Improved model-based deep learning for quantitative susceptibility mapping

JUAN LIU

Keywords: Quantitative susceptibility mapping, Self-supervised learning.

Abstract: Quantitative susceptibility mapping (QSM) is a magnetic resonance imaging (MRI) technique that estimates magnetic susceptibility of tissue from MR phase measurements. Recently, several supervised deep learning (DL) techniques have demonstrated impressive performance in solving the challenging ill-posed field-to-source inverse QSM reconstruction problem. To address the lack of the inherent non-existent ground-truth QSM references, a model-based method was recently proposed using the well-established physical model. However, it fails to perform well at the regions with large susceptibility variations. Here, we proposed uQSM+ with data augmentation techniques to improve the model-based learning. The proposed method was evaluated on a multi-orientation QSM datasets and 2019 QSM reconstruction challenge datasets. Quantitative and qualitative evaluation showed that uQSM+ and zero-shot uQSM+ was capable of reconstructing high quality QSM. The code is available at \inkhttps{https://github.com/juana313/uQSM-plus}.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: methodological development

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Unsupervised Learning and Representation Learning

Source Code Url: https://github.com/juana313/uQSM-plus