A comparison of four image reconstruction algorithms for detection of small lesions in brain phantom

Abstract

We compared two fully three-dimensional (3-D) image reconstruction algorithms and two 3-D rebinning algorithms followed by reconstruction with a two-dimensional (2-D) filtered backprojection algorithm. The two 3-D image reconstruction algorithms were ordered subsets expectation maximization (3D-OSEM) and 3-D reprojection (3DRP). The two rebinning algorithms were Fourier rebinning (FORE) and single slice rebinning (SSRB). The 3-D projection data used for this work were acquired with a high-resolution PET scanner (MDAPET) with an intrinsic transaxial resolution of 2.8 mm. The scanner has 14 detector rings covering an axial field-of-view of 38.5 mm. We scanned three phantoms: (1) a uniform cylindrical phantom with inner diameter of 20.5 cm, (2) a 11.5-cm cylindrical phantom with four embedded small lesions with diameters of 3, 4, 5, and 6 mm, and (3) the 3-D Hoffman brain phantom with three embedded small lesion phantoms with diameters of 3, 5, and 8.6 mm. We evaluated the different reconstruction methods by comparing the noise variance of images, contrast recovery and contrast-noise trade-off, lesion detectability, and by visually inspecting images. We found that overall the 3D-OSEM algorithm followed by post filtering produced the best results. Even though the MDAPET camera has a relatively small maximum axial acceptance (/spl plusmn/5 deg), the 3DRP algorithm produced slightly better images compared to the faster 2-D rebinning methods.