Related Experiment Video
Updated: Aug 6, 2026
Three-Dimensional Phase Resolved Functional Lung Magnetic Resonance Imaging
Published on: June 21, 2024
A projection-domain deep learning approach for respiratory motion correction of myocardial perfusion imaging using a
Dylan J Malenfant1,2, Terrence D Ruddy1, R Glenn Wells1,2
1Division of Cardiology, Department of Medicine, University of Ottawa Heart Institute, Ottawa, Ontario, Canada.
Background:
Myocardial perfusion imagining (MPI) is a nuclear medicine technique used in the assessment of coronary artery disease. Differences in the perfusion of the myocardium at rest and stress and regions of perfusion defects can be indicators of various cardiac disease states. Respiratory motion (RM) can result in the degradation of image quality and the appearance of artificial regions of perfusion deficit. While RM correction is possible, many methods involve time-consuming calculation or additional equipment for motion tracking. This has limited the clinical uptake of such methods. This work presents a deep learning model for data-driven RM correction on dedicated cardiac SPECT scanners.
Purpose:
An AI-based method for estimating RM from gated projection data for MPI SPECT using a dedicated pinhole scanner was developed. Accuracy of these estimates and impact on image characteristics were assessed.
Methods:
RM motion parameters were calculated for rest and stress MPI scans using a minimized root mean squared (RMS) alignment of reconstructed respiratory gates for 90 patients supplemented with 33 NCAT simulated patients. Using these parameters, an AI network was trained to estimate motion based on gated projection data prior to reconstruction. Accuracy of AI predictions relative to RMS data were assessed for clinical scans and down-sampled data to simulate noisy acquisitions. Scans were reconstructed with AI motion correction (MC), RMS MC, and no MC. Anterior myocardial wall thickness and its contrast relative to the interior of the ventricle were calculated in each case.
Results:
AI motion predictions were accurate to within 1.86 ± 0.05 mm at clinical noise levels, increasing to 2.07 ± 0.07 mm at 1/8 signal. MC with AI was shown to provide a statistically significant increase anterior wall contrast and decrease measured wall thickness relative to no MC, with comparable results to RMS MC.
Conclusion:
Direct AI-based MC parameter estimation using gated projection data was shown to be feasible in clinical settings. Corrections based on AI estimates showed improved contrast and reduced myocardial wall thickness when incorporated into image reconstruction.
