Curation of Training Data for Supervised Deep Learning Reconstruction of Speech Real-Time MRI

Motivation: Supervised deep learning (DL) reconstruction requires large training sets and computationally demanding training. Real-time MRI offers large temporal redundancy which yields high reconstruction performance from training on a subset of frames. Goal(s): To develop a method for curating small DL training datasets that capture the variance of the entire training set and provide performance non-inferior to the entire training set, with reduced training time. Approach: We use clustering for each training speech task followed by selecting a fraction of each cluster to train U-Nets for reconstruction. Results: We achieve improved image quality metrics with comparable image quality metrics with 10x improved training time. Impact: By using curated training data based on identification and clustering of vocal tract postures, we demonstrate supervised DL-reconstruction of speech RT-MRI with 10-fold training time reduction and comparable NRMSE, PSNR, and SSIM. This may be generalized to other dynamic reconstructions.