Top Takeaways from SIIM 2020
Last week, medical imaging informatics professionals from around the world gathered online for the Society for Imaging Informatics in Medicine Virtual Meeting. Topics related to artificial intelligence and machine learning in medical imaging continued to be a central point of discussion in many of the virtual presentations and Q&A sessions. While participating in these sessions, we noticed several common themes emerge that offer predictions for where the future of medical imaging AI is heading and what is needed to encourage progress.
1. Medical imaging AI has entered the “Trough of Disillusionment”
AI has been hyped in the imaging informatics community for the past four years. At the 2019 Radiological Society of North America Annual Meeting, over 120 vendors exhibited in the AI showcase. This was nearly double the number of AI exhibitors from the year prior. However, we have now moved beyond the hype of AI into a period that is met with waning interest and skepticisim.
Dr. Nina Kottler, VP of Clinical Operations at Radiology Partners, explained in a session on Artificial Intelligence for the Non-Academic Community that there is currently less than 5% adoption of AI solutions as we have entered what the Gartner Hype Cycle refers to as the “trough of disillusionment.” Dr. Kottler shared that we will likely see an increased amount of consolidation in the coming year as medical imaging AI vendors will start to fold in this market. AI vendors that make it out of this period successfully will likely be those who have formed strong partnerships with clinical collaborators.
2. AI solutions need to address radiologists’ needs
As we move beyond the hype of AI, Elizabeth Krupinski, Ph.D. of Emory University shares that solutions powered by machine learning and deep learning need to be evaluated in terms of the impact they have on users in clinical practice.
In the 2020 Dwyer Lecture – Imaging Informatics – Tech is Cool, but What About the Users? Dr. Krupinski discussed the need for AI solutions to bring real value to radiologists. According to Dr. Krupinski, many research papers fail to include a discussion of the clinical significance of research results and ignore how AI will change the way radiologists operate day-to-day. Vendors need to employ methods to evaluate and assess the real value of AI solutions in clinical settings to understand and unlock human benefits. As radiologists experience high rates of burnout and fatigue, algorithms will need to be evaluated using quasi-experimental designs that will reveal both internal and external validities of the studies.
3. Tools need to be developed to evaluate AI algorithms in clinical practice
AI algorithms used in radiology are impacted by common challenges including, poor generalizability, brittleness, low adherence to annotation standards, and a lack of appropriate tools to evaluate algorithms in clinical practice, according to Jayashree Kalpathy-Cramer, Ph.D. of Massachusetts General Hospital and Dr. Daniel L. Rubin of Stanford University.
Dr. Kalpathy-Cramer and Dr. Rubin expanded on the challenges of creating robust and unbiased algorithms in their session, Opportunities and Challenges to Developing Robust AI Algorithms. Dr. Rubin described the need for tools to be developed that would allow vendors to collect feedback on the performance of AI algorithms to gain insight into how well they work in clinical practice. Currently, Dr. Rubin is working alongside colleagues under an ACR Innovation grant to develop a prototype of a toolkit that would be used to collect AI performance metrics.
4. Global collaboration is necessary to accelerate progress
The COVID-19 pandemic has reminded many in the healthcare community that global collaborations are necessary to drive innovation in medical imaging AI. Aashima Gupta, Director of Global Healthcare Solutions at Google Cloud, discussed how COVID-19 is driving healthcare innovation by bringing together members in the healthcare community in a way that has never before been done in the SIIM opening session, What the Doers of Imaging Informatics Need to Know About Machine Learning.
Gupta shared how global collaborations and data sharing can be accelerated through the democratization of infrastructure, the democratization of patient health records and interoperability, and machine learning. As Gupta closed the opening general session:
“Progress is unstoppable, but it must be shaped for good. This is not something that can easily be solved by one entrepreneur, one tech company, government, or one healthcare organization, but rather is a collaborative ecosystem effort.”