Implementation and optimization of a new Super-Resolution technique in PET imaging

Abstract

Super-Resolution (SR) techniques are used in PET imaging to generate a high-resolution image by combining multiple low-resolution images that have been acquired from different points of view (POV). In this paper, we propose a new implementation of the SR technique (NSR) whereby the required multiple low-resolution images are generated by shifting the reconstruction pixel grid during the image-reconstruction process rather than being acquired from different POV. In order to reduce the overall processing time and memory storage, we further propose two optimized SR implementations (NSR-O1 & NSR-O2) that require only a subset of the low resolution images (two sides & diagonal of the image matrix, respectively). The objective of this paper is to test the performances of the NSR, NSR-O1 & NSR-O2 implementations and compare them to the original SR implementation (OSR) using experimental studies. Materials and Methods A point source and a NEMA/IEC phantom study were conducted for this investigation. In each study, an OSR image (256×256) was generated by combining 16 (4×4) low-resolution images (64×64) that were reconstructed from 16 different data sets acquired from different POV. Furthermore, another set of 16 low-resolution images were reconstructed from the same (first) data set using different reconstruction POV to generate a NSR image (256×256). In addition, two different subsets of the second 16-image set (two sides & diagonal of the image matrix, respectively) were combined to generate the NSR-O1 and NSR-O2 images respectively. The 4 SR images (OSR, NSR, NSR-O1 & NSR-O2) were compared with one another with respect to contrast, resolution, noise and SNR. For reference purposes a comparison with a native reconstruction (NR) image using a high resolution pixel grid (256×256) was also performed. Results The point source study showed that the proposed NSR, NSR-O1 & NSR-O2 images exhibited identical contrast and resolution as the OSR image (0.5% and 1.2% difference on average, respectively). Comparisons between the SR and NR images for the point source study showed that the NR image exhibited an average 30% and 8% lower contrast and resolution respectively. The NEMA/IEC phantom study showed that the three NSR images exhibited similar noise structures as one another but different from the OSR image. The SNR of the three NSR images were similar (2.1% difference) but on average 22% lower than the OSR image largely due to an increase in background noise, while the NR image had an average of 14.5% lower SNR versus the three NSR images. Conclusion The NSR implementation can potentially replace the OSR approach in current PET scanners while maintaining similar contrast and resolution, but at a relatively lower SNR. This NSR implementation can be further optimized as NSR-O1 & NSR-O2 implementations by using only a subset of low-resolution images which can achieve similar image contrast, resolution and SNR but require less processing time and memory storage. A major advantage of the NSR versus OSR implementation is its shorter overall scan duration which results in an increase in scanner throughput and a reduction of patient motion.