CMR-Analysis (including machine learning)

1357442 - A k-Space Sharpening Loss for Self-Supervised Physics-Guided Deep Learning Reconstruction of High-Resolution 3D LGE CMR

Wednesday, January 25, 2023

Location: E-POSTER

Presenting Author(s)

CZ

Chi Zhang

PhD Candidate
University of Minnesota, Minnesota, United States

Primary Author(s)

CZ

Chi Zhang

PhD Candidate
University of Minnesota, Minnesota, United States

Co-Author(s)

OD

Omer B. Demirel, MSc

PhD Candidate
University of Minnesota
Minneapolis, Minnesota, United States
BY

Burhaneddin Yaman, BSc

PhD
University of Minnesota
Saint Paul, Minnesota, United States
Hossam El-Rewaidy

MSc.
Harvard Medical School
Boston, Massachusetts, United States
Chetan Shenoy, MBBS, MS

Associate Professor of Medicine
University of Minnesota
Minneapolis, Minnesota, United States
RN

Reza Nezafat, PhD

Professor
Harvard Medical School
Boston, Massachusetts, United States
Mehmet Akcakaya, PhD

Associate Professor
University of Minnesota
Minneapolis, Minnesota, United States

Disclosure(s):

Reza Nezafat, PhD: No financial relationships to disclose

Background: Late gadolinium enhancement (LGE) cardiac MRI (CMR) is the clinical reference for assessment of myocardial scar [1]. Recently, physics-guided deep learning (PG-DL) has gained interest for accelerated CMR, but it traditionally requires fully-sampled references to train. Isotropic high-resolution 3D LGE CMR cannot be fully-sampled, but a recent self-supervised DL technique was shown to enable training of PG-DL networks with improved image quality [2,3] compared to compressed sensing (CS) [4]. However, PG-DL reconstruction at higher acceleration rates may still suffer from blurring. In this study, we propose a sharpening loss to improve self-supervised PG-DL reconstruction for 3D LGE CMR with better delineation of small structures.

Methods:

Improved Self-Supervised Deep Learning Reconstruction: Regularized MRI reconstruction solves min_x ||y_Ω-E_Ωx||² + R(x), where E_Ω is the multi-coil encoding operator for sampling pattern Ω, y_Ωis the acquired k-space, x is the image and R(·) is a regularizer. PG-DL is implemented using algorithm unrolling [2], alternating between data fidelity and a CNN-based R(·). Building on the self-supervision strategy [2], we propose a sharpening loss term L(Hy_Λ, HE_Λ(f(y_Θ,E_Θ; θ))) in addition to the loss function in [2] (Fig. 1), where H denotes a sharpening filter, Λand Θ are disjoint sub-sets of Ω, L(·) is the l₁-l₂loss in [2], and (y_,E; θ) denotes output of the network with parameters θ.

LGE Data and Training: 3D LGE was acquired axially at 1.5T using 32-channel coil array with imaging parameters: resolution=1.2×1.2×1.2 mm³, FOV=320×320×100 mm³, ACS =40×24, R=3 acceleration. The data was divided into smaller 20×270×102 sub-volumes by taking IFFT in k_x direction [2]. PG-DL training was performed on 198 slabs from 10 subjects. Conventional normalized l₁-l₂ loss, along with proposed sharpening loss were utilized. The data was further retrospectively subsampled to R=6 by keeping a 24×24 ACS region representing higher acceleration. Testing was performed on 6 new subjects.

Results:

Fig. 2 shows representative results. Both PG-DL strategies successfully reconstruct the R = 6 image. PG-DL with sharpening loss demonstrates improved recovery of fine-details, matching the R = 3 reconstruction.

Conclusion:

The proposed sharpening loss improves the recovery of fine-details of self-supervised PG-DL, especially in relatively high acceleration rates for high-resolution LGE.