Radiology Leaders on Artificial Intelligence: Get in Front
Key findings
- Artificial Intelligence's (AI) potential patient care value propositions are: more diagnostic certainty, faster diagnosis and better patient outcomes
- Opportunities for AI in radiology include image analysis that highlights points of concern and improved quality
- Major challenges for AI in radiology are to identify rigorous validation methods and adapt AI to diverse imaging protocols
Radiomic information is being used for prognosis and diagnosis and will grow coincident with the arrival of artificial intelligence (AI) in radiology. Yet, AI in radiology is in its relative infancy. Professionals must create standards, tools and methods for organizing, sharing, using and storing. James Thrall, MD, emeritus radiologist-in-chief, James Brink, MD, radiologist-in-chief and colleagues at Massachusetts General Hospital make the case for why radiologists should get out in front as leaders of this movement to steer it for the benefit of their patients in an American College of Radiology paper.
Subscribe to the latest updates from Cardiovascular Advances in Motion
The beginning of AI traces back to a 1956 conference on machines acting like humans. In recent years, the field has exploded with potential in the field of medicine. AI’s potential value in patient care includes increased diagnostic certainty, quicker diagnostic turnaround and improved patient outcomes. A key question is how to best steer to such value. The authors point to one obvious desire: to ensure AI practices are flexible in accommodating diverse protocols, patient populations and outcomes data.
Recognizing that radiologists must play a lead role, the team organized current thinking and, in particular, opportunities and challenges.
AI Opportunities
Major opportunities are that AI can improve diagnostic accuracy and increase the speed of image scanning and processing. For example, AI may:
- Guide to points of concern or radiomic features that might signify phenotypic disease expression
- Prioritize cases by optimizing work lists
- Improve the quality of reconstructed images
To accomplish such goals, the team recommends several actions. The first is establishing a common language, image sharing networks, reference datasets of proven AI cases, and standardization and optimization criteria for imaging protocols. The second is recruiting and training qualified scientists in AI for radiology. Radiology professional societies can play a lead role in providing educational programs.
AI Challenges
Among all the challenges with AI, the foremost are identifying how to best validate AI results and to accelerate and customize AI processes for clinical settings. Protocols naturally constrain such environments. As a result, scientists must devise ways for AI to respond to variations in patient populations.
As AI advances, the Mass General team stated that leaders in radiology must recognize and manage the tradeoffs. For example, lower accuracy for AI’s more robust use with generalized patient programs might be expected as opposed to a higher accuracy for less robust use of an AI program with smaller, but more narrowly defined patient pools.
A core challenge in radiology is to validate AI against a rigorous standard such as patient outcomes data or a standard recognized metric. Whether AI learning is supervised or unsupervised, with the former referring to learning coming from proven cases, AI is likely best applied to sets of well-defined tasks or patient groups.
Further challenges with AI include:
- Adapting to imaging contexts in which there might be obvious limitations of discernment. For example, a common request in AI is to know what is normal and abnormal in a continuous stream of variable data. While standard deviation is used to set normal ranges, when applied to real patients, it may inadvertently miss those who are actually normal
- AI will trigger more regulations and policies that ideally should be helpful and not hindering. For example, the FDA is likely to weigh in on AI’s interaction with clinical trial data or other areas within their reach. An institution’s legal teams will be charged with defining terms around AI-related data ownership and use rights
Shaping AI’s Adaptation
The Mass General team believes AI is going to be a key part of radiology. The authors encourage the radiology community to fear not.
By taking the lead in understanding AI and its issues, radiologists can steer AI’s adoption in radiology. In doing so, they create the right kind of value that radiologists and their patients need and desire.