Quantification Performance of a Gradual Point Spread Function Reconstruction Approach for Task-Based Optimization of Positron Emission Tomography Data

Abstract PURPOSE This study proposes a novel Maximum Likelihood Expectation Maximization (MLEM) based positron emission tomography (PET) image reconstruction approach, designed to provide improved signal recovery, specifically in small structures. This proposed technique, termed GPSF-MLEM, utilizes a gradually decreasing full-width-half-maximum (FWHM) point spread function (PSF), designed to yield increased spatial details in progressing iterations when measured and estimated projection data are in closer consensus. METHODS GPSF-MLEM was compared with the analytical filtered back projection (FBP) and the iterative reconstruction techniques of standard MLEM and PSF incorporated into MLEM (PSF-MLEM). Simulated phantoms as well as cylindrical and brain phantoms were used to evaluate performance, contrast recovery and for visual assessment of image quality. The performance of GPSF-MLEM and PSF-MLEM with various FWHM kernel sizes was also investigated. This work presents evidence that the choice of reconstruction technique and parameters for the best quantification of reconstructed PET images should be based on the purpose (task-based optimization). RESULTS Quantitatively, across all reconstructions tested, the image intensity estimated using GPSF-MLEM terminated at an intermediate FWHM was in closest agreement to the raw image intensity in the simulated phantom. Qualitative analysis of reconstructed images with physical phantoms suggested that irrespective of the FWHM at the terminal iteration, GPSF-MLEM provided robust image contrast and increased suppression of background hotspot artifacts compared to MLEM and PSF-MLEM. Further, across all FWHM combinations tested, GPSF-MLEM yielded fewer artifacts inherent to PSF-MLEM images reconstructed with an overestimated FWHM. CONCLUSIONS Emphasizing common task scenarios frequently encountered in medical imaging: A) GPSF-MLEM is recommended for tasks requiring a low background noise; B) PSF-MLEM is recommended for tasks that require good image segmentation; C) GPSF-MLEM with a larger terminal FWHM is recommended for obtaining high contrast recovery in small image structures.