Experimental phantom evaluation to identify robust positron emission tomography (PET) radiomic features

Area of Science:

• Medical Imaging
• Radiomics
• Quantitative Imaging

Background:

• Multicenter positron emission tomography (PET) studies face challenges due to variations in imaging protocols, leading to radiomic feature variability.
• Standardized measurements, such as European Association Research Ltd. (EARL) accreditation, can help mitigate these issues.
• Metal artifacts in PET/CT imaging also introduce significant challenges for radiomic analysis.

Purpose of the Study:

• To identify radiomic features that are robust to various degrading factors in PET studies.
• To evaluate the reliability of radiomic features derived from different segmentation methods (COA and 40% threshold) compared to ground truth.
• To present a reproducible method for identifying robust PET radiomic features, including those resistant to metal artifacts.

Main Methods:

• Utilized European Association Research Ltd. (EARL) accreditation measurements to assess feature robustness across different systems and reconstruction protocols.
• Developed and applied a procedure to identify PET radiomic features robust to PET/CT metal artifacts.
• Evaluated segmentation accuracy and radiomic feature reliability using 27 homemade phantoms with CT-defined ground truth, comparing contrast-oriented algorithm (COA) and 40% threshold PET segmentation.

Main Results:

• Identified 48 features robust to system variations, 22 to resolution, 102 to metal artifacts, and 42 to segmentation tools.
• Discovered only 4 radiomic features were simultaneously robust to all degrading factors.
• Found strong correlations between radiomic features derived from ground truth and those from COA (98 features) and 40% threshold (102 features) segmentation, despite volume underestimation.

Conclusions:

• Recommends using EARL accreditation measurements in multicenter studies to identify robust PET radiomic features retrospectively.
• Highlighted four specific radiomic features (area under the curve of cumulative SUV volume histogram, skewness, kurtosis, gray-level variance) as robust to all degrading factors.
• Demonstrated the feasibility of using both 40% threshold and COA segmentations for radiomics analysis.