Automated Segmentation and Annotation of Cardiac Diffusion Tensor Images

Motivation: Segmentation of cardiac diffusion tensor images (cDTI) is time-consuming and observer dependent thus increasing the overhead for broader research adoption. Standardized approaches are needed. Goal(s): To automate deep-learning segmentation in cardiac cDTI that match expert image segmentation. Approach: We trained three nnUNets with different cDTI parametric maps. Dice similarly coefficients (DSC) and Hausdorff Distances (HD) assessed performance. Results: High DSC scores (0.903) for LV, low HD for RVIPs (averages <3mm). No significant differences in mean diffusivity (MD) compared to the expert annotated data. Impact: Automated segmentations of cDTI enable streamlined analysis, reduce manual effort, and improve efficiency of cDTI research.