Abstract
270
Objectives To develop a method for objective and unsupervised selection of TV hyperparameter for improved uniformity of the reconstructed myocardium with preserved spatial resolution and optimized the tradeoff between noise and bias in parallel-beam collimator SPECT myocardial perfusion imaging (MPI).
Methods We used NCAT digital thorax phantom containing a normal heart and simulated MPI SPECT data via Monte Carlo technique with 10% and 20% noise levels. The SPECT data were reconstructed using maximum a posteriori (MAP) ordered subset expectation maximization (OSEM) algorithm with total variation (TV) regularization (MAP-OSEM-TV). We employed three methods to estimate the optimal value of hyperparameter: (1) the highest curvature point on the L-curve (log of TV norm vs. log of loglikelihood); (2) a point on the bias-noise curve for the ROI located inside the myocardium region that corresponded to a minimum of square root of sum of squares of bias and noise; (3) a point on the bias-noise curve for the ROI located inside the lungs that corresponded to a minimum of square root of sum of squares of bias and noise.
Results The hyperparameter selected using Method 3 resulted in the reconstructed images with best myocardium uniformity and best contrast-to-noise ratio (defined as contrast between the hart wall and the ventricle over square root of sum of standard deviations in these regions). Method 2 resulted in the lowest bias accompanied by the highest noise and the largest non-uniformity in the myocardium.
Conclusions Method 3 resulted in best quality of the reconstructed myocardium and could be easily implemented and used for objective and unsupervised selection of the optimized hyperparameter in MPI SPECT reconstruction