Room: Stars at Night Ballroom 2-3
Purpose: Spectral CT is an emerging modality that permits material density distribution estimation. Model-based material decomposition (MBMD) permits one-step density estimation directly from spectral measurements allowing for arbitrary sampling patterns and effective beam-hardening artifact elimination. However, the relationship between reconstructed image properties and regularization can be complicated with object- and data-dependence. Typical exhaustive sweeps for parameter tuning are computationally expensive and can fail to generalize to arbitrary imaging conditions. In this work, we present a prediction framework for generalized local impulse response prediction in MBMD, facilitating prospective and robust regularization design.
Methods: An explicit expression for local impulse response in MBMD with a quadratic penalty is derived. Both iterative and fast Fourier approaches were implemented to compute these responses. Prediction accuracy is validated through comparison with direct local impulse response measurements for a kVp-switching dual-energy CT system and a water-iodine phantom. The generalized impulse response includes cross-talk terms between bases allowing quantification and regularization customization. A new quadratic cross-basis penalty design is proposed for cross-talk reduction.
Results: Comparisons between response predictions and measurements show near-perfect agreement over different regularization strengths. Impulse responses within material bases and cross-talks between bases are accurately quantified through the proposed prediction framework. A cross-basis quadratic penalty design using a local 4-neighborhood achieved over 90% reduction in the cross-talk response, while increasing the in-basis responses by 40%.
Conclusion: We have proposed and validated a generalized local impulse response predictor for MBMD that quantifies spatial resolution within a single material basis as well as cross-talk between bases. Since these predictors account for the dependence on material distribution in the object, spectral CT system characteristics (including system geometry, spectral combination, and decomposition material bases), and reconstruction regularization, this framework has the potential to facilitate robust and prospective regularization design and control of MBMD image properties.
Funding Support, Disclosures, and Conflict of Interest: This work was supported, in part, by NIH grants R01EB025470 and R21EB026849.