The Impact of Image Reconstruction Bias on PET/CT ⁹⁰Y Dosimetry After Radioembolization

TO THE EDITOR: We read with great interest the paper entitled “The Impact of Image Reconstruction Bias on PET/CT ⁹⁰Y Dosimetry After Radioembolization” (1). In this paper, Tapp et al. showed that commercial software that truncates the negative pixels (resulting from random-coincidence correction) can display a significant positive bias in ⁹⁰Y PET imaging. We would like to bring some additional information to the attention of readers and PET system manufacturers.

We were also confronted by this problem while developing a correction method for the spurious-coincidence contamination occurring in ⁸⁶Y PET imaging (2): directly subtracting the estimated spurious-coincidence sinogram from the 511–511 keV true-coincidence sinogram (i.e., prompt minus delayed coincidences) resulted in numerous negative pixel values due to the high fraction of spurious coincidences generated by the multiple γ cascades present in ⁸⁶Y decay (3). Afterward, truncating these negative pixels before ordered-subsets expectation maximization reconstruction ended up in significant bias. Additional smoothing of the spurious-coincidence sinogram only partially improved the reconstruction.

Initially, we tried to take into account the estimation of spurious coincidences by adding it to the projection estimate in the denominator of the iteration step, which can be schematically written aswhere Aⁿ is the activity image estimate at step n, I the identity image, c the projection matrix, S the measured true-coincidence sinogram, and S_sc the estimation of the spurious-coincidence sinogram. The operation +,— is performed ray-sum by ray-sum, and the operation × is performed voxel by voxel. Although this method preserves the reconstructed voxel positivity in an elegant, natural way, we observed that Algorithm 1 no longer correctly converges when the estimated term S_sc becomes too large (data not published). In ⁸⁶Y PET imaging, this was especially the case for corpulent patients. This method is currently implemented in the Gemini TF PET system (Philips) for correction of scatter and random coincidences (4,5). Care should thus be taken when imaging low-⁹⁰Y specific activity with this lutetium yttrium oxyorthosilicate–based system (6).

Finally, we decided to remove the negative pixels from the subtracted sinogram by transferring to them an appropriate number of counts from neighboring positive pixels (a detailed description of the method has been published (2)). The rationale of this strategy is that Poisson noise is characterized mainly by high-spatial-frequency positive–negative fluctuations. This transfer of counts was performed in a special way that avoids artifact generation in the reconstructed image. Phantom and patient studies showed that this method prevents bias in ⁸⁶Y PET imaging (2). The method could also be evaluated in ⁹⁰Y imaging with PET systems, allowing separated prompt- and random-coincidence acquisitions such as the one used by Tapp et al. (1).

• © 2015 by the Society of Nuclear Medicine and Molecular Imaging, Inc.