Invariant Texture Features to Gray Level Discretization

Texture features are widely used in medical image analysis as quantitative descriptors of spatial intensity patterns. However, many texture features are strongly dependent on gray-level discretization, which limits their reproducibility and robustness across datasets. In this study, we propose a normalization framework to reduce gray-level dependence and construct discretization invariant texture features. Original and normalized invariant features were extracted from four medical datasets and their stability was evaluated using intraclass correlation coefficient (ICC) analysis. Downstream predictive performance was evaluated using AdaBoost models for binary classification tasks. Feature stability was further assessed under different sample size and class distribution to assess robustness to sampling variation. Statistical analyses were performed to compare feature-wise ICC values and stability distributions across experimental settings. Invariant normalization significantly increased ICC values and yielded more consistent feature stability across different sampling conditions. Overall, the proposed framework substantially improves feature stability and reproducibility while enhancing downstream predictive performance, thereby reducing the dependence of texture features on gray-level discretization.