January 24, 2019

Dreyer Explains What Radiologists Need to Know About AI

Keith Dreyer, DO, Ph.D., FACR, chief science officer of the American College of Radiology Data Science Institute (ACR DSI), explained to radiologists in two separate lectures at the recent 2019 ACR/RBMA Practice Leaders Forum that artificial intelligence may have fired the public’s imagination, but it also faces challenges that match the magnitude of its great promise for medical imaging

Artificial intelligence is a term that has been around for decades. It describes how a computer may be able to replicate some facets of human behavior. The various techniques used to develop AI algorithms are manifold.

Machine learning describes a specific way of teaching computers to learn these human behaviors and has been applied to problems such as image recognition. Deep learning describes the multiple computational layers of pattern recognition necessary for image analysis. At the core are artificial neural networks, which describe the automated interactions of multiple computational layers. It is a deep learning technique most often used for radiological imaging analysis. Data Science is the outer layer that encompasses all of computational techniques that will help AI improve healthcare for our patients.

AI is not a new phenomenon nor have the public’s expectations from it been consistently strong over its 70-year history, Dreyer noted. AI has suffered from the effects of overhype since the invention of the first modern computers and neural nets in the mid-1940s. When the expectations were not realized, enthusiasm for AI dropped sharply into what Dreyer called “AI winters” where there was little investment or research in AI until the next “breakthrough” development, which once again was overhyped and followed by a fall in the level of enthusiasm. Rapid advances in computing hardware platforms beginning in the about 1990 and continuing into the present day have once again created a surge of interest in artificial intelligence.

Dreyer wondered if this maturing data science will prevent the next “AI winter.” From January 2012 to November 2017, 32 companies invested nearly $556 million for the development of AI and deep-learning algorithms that interpret medical images, according to a study by Alexander, McGill and Tarasova, published on-line in December 2018 in the Journal of the American College of Radiology.

The same study found 14 companies spent about $365 million on precision-guidance intraoperative equipment, MRI guidance software and wire-guided catheters. Twelve spent $194.4 million in a nearly six-year period on developing cloud-based technologies for clinical data exchanges. Fifteen invested $71 million on “block chain” technologies for health data aggregators and big data analysis, and 19 companies dedicated $197 million to develop 3-D visualization and virtual reality technologies.

With Apple, IBM, Google and Microsoft each spending billions of dollars annually on AI-related research, their powerful AI applications are already thrilling consumers. Thousands of photos can be sorted in the blink of an eye with key-word searches on smart phones using AI-developed artificial neural networks. Searches for “mountains,” for example, are performed with 93.2 percent accuracy for all mountains that appear on photos stored on the phone or the vendor’s cloud, Dreyer said.

Though AI is an increasingly powerful force in electronic product design, there is no reason for radiologists to fret about its application to diagnostic imaging. Data science is still fallible – sometimes to a ludicrous degree – unless a skilled person intervenes to add human intelligence to improve its performance.

However, human perception and cognition are imperfect as well. Dreyer stressed that everything we experience is just our brain’s best guess at reality. The accuracy of those guesses can improve with the accumulation of experiences, just as the continued enhancement of AI enables it to increasingly augment human intelligence as a tool for general living or an instrument for diagnostic imaging practice.

At this still early stage, the American College of Radiology (ACR) has strategically positioned itself through the ACR Data Science Institute (DSI) to become a global leader in data science, Dryer noted. The College is guided by a strategic plan that calls upon radiology researchers to define the beneficial use of data science in radiology and to educate the medical imaging community about its appropriate use in clinical practice, he said.

ACR DSI’s most substantive action to date came in December 2018 with the publication of 50 use cases that are now guiding medical imaging developers and their selection of the potentially most useful imaging-related applications of AI.

Already a number of healthcare institutions have set up data science programs. For example, at the Center for Clinical Data Science (CCDS) at the Brigham and Women’s Hospital and the Massachusetts General Hospital, where Dreyer practices, they are working to create, promote and translate AI for health care into AI tools that will enhance clinical outcomes, improve efficiency and enhance patient-focused care.

One such tool, the CCDS’ DeepSPINE Project was developed to decrease variability in interpretations and enhance reporting of lumbar spine MRI. It used applied deep-learning techniques to a training set of 4,075 lumbar MRI studies and 22,795 discs in patients with suspected lumbar spine stenosis.

DeepSPINE calculates the probability of normal spines and mild, moderate, and severe spinal canal, and right and left foraminal stenosis. Early experience suggests accuracy rates of 89 percent to 99 percent for radiologists performing the DeepSPINE-assisted procedures.

Many similar examples where specific AI applications are showing promise for assisting radiologists in day-to-day practice are being reported. For example, the CheXNeXt algorithm for detecting significant thoracic disease on chest radiographs, developed at Stanford University’s Center for Artificial Intelligence in Medicine and Imaging, was able to achieve a performance similar to practicing radiologists.

However, to be effective in widespread clinical practice, these algorithms developed as single institution tools must be generalizable to the health care system at large. The ACR DSI is developing a series of tools that will facilitate training, testing and validating the performance of AI algorithms for the radiological sciences across multiple sites so that we can ensure these tools will perform as expected across multiple institutions.