Multi-Stage Harmonization for Robust AI across Breast MR Databases.

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning

Background:

• Radiomic features from medical images can exhibit batch effects due to varying data sources.
• These batch effects may hinder the generalizability of machine learning models in clinical practice.

Purpose of the Study:

• To investigate the impact of batch harmonization on the classification performance of radiomic features.
• To evaluate whether harmonized radiomic features can yield generalizable machine learning models for breast lesion classification.

Main Methods:

• Retrospective study using 32 radiomic features from DCE-MR images of breast lesions.
• Data from two databases (A: 944 lesions, B: 1986 lesions) were used for training and independent testing.
• ComBat batch harmonization was applied to minimize data source discrepancies.
• Linear discriminant analysis classifiers were trained and tested using both pre- and post-harmonization features.

Main Results:

• Four out of five training and testing scenarios showed statistically equivalent classification performance between pre- and post-harmonization features.
• The area under the receiver operating characteristic curve (AUC) was used to evaluate performance.
• No significant performance degradation was observed after applying batch harmonization.

Conclusions:

• Batch harmonization of radiomic features can effectively mitigate batch effects without compromising classification performance.
• Machine learning models utilizing harmonized radiomic data show potential for improved generalizability in clinical applications.
• This approach supports the translation of radiomics into robust, real-world diagnostic tools.