Fully automated artificial intelligence-based cine cardiac magnetic resonance image quantification

Abstract

Abstract Background Analyzing cine cardiac magnetic resonance (CMR) images is a task that demands substantial experience and time, and it is subject to both inter- and intra-observer variability. Existing automated tools in the clinics for image analysis often necessitate substantial editing to enable accurate quantification. Objective To tackle this issue, in this study, we developed and validated a fully automated artificial intelligence (AI)-based system for the analysis of cine CMR images, aiming to address these challenges. Method In this study, 1,038 patients with CMR images were enrolled. Out of these, 350 CMR images, along with corresponding segmentations for the left ventricle myocardium (LVM) and the left- and right ventricles (LV and RV, respectively), sourced from open access data from 5 centers and databases (with different cardiac diseases), were utilized for model development. The remaining images from two local registries were used to evaluate the model. A 2D U-Net network with deep supervision, enhanced with various image augmentation techniques, was developed for automatic segmentation. The developed model was subsequently utilized to segment the test dataset of 4D (3D+time) short-axis cine CMR images. End-diastolic volumes (EDV) and end-systolic volumes (ESV) were automatically detected from these segmentations and evaluated against the ground truth calculations performed by two experienced physicians using commercially validated software. Results In the test set, Dice scores of 0.84, 0.90, and 0.93 were achieved for the RV, LV, and LVM, respectively. The mean percentage error for EDV was 0.6% for LV and -6.32% for RV. Regarding ESV, the mean percentage error was 2.95% for LV and -6.32% for RV. A -6.95% error was achieved for LV mass calculation. The stroke volumes for LV and RV were calculated with a mean percent error of -2.52% and -8.43%, respectively. We achieved a mean percent error of -3.22% for LV and -1.64% for RV ejection fraction. Conclusion The fully automated AI-based quantitative analysis of cine cardiac MR images developed in this study demonstrated high agreement with the ground truth across large datasets, including those from external centers. This method can significantly reduce the time required for image analysis and enhance the accuracy and consistency of cardiac quantitative assessments in clinical settings.