A Target-Free Harmonization Method for MRI
No target domain data needed – this technique harmonizes MRI images while keeping patient data private.
In MRI, variations in scan parameters, sequences, or hardware cause domain shifts that degrade deep learning models. Existing harmonization methods need both source and target domain data, forcing data sharing between institutions and raising patient privacy concerns – a major barrier to clinical adoption.
TgtFreeHarmony removes this requirement by estimating the target domain style through a manifold search of MRI styles, using a disentanglement-based generator and Bayesian optimization guided by a downstream task model trained on target data. Evaluated on brain tissue segmentation across multiple institutes, it successfully harmonized source images to target style, boosting segmentation accuracy without any access to target-domain data. This privacy-preserving approach establishes a new direction for MRI harmonization, enabling hospitals to collaborate on AI models without breaching patient confidentiality.
- Eliminates need for target domain data at training or test time, enabling privacy-preserving harmonization
- Uses Bayesian optimization with a disentanglement-based generator to estimate target style from a learned manifold
- Improved brain tissue segmentation performance across multiple institutes without data sharing
Why It Matters
Enables hospitals to harmonize MRI scans without sharing patient data, improving AI diagnostic models across institutions.