LV cos(α) curves showed significant differences between healthy patients and those with DCM, HCM, or MINF, with U values ≥ 26 and p<0.001 in comparisons.
Does an automated AI-based vector approach for CMR myocardial motion curves differentiate between healthy individuals and those with cardiomyopathies or myocardial infarction?
An automated AI-based approach to extract 1D myocardial motion direction and magnitude curves from CMR can significantly differentiate healthy individuals from those with various cardiomyopathies independent of ejection fraction.
Absolute Event Rate: 0% vs 0%
Abstract Background Cardiac magnetic resonance imaging (CMR) is the gold standard for analysis of myocardial function. Volumetric parameters like ejection fraction (EF) give a general overview of cardiac performance whereas myocardial strain allows detection of early dysfunction. However, both methods focus mainly on contraction from end-diastole to end-systole whereby information about intermediate patterns of contraction and relaxation are not accounted for. For temporal analysis and visualization of cardiac motion, we propose an automated vector-based approach. By integrating two AI models - one for supervised segmentation and another for self-supervised registration our method characterizes cardiac motion as an angle α between motion direction of the myocardium and a focus point over a full cardiac cycle. Purpose To characterize left (LV) and right ventricular (RV) myocardial motion using the cosine of α (cos(α)) as an indicator of direction, and the vector norm |v| as a measure of motion magnitude, and to investigate differences in healthy patients (NOR) and those with hypertrophic (HCM), dilated (DCM) or right heart cardiomyopathy, as well as myocardial infarction (MINF). Methods A biventricular segmentation model was trained on an open-source CMR dataset (1) to generate LV and RV myocardial masks, and a second deformable image registration CNN model was trained to generate dense vector fields ϕ for subsequent direction calculation. Based on another method (2), a direction module was used to calculate the relative direction of the motion, αi, from the displacement vector vi in relation to a focus point derived from the respective segmentation mask, for each voxel xi in the masked LV and RV. By anatomical mapping and spatial aggregation, a 1-dimensional (D) direction curve cos(α) is generated over a full cardiac cycle, as well as a magnitude curve |v| from the norm of the vector field ϕ. LV and RV motion and magnitude curves of an independent dataset (3) were evaluated. Results The model generated distinct LV and RV cos(α) motion curves, showing peak motion towards focus point in systole, reversal of direction while transitioning into diastole and bimodal course in diastole representing early ventricular relaxation and late filling by atrial contraction. Corresponding |v| curves depicted myocardial deformation magnitude across time. Extrema of cos(α) (LV cos(α) minimum NOR vs. DCM: U = 26, p0,001; NOR vs. HCM: U = 27, p0,001; NOR vs. MINF: U = 30, p0,001) and several other motion features (Figure 2B) showed significant differences between NOR and patients with disease independent of LV-EF. Conclusion Myocardial motion—a complex 4D deformation process—was reduced to two interpretable 1D curves describing direction and magnitude, who can complement traditional strain analysis. While general differences between cohorts were observed, more comprehensive datasets are required to assess the relevance of the proposed method for improved diagnosis.Schematic illustration of cos(a)NOR vs. DCM diagramm and tests
Kiekenap et al. (Thu,) reported a other. LV cos(α) curves showed significant differences between healthy patients and those with DCM, HCM, or MINF, with U values ≥ 26 and p<0.001 in comparisons.