Spectral analysis using regularized non-negative least-squares estimation

The implementation of spectral analysis techniques involves solving a highly underdetermined linear system equation and is prone to the effect of measurement noise. We propose to use a regularized non-negative least-square estimator to stabilize the implementation of the technique. We introduce a penalty term in our formulation of the objective function to discourage disparities in tracer kinetics between neighboring pixels and use an iterative method to impose positivity constraints. We show results from analysis of FDG thorax images of patients suspected to have cancers and summarize our findings.