Texture analysis in 177Lu SPECT phantom images: Statistical assessment of uniformity requirements using texture features

Abstract

The purpose of this study was to apply texture analysis (TA) to evaluate the uniformity of SPECT images reconstructed with the 3D Ordered Subsets Expectation Maximization (3D-OSEM) algorithm. For this purpose, a cylindrical homogeneous phantom filled with 177Lu was used and a total of 24 spherical volumes of interest (VOIs) were considered inside the phantom. The location of the VOIs was chosen in order to define two different configurations, i.e. gravity and radial configuration. The former configuration was used to investigate the uniformity of distribution of 177Lu inside the phantom, while the latter configuration was used to investigate the lack of uniformity from center towards edge of the images. For each VOI, the trend of different texture features considered as a function of 3D-OSEM updates was investigated in order to evaluate the influence of reconstruction parameters. TA was performed using CGITA software. The equality of the average texture feature trends in both spatial configurations was assumed as the null hypothesis and was tested by functional analysis of variance (fANOVA). With regard to the gravity configuration, no texture feature rejected the null hypothesis when the number of subsets increased. For the radial configuration, the statistical analysis revealed that, depending on the 3D-OSEM parameters used, a few texture features were capable of detecting the non-uniformity of 177Lu distribution inside the phantom moving from the center of the image towards its edge. Finally, cross-correlation coefficients were calculated to better identify the features that could play an important role in assessing quality assurance procedures performed on SPECT systems.

Fig 1. SPECT/CT images of the homogeneous cylindrical phantom.

The top row shows CT images of the homogeneous phantom in A) coronal, B) sagittal and C) transaxial sections, together with D) VOI location used for texture analysis. The second and third rows show reconstructed SPECT images of the central slice of the phantom for E) 5S and F) 10S, and different I values.

Fig 2. Spatial configurations used for statistical analysis.

A) Gravity configuration: the horizontal black dashed line defines the midplane of the phantom, dividing it into two halves. B) Radial configuration: the location of the VOIs is the same as in A), but they are grouped in inner (dotted line), middle (dashed line) and fringe (continuous line) rings. The VOIs crossed by the aforementioned lines were considered for statistical analysis.

Fig 3. Workflow adopted for the statistical analysis.

Fig 4. Texture features for VOIs in radial configuration for 5S.

A) NC-correlation, B) TFC-coarseness and C) TFCC-code similarity. Dots represent individual texture values inside the VOIs in the considered rings, while thick lines are their average values. Colors indicate spatial groupings of the VOIs (inner, middle and fringe). Box plots are reported under each plot.

Fig 5. Texture features for VOIs in radial configuration for 10S.

A) NID-busyness, B) TFC-coarseness, C) TFCC-homogeneity and D) TFCC-code similarity. Dots represent individual texture values inside the VOIs in the considered rings, while thick lines are their averaged values. Colors indicate spatial groupings of the VOIs (inner, middle and fringe). Box plots are reported under each plot.

Fig 6. Heat-maps of cross-correlation coefficients between texture features rejecting H₀ in the radial configuration.

In each heat-map the horizontal and vertical axes refer to the value of the texture features obtained at a specific I for A) 5S and B) 10S. The legend reported in both maps refers to the Pearson’s correlation coefficient, and the color associated with a legend value is reported on the square regions of the maps, representing the level of correlation between the texture features.