RETURN TO ABSTRACT LISTING


Objective:
Poor reproducibility of radiomics analysis stemming mainly from variations in image acquisition and processing has hampered its clinical translation. Here, we systematically evaluate the robustness of CT radiomic metrics under different imaging conditions, including varying slice thickness, dose, tube voltage, and reconstruction algorithm using intraclass correlation coefficients (ICC).

Materials and Methods:
Scans of an anthropomorphic liver phantom with seven different texture inserts for simulating clinically observed radiological textures were acquired on a 16-cm detector GE Revolution Apex Ed. CT scanner. Robustness was assessed under four conditions: 1) nominal slice thicknesses: 0.625 mm, 1.25 mm, and 2.5 mm; 2) dose levels: CTDIvol of 13,86 mGy for the standard dose, 40% reduced dose and 60% reduction; 3) tube voltages i.e., 100 kVp and 120 kVp; and 4) reconstruction algorithms: a deep learning image reconstruction (DLIR-high) algorithm and a hybrid iterative reconstruction (IR) algorithm ASiR-V50% (AV50). The assessment was conducted with one condition change at a time. Opensource CapTk software was used for texture analysis. The texture panel comprised of 265 metrics belonging to different texture families, including intensity, histogram, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighboring gray-tone difference matrix (NGTDM). ICC two-way-mixed with absolute agreement was used to evaluate robustness of the radiomics metrics based on their families under different imaging conditions.

Results:
In general, considering the cumulative effect of all the changes in the imaging conditions, 56% of intensity metrics have an ICC greater than 0.8, indicating high robustness; 28% of NGTDM metrics and 24% of GLCM metrics also displayed high robustness. A similar trend, albeit with different percentage of radiomic metrics per texture family, was observed with changes in slice thickness, dose, and tube voltage. In comparison, with changes in reconstruction algorithm, intensity metrics (41.6%), followed by GLRLM (22%), and histogram (18.8%) showed high robustness. Compared to changes in slice thickness, dose and tube voltage, changes in reconstruction algorithm caused the largest inconsistency in radiomic metrics across all texture families, with a maximum of 41.6% metrics (intensity metrics) that remained robust.

Conclusion:
Systematically evaluating the behavior of the radiomic metrics under different imaging conditions can optimize the use of CT radiomics. The robust radiomic metrics panel identified in this study can serve as an application specific feature-selection step to reduce data dimensionality issues prevalent in radiomics studies. It can also improve the robustness of the prediction models derived from CT radiomics. Specifically, first-order texture metrics: intensity metrics showed the highest robustness compared to other higher-order texture metrics.