The purpose of this study is to investigate how well model observer can correlate with human observer in the lesion detection and localization task when the location of lesion is uncertain in CT imaging. A 35 x 26 cm oblong-shaped water phantom was scanned with and without two cylindrical rods (3 mm and 5 mm diameters) to simulate lesions with - 15HU contrast. Scans were repeated 100 times with the rods and 100 times without for each of 4 dose levels. Signal and background images were generated by selecting ROIs with 128x128 pixels, with the location of signal in each ROI randomly distributed. Human observer studies were conducted as three medical physicists identified the presence or absence of lesion, indicated the lesion location in each image and scored confidence level with a 6-point scale. ROC curves were fitted and area under curve (AUC) was calculated. The same data set was also analyzed using a Channelized Hottelling model observer with Gabor channels. Internal noise was added to the test variables for model observer study. AUC of ROC and LROC curves were calculated using non-parametric approach. The performance of human observer and model observer was compared. The Peason's product-moment correlation coefficients were 0.994 and 0.998 for 3mm and 5mm diameter lesions in ROC analysis and 0.987 and 0.999 in LROC analysis, indicating that model observer performance was highly correlated with the human observer performance for different size of lesions and different dose levels when signal location is uncertain. These results provide the potential of using model observer that correlates with human observer to assess CT image quality, optimize scanning protocol and reduce radiation dose.