Probabilistic multi-site MR image harmonization via feature preserving conditional generative adversarial networks

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Brain magnetic resonance imaging (MRI) is crucial for neurological disorder diagnosis.
• Inconsistent image intensity distributions across different acquisition sites pose challenges for multi-site studies.
• Existing image harmonization methods have limitations, especially in preserving anatomical features.

Purpose of the Study:

• To develop a novel deep learning-based image harmonization technique for multi-site brain MRI.
• To ensure anatomical feature preservation during image harmonization.
• To integrate probabilistic generative models for improved harmonization.

Main Methods:

• Introduced a feature preserving conditional generative adversarial network (FP-cGAN) with a novel regularizing constraint.
• Trained models on unpaired MRI data and evaluated on paired (traveling subjects) data from the SRPBS dataset.
• Compared FP-cGAN against histogram matching, ImUnity, and CycleGAN using distributional measures and heterogeneous datasets.

Main Results:

• The proposed FP-cGAN method demonstrated superior performance compared to existing harmonization techniques.
• The method effectively converted images from multiple origins to a target site format while preserving anatomical details.
• Probabilistic formulation provided meaningful uncertainty estimates, analyzed through sample variance and uncertainty maps.

Conclusions:

• FP-cGAN offers a robust and generalizable solution for harmonizing multi-site brain MRI data.
• The approach effectively addresses the challenge of intensity inconsistency in neuroimaging.
• The integration of probabilistic generative models enhances harmonization quality and provides uncertainty quantification.