Exploring domain adaptation for deep neural network trained QSM

TL;DR

This paper proposes a domain adaptation method using domain-specific batch normalization to improve deep learning-based quantitative susceptibility mapping (QSM) in MRI, addressing domain shift issues between synthetic and in-vivo data.

Contribution

It introduces a simple domain adaptation technique with domain-specific batch normalization to enhance deep neural network performance for QSM.

Findings

Improved QSM accuracy on in-vivo data

Effective domain adaptation with minimal additional complexity

Addresses domain shift between synthetic and real data

Abstract

Quantitative susceptibility mapping (QSM) is a MRI technique that estimates tissue magnetic susceptibility. The generation of QSM requires solving a challenging ill-posed field-to-source inversion problem. Recently, several deep learning (DL) QSM techniques have been proposed and demonstrated impressive performance. Due to the inherent non-existent ground-truth QSM references, these techniques used either COSMOS maps or synthetic data for network training. Synthetic data is easy to generate but often adapt poorly to in-vivo data due to domain shifts. Here, we introduce an easy domain adaptation technique using domain-specific batch normalization to address this problem.