Learning from scanners: Bias reduction and feature correction in radiomics

Area of Science:

• Medical Imaging Analysis
• Quantitative Feature Extraction
• Radiomics

Background:

• Radiomics, quantitative features from medical images, are sensitive to non-tumor factors like scanner signal-to-noise ratio (SNR).
• This variability in radiomic features degrades the performance of predictive models.
• Standardized radiomic feature extraction is crucial for reliable clinical applications.

Purpose of the Study:

• To investigate the use of phantom measurements for characterizing and correcting scanner SNR dependence in radiomic features.
• To assess the impact of image acquisition settings on radiomic feature stability.
• To develop a method for reducing undesirable value variations in radiomic features.

Main Methods:

• Utilized a phantom with 17 regions of interest (ROIs) to study SNR influence.
• Acquired CT scans across 9 different exposure settings.
• Developed an additive correction model to mitigate scanner SNR effects on radiomic features.

Main Results:

• 62 out of 92 radiomic features exhibited high variance due to scanner SNR.
• The additive correction model reduced standard deviation by at least a factor of 2 for 47 of these features.
• Clinical relevance was assessed using a non-small cell lung cancer (NSCLC) patient cohort.

Conclusions:

• Approximately two-thirds of radiomic features are dependent on scanner exposure settings.
• The developed model significantly reduces feature value variation (at least factor 2).
• Scanner SNR correction enhances the reliability of radiomics predictions, particularly in NSCLC.