Abstract
BACKGROUND: Measurements of cardiac size and function drive clinical decisions. Left ventricular (LV) metrics can be derived from cardiac MR images by delineating the blood pool and myocardium, by either drawing a rounded contour to approximate the compacted myocardial border, or by delineating the papillary muscles and trabeculae (trabecular segmentation). There is no consensus as to which is best, particularly in the emergent AI era. We developed machine-learning (ML) approaches for both and compared them for clinically important metrics (error rate, precision, and prognosis).
METHODS: Separate ML models were developed for rounded and trabecular segmentation, using U-net models trained on 1923 subjects (mixed pathology, multiple scanners, multiple centers). Blood and myocardial volumes for each segmentation method were compared on 4118 healthy UK biobank subjects. Model segmentation quality was evaluated subjectively on a real-world clinical dataset of 1594 consecutive CMR scans, with all scans included regardless of image quality and artifacts. Scan-rescan precision was measured on a multi-center, multi-disease dataset of 109 subjects scanned twice and compared to human performance. Finally, prognostication ability was evaluated on 1215 clinical patients, using a primary outcome of all-cause mortality and hospitalization with heart failure.
RESULTS: Error rates (where a human disagreed by >1 mL) were the same, occurring in 0.6% (184/29680) of images and 3.6% (60/1594) of patients. In health, the mean EF was 4% higher for trabecular vs rounded segmentation. On test-retest data, there was no difference between rounded and trabecular ML models for precision, apart from end-diastolic and end-systolic volume, which was better for rounded segmentations. ML rounded and trabecular precision exceeded clinician performance for EF. There were marginal differences in prognostication between rounded and trabecular models.
CONCLUSION: We developed an automated method for annotating papillary muscles and trabeculae from cardiac MR images with low error rates. We found higher precision than clinicians in ejection fraction. There was similar precision and prognostication to an ML rounded model with similarly low error rates. Findings support the feasibility of automated trabecular segmentation in clinical care and clinical trials.