Harmonizing the pixel size in retrospective computed tomography radiomics studies

Area of Science:

• Medical Imaging
• Radiomics
• Computational Pathology

Background:

• Consistent pixel sizes are crucial for reliable radiomics texture feature analysis.
• Variable pixel sizes in computed tomography (CT) scans introduce significant intra-patient variability.
• This variability can confound the interpretation of radiomics features in clinical studies.

Purpose of the Study:

• To develop and evaluate a correction method for variable pixel sizes in CT radiomics.
• To assess the impact of this correction on intra-patient and inter-patient agreement.
• To determine the generalizability of the correction method using a radiomics phantom.

Main Methods:

• Combined image resampling with Butterworth low-pass filtering in the frequency domain.
• Applied correction to CT scans of lung cancer patients with varying pixel sizes (0.59–0.98 mm).
• Calculated 150 radiomics features, compared intra- and inter-patient agreement using OCCC, and used hierarchical clustering for evaluation.
• Validated the method on 17 CT scans of a radiomics phantom.

Main Results:

• Correction significantly reduced intra-patient variability from 79% to 10% of features.
• Hierarchical clustering correctly identified 8/8 patients post-correction versus 2/8 without.
• Phantom study showed reduced variability in 61% of features post-correction, with resampling alone increasing variability.
• Resampling and Butterworth filtering effectively mitigated pixel size-induced variability.

Conclusions:

• The proposed correction method (resampling + Butterworth filtering) effectively reduces CT radiomics feature variability due to pixel size differences.
• This approach enhances the reliability of radiomics analysis and may generalize to other CT acquisition parameter variations.
• The findings support the use of this correction for more robust radiomics studies.