Accounting for reconstruction kernel-induced variability in CT radiomic features using noise power spectra

Area of Science:

• Medical Imaging
• Radiomics
• Image Analysis

Background:

• Computed tomography (CT) radiomics features exhibit significant variability due to imaging parameters, potentially affecting their prognostic and predictive value.
• Understanding and mitigating this variability is crucial for reliable clinical application of radiomics.

Purpose of the Study:

• To investigate the impact of pitch, dose, and reconstruction kernel on CT radiomic features.
• To develop and validate correction factors for reducing feature variability introduced by reconstruction kernels.

Main Methods:

• Utilized credence cartridge and American College of Radiology (ACR) phantoms scanned on five different CT scanners.
• Quantified correlated noise using 3-D noise power spectrum (NPS) measurements from the ACR phantom.
• Assessed feature variability using the coefficient of variation (COV) and applied NPS peak frequency and ROI maximum intensity as correction factors.

Main Results:

• CT texture features demonstrated strong dependence on reconstruction kernels, evidenced by shifts in NPS peak frequency, but were largely dose-independent.
• Correction factors based on NPS peak frequency and ROI maximum intensity successfully reduced texture feature variability.
• Reported percentage improvements in robustness ranging from 30% to 78% for 19 features after applying corrections.

Conclusions:

• NPS peak frequency and ROI maximum intensity serve as effective correction factors to enhance the robustness of CT texture features.
• Mitigating reconstruction kernel-induced variability improves the reliability of CT radiomics for clinical applications.