Enhancing 3D heart model reconstruction via cardiac MRI interslice labels interpolation

Visualization of the heart using 3D models is important for analyzing cardiac structure and conducting comprehensive cardiac disease studies. Traditional cardiac magnetic resonance imaging (CMRI) 3D reconstruction relies on 2D expert-traced labels from slices, but gaps between slices limit resolution. This resolution could be improved by increasing the number of CMRI scans. However, doing so is costly, requires long acquisition times, and demands high patient cooperation, as scans are typically acquired during breath holds. In this study, we developed an image processing algorithm to interpolate between labeled slices, which were then used to generate enhanced 3D models. Short-axis (SAX) CMRI scans with an axial resolution of 5 mm were obtained from three subjects and labeled by experts at end-diastole (ED) and end-systole (ES) phases. On the ED and ES images, the structures of the left ventricle (LV), left myocardium (LM), right ventricle (RV), right myocardium (RM), and cardiac adipose tissue (CAT) were labeled on the slices. Ground truth labels were taken from mid-slices between the existing labeled slices. The algorithm was validated using the Dice similarity coefficient (DC) and the Hausdorff distance (HD) as performance metrics. The global mean across all cardiac structures was 0.81 for DC and 4.56 for HD. The best-performing structure was the LV, with a DC of 0.88 and an HD of 1.77. These results suggest the potential of the interslice label interpolation algorithm for improving 3D heart model reconstruction. The algorithm was first used to interpolate between the 5 mm slices, generating labeled slices every 2.5 mm. The original labeled slices were then interpolated again to achieve an inter-slice resolution of 1.25 mm. Qualitative visualization of the reconstructed 3D models indicates improved anatomical representation.