speed and accuracy and human knowledge and interaction.

— Chris McIntyre is the director of product management for Calgary Scientific, Inc.