Artificial Intelligence Performance in Cardiac Magnetic Resonance Strain Analysis for Aortic Stenosis: Validation with Echocardiography and Healthy Controls

TL;DR

AI improves heart strain analysis in aortic stenosis using MRI, but errors are higher in patients than in healthy controls.

Contribution

Validates AI-based CMR strain analysis in aortic stenosis patients and identifies higher error rates compared to healthy controls.

Findings

AI-based CMR GLS strongly correlates with echocardiographic GLS (r = 0.694) and shows lower variability.

AI-derived GLS is significantly lower in aortic stenosis patients compared to healthy controls.

AI strain analysis had higher error rates in AS patients (18.6%) than in controls (2.44%).

Abstract

Background and Objectives: Aortic stenosis (AS) leads to progressive left ventricular (LV) dysfunction, making early detection crucial. Global longitudinal strain (GLS) is an echocardiographic marker of subclinical LV dysfunction; however, echocardiography has limitations, including operator dependency and acoustic variability. Cardiac magnetic resonance (CMR) is a valuable complementary tool, and artificial intelligence (AI) may enhance strain measurement accuracy, though its role in AS remains underexplored. To evaluate the performance of an AI-based CMR feature tracking tool for the assessment of LV global and segmental GLS in AS patients and compare results with the respective measurements from healthy volunteers (control group), as well as with the GLS obtained using the echocardiographic speckle tracking technique. Materials and Methods: This retrospective study analysed 111 CMR exams (70 AS patients, 41 healthy controls) from a single centre. AI-derived GLS values from gradient echo 2-, 3-, and 4-chamber CMR views were manually reviewed for accuracy. Error rates, segmental, and global myocardial strain differences were assessed between AS patients and the control group. Results: AI-based CMR GLS strongly correlated with echocardiographic GLS (r = 0.694, p < 0.001) and showed lower variability. The AI-derived GLS from CMR was significantly lower in aortic stenosis patients compared to controls (−17.86 ± 3.47 vs. −20.70 ± 1.98). However, AI-based strain analysis had an overall error rate of 6%, which was significantly higher in AS patients (18.6%) compared to healthy controls (2.44%) (p = 0.0088). The 3-chamber CMR view was the most error-prone (50% of isolated errors). Segmental strain variability between AS patients and controls was most pronounced in basal segments, with smaller differences in middle and apical segments. CMR demonstrated greater precision than echocardiography, as indicated by a smaller standard deviation in GLS measurements (3.47 vs. 4.98). Conclusions: The AI-based CMR feature tracking technique provides accurate and reproducible GLS measurements, showing strong agreement with echocardiographic speckle tracking-based GLS. However, the higher error rates in AS patients compared to controls underscore the need for more advanced AI algorithms to improve performance in cardiac pathology.

Figures