Region-based maximum likelihood reconstruction in positron emission tomography for quantitative oncological analysis

We propose a maximum likelihood (ML) reconstruction strategy for the quantification of uptake and volume of oncological lesions in Positron Emission Tomography. The ML is applied on volumetric regional basis functions identified on smooth standard clinical images (STD-CI). The lesion to be quantified is segmented with a k-means algorithm in three areas to take into account and recover the blurring introduced by the scanner Point Spread Function (PSF) and by the algorithm used to obtain STD-CI. For each of the three areas, a constant regional basis function initialized to the average inside the area is defined. The volume outside the lesion is handled as an unique basis function initialized to the activity spatial distribution estimated by STD-CI and is "frozen" in the reconstruction process except for a gain coefficient. The coefficients of the regional basis functions are iteratively estimated with an AWOSEM algorithm containing a 3D anisotropic and spatially variant model of the scanner PSF. The process of segmentation-reconstruction is iterated until the lesion segmentation converges. At each iteration the reconstructed volume is locally filtered with the scanner PSF before the segmentation step. We report results obtained on sphere phantom studies with activity contrasts of 7.5 and 4. The improvement in quantification obtained is remarkable thanks to the recovery of the scanner blur and to the possibility to locally better approach the maximum likelihood solution. The strategy appears also able to provide a reliable quantification of volume.