Joint Spectral Reconstruction for Y90 Bremsstrahlung SPECT using Guided Filtering

Objectives: Quantitative yttrium-90 (Y90) bremsstrahlung SPECT imaging in radionuclide therapy is challenging. Recently proposed joint spectral reconstruction (JSR) with wide energy windows yielded promising improvements in quantitative image qualities with low noise over single spectral reconstruction (SSR). In this work, based on our observations of the contrast and noise trade-offs over iterations, we propose guided filtering (GF)-aided reconstruction methods to reduce noise level while maintaining other image qualities and details.

Methods: Reconstructed images at early iterations often contain low noise and rough image details (e.g., low contrast), while reconstructed images at later iterations usually contain high noise and better image details. The guided filter yields the output by considering the content of a guidance image based on a local linear model and can be used as an edge-preserving smoothing operator. We conjecture that images at early iterations are perfect candidates to be guide images of GF and proposed GF-aided image reconstruction methods for both SSR and JSR in Y90 bremsstrahlung SPECT. Assuming that one cycle consists of k iterations, GF was performed every cycle (k = 4 in our case) using the guided filtered image at the previous cycle as a guide image. The output of GF was used for both an initial image for the next k iterations of 3D OS-EM as well as a guide image for the next cycle. Our proposed scheme was applied to both SSR (called SSR-GF) with a 105-135keV energy window and JSR (called JSR-GF) with six 30keV-width energy windows on 105-285keV with phantom simulation and experimental measurement using high-energy collimators on SPECT/CT. MC scatter correction was used for reconstructions with experimental measurement.

Results: Qualitatively, in both simulation and phantom experiment (2 to 100 mL spheres in warm background), our proposed SSR-GF and JSR-GF yielded significantly reduced noise levels compared to SSR and JSR. Quantitatively, for simulation, SSR-GF, JSR-GF yielded 126.5%, 71.5% higher contrast-to-noise ratio (CNR) than SSR, JSR, respectively. Note that JSR-GF still yielded 15.2% better CNR than SSR-GF. Similarly, in phantom measurements, SSR-GF, JSR-GF yielded 60.9%, 46.1% higher CNR than SSR, JSR, respectively. Note also that JSR-GF still yielded 31.2% better CNR than SSR-GF. Our proposed GF scheme yielded substantially lower root mean square error (RMSE) as follows: for simulation from 4.41 to 3.80, from 4.00 to 3.75 for SSR, JSR, respectively, and for measurement from 3.94 to 3.74, from 3.63 to 3.54 for SSR, JSR, respectively. For given average recovery coefficient, JSR-GF yielded the lowest noise level among all methods for both simulation and experiment.

Conclusions: We propose Y90 single and joint spectral reconstructions using guided filtering with reconstructed images at lower iterations as guidance images. Our proposed methods yielded the highest CNR, the lowest RMSE, the best averaged recovery coefficient - noise level trade-offs, and enhanced visual quality images. Research support: This work was supported in part by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2017R1D1A1B05035810) and in part by grant R01 EB022075, awarded by National Institute of Biomedical Imaging and Bioengineering, National Institute of Health, U.S. Department of Health and Human Services.