Integration of AI Decision Support Software within a Community Radiology Practice: Impact on Nodule Characterization and Decision to Biopsy
Tuesday, April 9, 2024
12:22pm – 12:29pm
Location: 412
Authors: Louis Mazzarelli, 385003213975 Kevin Hricko, Lawrence and Memorial, Yale New Haven Ira Sitko, Lawrence and Memorial Hospital David Zucker, Ocean Radiology Associates Ting Chen, YNH Lawrence & Memorial Hospital Lev Barinov, University of Pennsylvania ,
Thyroid nodules are commonly incidentally identified in diagnostic imaging and while TI-RADS has been broadly adopted in the United States to characterize and risk stratify thyroid nodules, continued challenges to successful implementation remain including those related to interpretation, reader variability and over biopsy. Decision support software powered by artificial intelligence has been proven to improve reading efficiency but similarly decrease interpretive variability while decreasing the rate of unnecessary biopsy amongst skilled ultrasound radiologists. Within the context of a community practice composed of general radiologists of varying skill levels, the impact of decision support would be anticipated to be even greater. Through a retrospective analysis of all FNAs performed at our institution and a comparison of radiologists TI-RADS interpretation compared to AI decision support, this study demonstrated that incorporation of AI DS would have facilitated a 30% reduction in unnecessary biopsy, greater than the 25% biopsy avoidance rate reported amongst skilled ultrasound radiologists. Moreover, AI DS can help decrease assessment variability amongst a range of readers facilitating more cohesive lesion characterization. Additive gains of efficiency, decreased variability and biopsy avoidance will allow more focused assessment of indeterminate biopsies and as Bethesda III nodules which remain a diagnostic challenge.