September 12, 2024
Adam E. Flanders, MD, FSIIM, CIIP | Jefferson University Hospital
As we inch toward the precipice of ~1000 FDA cleared AI tools (the majority of which are in the imaging space), is it time to consider whether we are better off with AI in our toolbox than without? What tools will augment the specialty sufficiently to make a difference for patient care? Are we addressing bias/fairness sufficiently and have we been using the correct datasets to accurately gauge clinical performance? Recent federal legislation has put bias in the forefront and the users accountable for bias mitigation. We will discuss some of the lessons learned through curation of high-quality datasets via the NIBIB Medical Imaging and Data Resource Center (MIDRC), the Radiological Society of North America (RSNA) Data Science Challenges and the Advanced Research Project Agency for Healthcare’s Biomedical Data Fabric (ARPA-H BDF). Multi-dimensional data may provide a more robust resource for understanding the true generalizability of clinical AI models.
