Radiomic Features in Hepatocellular Carcinoma: Stability with Regard to the Lesion Segmentation

Abstract

[eng] —This thesis investigates the impact of spatial mask transformations on radiomic feature importance and stability in the context of Hepatocellular Carcinoma (HCC) imaging. Radiomic features, derived from medical images, have become critical tools in predicting patient outcomes, aiding in diagnosis, and personalizing treatments. However, variations in segmentation masks, such as those introduced by different radiologists or segmentation protocols, can alter the calculated features and affect the performance of predictive models. The study applies several spatial transformations, including dilation, translation, rotation, and others, to simulate variations in segmentation practices and assesses how these transformations influence radiomic feature importance, model performance, and feature stability. Through a detailed analysis involving Random Forest models, we observe significant shifts in feature importance, with texture-based features becoming more dominant in transformed models. Furthermore, the study reveals that shape-based features exhibit the highest instability under dilation, while intensitybased features are particularly sensitive to translation. These findings suggest that radiomic feature calculations are highly dependent on the consistency of segmentation criteria, and any differences in segmentation approaches must be considered when developing predictive models. The study highlights the need for robust feature selection methods and model generalization strategies to ensure reliable clinical outcomes across different imaging scenarios