SU‐GG‐I‐136: Improving SNR by Using Super‐Resolution (SR) Incorporated Image Reconstruction in PET Imaging

Purpose: Two SR Implementations (original and new) have been recently proposed to improve PET image resolution and SNR (IEEE Med. Img., 2006, 25, p137–147 & J Nucl Med. 2007; 48:411P). These implementations however, are performed retrospectively following image reconstruction. The objective of this abstract is to propose a prospective SR‐Incorporated reconstruction (SRIR) method and compare it to native PET reconstruction (NR). Method and Materials: A 2D digital phantom consisting of 10⁷ counts and containing various spheres (3.6–32mm diameter) was simulated in a PET scanner. The digital phantom was forward‐projected and poisson noise was added to generate the corresponding sinogram. Four such simulations/sinograms were generated each time by shifting the phantom along X or Y axis by 1.8 mm. These sinograms were then processed using a SRIR algorithm (supporting document) to produce a new (New‐SRIR) and original (Original‐SRIR) SR image. The images were reconstructed in a 46 cm FOV using OSEM (21 subsets, 2 iterations). In addition, a 256*256 NR was also generated from the first sinogram using the same FOV. The three images (Original‐SRIR, New‐SRIR, NR) were compared using SNR and line profiles drawn across the two smallest spheres. To assess noise content, a 2‐D FFT was also applied to a square ROI drawn in the background of the three images. Results: Line profiles of all images showed similar resolution and contrast. However, the original and new‐SRIR showed on average 17.5 and 10.3% higher SNR compared to NR respectively. The Original‐SRIR had a 6.5% higher SNR versus new‐SRIR. FFT analysis indicated that NR contains more high‐frequency noise compared to the two SRIRs. Noise content in the original‐SRIR was less compared to the new‐SRIR. Conclusion: SRIR processing has similar resolution and contrast but higher SNR compared to NR of PET images.