Prevalence of patient motion in dynamic PET

Objectives Quantification of physiologic function, such as myocardial perfusion, is achievable with dynamic PET imaging and is of increasing clinical importance. However, image processing using either region of interest derived functions or image decomposition assumes that the patient is stationary throughout the dynamic image sequence. Patient motion can lead to quantification artifacts and misdiagnosis. The goal of this project was to measure the prevalence of motion in a typical clinical population and to develop an automated method for detection of the patient motion.

Methods Patients (n=236, 120 Male, 63.6±11.4 yo) that were referred for routine myocardial perfusion imaging were imaged using 82Rb PET/CT at rest and pharmacologic stress with a dynamic imaging protocol (15 variable length time frames over 10 min). All rest and stress images (n=472 total) were manually reviewed by a single observer for the presence of patient motion. Any image having >6 mm shift over the dynamic sequence was classified as having motion. The images were also processed using an in-house software package for myocardial blood flow quantification which includes a new patient motion warning. Sequential images time frames, in the last 8 min of the scan (post blood peak activity), were shifted with respect to each other in three dimensions and the cross-correlation vs. shift was recorded. The shift with highest correlation was assumed to correspond to motion between frames, and total motion >4 mm was assumed significant. The agreement between software detection and manual detection of motion was compared.

Results Of the 472 images, 134 (29%) had significant motion as determined by the observer. Of these images 126 (94.0%) were correctly identified by the software. Of the images that were not identified correctly, motion was gradual over several time frames.

Conclusions Patient motion is common in a clinical setting and the need for motion correction exists. Cross-correlation between neighboring time frames can be used to detect patient motion