Low latency in-line segmentation of cardiac structures for real-time cardiac MRI

Motivation: There is an unmet need for real-time segmentation of anatomic structures and interventional tools during MRI-guided cardiovascular interventions. Goal(s): To develop a low-latency method for left ventricle (LV) and myocardium (MYO) segmentations in real-time (RT) 2D cardiac MRI. Approach: We train a semantic segmentation network using short-axis end-diastolic (ED) and end-systolic (ES) images with reference segmentations. Results: We achieved 0.93 LV and 0.84 MYO Dice score with latency <10ms, which is 10-fold lower than state-of-the-art methods (nnUNET, 0.89 LV and 0.84 MYO Dice score) Impact: We demonstrate in-line low-latency segmentation with comparable LV and MYO Dice score to state-of-the-art-methods using a semantic segmentation model.