Assessing reproducibility and robustness in 2D CINE CMR volumetrics using simulation on 3D CINE AutoCMR data

Background: Conventional cardiac magnetic resonance (CMR) imaging is reliant on 2D slices to assess cardiac function^1,2. However, measurements based on 2D imaging are vulnerable to slice misalignment and varied slice thickness and slice gap values, leading to a decrease in reproducibility in the ejection fraction (EF), particularly in chambers other than the left ventricle (LV)^3–5. In this work, we utilize the 3D CINE AutoCMR data as a baseline to perform simulations of possible errors in 2D CINE CMR and assess the resulting impact on functional features of the heart.

Methods: For the simulation of 2D CINE CMR, 10 patients with cardiovascular diseases were scanned using the 3D CINE AutoCMR sequence⁶. The 4 chambers of the heart were manually/automatically segmented. Standard views were defined based on combinations of the centroids of the segmented chambers. The slices were perturbed by randomly moving the centroids by up to 7 pixels. We perturbed randomly perturbed slices 100 times in each example study to measure the potential distribution in EF error. The SAX stacks were also defined with slice thicknesses of 6, 8, and 10mm and slice gaps of 0 and 2mm (Slice thickness 8mm and slice gap 0mm was used as the standard) to simulate different slice gaps and thicknesses. For 4CH simulations, EF was calculated in all 4 chambers using the area-length method⁷. For SAX simulations, EF was calculated in the ventricles using the SAX stack method⁸. The distance between the centroid on the original slice and the centroid on the perturbed slice of each chamber were calculated to quantify the degree of perturbation (Figure 1).

Results: In figure 1, we show how a 3D manual segmentation of an AutoCMR scan from one patient was sliced into 2D planes and perturbed from the original landmarks, resulting in the changes in centroid distance and EF. Figure 2 is an amalgamation of all random simulations of perturbed 4CH and SAX slices in 10 patients, where the EF of the atria can have up to a 100-150% difference from that of the unperturbed 4CH slice. Additionally, the scatterplots also show a generally increasing relationship between the average centroid distance and the percent difference in EF of all four chambers using the area length method. Finally, figure 3 demonstrates that variations in slice gap and slice thickness affect RVEF more than LVEF, with up to a 40% and 15% absolute percent difference respectively.

Conclusion: The simulations performed in this work showed the extent to which common errors in conventional 2D CINE CMR acquisition, such as small variations in slice alignment and slice thickness/gap variation, can affect the resulting clinical values. The impact of slice variation is much more severe in the other chambers besides the LV. As such, to obtain more accurate functional measurements of the whole heart, we should make an advance towards 3D CINE CMR.

An example of the slice plane perturbation experiment. A patient’s 3D segmentation was sliced twice: once with the manually defined landmarks as anchors, and the other with shifted landmarks to replicate human variability in conventional 2D CMR image acquisition. The perturbation resulted in the offset of each chamber’s centroids as well as a change in the EF of the cardiac chambers
Histogram plots showing the distribution of percent difference of perturbed EF compared to the original EF in four chambers using SAX and 4CH data. Scatter plots with a fitted linear regression line depicting the increasing relationship between centroid distance and percent difference of perturbed EF compared to the original EF in four chambers using two different volume calculation methods. For the area-length method (4CH), the average centroid distance between four chambers was used. For the SAX stack method, the individual centroid distance of the RV and the LV were used.
Violin plot showing the distribution of percent difference in RV and LV ejection fraction when the slice thickness and slice gap were varied by different values.