Radiomics robustness assessment and classification evaluation: A two-stage method demonstrated on multivendor FFDM

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Radiomic texture analysis typically relies on homogeneous imaging conditions, which differ from real-world clinical variability.
• This limitation impacts the generalizability of radiomic signatures across different mammography equipment vendors.

Purpose of the Study:

• To develop a two-stage radiomic texture analysis method to create robust texture signatures applicable across various mammography unit vendors.
• To assess the impact of feature reproducibility on the generalizability of radiomic signatures for breast cancer risk classification.

Main Methods:

• A two-stage method, Robustness Assessment, Classification Evaluation (RACE), was developed using full-field digital mammograms from two vendors (Hologic, GE).
• Stage one identified reproducible features across vendors using hierarchical clustering and robustness metrics.
• Stage two constructed radiomic signatures using stepwise feature selection and QDA for breast cancer risk classification, evaluating generalizability via inter- and intravendor performance.

Main Results:

• Generalizability of radiomic signatures decreased monotonically with increased feature robustness criteria (fewer reproducible features).
• Intervendor classification performance was significantly higher with the RACE method compared to ComBat harmonization alone.
• No significant difference in performance was observed between ComBat followed by RACE and either ComBat or RACE alone.

Conclusions:

• The proposed RACE method effectively constructs robust radiomic signatures for breast cancer risk assessment, demonstrating improved generalizability across mammography vendors.
• Feature robustness is crucial for enhancing classifier generalizability in diverse mammography datasets.
• Harmonization methods like ComBat show potential utility in classification schemes and warrant further investigation.