AI-based quantification of strain in all cardiac chambers

Background: Cardiac strain is a sensitive indicator of cardiac function and can detect early heart dysfunction before structural changes may occur. Despite its clinical value, strain analysis is not widely used in practice due to the time-consuming manual segmentation required. To address this, we developed an AI-based method for automatic segmentation of all four cardiac chambers and subsequent automated strain quantification, reducing reliance on manual contouring and improving clinical efficiency.

Methods: This study included 300 anonymized CMR scans comprising 70 healthy individuals and 230 patients with previous myocardial infarction. Cine long-axis 2-, 3- and 4-chamber views were manually segmented. Manual segmentation of all cardiac chambers was performed in 4CH and left ventricle and atria additionally in 2CH and 3CH views. Segmentation was performed in end-systole and end-diastole. The data set was divided into a training set (N=250) and a test set (N=50). 14 studies were excluded based on poor image quality or incorrect imaging planes. A convolutional neural network (U-Net architecture) was trained on 236 subjects. The test set included ten healthy subjects, and the rest was chosen randomly from the data sets including all chambers. A. The segmentation was integrated into the CMR analysis software Segment software (Medviso, AB, Lund, Sweden). After automatic AI-based segmentation, strain was quantified by using the strain tracking methods in the software. Tracking from end-diastole was propagated throughout the whole cardiac cycle.

Results: The automatic algorithm was successful in all, but one imaging-plane (1/150) and the tracking worked in all but one imaging-plane and these two datapoints were excluded from the analysis. Figure 1 shows scatter plots comparing strain derived from manual vs automated segmentation was excellent for left ventricle and atrial, and right ventricle. The performance for right atrial strain measures was good. Biases of strain were 0.1 ± 1.3% (SD) for left ventricle, 0.7 ± 1.8% for right ventricle, 0.02 ± 1.6% for left atria, and 0.7 ± 2.9% for right atria (Figure 2).

Conclusion: The results demonstrate that the automated segmentation tool in combination with the tracking method can be used to provide accurate and reproducible biatrial and -ventricular strain measurements with low bias and variability compared to manual delineations.

Figure 1: Strain derived from automatic segmentations versus strain derived from automated segmentation. Solid orange line indicates the line of identity (manual = automated), while the blue dotted line represents regression line.
Figure 2. Bland–Altman plots comparing strain derived from automated and manual segmentation, respectively.