Machine Learning and Artificial Intelligence
Raymond Y. Kwong, MD, MPH, FSCMR
Director of Cardiac Magnetic Resonance Imaging
Brigham and Women's Hospital
Boston, Massachusetts, United States
Raymond Y. Kwong, MD, MPH, FSCMR
Director of Cardiac Magnetic Resonance Imaging
Brigham and Women's Hospital
Boston, Massachusetts, United States
Michael Jerosch-Herold, PhD
Associate Professor
Harvard Medical School, United States
Tuan Luu, BSc, RT
Chief Technologist
Brigham and Women's Hospital, Massachusetts, United States
Jon Hainer, MSc
Informatics Specialist
Brigham and Women's Hospital, Massachusetts, United States
Rhanderson Cardoso, MD
Instructor of Medicine
Brigham and Women's Hospital, Massachusetts, United States
Ravi Teja. Seethamraju, PhD
Research Scientist
Brigham and Women's Hospital
Malden, Massachusetts, United States
Okai Addy, PhD
Professor
Stanford University
Los Altos, California, United States
Lori Powell, BSc, RT
Chief Technologist
Brigham and Women's Hospital, California, United States
Juan Santos, PhD
Professor
Stanford University
palo alto, California, United States
Bob Hu, MD, PhD
Professor
Stanford University, California, United States
CMR provides effective guidance to clinical decisions, but its widespread adaptation is limited by complex imaging protocols and long scan times. Artificial intelligence may improve throughput and consistency of clinical CMR imaging.
Methods:
We implemented an AI-assisted platform (HeartVista, Palo Alto, USA) for automated CMR imaging on a 3T system, with integrated scan planning and quality optimization using pulse sequences provided by the MRI manufacturer (Siemens Medical). Typical CMR protocols of all studies included cines in short and long axis, T1 and T2 mapping, T2-weighted fast spin-echo, resting perfusion, and late gadolinium enhancement (LGE). Studies were grouped as fully AI-assisted, partially AI-assisted, and non-AI-assisted, where partial AI-assist referred to using a combination of AI-assisted and non-AI assisted pulse sequences in a study. Total scan time and image quality (graded by a 5-point scale) were recorded. Technologists’ experience in CMR (< 1, 1-5, 5-10, >10 yrs) was assessed as a modifier to scan time and quality.
Results:
In the 1,178 consecutive CMR studies were protocolled to assess cardiomyopathy (559, 49%), arrhythmia (94, 8.2%), suspected myocarditis (66, 5.8%), and other (427, 37%), 262 (22%) using AI-guidance. Comparing fully AI-assisted to partial and non-AI-assisted exams, a 17% and 30% reduction of scan time were observed (Figure 1: 38±7 vs 46±11 vs 54±17 min, p< 0.0001), respectively. Image quality scores for cine (4.5±0.5 vs 4.2±0.7 vs 4.0±0.6, p< 0.0001) and LGE images (4.3±0.5 vs 4.2±0.7 vs 4.0±0.6, p< 0.0001) were higher. Reduction of scan time and improvement of image quality scores for cine and LGE were similar across all CMR protocols. Technologists’ CMR experience did not modify the above results.
Conclusion:
With widely-used, routine CMR protocols, AI-assisted guidance was associated with a 30% reduction of scan time, higher consistency in patient throughput, and improved image quality of key protocol components.