Automated Cine CMR Contouring and Strain Analysis in Coronary Artery Disease

Background: Cine CMR is the gold standard for quantification of cardiac chamber volumes and function. Myocardial strain can be calculated from standard cine images, providing added sensitivity for detecting functional impairment and added prognostic value over LVEF alone. Robust, automated strain analysis using artificial intelligence (AI) has the potential to reduce labor, expedite processing, and minimize variability – thereby facilitating incorporation into routine clinical care. This study aimed to validate a pre-trained AI tool for automated volumetric and strain measurement in a multicenter, multivendor cohort of patients with established CAD.

Methods: Patients with prior CAD enrolled in the DETERMINE (NCT00487279) and PRE-DETERMINE (NCT01114269) studies underwent clinical CMR at 66 U.S. sites using scanners from three vendors (GE, Siemens, Philips) (1). A pre-trained UNet-like convolutional neural network with 16 convolutional layers and residual learning units, implemented in Python using TensorFlow (MASS, version V2025-EXP, LUMC, Netherlands) (2), was used to generate automated contours from short-axis and 2- and 4-chamber cine images. This method was previously validated for volumetric and EF measurement and had successfully predicted adverse outcomes (2, 3). Derived parameters included left ventricular (LV) and left atrial (LA) end-diastolic volume (EDV), end-systolic volume (ESV), and ejection fraction (EF), along with strain measures: global longitudinal strain (GLS), global circumferential strain (GCS), and global radial strain (GRS) for the LV, and GLS for the LA. Manual contouring by expert readers using QMass MR version 7.5 (Medis, Leiden, the Netherlands) served as the reference. Spearman correlation assessed agreement between AI-derived and manual measurements.

Results: A total of 761 patients (mean age 63 years, 79% male) were included. The AI algorithm successfully generated LV contours in 93.6%, 97.1%, and 94.7% of 2-chamber, 4-chamber, and short-axis views based on expert review. Most contour errors were partial frame defects, defined as localized contouring errors where the AI boundary deviated or was incomplete. Correlations between AI-derived and manual LV EDV, ESV, and EF in 690 patients were 0.94, 0.93, and 0.85, respectively (all p < 0.001). For the LA (n = 731), correlations were 0.93, 0.91, and 0.79 (all p < 0.001) (Figure 1). LV strain correlated strongly with LVEF for GCS (r = -0.74), 2-chamber GLS (r = -0.70), 4-chamber GLS (r = -0.67), and with GRS (r = 0.54) (all p < 0.001). LAEF correlated with LA GLS in the 2-chamber (r = -0.69) and 4-chamber (r = 0.67) (all p < 0.001) (Figure 2).

Conclusion: Fully automated contouring and quantification of cardiac volumes and strain can be performed with high accuracy when compared to expert manual analysis Strong correlations highlight its potential for clinical care and risk assessment in CAD, though further validation against manual strain and outcomes is needed.

Figure 1. Spearman correlation between AI-derived volumes and EF vs. manual volumes and EF
Figure 2. Spearman correlation between AI-derived strain values vs. manual volumes and EF in A) left ventricle and B) left atrium