The Next Technology Wave
How will emerging tools like machine learning affect the specialty?
At this year's intersociety summer conference, the theme was "Big Data and Machine Learning — Strategies for Driving our Value Bus." Jonathan B. Kruskal, MB, ChB, PhD, chair of the Intersociety Commission, led an all-star executive committee to plan an absolutely terrific program.
With strong contributions from ad hoc committee member Keith J. Dreyer, DO, PhD, FACR, chair of the ACR Commission on Informatics, a very informative and stimulating program came together.
The opening keynote and counter-keynote lectures by Dr. Dreyer and Don Dennison, president of Don K. Dennison Solutions Inc., encapsulated the controversy central to this field. Dr. Dreyer took the pro-informatics portion, with his lecture entitled, "Deep Learning and Big Data — We Will Rule the Machines." Meanwhile, Mr. Dennison presented the counter-point with, "We Will Crush You — Machine Learning, Real Implications for Radiologists." Dr. Dreyer and Mr. Dennison provided two perspectives on how the advancement of this burgeoning field can be managed by radiology as opposed to entities outside of the current health care delivery system.
While many chronicles of this meeting are sure to follow, I thought I would share some of my takeaways.
First, I remain optimistic that future efficiency gains achieved through machine learning will enable opportunities for radiologists to extend the benefits of their services to a greater percentage of the population. As we transition to population health management and risk-based contracts, small efforts to help primary care physicians bring patients into compliance with recommended screening and diagnostic imaging examinations will go a long way toward integrating radiologists in the care teams responsible for ensuring the health of managed populations.
While clinical decision support (CDS) for ordering and reporting of radiology exams will help reduce over- an under-utilization of our services, integrating machine learning into these tools will reduce reliance on pick lists to define the attributes required to trigger relevant guidance. By mining patient data in the electronic health record and image data in the PACS, machine-learning algorithms may pre-populate relevant fields in CDS algorithms to streamline the workflow and reduce the administrative burden that may otherwise encumber use of clinical guidance at the point of care. Moreover, machine learning for repetitive mechanical tasks may streamline the technical execution of our imaging exams, potentially increasing throughput and efficiency. If properly applied, these efforts may improve our reach to segments of the population that might not have benefited from radiology services previously.
The introduction of machine learning in image interpretation may transform the way in which we interpret and report on imaging examinations. Combining machine learning with CDS for reporting may enable complete documentation of all abnormal findings, including those that are both relevant and incidental to the clinical question at hand. While some controversy exists as to which incidental findings should be included in radiology reports, the ability to document all findings discovered throughout the body may result in the opportunity to execute an imaging "review of systems" in every imaging exam performed. With each finding represented as a discrete data element, individual practitioners may review those findings germane to their particular subspecialty or to clinical questions at hand. Current arguments against reporting incidental findings include cluttering conventional imaging reports with information that is irrelevant to the clinical question at hand, as well as the time and expense of reporting these findings. If machine learning can help document and catalog all primary and incidental findings, our value as radiologists will be to ensure that this information has been captured correctly and, more importantly, to synthesize and prioritize differential diagnostic possibilities. As keynote speaker Paul J. Chang, MD, an abdominal radiologist at the University of Chicago with extensive expertise in imaging informatics, opined, our job will be to reduce the massive amounts of primary and secondary data that are accrued with each imaging examination to meaningful synthesized elements and conclusions.
Another important takeaway relates to the education that will be required for radiologists to have an active role in shaping radiology's future in the era of big data and machine learning. All meeting participants agreed that radiologists must be actively involved in guiding the introduction and evolution of machine learning in our profession. Our roles will not be to clean and curate image data required for deep learning or to develop machine-learning algorithms.
Instead, we must be facile with issues surrounding this technology and insightful enough to pose the correct questions and translate the work products to clinical practice. We spend a tremendous amount of energy to ensure that our clinical trainees are well versed in radiation physics, and we must commit a similar effort to education the next generation of physicians on issues surrounding medical informatics and information technology.
Finally, issues related to data security and ownership come to mind immediately when big data and machine learning projects are proposed. It is unlikely that a single health care institution will have the necessary talent base to execute state-of-the-art big data and machine learning projects exclusively. This brings with it a host of questions. As data are shared between clinical practitioners and non-clinical data scientists, who should make sure the data are devoid of all patient identifiers? How should shared data be controlled to guarantee they are used exclusively for the intended purpose? Who owns that vast collection of image and other patient data that may be used to produce potentially lucrative patient care algorithms? Do patients have some right to the derivative value of their image data? The answers to these questions are complex and will require substantial effort to define best practices in data sharing and to build policies that promote the dissemination and adoption of this emerging technology.
Lawrence R. Muroff, MD, FACR, is fond of saying, "The future for radiology is bright; the future for radiologists is far less certain." I am confident that by engaging radiologists in the development of data science tools for our profession, we will ensure these tools improve our ability to avail the population at large with the benefits of medical imaging and image-guided intervention. Our future is bright so long as we embrace and participate in the changes ahead.
By James A. Brink, MD, FACR, Chair