P14483D shape assessment from 2D echocardiography using machine learning

Abstract

In the context of cardiac remodelling, shape assessment of the heart is important for diagnosis and follow up. Despite offering a full view of the heart, 3D imaging presents some disadvantages, such as the poor acoustic window of 3D echography, ionisation in the case of Computed Tomography, and the cost in the case of Magnetic Resonance. Given that 2D echocardiographic data are widely available, we investigated the feasibility of machine learning for assessing the 3D left ventricular (LV) end-diastolic shape from 2D measurements enriched with clinical information. We used a dataset consisting of 116 preadolescents, 45 with intrauterine growth restriction (IUGR) and 71 controls. A complete 2D echocardiographic study was performed and 3D LV shape was extracted using 3D echocardiography. The measurements related to the LV shape were long axis, end-diastolic volume, basal diameter and internal dimension. As a clinical variable, the fact of having IUGR or not during fetal life was used. A statistical shape model was computed from the 3D echocardiography shapes through Principal Component Analysis (PCA), leading to a template shape and deformations from that template. In the IUGR dataset, most of the shape variation was explained by (1) overall size variability, (2) radial scaling, (3) an inclination of the apex with respect to the base. Those shape variations can be seen in the Figure. Linear Regression was performed to predict each deformation coefficient from the available 2D measurements. We could predict most of the overall size (R2 =0.75, p value <1e-5) and some of the radial scaling (R2 = 0.3, p value<1e-5) while the other deformations could not be predicted. Introducing the IUGR label to the regression improved the quality of the radial scaling prediction to R2=0.4. This confirms that IUGR is related to LV sphericity, and that clinical and functional data can indirectly assess the 3D ventricular shape. The effect of (Gaussian) noise on the results was also investigated using a synthetic dataset, which includes non-symmetrical bulging of the septal wall. As with the real dataset, we could not recover the asymmetrical modes even with total absence of noise. Comparing the results of the size and radial scaling R2 coefficients, we estimated the level of noise in the real dataset to be around 10%. The plot in the figure shows the effect of noise in the regression quality. Adding new 2D measurements that are not in the guidelines allowed us to recover the asymmetrical shape. In conclusion, using only the recommended 2D measurements limited ventricular shape prediction to the symmetric components. The prediction deteriorated with increased measurement noise. However, adding extra measurements, or relevant clinical information, some asymmetrical components can be regressed. We believe that this work is a promising first step towards deriving more complete shape measurements from routinely available 2D data. (Figure Presented).