Deep-learning-based automatic patient motion correction for 18F-flurpiridaz PET-MPI

Accurate quantification of myocardial blood flow (MBF) and flow reserve (MFR) myocardial perfusion imaging (MPI) is critically dependent on effective inter-frame motion correction (MC) in the dynamic stress and rest scans. Manual frameby- frame correction is both time-consuming and prone to significant interobserver variability. Methods: We propose a deep learning (DL) framework for MC of 18F-flurpiridaz PET-MPI using 3D Resnet based architecture with added positional embeddings. Positional embedding gives the network context regarding the frame number, adding temporal dimension to the network. We trained the model on 116 stress and rest studies from the Phase III Flurpiridaz trial (NCT01347710). All patient scans were dynamic, and each frame was treated as an independent training and test data point. Manual MC was performed by two experienced operators with clinical software and their corrections served as ground truth data for training the network. The output of the model was the predicted translation vector (tx, ty, tz) to correct for motion. The model's performance was assessed through inference on a separate unseen test dataset comprising 117 patient scans. Results: DL MC takes 1-2 seconds for inference per scan which is significantly faster than manual MC which takes about 10 minutes per scan with 26 frames with comparable results. The root mean square error between the predicted and targets of the test set was 1.2mm, 1.04mm and 2.28mm in x, y and z directions respectively. Conclusion: DL model trained on user MC vectors provides faster method for automatic inter-frame MC for dynamic PET-MPI 18F-flurpiridaz and may automate and objectify the analysis of the myocardial blood flow.