Neural Implicit Functions for 3D Shape Reconstruction from Standard Cardiovascular Magnetic Resonance Views.

In cardiovascular magnetic resonance (CMR), typical acquisitions often involve a limited number of short and long axis slices. However, reconstructing the 3D chambers is crucial for accurately quantifying heart geometry and assessing cardiac function. Neural Implicit Representations (NIR) learn implicit functions for anatomical shapes from sparse measurements by leveraging a learned continuous shape prior, without the need for high-resolution ground truth data. In this study, we utilized coronary computed tomography (CCTA) images to simulate CMR sparse label maps of two types: standard (10mm spaced short axis and 2 long axis slices) and 3-slice (single short and 2 long axis slices). Whole heart NIR reconstructions were compared to a Label Completion U-Net (LC-U-Net) network trained on the dense segmentations. The findings indicate that the LC-U-Net is not robust when tested with fewer slices than those used during training. In contrast, the NIR consistently achieved Dice scores above 0.9 for the left ventricle, left ventricle myocardium, and right ventricle labels, irrespective of changes in the training or test set. Predictions from standard views achieved average Dice scores across all labels of 0.84 +/- 0.03 and 0.88 +/- 0.03, when training on 3-slice and standard data respectively. In conclusion, this study presents promising results for 3D shape reconstruction invariant to slice position and orientation without requiring full resolution training data, offering a robust and accurate method for cardiac chamber reconstruction in CMR.