Publication date: Aug 25, 2026
This study aims to quantify the sensitivity of radiomics features to variations in NPS peak frequency and to develop a robust, physics-driven non-linear correction method to mitigate kernel-induced variability in extracted features.
Methods: A homogeneous computational phantom was generated to simulate CT images with varying noise textures. Noise texture was synthesized with NPS peak frequencies (f_peak) ranging from 0. 3 to 1. 0 mm-1 across three noise levels (25, 50, and 75 HU). Radiomics features, specifically First Order and Gray Level Cooccurrence Matrix (GLCM) were extracted. Four mathematical models were evaluated to characterize the feature dependency on f_peak. The second-order polynomial model was selected based on the optimization of Akaike Information Criterion (AIC), RMSE, and Adjusted R-Squared. Radiomics features sensitivity was assessed using the coefficient of determination (R2) and the Coefficient of Variation (COV). A robust correction algorithm was developed to normalize feature values to a reference frequency (f_(peak,ref) = 0. 6 mm-1). The efficacy of the correction was quantified using the Percentage Improvement (PI) in COV.
Results: GLCM features exhibited significantly higher sensitivity to f_peak shifts compared to First Order features, with a large proportion showing a high coefficient of determination (R2 ≈ 1) and statistical significance (p
| Concepts | Keywords |
|---|---|
| computed tomography | |
| GLCM | |
| noise power spectrum | |
| radiomics | |
| texture analysis |
Semantics
| Type | Source | Name |
|---|---|---|
| disease | MESH | NPS |
| disease | MESH | AIC |