Improved Image Quality Using Joint Image Reconstruction and Non-Local Means Filtering for Multi-Spectral SPECT

Abstract

Single-photon emission computed tomography (SPECT) is one of the major imaging modalities in medical imaging, including quantitative imaging for the evaluation of efficacy and toxicity in radionuclide therapy. Choosing optimal SPECT image reconstruction strategy for radionuclides with wide energy spectrum affects resulting image quality due to energy-dependent attenuation information in forward projection models and energy-dependent scatter information. A post-reconstruction filtering is also important to suppress noise propagated during reconstruction process.

Yttrium-90 (Y-90) is a commonly used radionuclide in targeted radionuclide therapy. Recently, bremsstrahlung in Y-90 has been successfully imaged for good quantification of radioactivity to predict therapy response more accurately. However, wide continuous energy spectrum of bremsstrahlung photons is challenging in Y-90 SPECT image reconstruction. Previously, forward projection models with narrow single-energy window were used for image reconstruction from a single acquisition energy window. We propose a new Y-90 SPECT joint image reconstruction method from multiple acquisitions windows, referred to as joint spectral reconstruction (JSR) using multi-energy window forward models. Our proposed method yielded significantly higher recovery coefficient and lower standard deviation than other methods that use a single acquisition window and single energy window for projection model with narrow and wide energy spectra.

We also investigated parameter selection methods for non-local mean (NLM) filter with SPECT. Self-weight estimation is an important factor to influence denoising performance of NLM. Recently introduced local James-Stein type center pixel weight method (LJS) outperformed other existing self-weight estimation methods in determining the contribution of the self-weight to NLM. However, the LJS method may result in excessively large self-weight estimates since no upper bound for self-weights was assumed. It also used relatively large local area for estimating self-weights, which may lead to strong bias. We propose novel local minimax self-weight estimation methods with direct bounds (LMM-DB) and re-parametrization (LMM-RP) using Baranchik’s minimax estimator. Our proposed methods yielded better bias-variance trade-off, higher peak signal-to-noise (PSNR) ratio, and less visual artifacts than the classical NLM method and the original LJS method. Our proposed methods also provide a heuristic way of choosing global smoothing parameters of NLM to yield PSNR values that are close to the optimal values without knowing the true image.