Hybrid optimization between iterative and network fine-tuning reconstructions for fast quantitative susceptibility mapping

Jinwei Zhang, Hang Zhang, Pascal Spincemaille, Thanh Nguyen, Mert R. Sabuncu, Yi Wang

Keywords: convolutional neural network, alternating direction method of multiplier, domain adaptation, quantitative susceptibility mapping

Abstract: A Hybrid Optimization Between Iterative and network fine-Tuning (HOBIT) reconstruction method is proposed to solve quantitative susceptibility mapping (QSM) inverse problem in MRI. In HOBIT, a convolutional neural network (CNN) is first trained on healthy subjects’ data with gold standard labels. Domain adaptation to patients’ data with hemorrhagic lesions is then deployed by minimizing fidelity loss on the patient training dataset. During test time, a fidelity loss is imposed on each patient test case, where alternating direction method of multiplier (ADMM) is used to split the time consuming fidelity imposed network update into iterative reconstruction and network update subproblems alternatively in ADMM, and only a subnet of the pre-trained CNN is updated during the process. Compared to the method FINE where such fidelity imposing strategy was initially proposed to solve QSM, HOBIT achieved both performance gain of reconstruction accuracy and vast reduction of computational time.

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: both

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Transfer Learning and Domain Adaptation