Introduction: Myocardial motion estimation is critical for studying regional heart function and entails finding an optical flow representation between consecutive 3D frames of a 4D cine cardiac magnetic resonance imaging (MRI) dataset. Here we describe an unsupervised deep learning framework for deformable registration of 3D CMRI images.
Methods: Our method leverages the VoxelMorph framework, which uses a 3D image pair as input to a convolutional neural network (CNN). The CNN outputs a deformation field, which, along with one of the input 3D images, serves as input to a spatial transformer network that yields a warped image. Before registration, the input 3D images are corrected for slice misalignment by regressing the centers of the left ventricle (LV) blood-pool from all image slices, from which the myocardium (MC) and right ventricle (RV) were also segmented using a U-Net model. The manual segmentations of LV, MC, and RV at end-diastole and end-systole are used to train the U-Net model to segment the other frames in the dataset.
Results: We conducted experiments using 100 4D CMRI data (75: training, 10: validation and 15: testing). We used the registration deformation field to warp the manually segmented LV, MC and RV labels from end-systole to end-diastole and estimated their similarity. We achieved a mean Dice score of 91.01% (LV), 73.63% (MC), and 80.72% (RV) compared to 83.52%, 61.51% and 71.88%, respectively, before registration. The LV, MC and RV labels from the remaining cardiac phases were also warped to end-diastole and yielded a mean Dice score of 93.29% (LV), 79.19% (MC) and 86.32% (RV) compared to 88.78%, 73.28% and 81.14%, respectively, before registration.
Conclusion: We described a CNN-based deformable registration to estimate heart motion from cine cardiac MRI. We will use this technique to build dynamic patient-specific myocardial models with associated fiber architecture for biomechanical cardiac simulations.