The effects of compensation for scatter, lead X-rays, and high-energy contamination on tumor detectability and activity estimation in Ga-67 imaging

Compton scatter, lead X-rays, and high-energy contamination are major factors affecting image quality in Ga-67 imaging. Scattered photons detected in one photopeak window include photons exiting the patient at energies within the photopeak, as well as higher energy photons which have interacted in the collimator and crystal and lost energy. Furthermore, lead X-rays can be detected in the main energy photopeak (93 keV). We have previously developed two energy-based methods, based on artificial neural networks (ANN) and on a generalized spectral (GS) approach to compensate for scatter, high-energy contamination, and lead X-rays in Ga-67 imaging. For comparison, we considered also the projections that would be acquired in the clinic using the optimal energy windows (WIN) we have reported previously for tumor detection and estimation tasks for the 93, 185, and 300 keV photopeaks. The aim of the present study is to evaluate under realistic conditions the impact of these phenomena and their compensation on tumor detection and estimation tasks in Ga-67 imaging. ANN and GS were compared on the basis of performance of a three-channel Hotelling observer (CHO), in detecting the presence of a spherical tumor of unknown size embedded in an anatomic background as well as on the basis of estimation of tumor activity. Projection datasets of spherical tumors ranging from 2 to 6 cm in diameter, located at several sites in an anthropomorphic torso phantom, were simulated using a Monte Carlo program that modeled all photon interactions in the patient as well as in the collimator and the detector for all decays between 91 and 888 keV. One hundred realistic noise realizations were generated from each very-low-noise simulated projection dataset. The presence of scatter degraded both CHO signal-to-noise ratio (SNR) and estimation accuracy. On average, the presence of scatter led to a 12% reduction in CHO SNR. Correcting for scatter further diminished CHO SNR but to a lesser extent with ANN (5% reduction) than with GS (12%). Both scatter corrections improved performance in activity estimation. ANN yielded better precision (1.8% relative standard deviation) than did GS (4%) but greater average bias (5.1% with ANN, 3.6% with GS).