Computer-aided lesion detection with statistical model-based features in PET images

Positron emission tomography (PET) with the glucose analog [¹⁸F] fluorodeoxyglucose is proving to be useful in cancer diagnosis and treatment. However, as in all nuclear medicine imaging technologies, lesion detection with PET is often hindered by limited spatial resolution and low signal-to-noise ratios. Under such conditions, conventional diagnosis by visual inspection usually becomes difficult and potentially inaccurate. In this paper, we propose use of computer-aided lesion detection methods for PET imaging by applying a maximum likelihood ratio test and a composite hypothesis test, assuming that the mean positron emission rate is deterministic or random, respectively. In our approach, different statistical models characterizing the mean positron emission rate, the raw sinogram data and the filtered backprojection (FBP) reconstructed image are used to derive the test criteria. Three methods to estimate the unknown parameters of the test functions from observations are presented. The performance of one of the proposed methods is evaluated and compared with both simulated and experimental phantom data. In the preliminary trials, the methods detect correctly (with a high probability >0.9) lesions of diameter⩾15 mm with lesion-to-background contrast 1.1:1. Under the same conditions, the test lesion could not be detected by visual inspection alone in the images reconstructed by either the FBP or the maximum likelihood iterative algorithms. The methods may also be used for the objective assessment of the quality of images reconstructed from different algorithms