Adaptive PSF Modeling for Enhanced Heterogeneity Quantification in Oncologic PET Imaging

Objectives Point-spread function (PSF) modeling, also known as resolution Modeling (RM), is a reconstruction-based partial volume correction (PVC) method. Unlike post-reconstruction PVC, it models resolution degradation phenomena within the reconstruction step. Past and ongoing studies have overwhelmingly focused on the impact of PSF modeling on common measures such as SUVmean and SUVmax. This study, by contrast, aims to investigate the impact of incorporating varying PSF kernels, enabling under- and overestimation of the true PSF, and study their effect on heterogeneity quantification.

Methods Images from twelve head and neck cancer patients with baseline 18F-FDG PET images (14.9±4.1 mCi) with lesions larger than 8mL (of varying heterogeneity extents) were considered. We subsequently used these images as the reference truth, and simulated clinically realistic data (100 noise realizations) for clinical scanner geometry. This incorporated normalization, attenuation (using patient’s CT image), decay, and resolution degrading phenomena (including photon non-collinearity, inter-crystal penetration and scattering). We introduced controllable scaling factors to generate 10 analytically modeled asymmetric blurring kernels (incorporating abovementioned physical phenomena with additional scaling in extent), including no/full PSF modeling, as well as 4 underestimated and 4 overestimated kernels. Twelve tumor contours were delineated by a trained nuclear medicine physician and used for threshold-based tumor segmentation. For our proposed investigation, we performed Haralick analysis, which has increasing utility in the field of heterogeneity quantification. We extracted the gray-level co-occurrence matrix, utilizing 64 gray-level quantization and a distance of 1 (which were seen to results in optimized performance). We computed a number of Haralick measures, including: (1) energy, (2) entropy, (3) correlation, (4) contrast, (5) variance, (6) sum mean, (7) agreement, (8) cluster tendency, (9) homogeneity, (10) inverse variance and (11) dissimilarity, along with conventional mean uptake. For each metric, bias vs. noise (reproducibility) trade-off performance analysis was performed.

Results Entropy was the only metric depicting reproducibility consistently lower than conventional mean uptake (by 50-80%). The various metrics were found not to be overly sensitive to the range of no->underestimated>full PSF modeling, but responded to PSF overestimation. Specifically, energy and entropy resulted in poorer bias vs. reproducibility performance for PSF overestimation (50-100% degradation in bias for matched reproducibility). At the same time, and interestingly, dissimilarity, contrast, correlation, homogeneity and variance were often quantitatively enhanced due to PSF overestimation (20-40% improvement in reproducibility for matched bias). PSF modeling with increasing kernels widths results in two phenomena: (1) reduction of spatial noise (due to reduced voxel variance and increased inter-voxel covariance), and (2) increased edge artifacts. These two phenomena both impact texture quantification, and different texture measures are optimized differently as they quantify heterogeneity in different ways.

Conclusions In terms of quantitative bias vs. noise (reproducibility) trade-off performance, different texture metrics responded differently whether or not PSF overestimation was performed. ‘Entropy’ consistently outperformed conventional mean uptake analysis in terms of reproducibility. It along with ‘energy’ performed optimally when overestimation was not performed. This was in contrast to some other metrics that performed optimally when overestimation was utilized. These results are also linked to the performance of vendor PET images, as some parameterize the PSF kernels using Ge-68 point sources (with higher positron range) instead of 18F, thus implicitly implementing overestimated PSF kernels.