Comparing lesion detection performance for PET image reconstruction algorithms: a case study

We present a comparison of the performance of filtered backprojection (FBP) and maximum a posteriori (MAP) reconstruction of PET images for the task of hot lesion detection. The comparison is performed on data generated by combining FDG chest scans of normal patients (i.e., without lesions) with pseudo-Poisson “lesion” data generated from appropriately scaled sinograms collected using a separately scanned 1.25 cm³ spherical source. Scaling factors were used to achieve approximately 2.5:1 lesion-to-background activity ratios. A total of 60 “abnormal” cases were generated from their normal counterparts. A 3D non-prewhitening (NPW) observer model based on a matched filter was used to test for the presence of the lesion in the vicinity of the known lesion location. ROC curves were generated for several choices of cut-off frequency for ramp-filtered FBP and the smoothing parameter for the MAP reconstructions. The NPW detector was matched to the algorithm and smoothing parameter in each case. Our experiments show that MAP reconstruction over a range of smoothing parameter values results in statistically significant improvements, compared to FBP, for the task of lesion detection using a NPW observer