Artificial neural network as a tool to compensate for scatter and attenuation in radionuclide imaging

This study investigates the ability of artificial neural networks (ANN) to simultaneously correct for attenuation and Compton scattering in scintigraphic imaging.

Three sets of experiments are conducted using images of radioactive sources with various shapes and distributions in a homogeneous medium. Numerical Monte Carlo simulations and physical phantom acquisitions of radioactive geometric sources provide the basic material for correction. Our method is based on the following assumptions: information needed to correct for scattering can be extracted from the energy spectrum at each pixel without any assumption concerning the source distribution, and two diametrically opposed energy spectrum acquisitions yield enough information on the source location in the diffusing medium for simultaneous correction for attenuation and scattering.

Qualitative and quantitative evaluations of scatter correction by ANN demonstrate its ability to perform scatter correction from the energy spectra observed in each pixel. By using the energy spectra of incident photons detected in two diametrically opposed images, multilayer neural networks are able to perform a proper restitution of projection images without any assumption on geometry or position of radioactive sources in simple geometric cases. ANN corrections compare favorably to those provided by five of the most popular methods. A satisfying correction of both scatter and attenuation is observed for a human pelvis scan obtained during routine clinical practice.

An ANN is an efficient tool for attenuation and Compton scattering in simple model cases. The results obtained for routine scintigrams in a much more complex situation are strong incentives for performing further studies.