Multi-centre validation of an automatic algorithm for fast 4D myocardial segmentation in cine CMR datasets.

AIMS: Quantitative analysis of cine cardiac magnetic resonance (CMR) images for the assessment of global left ventricular morphology and function remains a routine task in clinical cardiology practice. To date, this process requires user interaction and therefore prolongs the examination (i.e. cost) and introduces observer variability. In this study, we sought to validate the feasibility, accuracy, and time efficiency of a novel framework for automatic quantification of left ventricular global function in a clinical setting. METHODS AND RESULTS: Analyses of 318 CMR studies, acquired at the enrolment of patients in a multi-centre imaging trial (DOPPLER-CIP), were performed automatically, as well as manually. For comparative purposes, intra- and inter-observer variability was also assessed in a subset of patients. The extracted morphological and functional parameters were compared between both analyses, and time efficiency was evaluated. The automatic analysis was feasible in 95% of the cases (302/318) and showed a good agreement with manually derived reference measurements, with small biases and narrow limits of agreement particularly for end-diastolic volume (-4.08 ± 8.98 mL), end-systolic volume (1.18 ± 9.74 mL), and ejection fraction (-1.53 ± 4.93%). These results were comparable with the agreement between two independent observers. A complete automatic analysis took 5.61 ± 1.22 s, which is nearly 150 times faster than manual contouring (14 ± 2 min, P < 0.05). CONCLUSION: The proposed automatic framework provides a fast, robust, and accurate quantification of relevant left ventricular clinical indices in 'real-world' cine CMR images.