A new scatter compensation method for Ga-67 imaging using artificial neural networks

Abstract

A new scatter correction method for Ga-67 based on artificial neural networks (ANN) with error back-propagation was designed and evaluated. The ANN consisted of a 37-node input layer (37 energy channels in the range 60-370 keV), an 18-node hidden layer, and a 3-node output layer to estimate the scatter-free distribution in the 93-, 185-, and 300-keV photopeaks. Two separate activity and attenuation distribution sets, based on a segmented realistic anthropomorphic torso phantom, were simulated. The first set was used for ANN learning and the second to evaluate the scatter correction. The authors' Monte Carlo simulation modeled all photon interactions in the patient, collimator, and detector. Interactions simulated in the collimator included Compton and coherent scatter and photoelectric absorption with forced production of lead K-shell X-rays. Ninety very high count projections were simulated and used as a basis for generating 15 Poisson noise realizations for each angle; noise levels were characteristic of 72-h post-injection Ga-67 studies. The energy window images (WIN) used clinically were also generated for comparison. Bias and variance were computed with respect to the primary distributions over reconstructed volumes of interest in the lungs, abdomen, liver, and tumors. ANN overall bias in all structures was less than 16% (8% in the abdomen) as compared to 85% with WIN. The variance of the activity estimates was systematically greater with WIN than ANN. ANN is a promising approach to scatter correction in Ga-67 studies.