Original articleArtificial Intelligence and Machine Learning in Radiology: Opportunities, Challenges, Pitfalls, and Criteria for Success
Introduction
The invention of the programmable digital computer in the 1940s stimulated mathematicians and philosophers to speculate on the limits of what machines could do. Could machines learn to think? How close could machine capabilities come to those of human beings? A conference at Dartmouth University in 1956 explored these questions and led to the coining of the term artificial intelligence (AI) [1]. The race was on.
In the ensuing 60 years, enthusiasm for AI has waxed and waned but has reignited recently with the availability of ever less expense, massively parallel computing systems. The term deep learning was added to the AI lexicon to reflect the ability to harness new computing power to develop more powerful AI approaches with more layers of analysis than heretofore possible. The successes of AI programs from IBM (Armonk, New York) in the games of chess and the quiz show Jeopardy! (Deep Blue) and from Google (Mountain View, California) in the game of Go (DeepMind) [2] were exciting milestones that made people outside of the scientific community aware of AI and its potential.
Major corporations and governments around the world have embraced AI technology as one of the important strategies for dealing with the enormous amounts of digital data being generated in the information age—the age of “big data.” AI is also on the doorstep of medical practice. The trickle of publications now appearing in journals will soon turn into a flood.
The radiology community has played a leading role in propelling medicine into its digital age and now has the opportunity to become a leader in exploring medical applications of AI. The tens of millions of radiology reports and billions of images now archived in digital form exemplify the concept of “big data” and constitute the required substrate for AI research.
The fundamental question is whether AI applications in radiology can add value. Adding value includes the discovery of new knowledge and extraction of more and better information from imaging examinations to achieve better outcomes for patients at lower cost. For radiologists, adding value includes establishment of more efficient work processes and improved job satisfaction.
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.
Section snippets
Opportunities
Two areas of opportunity that can help provide a framework for approaching AI in imaging deserve discussion: 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 to help identify and understand their respective value propositions.
Challenges
The challenges in exploiting the potential of AI in medicine may be thought of as circumstantial, relating to human societal behaviors, and intrinsic, relating to the capabilities of the underlying science and technology. None of the challenges alone will be a showstopper but all may slow progress and need to be addressed.
Pitfalls
AI programs typically require substantial numbers of cases for training. Institutional xenophobia and other proprietary interests may restrict access to image data between institutions. Failure to assemble a sufficiently large enough training set is a potential pitfall that could have the effect of making the results less accurate or generalizable [19]. The risk of overfitting was noted previously 6, 13.
The tolerance of using AI programs in imaging between different patient populations is not
Criteria for Success
The most important criteria for success follow directly from the discussion of opportunities. 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 [15]. As with any new technology, substantial time and experience will be required to establish whether these benefits apply and
Conclusion
It is not yet clear what the full or final role of AI methods will be in imaging or their impact on radiologists. What is clear is that AI provides a promising new set of tools for interrogating image data that should be explored with vigor. The growing interest in AI in the imaging community bodes well for its potential leadership role [6].
Take-Home Points
- ■
Worldwide interest in AI applications, including imaging, is high and growing rapidly.
- ■
The large amount of image and report data now in digital form (“big data”) provides a substrate for development of AI applications.
- ■
Development of AI applications in imaging would benefit from the development of standards and infrastructure—acquisition protocols, validation criteria, lexicon for communication.
References (19)
- et al.
The end of radiology? Three threats to the future practice of radiology
J Am Coll Radiol
(2016) - The Dartmouth Artificial Intelligence Conference. The next fifty years. Available at:...
- McAfee A, Brynolfsson E. Where computers can defeat humans and where they can’t. Available at...
- et al.
2014. DeepFace: closing the gap to human-level performance in face verification. Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition
(2014) - et al.
The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository
J Digit Imaging
(2016) - Russakovsky O, Deng J, Su H, et al. 2015. ImageNet large scale visual recognition challenge. Int J Comput Vis...
- et al.
Deep learning for health informatics
IEEE J Biomed Health Inform
(2017) - Balogh EP, Miller BT, Ball Jr eds. Improving diagnosis in health care. Washington DC, USA: National Academies Press;...
- et al.
Lung pattern classification for interstitial lung diseases using a deep convolutional neural network
IEEE Trans Med Imaging
(2016)
Cited by (0)
Dr Li has an institutional disclosure for the system: “Automatic pre-screening method for pneumothorax detection.” Dr Thrall has an institutional disclosure on analysis of liver lesions. No patent has been applied for. The other authors have no conflicts of interest related to the material discussed in this article.