Room: AAPM ePoster Library
Purpose: Recently developed CT deep learning iterative reconstruction (DLIR) algorithms aim to improve image quality over currently established algorithms such as filtered back projection (FBP) and adaptive statistical iterative reconstruction (ASIR). However, the impact of the new DLIR algorithms on quantitative image metrics in a wide variety of media is not well established. Here, we evaluated the variations in radiomics features between FBP, ASIR, and three levels of DLIR within a texture phantom.
Methods: The updated Credence Cartridge Radiomics phantom was scanned using both a lung and a standard axial kernel, each at pixel sizes of 0.586 mm and 0.293 mm. All scans were reconstructed using FBP, ASIR, and three levels of DLIR. A cylindrical region of interest, 7.5cm diameter and 1.5cm thick, was defined in the center of each of the six texture regions in the phantom and radiomics analysis was performed using open-source software PyRadiomics. Differences in features between the two pixel sizes for each of the reconstruction algorithms were calculated for all textures.
Results: Feature values from FBP and ASIR or the three levels of DLIR tended to cluster together, and trends were similar across all texture regions. For the standard axial kernel, variations due to the reconstruction algorithm tended to be larger than the variation of changing the pixel size by a factor of two, for both first-order and gray level co-occurrence (GLCM) features. For the lung kernel, feature value variations due to changing the pixel size by a factor of two tended to be larger than variations due to the reconstruction algorithm.
Conclusion: Reconstruction algorithms have the potential to cause variations of radiomics features larger than those caused by doubling the pixel size in a variety of textures. The impact of these discrepancies should be considered in the construction and standardization of automated lesion classifiers.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the NIH T-32 Training Grant T32 EB002103. H.A.A. received royalties and licensing fees for computer-aided diagnosis technology through the University of Chicago.