Ana Cismaru

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-98

May 13, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-98.pdf

Zero Echo Time (ZTE) MRI is a rapid and near-silent 3D radial MRI sequence that is based on constant gradients incremented by small amounts between TRs. Because the readout gradients are active during the RF excitation, the initial readout points for each radial spoke are missed, resulting in a spherical Dead Time Gap (DTG) at the center of k-space. This gap results in incorrect image contrast if left uncorrected. Traditional approaches to correcting the inaccurate contrast from the DTG often require additional scan time or complex hardware adjustments. This limits ZTE’s practical applicability, especially in settings such as dynamic contrast-enhanced where image contrast changes rapidly. This work presents multiple deep-learning frameworks designed to reconstruct large DTG solely from multi-coil k-space data, without the need for explicit coil sensitivities or additional acquisitions. A large ZTE dataset is simulated by applying the ZTE forward model to a Cartesian fully sampled MRI dataset. Various learning approaches were explored to assess their abilities to recover the missing center of k-space, including UNet, Interlacer, and Diffusion models. The proposed methods demonstrated strong 2D reconstruction results on simulated datasets with the Diffusion model performing best followed closely by the UNet approach. Due to the significant memory and computational demands of 3D Diffusion models, a 3D UNet model was chosen for reconstructing in-vivo data. This 3D UNet model demonstrated strong performance in reconstructing simulated 3D ZTE data. However, application to in-vivo data revealed challenges such as the model hallucinating auras that correlate with the field of view of the training data. This work contributes to the ongoing development of efficient and accurate reconstruction techniques for ZTE MRI, aiming to extend this fast and silent sequence to applications such as DCE, where image contrast is dynamically changing.

Advisors: Michael Lustig

---

BibTeX citation:

@mastersthesis{Cismaru:EECS-2024-98,
    Author= {Cismaru, Ana},
    Title= {DL-ZTE: Towards Deep Learning-based Methods for Dead Time Gap Recovery in Zero TE MRI},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-98.html},
    Number= {UCB/EECS-2024-98},
    Abstract= {Zero Echo Time (ZTE) MRI is a rapid and near-silent 3D radial MRI sequence that is based on constant gradients incremented by small amounts between TRs. Because the readout gradients are active during the RF excitation, the initial readout points for each radial spoke are missed, resulting in a spherical Dead Time Gap (DTG) at the center of k-space. This gap results in incorrect image contrast if left uncorrected. Traditional approaches to correcting the inaccurate contrast from the DTG often require additional scan time or complex hardware adjustments. This limits ZTE’s practical applicability, especially in settings such as dynamic contrast-enhanced where image contrast changes rapidly. This work presents multiple deep-learning frameworks designed to reconstruct large DTG solely from multi-coil k-space data, without the need for explicit coil sensitivities or additional acquisitions. A large ZTE dataset is simulated by applying the ZTE forward model to a Cartesian fully sampled MRI dataset. Various learning approaches were explored to assess their abilities to recover the missing center of k-space, including UNet, Interlacer, and Diffusion models. The proposed methods demonstrated strong 2D reconstruction results on simulated datasets with the Diffusion model performing best followed closely by the UNet approach. Due to the significant memory and computational demands of 3D Diffusion models, a 3D UNet model was chosen for reconstructing in-vivo data. This 3D UNet model demonstrated strong performance in reconstructing simulated 3D ZTE data. However, application to in-vivo data revealed challenges such as the model hallucinating auras that correlate with the field of view of the training data. This work contributes to the ongoing development of efficient and accurate reconstruction techniques for ZTE MRI, aiming to extend this fast and silent sequence to applications such as DCE, where image contrast is dynamically changing.},
}

EndNote citation:

%0 Thesis
%A Cismaru, Ana 
%T DL-ZTE: Towards Deep Learning-based Methods for Dead Time Gap Recovery in Zero TE MRI
%I EECS Department, University of California, Berkeley
%D 2024
%8 May 13
%@ UCB/EECS-2024-98
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-98.html
%F Cismaru:EECS-2024-98