Inter-Frame distance metric-based Auto-Inversion-Time prediction for Cardiac MR

Motivation: Optimal inversion time (TI) is critical for Late Gadolinium Enhancement (LGE) cardiac imaging but currently TI estimation done visually on TI-Scout which introduces variations based on operator skill. Goal(s): Completely automated TI prediction (Auto-TI) on post-contrast TI scout data opens access to cardiac imaging. Approach: We abstract cropped TI-Scout frame data into inter-frame statistical distance space (Hellinger and Jensen-Shannon distance metrics) and do optimal TI prediction using a deep-learning network operating on distance maps. Results: We obtained 82% accuracy in optimal TI prediction. We notice excellent performance in cases with motion, metal, pediatric and different cardiac views, even though sparse in training pool Impact: Abstraction of time-series based TI scout data to a metric space and consequent Auto-TI prediction offers resilience to artifacts with a smaller footprint model, reduces training data variety needs, and offers explainability for prediction