Introduction: Automation is critical for quantitative image analysis and quality control in large cardiac magnetic resonance (CMR) datasets. We developed a pipeline for processing short-axis cine images from the prospective multi-center NAKO Health Study.
Methods: A deep learning algorithm based on nnU-Net architecture was trained for semantic segmentation on an independent dataset of expert-annotated short-axis CMR cine images including standard contours for both ventricles. The trained model was applied to CMR data from the NAKO Health Study baseline cohort, comprising 29,491 examinations from individual participants. The model inference was used to calculate ventricular volumes at all time points and to subsequently determine parameters of left and right ventricular (LV, RV) morphofunction, including volumes at end-diastole and end-systole (EDV, ESV), and myocardial mass (LVM). A visual quality control for image and segmentation quality included cases with (1) outliers in individual morphofunctional parameters, also considering the difference between left and right ventricular stroke volumes (≥2.5 standard deviations from the cohort mean), or (2) outliers in timepoint differences between EDV and ESV, or (3) abnormal LV time-volume curves. The ratings used a five-point Likert scale (5: good, 4: acceptable, 3: moderate, 2: poor, 1: non-diagnostic).
Results: All 29,491 CMR examinations were processed. The outlier analysis flagged 5,180 (17.6%) cases, of which 1,936 (6.6%) received a moderate or lower rating for image or segmentation quality. Most of the latter were associated with outliers in morphofunctional parameters and abnormal time-volume curves. Excluding these cases, the automated segmentation produced the following mean values across all participants: LVEDV 141.5ml, LVESV 52.3ml, LVM 113.3g. Excluded cases showed significantly higher means by 4.6ml (3.3%) for LVEDV, 13.8ml (26.4%) for LVESV, and 8.1ml (7.2%) for LVM compared to the included cases (all p<0.001).
Conclusion: The presented pipeline enabled automated segmentation of a large CMR dataset and facilitated quality control.