Dynamic Mode Decomposition ( DMD ) for Low‐Latency Real‐Time Cardiac MRI

ABSTRACT Purpose To demonstrate dynamic mode decomposition ( DMD ) for high spatiotemporal low‐latency online reconstruction in 2D real‐time cardiac MRI. Methods DMD was applied to 2D spiral balanced steady state free precession ( bSSFP ) real‐time adult and fetal cardiac MRI at 0.55 T, with data from 10 healthy adult volunteers (3F/7M; age: 21–49; BMI: 20–34) and 6 pregnant females (maternal age: 30–41; maternal BMI: 22–47; gestational age: 23 weeks 6 days–37 weeks 5 days). DMD model appropriateness was assessed against off‐line spatiotemporally constrained reconstruction ( STCR ) as the reference. We retrospectively evaluated DMD‐based low‐latency online reconstruction at two temporal resolutions (21 and 42 ms/frame). DMD modes were estimated from the most recently acquired frames and used to remove aliasing while preserving underlying physiological motion. Results DMD represented cardiac dynamics with normalized root‐mean‐square error ( NRMSE ) less than 7% when all modes retained. Low‐latency DMD‐based online reconstruction performed de‐aliasing while preserving the physiological motion, supporting framerates (21 and 42 ms/frame). Conclusion We have demonstrated that the DMD framework is applicable to 2D real‐time cardiac MRI and for low‐latency de‐aliasing for better online reconstruction.