RT Journal Article SR Electronic T1 A texture analysis approach to evaluating SPECT system quality assurance JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 1313 OP 1313 VO 58 IS supplement 1 A1 Kenneth Nichols A1 Fritzgerald Leveque A1 Christopher Palestro YR 2017 UL http://jnm.snmjournals.org/content/58/supplement_1/1313.abstract AB 1313Objectives: Routine quarterly quality assurance (QA) assessment of SPECT systems includes analysis of multipurpose phantoms constructed of radioactivity added to cylindrical water baths, along with Plexiglas inserts of 6 different solid sphere & rod sizes. When evaluated by accreditation agencies, criteria applied to assess image quality consist of the minimum number of spheres & rod sectors deemed visible. However, this is a subjective, not quantitative, assessment. While measured contrast may seem a reasonable metric to associate with sphere visibility, associating rod appearance with contrast is less straightforward due to rod visibility varying with depth from the phantom surface due to worsening spatial resolution with increasing distance from the collimator. Our investigation was conducted to ascertain whether texture analyses methods applied to PET oncology scans (Nuc Med Commun 2012;34:40-6) could prove helpful in linking qualitative statements of phantom sphere & rod visibility to quantified image texture analyses metrics.Methods: Data were processed retrospectively for 40 different SPECT scans acquired of standardized phantoms for 7 gamma cameras for routine quarterly QA. Projection data were acquired as 128x128 matrices for 120-128 projections over 360º using 666-814 MBq 99mTc in water for 40±8 Mcts. Algorithms were written in IDL v8.4 to compute 3 classes of texture analysis metrics: quantile (Q) curve metrics, gray-level co-occurrence matrix (GLCM) energy & entropy metrics, and contrast (C) metrics. For Q metrics, sphere & rod count curves were sorted from minimum to maximum values & plotted against minimum to maximum uniform section counts to form Q curves, to which linear regression was applied; deviations from unity for Q curves were taken as evidence that count distributions of sphere & rod regions were significantly different from uniform count distributions. For GLCM metrics, the entropy metric value was computed as - Σ P(i,j) log(P(i,j)) & the energy metric was computed as Σ P(i,j)2, where P(i, j) is the number of times a count level i co-occurred with count level j within a 1-pixel 2-dimensional radius. C was computed as (maximum-minimum)÷(maximum+minimum) for raw counts for sphere, rod & uniform phantom sections; for spheres, it was also feasible to perform fitting of radial count profiles to 2nd order polynomial curve fits, from which fitted contrast also was computed. For the qualitative image scores, a physicist graded sphere & rod visibility on a 5-point scale (0 = “definitely not visible” to 4 = ”definitely visible”), as well as assigning a dichotomous visibility score, without knowledge of the quantified texture analysis metrics.Results: For spheres, the metric with the strongest rank correlation with 5-point scale readings was the Q curve slope (ρ = 0.873, p < 0.0001), while for rods the metric with the strongest rank correlation was entropy (ρ = -0.819, p < 0.0001). Simultaneous ROC comparisons against dichotomous readings indicated that the metric with the highest area under the ROC curve (AOC) for spheres was also for Q curve slopes (AOC = 97±1%, sensitivity = 88%, specificity = 97%), & for rods was also for entropy (AOC = 92±2%, sensitivity = 86%, specificity = 88%). By ROC analyses, these metrics agreed with dichotomous qualitative visibility scores for both spheres & rods significantly more strongly than did the straightforward C measurements (p < 0.001).Conclusion: When anticipating submitting standardized quality assurance images to accrediting agencies, a reliable gauge of sphere & rod qualitative visibility judgements can be accurately predicted through the use of quantified texture analysis metrics. Research Support: None