Charting a Course
Keeping up and moving forward in machine learning means having a roadmap for success.
Gauging the practical implementation of day-to-day, timesaving AI in medical imaging is no simple calculation. Setting your group’s compass to ensure quality care doesn’t have to be an arduous endeavor.
According to Bibb Allen Jr., MD, FACR, chief medical officer of the ACR Data Science Institute™ (DSI), “AI translation to routine clinical practice has been slower
than many expected because we have to ensure that AI in medical imaging is useful, safe, effective, and easily integrated into existing radiology workflows.” To move the needle, DSI created well-defined use cases for AI in 2018. These include implementing and streamlining existing ACR programs, including ACR Clinical Guidelines and ACR Assist™ modules. DSI is also working to engage other radiology societies to build models that support clinical guidelines and care pathways.
As part of a multi-stakeholder approach, the National Institute of Biomedical Imaging and Bioengineering held a workshop last year at the National Institutes of Health — from which a two-part roadmap to AI was published by RSNA and ACR. The most recent ACR report (available at bit.ly/JACR_AIReport) details the real-world AI challenges facing radiologists, considering translational research in AI that will help medical imaging speed up the best uses of AI in clinical practice.
AI in Practice
To boost that transition, the College launched the ACR AI-LAB™ — a constellation of software tools designed to help radiologists learn the basics of AI and participate directly in the creation, validation, and use of healthcare AI. The ability to improve on an algorithm, while keeping patients’ personal information secure, protects them, your practice, and makes the adoption of AI more practical, Allen says.
“DSI is actively pursuing all of the aspects of translating AI to clinical practice,” Allen says. “The AI-LAB allows for sites to participate in algorithm validation and
collaboration with other sites through transfer learning. Modifying algorithms using local data improves diversity and generalizability of the model.”
The four key priorities set out in ACR’s roadmap are:
• Creating structured AI use cases that define and highlight clinical challenges that may be solved by AI
• Establishing methods to encourage data sharing for training and testing AI algorithms to promote generalizability (spreading widespread clinical practice and mitigating bias)
• Establishing tools for validation and performance monitoring for AI algorithms to facilitate regulatory approval
• Developing standards and common data elements for the seamless integration of AI tools into existing clinical workflows
The 2019 Data Science Summit in June, hosted by DSI (in conjunction with the Society for Imaging Informatics in Medicine’s annual meeting), focused on the current state of data access and liquidity; ethical issues regarding data ownership, privacy, sharing, bias, and stewardship; and the democratization of AI. The event brought together thought leaders to discuss and evaluate where radiology is today and what the field can expect in the future.
“We know algorithms can underperform when deployed at sites where they weren’t trained,” Allen says. “Now radiologists in the new AI-LAB pilot program will have access to AI algorithms developed outside their institutions to evaluate a model’s performance using their own data and, as necessary, refine and tune the algorithm using their local data to enhance its performance.”
AI tools will be a critical driver of the future of radiology. They hold the potential to deliver a wealth of information to inform accurate diagnoses and to identify patients most at risk of serious illnesses.1 If you’re interested in AI’s potential to bolster patient-centered care, attend the ACR Annual Conference on Quality and Safety (acr.org/QSMeeting) to learn about patient-centered quality improvement initiatives in a scholarly environment conducive to networking and discussion among peers. The conference will be held in Denver, Oct. 11–12, with intensive pre-conference workshops available on Oct. 10 at no additional charge.
The ACR has also brought back the Imaging Informatics Summit for radiologists, practice leaders, industry partners and policymakers to explore strategies for putting AI to good use. This year, the summit is being held Oct. 5–6 in Washington, D.C. Sign up to participate in the hands-on, interactive event — including an onsite demonstration of the AI-LAB to inform AI implementation at acr.org/IISummit.
By Chad Hudnall, senior writer, ACR Press
1. Allen B, Seltzer SE, Langlotz CP, Dreyer K, et al. A road map for
translational research on AI in medical imaging [published online ahead of
print May 28, 2019]. J Am Coll Radiol.