Matching and Homogenizing Convolution Kernels for Quantitative Studies in Computed Tomography.

Area of Science:

• Medical Imaging
• Radiomics
• Quantitative Imaging

Background:

• Kernel selection in computed tomography (CT) reconstruction significantly influences quantitative image features.
• Variations in kernels across different scanner manufacturers hinder direct comparison of radiomics data.
• Standardizing radiomics features is crucial for reliable multi-center and longitudinal studies.

Purpose of the Study:

• To identify CT reconstruction kernels that yield comparable radiomics feature values across different scanner vendors.
• To evaluate a novel image filtering technique for harmonizing kernel-induced differences in radiomics features.
• To facilitate more effective comparison of CT images acquired on diverse equipment.

Main Methods:

• A radiomics texture phantom was scanned on GE, Philips, Siemens, and Toshiba CT scanners.
• Images were reconstructed using various kernels (smooth to sharp), and 38 radiomics features were extracted.
• Kernel-homogenization filters were developed based on noise power spectra to minimize feature discrepancies.

Main Results:

• Philips C, Siemens B30f, and Toshiba FC24 kernels demonstrated feature values most similar to the GE Standard baseline.
• Kernel homogenization filters effectively reduced median feature differences to less than one coefficient of variation for most tested kernels.
• GE Edge and Toshiba kernels remained less comparable even after filter application.

Conclusions:

• Specifying consistent kernels is recommended for prospective CT radiomics studies.
• Kernel homogenization filters offer a viable solution for retrospective studies to mitigate cross-vendor kernel variations.
• Standardization of kernels or application of harmonization filters is essential for robust CT radiomics analysis.