Abstract
2235
Objectives Our aim was to correct 18FDG-PET images for Partial Volume (PV) using a modified Recovery Coefficient approach (RCV,C), based on lesion local contrast (C) and functional volume (V), automatically delineated with an adaptative thresholding approach (Vauclin 2009).
Methods Modified Jaszczak and thoracic phantoms (Data Spectrum) containing respectively 9 and 17 spheres (0.40 to 97.8 mL) filled with various local sphere-to-background contrasts (1.7 to 22.9 and 2.5 to 8.5 resp.) were acquired on a Biograph PET/CT (Siemens). RCV,C was compared to conventional RCV, pixel-by-pixel (PbP, Müller-Gärtner 1992) and geometry transfer matrix (GTM, Rousset 1998) methods. Contrary to RCV,C and RCV, PbP and GTM are based on the lesion volume defined on anatomical images. RCV,C and RCV were calibrated from 10 min acquisitions of the Jaszczak phantom. On the thoracic phantom data, we evaluated the 4 correction methods, comparing by a paired t-test, the error of the corrected activity concentration to the true activity concentration.
Results For all spheres and contrasts, the mean absolute error with RCV,C is statistically lower than with PbP and GTM (14.2±20.9 vs 19.1±11.7 and 22.4±15.3% resp.) and slightly but significantly lower than with RCV, (15.9±25.2%) both. For small spheres (V<2mL), PbP yield errors comparable to GTM (33.8±11.3 vs 38.7±19.1%, NSS) and statistically lower than RCV,C and RCV (34.3±33.5 and 37.6±42.2% resp.). For spheres >2mL, RCV,C remains statistically better compared to the other methods (7.1±4.5 vs 8.3±5.7, 14.0±6.1 and 16.6±8.1% for RCV, PbP and GTM resp.). Results for spheres with C<5, shows the interest of integrating the lesion contrast into the RCV,C model (12.7±16.8 vs 17.1±27.3% for RCV).
Conclusions PV correction with RCV,C is statistically better than with RCV, GTM and PbP excepting for spheres <2mL for which functional volume can be overestimated by our automatic segmentation tool