The Technome - A Predictive Internal Calibration Approach for Quantitative Imaging Biomarker Research

Area of Science:

• Radiology
• Medical Imaging
• Data Science

Background:

• Radiomics extracts quantitative imaging features from medical scans for clinical analysis, such as predicting patient survival.
• Technical variations in computed tomography (CT) scans, including reconstruction kernel changes and inter-patient differences, can impact radiomics analysis.
• Existing radiomics stability analyses often overlook inter-patient technical variations.

Purpose of the Study:

• To develop an automated internal calibration method for radiomics using control regions (CRs) to enhance prediction performance.
• To derive general rules for internal calibration based on analyzed features and selected CRs.
• To improve the robustness and accuracy of radiomics in clinical applications.

Main Methods:

• Measurements within 3D regions-of-interest (ROIs) were calibrated using additional ROIs (air, adipose tissue, liver) as control regions.
• Qualification criteria, based on radiomics stability analysis, were defined to select relevant information from CRs.
• An optimization process used these criteria to automatically derive internal calibration suitable for prediction tasks.

Main Results:

• The internal calibration method was applied to two distinct studies: prediction of centrilobular emphysema in Chronic Obstructive Pulmonary Disease (COPD) and prediction of one-year survival in cancer patients.
• The calibration approach demonstrated enhanced performance in both prediction tasks.
• The study successfully derived general rules for automated internal calibration.

Conclusions:

• Automated internal calibration using control regions can significantly improve radiomics prediction performance.
• This method addresses the challenge of technical variations in medical imaging, leading to more reliable radiomics analyses.
• The findings suggest a pathway for more robust and accurate clinical decision-making based on radiomics data.