A new study showed the efficiency and speed of a new AI model for a heart scan.
Pankaj Garg, MD, PhD
Credit: UEA University of East Anglia
A new artificial intelligence (AI) model accelerates heart scan diagnoses from about 45 minutes to a few seconds, a recent study found. Using AI can not only save a doctor’s time but can improve patient outcomes and save lives.
“The AI model precisely determined the size and function of the heart's chambers and demonstrated outcomes comparable to those acquired by doctors manually but much quicker,” said lead investigator Pankaj Garg, MD, PhD, from the University of East Anglia’s Norwich Medical School and a consultant cardiologist at the Norfolk and Norwich University Hospital, in a press release.
Garg and colleagues conducted a multicenter, multivendor, and retrospective observational study to train and develop a deep learning AI model to examine heart scans over time with the specific view of the 4-chamber cine. The cardiac magnetic resonance (CMR) in the 4-chamber plane provides a thorough insight into the volumetrics of the heart.
The second objective was to confirm the consistency and reliability of the AI model’s results by comparing them with both manual expert analyses on the 4-chamber cine and existing automated methods on short-axis segmentation. Lastly, the third objective was to assess the model’s prognostic value.
The team trained the AI model using the data of 814 participants, leveraged from the 2 studies: the ASPIRE register from Sheffield Teaching Hospital (n = 624; 367 scans with GE Healthcare equipment and 257 scans with Siemens Healthineers equipment) and Leeds Teaching Hospitals NHS Trust (n = 190; Phillips Healthcare equipment).
An independent cohort of 101 patients (studied with Siemens Healthineers equipment) from the PREFER-CMR register in Norfolk and Norwich University Hospitals was created to evaluate validation, reproducibility, and mortality prediction with the AI model. The mean age of the validation cohort was 54 years, and 65% were males.
Participants were included in the derivation and validation cohorts if they were ≥ 18 years old, had a clinical indication for cardiac magnetic resonance, had a good quality scan for segmentation, and provided written informed consent.
Investigators found the left and right heart measurements taken by automated methods closely matched those taken manually, with strong correlations (P = 0.91 – 0.98 and P = 0.89 – 0.98, respectively). All 4-chamber volumetrics were highly repeatable and showed no bias and high correlations with automated and manual analyses (P = 0.99 – 1.00).
They did observe the automated method underestimated the volumes of both the left and right ventricles compared to the manual method, and the team suggested 2 correction factors for left ventricular and right ventricular 4-chamber analyses to adjust for this underestimation. After applying those corrections, the results were more accurate, and there were no significant differences and minimal bias between automated 4-chamber and short-axis measurements for left ventricular end-diastolic volume (bias = 2.04 mL, P = 0.493) and left ventricular end-systolic volume (bias = -0.26 mL, P = 0.903). The results stayed similar for right ventricular measurements, with a reduced bias from -69 mL to -18.7 mL for right ventricular end-diastolic volume and from -37.6 mL to 13.2 mL for right ventricular end-systolic volume
During an average follow-period of 6.75 years, 16 patients died. A stepwise multivariable analysis demonstrated a specific measure of heart function—left atrial ejection fraction—was independently linked to the mortality risk in the manual (hazard ratio [HR], 0.96; P = .003) and AI analyses (HR, 0.96; P < .001).
Ultimately, an automated 4-chamber cardiac magnetic resonance offers an efficient and accurate prognostic tool for identifying heart issues. In the press release, Garg said the AI takes only a few seconds, unlike a standard manual MRI analysis which can take 45 minutes more.
“Automating the process of assessing heart function and structure will save time and resources and ensure consistent results for doctors,” said PhD student Hosamadin Assadi, from UEA’s Norwich Medical School, in a press release. “This innovation could lead to more efficient diagnoses, better treatment decisions, and ultimately, improved outcomes for patients with heart conditions. Moreover, the potential of AI to predict mortality based on heart measurements highlights its potential to revolutionize cardiac care and improve patient prognosis.”
References