Frederic Wang

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-170

August 9, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-170.pdf

Although deep learning (DL) offers powerful tools for image reconstruction from undersampled MRI measurements, a major hurdle is the need for high-quality training data, which are difficult to acquire due to practical constraints. To address this, we introduce the k-band framework, which enables training DL models using only partial, limited-resolution data, which can be acquired easily. Moreover, k-band offers test-time generalization to high-resolution reconstructions. The k-band framework designs the acquisition and training in-tandem. We propose the acquisition of k-space bands, with limited resolution in the phase-encoding dimension; this is fast, practical, and easy-to-implement. To enable training using only these k-space bands, we also propose an optimization method dubbed stochastic gradient descent over k-space subsets. We prove analytically that this method stochastically approximates fully supervised training when the bands’ angles are randomized across subjects and a k-space loss-weighting mask is applied. Experiments with raw MRI data demonstrate that k-band achieves performance on-par with that of fully supervised and self-supervised methods trained on high-resolution data, with the benefit of training on limited-resolution data. The proposed k-band framework offers practical strategies for fast acquisition and self-supervised training using limited-resolution data, with theoretical guarantees. It is easy-to-implement and agnostic to the pulse sequence and DL architecture. It can thus facilitate curation of new datasets and development of DL models for data-challenging regimes.

Advisors: Michael Lustig

---

BibTeX citation:

@mastersthesis{Wang:EECS-2024-170,
    Author= {Wang, Frederic},
    Editor= {Lustig, Michael},
    Title= {A Novel Self-Supervised Deep Learning Method for MRI Reconstruction},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-170.html},
    Number= {UCB/EECS-2024-170},
    Abstract= {Although deep learning (DL) offers powerful tools for image reconstruction from undersampled MRI measurements, a major hurdle is the need for high-quality training data, which are difficult to acquire due to practical constraints. To address this, we introduce the k-band framework, which enables training DL models using only partial, limited-resolution data, which can be acquired easily. Moreover, k-band offers test-time generalization to high-resolution reconstructions. The k-band framework designs the acquisition and training in-tandem. We propose the acquisition of k-space bands, with limited resolution in the phase-encoding dimension; this is fast, practical, and easy-to-implement. To enable training using only these k-space bands, we also propose an optimization method dubbed stochastic gradient descent over k-space subsets. We prove analytically that this method stochastically approximates fully supervised training when the bands’ angles are randomized across subjects and a k-space loss-weighting mask is applied. Experiments with raw MRI data demonstrate that k-band achieves performance on-par with that of fully supervised and self-supervised methods trained on high-resolution data, with the benefit of training on limited-resolution data. The proposed k-band framework offers practical strategies for fast acquisition and self-supervised training using limited-resolution data, with theoretical guarantees. It is easy-to-implement and agnostic to the pulse sequence and DL architecture. It can thus facilitate curation of new datasets and development of DL models for data-challenging regimes.},
}

EndNote citation:

%0 Thesis
%A Wang, Frederic 
%E Lustig, Michael 
%T A Novel Self-Supervised Deep Learning Method for MRI Reconstruction
%I EECS Department, University of California, Berkeley
%D 2024
%8 August 9
%@ UCB/EECS-2024-170
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-170.html
%F Wang:EECS-2024-170