AI and machine learning in radiology: 4 things to know

As industry experts continue to explore artificial intelligence (AI) applications in radiology, the question remains of whether AI applications can and will add value, including in new knowledge and information to provide patients with better outcomes at lower costs.

In a new editorial published in JACR by a team of researchers from the department of radiology at Massachusetts General Hospital and Harvard Medical School in Boston, the "big data" consuming technologies of AI and machine learning are evaluated in terms of opportunities, challenges, pitfalls and criteria for success.  

"For radiologists, adding value includes establishment of more efficient work processes and improved job satisfaction," said lead author of the study James H. Thrall, MD, chairman emeritus of the department of radiology at Massachusetts General Hospital. "The goal of this perspective is to help create a framework, apart from a discussion of AI technology per se, for developing strategies to explore the potential of AI in radiology and to identify a number of scientific, cultural, educational and ethical issues that need to be addressed."  

The researchers note that although the ultimate role of AI in medicine is not yet clear, AI will provide advanced tools to more thoroughly analyze imaging data.  


Thrall and his colleagues explain two areas of opportunity that may help in the inclusion of AI and machine learning in medical imaging: "the desirability of establishing standards and infrastructure" and "the opportunity to establish a categorical model for approaching the spectrum of clinical and research applications of AI."  

In terms of standards and infrastructure, AI imaging research would greatly benefit from image-sharing networks, reference data sets to compare and test AI programs, criteria for AI application standardization and optimization and a common terminology for describing and reporting AI applications, according to the researchers. Additional opportunities regarding AI applications may include optimization of work lists to prioritize cases, pre-analysis of cases at high volumes, extracting information undetectable to the naked eye and improving the overall quality of reconstructed images.  

"AI applications offer an important new way to extract heretofore unavailable information from images and are a new portal for imaging contributions to the era of precision medicine," Thrall said. "The age of big data and deep learning has spawned the concept of 'radiomics' wherein hundreds of abstract mathematical features of images can be defined or detected and through AI programs correlated with other data on genomics or response to therapy."  


Circumstantial challenges, relating to human societal behaviors, and intrinsic challenges, regarding the capabilities of underlying AI and machine learning, do not cease to exist, according to researchers.  

Radiologists worrying about AI and machine learning taking away the need for a human workforce remains a concern in the medical community. Also, the lack of data to feed AI and machine learning systems risks "overfitting" the data with loss of generalizability, researchers wrote. Lastly, the limited number of radiologists currently trained in AI methods and necessary financial investments to install AI and machine learning also loom.  

"Historically, once an area is recognized as important, capable people quickly populate it, so this is not likely to be a long-term issue," the researchers wrote. "Practicing radiologists will need to learn about AI but will not need to become experts in AI research or design of AI programs to beneficially use AI-based results."  

On the other hand, intrinsic challenges for AI in imaging include how to best establish honest, validate results, develop "protocol tolerant" AI programs and establish criteria to validate patient populations in a given program.  


Thrall and his colleagues explained that because AI requires substantial cases for training, "institutional xenophobia" may restrict access to image data between institutions, which could contribute to a failure in assembling a large enough data training set. Additionally, they explain that the tolerance of using AI programs in imaging between different populations has not been explored yet. However, the biggest limitation for AI involves defining normal versus abnormal continuously variable biologic data, according to Thrall.  

"Ranges for normal are set as a certain number of standard deviations from the mean of a supposedly 'normal' population," Thrall said. "This means for any test or measurement, a given percentage of truly normal people will have 'abnormal' results. Investigators in AI will face this conundrum where nominal criteria for normal versus abnormal can be difficult to define when, for example, setting limits for organ sizes." 

Criteria for Success  

"The name of the game is to create value in the delivery of medical care and delivery of radiology services: increased diagnostic certainty, decreased time on task for radiologists, faster availability of results and reduced costs of care with better outcomes for patients," Thrall said.  

Time and experience will be needed to establish whether the benefits Thrall described can be applied and understood by those who utilize AI technology. According to the researchers, success is possible if AI programs can be made tolerant of different data acquisition protocols and work with diverse patient applications.