Global heart function: Automated characterization of cardiac spatiotemporal features from isotropic 4D whole thoracic automated cardiac MRI (AutoCMR)

For technical development and validation of the proposed pipeline, 80 patients with a variety of cardiomyopathies were scanned using a novel ungated, free-breathing, isotropic 4D whole thoracic automated cardiac MRI sequence (AutoCMR: GRE, FA=12deg, TR = 4ms, TE= 2.9ms, 1.6 mm isotropic resolution, 192x192x192 matrix, FOV=307x307x307mm³, 30 cardiac frames). With manual segmentations of 48 patients for the training cohort, 4 patients for the validation cohort, and 12 patients for the testing cohort, an adapted version of the Swin-UNETR segmentation algorithm was trained to produce automatic segmentations of the four cardiac chambers for each cardiac frame over each patient’s R-R interval. Automatic spatiotemporal features, which include the ejection fraction (EF), volume time delay (VTD), rate time delay (RTD), and ejection to filling fraction (EtoF) were calculated and visualized across 80 patients to investigate their distribution for further analysis. 

Results: In figure 1, we show the specific input and output of our pipeline, how our auto-segmentation model achieves a Dice score of 0.8566 ± 0405 and a Hausdorff distance of 12.5787 ± 4.7434 compared to manual segmentations for 12 patients of the testing cohort, and an example of volumetric change of the cardiac chambers over 30 cardiac frames. The Hausdorff distance specifically outperformed previous algorithms in the whole heart segmentation challenge. Figure 2 utilizes Bland-Altman plots to match the AutoCMR pipeline’s volumetric measurements with the standard clinical pipeline. Figure 3 demonstrates the distribution of the four spatiotemporal features across 80 patients, with EF averaging around 0.46, VTD averaging 25around 0.46 seconds, RTD averaging around 0.41 seconds, and EtoF averaging around –1.25.

Conclusion: Based on our novel imaging sequence and adapted auto-segmentation algorithm, we were able to non-invasively capture spatiotemporal features of the four chambers of the heart over time. The proposed pipeline has the potential to expand the research scope for spatiotemporal metrics and how they affect heart functions in patients with cardiac disease. Future studies will focus on how to leverage this pipeline for deeper phenotyping of cardiac function.

A) Auto-segmentation pipeline utilizing the adapted Swin-UNETR segmentation pipeline to create 6 different labels from 4D whole thoracic AutoCMR: Right atrium (RA), Right ventricle (RV), Left atrium (LV), Left ventricle (LV), Left Myocardium, and the Aorta. B) The total calculated volume of five chambers (RA, RV, LA, LV, and aorta) for one representative patient across 30 cardiac frames over the patient’s R-R interval. C) The performance metric of the segmentation algorithm on 12 patients in the testing, demonstrated using Dice score and Hausdorff distance

Bland-Altman plots comparing the volume in the four cardiac chambers using standard clinical methods versus using the AutoCMR pipeline. For each plot, the R-squared correlation coefficient is attached to the y-axis.

Four spatiotemporal features extracted from the automated segmentation of 4D whole heart MRI data in 80 patients. A) The ejection fraction is calculated by subtracting the minimum volume of from the maximum volume of one chamber over time and dividing it by the maximum volume. B) The volume time delay is calculated by measuring the time distance between the maximum volume and the minimum volume of one chamber. C) The rate time delay is calculated by measuring the time distance between the maximum filling rate and the maximum ejection rate of one chamber. D) The ejection to filling fraction is calculated by dividing the maximum ejection fraction by the maximum filling fraction of one chamber