This article is more than 5 years old. Information may no longer be current.
AI 186 times faster reading cardiac MRI scans than human expert counterparts
Click Here to Manage Email Alerts
A machine learning network, a form of artificial intelligence, was able to consistently measure left ventricular ejection fraction as well as LV mass in cardiac MRI scans faster, with comparable reliability to experienced clinicians and a significantly shorter training period.
According to the findings published in Circulation: Cardiovascular Imaging, compared with expert observers and trainees, machine learning showed similar precision in detecting a 9% change in LVEF (coefficient of variation = 6.1% [95% CI, 5.2-7.1] for expert observers; 8.3% [95% CI, 5-6-10.3] for trainees; 8.8% [95% CI, 6.1-11.1] for automated).
Additionally, the human analysis of the cardiac scans took about 13 minutes per scan (interquartile range, 9-19) while the machine learning analysis of 25-phase short-axis stack was approximately 4.2 seconds, making it 186 times faster, according to the researchers.
Moreover, the trainees, who previously had initial Society for Cardiovascular Magnetic Resonance level 1 accredited training in scan analysis, had undergone a 1-month training program, whereas the experts assessed in this study had more than 15 years of experience. In comparison, machine learning was trained from scratch using 599 multicenter, multi-scanner data sets of patients with severe aortic stenosis during 8 hours and 40 minutes.
“Cardiovascular MRI offers unparalleled image quality for assessing heart structure and function; however, current manual analysis remains basic and outdated. Automated machine learning techniques offer the potential to change this and radically improve efficiency, and we look forward to further research that could validate its superiority to human analysis,” Charlotte Manisty, MD, PhD, of the department of cardiac imaging at Barts Heart Centre, London, said in a press release. “Our data set of patients with a range of heart diseases who received scans enabled us to demonstrate that the greatest sources of measurement error arise from human factors. This indicates that automated techniques are at least as good as humans, with the potential soon to be ‘super-human’ — transforming clinical and research measurement precision.”
In other findings, machine learning was also consistent in reliability and speed in measuring LV mass in these scans, with the average coefficient of variation being 4.6% with experts, 7.6% with trainees and 5.5% with machine learning.
Researchers and machine learning analyzed the cardiac MRI scans of 110 patients (32 with machine learning; 17 with LV hypertrophy; 17 with cardiomyopathy; 14 with other pathology; 30 healthy volunteers) across five institutions, using two different scanner manufacturers and two field strengths.
“Despite reliance on measurements of LVEF and LV mass for clinical decision-making and as endpoints in research studies, analysis is often not standardized, and the relative contributions of error sources are imperfectly known,” the researchers wrote. “These data show that using current standardized image acquisition and multicenter, multi-vendor, multi-field strength, multi-disease, scan-rescan data at scale, measurement error was largely due to inconsistency in the human observer rather than variation in modifiable factors — clinician experience, scan acquisition or human contour strategy.” – by Scott Buzby
Disclosures: Manisty reports no relevant financial disclosures. One author reports consulting for Circle Cardiovascular Imaging. The other authors report no relevant financial disclosures.
Perspective
Back to Top
Joao A.C. Lima, MD, MBA
I am confident that cardiac MRI data acquisition and analysis will use machine learning methods. The paper represents a wonderful demonstration of the power of these AI methods to tackle one of if not the most important problems in MRI analysis, which is inter-reader reproducibility. I think these methods will be used clinically soon.
I think it will greatly enhance the use of cardiac MRI by making data analysis easy and straightforward. The latter tends to be an important bottleneck in the adoption of cardiac MRI clinically, and these techniques are likely to greatly reduce the need for laborious analysis of MRI findings.
A lot more research is needed on the implementation of these methods, and extension to other types of clinical phenotypes are commonly determined by MRI.
Click Here to Manage Email Alerts