Program Information
Real-Time Prediction in Free-Breathing Perfusion MRI
H Song1 , W Liu2 , D Ruan2,3*, S Jung4 , M Gach1,5 , (1) Department of Radiology, University of Pittsburgh, Pittsburgh, PA, (2) Department of Bioengineering, UCLA, Los Angeles, CA, (3) Department of Radiation Oncology, UCLA, Los Angeles, CA, (4) Department of Statistics, University of Pittsburgh, Pittsburgh, PA, (5) Department of Bioengineering, University of Pittsburgh, Pittsburgh, PA
Presentations
MO-G-18C-5 Monday 4:30PM - 6:00PM Room: 18CPurpose: The aim is to minimize frame-wise difference errors caused by respiratory motion and eliminate the need for breath-holds in magnetic resonance imaging (MRI) sequences with long acquisitions and repeat times (TRs). The technique is being applied to perfusion MRI using arterial spin labeling (ASL).
Methods: Respiratory motion prediction (RMP) using navigator echoes was implemented in ASL. A least-square method was used to extract the respiratory motion information from the 1D navigator. A generalized artificial neutral network (ANN) with three layers was developed to simultaneously predict 10 time points forward in time and correct for respiratory motion during MRI acquisition. During the training phase, the parameters of the ANN were optimized to minimize the aggregated prediction error based on acquired navigator data. During real-time prediction, the trained ANN was applied to the most recent estimated displacement trajectory to determine in real-time the amount of spatial adjustment required for each slice excitation during 2D MRI acquisitions.
Results: The respiratory motion information extracted from the least-square method can accurately represent the navigator profiles, with a normalized chi-square value of 0.037± 0.015 across the training phase. During the 60-second training phase, the ANN successfully learned the respiratory motion pattern from the navigator training data. During real-time prediction, the ANN received displacement estimates and predicted the motion in the continuum of a 1.0 s prediction window. The ANN prediction was able to provide corrections for different respiratory states (i.e., inhalation/exhalation) during real-time scanning with a mean absolute error of < 1.8 mm.
Conclusion: A new technique enabling free-breathing acquisition during MRI is being developed. A generalized ANN development has demonstrated its efficacy in predicting a continuum of motion profile for volumetric imaging based on navigator inputs. Future work will enhance the robustness of ANN and verify its effectiveness with human subjects.
Funding Support, Disclosures, and Conflict of Interest: Research supported by National Institutes of Health National Cancer Institute Grant R01 CA159471-01
Contact Email: