Deep Learning Processing for Non-Contrast Non-Subtraction MR Angiography of Arteries of the Lower Extremities

Motivation: Traditional FBI is effective for evaluating peripheral artery disease but is time-intensive, requiring image subtraction for clarity between arteries and veins. Goal(s): Develop a deep learning (DL) method to depict only arteries in non-subtraction MRA (nsMRA) images for faster, clearer diagnostics. Approach: A U-Net-based DL model trained on standard FBI images removes veins from nsMRA images, creating artery-only depictions, and fusion images retain the original appearance while highlighting arteries. Results: nsMRA saved ~64% scan time. DL processing removed veins effectively, enhancing readability for inexperienced readers and enabling simultaneous artery-vein evaluation through fusion images. Impact: Our deep learning model effectively depicted arteries in non-subtraction MRA images, with a potential for enhancing vascular disease assessment and reducing confusion for untrained readers by clearly distinguishing arteries from veins. Further testing is needed for clinical generalizability.