ImUnity: A generalizable VAE-GAN solution for multicenter MR image harmonization.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Multi-center neuroimaging studies require harmonized data to mitigate scanner and site-specific biases.
• Existing harmonization methods often struggle with flexibility and preserving biological information.

Purpose of the Study:

• To introduce ImUnity, a 2.5D deep learning model for efficient and flexible MR image harmonization.
• To evaluate ImUnity's performance against state-of-the-art methods across diverse datasets.

Main Methods:

• Developed a Variational Autoencoder Generative Adversarial Network (VAE-GAN) model incorporating a confusion module and optional biological preservation module.
• Trained the model using 2D slices from various anatomical locations and contrast transformations across three open-source databases (ABIDE, OASIS, SRPBS).
• Tested ImUnity on T1-weighted images from multiple scanner types and age ranges.

Main Results:

• ImUnity demonstrated superior performance in generating high-quality harmonized images compared to existing methods, as evidenced by traveling subjects analysis.
• The model successfully removed site and scanner biases, leading to improved patient classification accuracy.
• ImUnity harmonized data from new, unseen sites/scanners without requiring additional fine-tuning.
• The model offers flexibility in selecting multiple MR reconstructed images for different applications.

Conclusions:

• ImUnity provides an effective solution for MR image harmonization, crucial for large-scale, multi-center population studies.
• The model's adaptability and performance suggest its potential for harmonizing various medical imaging modalities beyond T1-weighted images.