Reproducibility of Cold Uptake Radiomics in 99m Tc-Sestamibi SPECT Imaging of Renal Cell Carcinoma

^99m
Tc-Sestamibi SPECT/CT imaging of renal cell carcinoma (RCC) has recently shown significant promise to distinguish benign oncocytomas from malignant RCC, where the former appears as high uptake in SPECT images and the latter as cold uptake. We aim towards radiomics analysis of cold uptake in SPECT images for another significant yet more daunting task of discriminating between RCC subtypes. The first step we pursue in the present work is feature selection by assessing the reproducibility of radiomic features.
^99m
Tc-Sestamibi SPECT/CT images of 50 patients with renal mass, plus a contrast-enhanced CT or MRI, were used by a trained radiologist to segment the region of interest (ROI) on the SPECT image. The impact of segmentation on reproducibility was studied by generating three ROIs via removing one voxel-layer from, or appending 1 and 2 voxel-layers to the manual ROIs (ROI
₀
, ROI
_-1
, ROI
₊₁
, ROI
₊₂
). Images were then uniformly quantized with eight grey-levels (GL) (2
²
-2
⁹
). A total of 363 radiomic features, verified within the Image Biomarker Standardization Initiative, were generated for all 50 patients/four ROIs/eight GLs. Voxel intensity in SPECT images is the number of counts, which is a patient-dependent non-normalized quantity. Thus, (i) fixed number of bin quantization must be chosen over fixed bin-width, and (ii) non-normalized statistical features should be excluded. The intra-class correlation (ICC) type C-1 calculated for all features between (a) all four ROIs and (b) between all except ROI
_-1
showed 204 features with (a) ICC>0.7 and (b) ICC>0.85, respectively. However, a subset of 204 features with emphasis on high GLs across higher-order feature classes should be further excluded due to the scarce presence of voxels with high intensities that can appear due to mis-segmentation. Moreover, reproducibility analysis of quantization using Spearman rank-correlation showed GLs below 32 should be avoided due to high variation. The resulting features and settings are recommended for further investigation of predictive value.