Effects of nonstationary system blur on radiomic texture features of trabecular bone in normal and ultra-high resolution CT.

Background: CT image texture features (TFs) of trabecular bone are being investigated as predictive markers in osteoporosis (OP) and osteoarthritis (OA). To support the development of such radiomic approaches, the reproducibility of bone TFs to CT system factors needs to be characterized. We hypothesize that, in addition to the well-established inter- and intra-scanner effects (e.g., changes in reconstruction algorithm), the non-stationarity of CT spatial resolution may introduce inconsistencies in TFs of morphologically similar bone regions placed in different locations within the scanner field-of-view (FOV).

Purpose: To characterize the impact of spatially-variant CT blur on the reproducibility of trabecular bone TFs to radial shift within the FOV for a conventional normal resolution (NR) protocol with ∼0.5 mm slice thickness, and an ultra-high resolution (UHR) protocol representative of the new generation of high-resolution CT (0.25 mm slice thickness).

Methods: Canon Aquilion Precision CT (Canon Medical Systems, Japan) was used to scan four human femora placed at 0, 9, and 18 cm radial shifts from the isocenter. NR (∼0.26 mm in-plane voxel size) and UHR (∼0.13 mm in-plane voxel) images were obtained at 1.5 s rotation. At each CT resolution, a total of 377 spherical regions of interest (ROIs) of 2.5 mm radius were seeded within the trabecular bone of the femora at 0 cm shift and transported via rigid registration to the images at 9 cm and 18 cm shift. TFs of the Gray Level Co-Occurrence and Run Length matrices (GLCM and GLRLM) were obtained for all ROIs. Reproducibility across radial shifts was evaluated (separately in NR and UHR) in terms of the concordance correlation coefficient (CCC) between the registered ROIs. Support vector machine (SVM) classifiers were applied to evaluate whether the radial shift of an ROI can be predicted from its TFs.

Results: In NR, the median CCCs for 0 cm vs. 9 cm radial shifts were ∼0.9 for both GLCM and GLRLM TFs, dropping to ∼0.6 for 0 cm vs. 18 cm shifts. In UHR CT, the median CCCs of GLCM TFs were 0.92 for 0 cm vs. 9 cm and 0.52 for 0 cm vs. 18 cm; for GLRLM, the median CCCs were 0.88 for 0 cm vs. 9 cm and 0.53 for 0 cm vs. 18 cm. In a separate analysis on only high contrast ROIs (upper quartiles of voxel value variance at 0 cm), we found an additional 20% (for 0 cm vs. 9 cm) to 40% (for 0 cm vs. 18 cm) reduction in CCC in both UHR and NR. The somewhat worse reproducibility in UHR CT is attributed to more pronounced effects of gantry motion blur. The classification models were able to discriminate ROIs at 0 cm from the other radial shifts with median accuracy of 0.54 (0.67 in high contrast ROIs) in NR and 0.46 (0.67 in high contrast ROIs) in UHR. The ROIs at 9 cm were identified with 0.31 (0.5) accuracy in NR and 0.46 (0.58) accuracy in UHR; the ROIs at 18 cm were identified with 0.60 (0.83) accuracy in NR and 0.77 (0.83) accuracy in UHR.

Conclusions: Nonstationary CT spatial resolution leads to a loss of reproducibility of trabecular bone TFs between different regions of the scan FOV, potentially confounding radiomic predictors in clinical data where the lateral shift of the skeletal site of interest changes with patient size.