Performance of maximum likelihood based image reconstruction for quantitative imaging

Potential improvements in accuracy and precision from a maximum-likelihood-based iterative image reconstruction algorithm (ML) compared to filtered backprojection (FBP) are examined. The contrast recovery coefficient (CRC) and noise-to-signal ratio (N/S) from ML images are compared to those from FBP images as a function of total image counts, resolution, region-of-interest (ROI) size, object contrast, and the incorporation of a positron emission tomography (PET) attenuation correction. The results are expressed in terms of relative ML to FBP accuracy and precision measured as ratios of CRC and N/S, respectively, because these ratios are found to be independent of total image counts. If resolution is matched based on a mean image resolution, for ML relative to FBP (1) large decreases (>50%) in N/S only exist for very small ROIs in high-resolution images; (2) the possible improvement in N/S varies inversely with object contrast, reaching a maximum for large interior holes with signal =0; and (3) compared to images without attenuation the incorporation of attenuation has little effect on the relative accuracy and precision performance of ML compared to FBP. A preliminary result from FBP with a Metz reconstruction filter demonstrates that modified, linear algorithms can achieve similar N/S improvements to those obtained using ML.<>