Computerized detection of Low SNR cases in NSECT: An ROC-based sensitivity analysis

Abstract

Neutron stimulated emission computed tomography (NSECT) is an imaging technique that uses gamma energy spectra emitted from inelastic scattering of fast neutrons to extract quantitative elemental information from tissue. The NSECT acquisition system consists of a neutron source and one or more gamma detectors. For the NSECT system to be adequately sensitive to low elemental concentrations it is important to accurately extract the relevant gamma counts despite low signal to noise ratio (SNR) conditions. One technique to improve the sensitivity of the system and lower the minimum detectable level is to use computerized post processing of the gamma spectra. In this project, we describe a method of improving the sensitivity of the NSECT system through computerized post processing of the NSECT signal. The signal and noise in NSECT are photon-counting systems and hence are Poisson distributed. Modifying the Gaussian signal known exactly (SKE) case of signal detection theory to incorporate Poisson distributions, a likelihood based optimum detector was designed for each gamma detector in the NSECT acquisition system. This detector was implemented in MATLAB for the simulated iron concentrations and was followed by ROC analysis to study the detection sensitivity of the designed detector. In this project a GEANT4 simulation of a tissue sample with a 2 cm lesion (at a fixed location) was used to generate an NSECT spectrum from a single projection for different iron concentrations in the lesion. The iron concentration values were set to represent clinical liver iron overload. The gamma signal corresponding to iron and the background noise from Compton scattering of high-energy gamma photons were estimated using Poisson distributions. The results showed that for the simulated lesion position, the area under the curve (AUC) increased with increasing iron concentration, and 4 of the 6 gamma detectors were able to detect the lowest simulated iron concentration (1 mg/g). These results demonstrate that NSECT combined with computerized post processing has the potential to detect clinically relevant concentrations of iron to diagnose and quantify liver iron overload.