Registration-weighted motion correction for PET

Abstract

Patient motion during a positron emission tomography (PET) scan can lead to significant resolution loss. Two of the main motion correction techniques employed to ameliorate the loss of image quality due to respiratory motion in the torso are postreconstruction registration (PRR) and motion-compensated image reconstruction (MCIR). In this study, the authors investigated whether versions of these methods that utilize registration-based weighting of the constituent respiratory gated data offer any advantage over the standard versions that use equal weighting. The registration-based weights were designed to penalize gates that were poorly registered to the reference gate. SimSET was used to simulate data from the NCAT phantom with six lesions added in the lung and liver. Images were reconstructed using registration-weighted PRR and MCIR algorithms, where the registration weighting was based on the mutual information (MI) of each registered gate and the reference gate. More specifically, relative to equal weighting, for which the weight for each gate is the inverse of the number of gates, the weights were increased for MI greater than the average MI and reduced for gates with MI less than the average MI. A scale factor was used to increase the range of the weights, and PRR and MCIR images were produced for a range of scale factor values. At the optimal values of the scale factor, registration-weighted PRR produced significantly higher contrast-to-noise ratio (CNR) for each lesion than PRR (p < 0.001), with average lesion CNR increasing significantly from (2.10 ± 0.05) to (2.70 ± 0.06) for 3 mm postsmoothing (p < 0.001) and from (2.03 ± 0.06) to (2.77 ± 0.05) for 6 mm postsmoothing (p < 0.001). Likewise, for MCIR registration weighting significantly increased the average CNR from (2.38 ± 0.04) to (2.62 ± 0.07) for 3 mm postsmoothing (p < 0.001) and from (2.56 ± 0.05) to (2.84 ± 0.08) for 6 mm postsmoothing (p < 0.001). These gains in lesion CNR were obtained despite corresponding reductions in signal-to-noise ratio, as expected from the use of unequal weighting of gated data with comparable variance. Registration weighting can significantly improve lesion CNR in motion corrected images, especially for PRR.