Acquisition of Cardiac Point-of-Care Ultrasound Images With Deep Learning
Evan Baum, Megha D. Tandel, Casey Ren, Yingjie Weng, Matthew Pascucci, John Kugler, Kathryn Cardoza, Andre Kumar
Abstract
Background Point-of-care ultrasonography (POCUS) machines may use deep learning, a subfield of artificial intelligence (AI), to improve image interpretation and acquisition in real time. The impact of AI on POCUS learning is unknown. Research Question Do AI-enhanced devices equipped with deep learning aid in cardiac image acquisition and interpretation among POCUS novices? Study Design and Methods We conducted a single-center investigation from 2021 through 2022. Internal medicine trainees (N = 43) with limited POCUS experience were randomized to receive a POCUS device with (Echonous; n = 22) or without (Butterfly; n = 21) AI functionality for 2 weeks while on inpatient rotations. AI device functionality included guidance for optimal probe placement to acquire an apical four-chamber (A4C) view and ejection fraction estimations based on deep learning. Participants used the devices at their discretion for patient-related care after randomization. The primary outcome was the time to acquire A4C images on a standardized patient. Secondary outcomes included A4C image quality using a validated scale, image quiz performance, and attitudes. Measurements were performed at randomization and at 2-week follow-up using the same standardized patient. Results Both AI and non-AI groups showed similar scan times and image quality scores at baseline. At follow-up, the AI group showed faster scan times (57 s [interquartile range (IQR), 32-75 s] vs 85 s [IQR, 50-172 s]; P = .01), higher image quality scores (4.5 [IQR, 2-5.5] vs 2 [IQR, 1-3]; P < .01), and more accurately identified reduced systolic function on the image quiz (85% vs 50%; P = .02) vs the non-AI group. The AI group used the devices more than the non-AI group (median, 5.5 times [IQR, 4-10 times] vs 2 times [IQR, 0-4 times]; P < .01). Trust in the AI features did not change during the intervention. Interpretation POCUS devices with deep learning functionality may improve A4C image acquisition and interpretation by novices. Future studies are needed to determine the extent that AI impacts POCUS learning.