Instance Segmentation Based approach for Robust automatic 3D Multi-View Planning for Cardiac MRI

Motivation: Manual planning for cardiac MRI (CMR) involving complex oblique views is time-consuming and introduces intra- and inter-technologist variability. Goal(s): To develop a fast and robust model that automatically predicts the standard CMR views from a 3D survey scan (SS) image. Approach: We use a deep learning (DL) based approach to map the 3D-SS image to four standard CMR using an instance segmentation strategy that identifies all planes followed by a finetuning strategy, which potentially offers improved robustness. Results: We achieved mean angulation and offset errors of 3.72±2.1 degrees and 2.36±1.8 mm, respectively across four standard CMR views averaged across 100 test subjects. Impact: Plane estimation as instance segmentation problem understands the spatial conformations of the planes and can adapt to various heart shapes and sizes and can go a long way in reducing intra- and inter-technologist variability and more importantly reducing scanning time.