Optimization of super-resolution processing using incomplete image sets in PET imaging

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 (POVs). The number of low-resolution images used defines the processing time and memory storage necessary to generate the SR image. In this paper, the authors propose two optimized SR implementations (ISR-1 and ISR-2) that require only a subset of the low-resolution images (two sides and diagonal of the image matrix, respectively), thereby reducing the overall processing time and memory storage. In an NxN matrix of low-resolution images, ISR-1 would be generated using images from the two sides of the NxN matrix, while ISR-2 would be generated from images across the diagonal of the image matrix. The objective of this paper is to investigate whether the two proposed SR methods can achieve similar performance in contrast and signal-to-noise ratio (SNR) as the SR image generated from a complete set of low-resolution images (CSR) using simulation and experimental studies. A simulation, a point source, and a NEMA/IEC phantom study were conducted for this investigation. In each study, 4 (2x2) or 16 (4x4) low-resolution images were reconstructed from the same acquired data set while shifting the reconstruction grid to generate images from different POVs. SR processing was then applied in each study to combine all as well as two different subsets of the low-resolution images to generate the CSR, ISR-1, and ISR-2 images, respectively. For reference purpose, a native reconstruction (NR) image using the same matrix size as the three SR images was also generated. The resultant images (CSR, ISR-1, ISR-2, and NR) were then analyzed using visual inspection, line profiles, SNR plots, and background noise spectra. The simulation study showed that the contrast and the SNR difference between the two ISR images and the CSR image were on average 0.4% and 0.3%, respectively. Line profiles of the point source study showed that the three SR images exhibited similar signal amplitudes and FWHM. The NEMA/IEC study showed that the average difference in SNR among the three SR images was 2.1% with respect to one another and they contained similar noise structure. ISR-1 and ISR-2 can be used to replace CSR, thereby reducing the total SR processing time and memory storage while maintaining similar contrast, resolution, SNR, and noise structure.