Image-based motion correction of the blood pool phase of dynamic PET data using blood pool isolation

476 Objectives Misalignment of the left ventricle (LV) blood pool images due to respiratory and patient motion can produce inaccurate estimates of myocardial blood flow (MBF) and coronary flow reserve (CFR) from dynamic PET myocardial perfusion images. The objective of this study was to evaluate an image-based automatic algorithm that detects and estimates respiratory and patient motion during the blood pool phase compared to expert visual assessment. Methods This study used 53 randomly selected patients referred for dynamic stress/rest Rb-82 PET imaging. Dynamic image series (16x5sec, 6x10sec, 3x20sec, 4x30s, 1x80s) were reconstructed using 3D-OSEM (24 subsets, 3 iterations) with point-spread-function (PSF) modeling, standard corrections including randoms, attenuation, scatter, and prompt gamma, and 7mm Gaussian post-filtering. LV myocardial surfaces were automatically generated from the tissue phase frames using the Corridor4DM software (INVIA). An automated algorithm then created an isolated LV blood pool image by identifying the right ventricle (RV) blood pool, LV blood pool, and tissue phase frames and then subtracting the summed RV blood pool image from the summed LV blood pool image. The algorithm then estimated the in-plane short-axis motion of the LV blood pool by finding the 2-D shifts which centered the LV blood pool within the LV endocardial surface. Detection and direction of motion by the automated algorithm were compared to visual assessment and manual motion correction by an expert user. Results Noticeable motion of pixel size 3.18mm or greater was detected in 60% of the studies by the automated algorithm compared to 49% of the studies by the expert. Compared to the expert, the automated algorithm had a sensitivity of 98%, a specificity of 76%, and an accuracy of 87% for the detection of motion. In studies with detected motion, the average amount of motion in any direction was 7.4mm ± 2.8mm for stress and 6.0mm ± 3.6mm for rest. The average motion shifts were in the craniocaudal direction, specifically 223 degrees for stress and 184 degrees for rest, with angles increasing counter-clockwise from the lateral direction in the short-axis plane. Figure shows stress study with LV blood pool shifted 3.5mm in the direction of 235 degrees before and after motion correction and affected vascular time activity curves. Conclusions Respiratory motion is problematic in the blood pool phase due to fast kinetics of the tracer and patient motion is problematic due to shifting of the patient between the blood pool and tissue phases. Motion was automatically detected with excellent sensitivity and good specificity. Automatic motion correction can be performed by applying interpolated motion shifts to each frame in the blood pool phase. $$graphic_AF4DE0DF-85D8-4F56-A3A6-3D86CA4999E8$$