Radiomics-informed CT protocol optimization through feature-level trade-off analysis in a phantom study

Area of Science:

• Medical Imaging
• Radiology
• Quantitative CT Analysis

Background:

• Reproducibility of radiomic features in quantitative CT analysis is challenged by acquisition protocol variations.
• Identifying robust features is crucial for reliable diagnostic performance.

Purpose of the Study:

• To identify radiomic features predicting low-contrast detectability.
• To quantify feature sensitivity to CT acquisition parameters.
• To develop a framework for CT protocol optimization based on feature trade-offs.

Main Methods:

• Extraction of 107 radiomic features from phantom scans across multiple CT protocols.
• Assessment of feature robustness using coefficient of variation (CV) and intraclass correlation coefficient (ICC).
• Development of a random forest model for detectability prediction and sensitivity analysis using Spearman's correlation.

Main Results:

• ComBat harmonization significantly improved feature reproducibility.
• First-order features demonstrated the highest predictive power for detectability (AUC = 0.94, accuracy = 0.90).
• Texture and shape features showed varying sensitivities to protocol parameters and dose, respectively. A trade-off exists between detectability and feature sensitivity.

Conclusions:

• A quantitative framework for radiomics-informed CT protocol optimization was established.
• Moderate acquisition settings offer a more balanced and robust performance.
• The findings provide a foundation for optimizing CT protocols in clinical settings.