Purpose: Sinogram-domain material decomposition (SDMD) is widely assumed to support more accurate multispectral CT imaging because it models beam hardening. We demonstrate that SDMD-based reconstruction from dual-energy CT sinograms results in biased monoenergetic CT images whereas statistical iterative reconstruction (SIR) results in unbiased images. The SDMD bias exhibits a systematic inverse dependence on imaging dose.
Methods: A 215 mm cylindrical water phantom containing samples of water, propanol, ethanol, and butanol, was fabricated and imaged on a Philips Big Bore CT scanner with 90 and 140 kVp known spectra. Raw transmission sinograms, with the vendorâ€™s beam-hardening corrections disabled, were exported. The (90/140 kVp) scans were performed at four dose levels: high (400 mAs/200 mAs), medium (200 /100), low (50 /25), and ultra-low (15/15). Monoenergetic CT numbers as a function of energy were estimated by a SIR engine (JSIR) operating jointly on the dual-energy transmission sinograms, which directly reconstructs the basis-material images, and filtered back projection of the SDMD basis-material sinograms. Both methods assumed a two-component basis-vector model of linear attenuation coefficients (LAC).
Results: Both high- and medium-dose scans resulted in bias less than 1% relative error for both JSIR and SDMD estimation. However, the low- and ultralow- dose scans, SDMD reconstruction revealed bias as large as 6% and 15%, respectively, compared to 1% and 2.5% for JSIR. Stochastic and simplified analytical simulations reveal that SDMD errors are not due to random noise amplification but arise from the fact that the log of the mean is a biased estimator of mean of logarithmically transformed noisy data.
Conclusion: Our work demonstrates that the SDMD reconstruction leads to inherently biased estimates of radiological quantities, and that these errors can be large in practical low-dose scans. In contrast, JSIR, which models the underlying nonlinear signal formation process, is immune to this class of systematic uncertainties.
Funding Support, Disclosures, and Conflict of Interest: This study was supported by NIH R01 CA 149305, NIH R01 CA 212638, and NIH 5T32EB01485505.