Texture analysis for automated evaluation of Jaszczak phantom SPECT system tests.

Abstract

Routine quarterly quality assurance (QA) assessment of single photon emission computed tomography (SPECT) systems includes analysis of multipurpose phantoms containing spheres and rods of various sizes. When evaluated by accreditation agencies, criteria applied to assess image quality are largely subjective. Determining a quantified image characteristic metric that emulates human reader impressions of image quality could be quite useful. Our investigation was conducted to ascertain whether image texture analysis metrics, such as those applied to PET scans to detect neoplasms, could prove helpful in linking qualitative statements of phantom sphere and rod visibility to quantified parameters. Because it is not obvious whether it is preferable to submit reconstructions to accrediting agencies performed using typical clinical (CLIN) protocol processing parameters or to follow agencies' filtered backprojection (FBP) suggestions, we applied texture analysis metrics to determine the degree to which these choices affect equipment capability assessment. Data were processed retrospectively for 125 different Tc-99m SPECT scans of standardized phantoms for 14 rotating Anger detector systems as part of routine quarterly QA. Algorithms were written to compute several classes of image metrics: quantile curve metrics, image texture analysis gray-level co-occurrence matrix (GLCM) metrics, contrast metrics, and count histogram metrics. For qualitative image scores, two experienced physicists independently graded sphere and rod visibility on a 5-level scale and assigned dichotomous visibility scores, without knowledge of quantified texture analysis metrics or each other's readings. The same phantom was used to collect 15 additional data sets with two dual-detector SPECT/CT systems, reconstructed both by FBP parameters that have been suggested by accrediting agencies and by manufacturers' default settings for CLIN SPECT/CT bone imaging protocols by ordered subsets expectation maximization (OSEM), incorporating attenuation correction using the CT scan. Image characteristics metrics were compared for FBP and CLIN reconstructions. For spheres, the metric with the strongest rank correlation with 5-level scale readings was the quantile curve slope (ρ=0.83, P<0.0001), while for rods it was GLCM Energy normalized to the maximum GLCM Energy value (EnergyNorm) (ρ=-0.88, P<0.0001). Compared to dichotomous readings, the metric with the highest ROC area under curve (AUC) for spheres was the quantile curve slopes (AUC=96±1%, sensitivity=91%, specificity=90%), and for rods was EnergyNorm (AUC=98±1%, sensitivity=92%, specificity=95%). Image contrast was higher for all sphere sizes and rod EnergyNorm was lower for sectors of intermediate-sized rods for FBP compared to CLIN reconstructions, in agreement with more rods judged to be visible from FBP than CLIN reconstructions (47% vs 33%, P=0.002). When preparing to submit quality assurance images of standardized phantoms to accrediting agencies, a reliable gauge of sphere and rod visibility can be predicted accurately using quantified reader-independent image texture analysis metrics, which also provide a useful basis for choosing among alternative image reconstruction options.