Image Reconstruction (including machine learning)

1348917 - Highly-Accelerated Free-Breathing Perfusion CMR Reconstruction Using Subject-Specific Zero-Shot Deep Learning

Wednesday, January 25, 2023

Location: E-POSTER

Presenting Author(s)

OD

Omer B. Demirel, MSc

PhD Candidate
University of Minnesota
Minneapolis, Minnesota, United States

Primary Author(s)

OD

Omer B. Demirel, MSc

PhD Candidate
University of Minnesota
Minneapolis, Minnesota, United States

Co-Author(s)

CZ

Chi Zhang

PhD Candidate
University of Minnesota, Minnesota, United States
BY

Burhaneddin Yaman, BSc

PhD
University of Minnesota
Saint Paul, Minnesota, United States
SM

Steen Moeller, PhD

Assistant Professor
University of Minnesota
Minneapolis, Minnesota, United States
Chetan Shenoy, MBBS, MS

Associate Professor of Medicine
University of Minnesota
Minneapolis, Minnesota, United States
Sebastian Weingartner, PhD

Assistant Professor
Delft University of Technology
Delft, Zuid-Holland, Netherlands
Tim Leiner, MD, PhD

Professor of Radiology
Mayo Clinic
Rochester, Minnesota, United States
Mehmet Akcakaya, PhD

Associate Professor
University of Minnesota
Minneapolis, Minnesota, United States

Disclosure(s):

Sebastian Weingartner, PhD: No financial relationships to disclose

Tim Leiner, MD, PhD: No relevant disclosure to display

Background: Perfusion CMR is used to functionally assess coronary artery disease. Its resolution and coverage remain limited, despite advances in rapid MRI [1,2]. Existing strategies for accelerated perfusion CMR use conventional methods, such as parallel imaging or compressed sensing [3]. Recently, physics-guided deep learning (PG-DL) has gained interest for rapid MRI, but several challenges remain for its use in perfusion CMR: 1) Lack of high-quality training databases, due to substantial differences of time-varying physiological processes such as breathing or contrast uptakes across subjects, 2) Need to train from sub-sampled data since fully-sampled references cannot be acquired [4]. In this work, we tackle these challenges and improve perfusion CMR using subject-specific zero-shot PG-DL without any training data.

Methods:

Theory: The inverse problem for simultaneous multi-slice (SMS) imaging [1] is given as x^SMS = arg min_x^SMS ||y^SMS_Ω-E^SMS_Ωx^SMS||² + ℜ(x^SMS) where y^SMS_Ωis acquired SMS k-space, Ω is the undersampling pattern, E^SMS_Ω is MR encoding operator, x^SMS is readout concatenated slices of images, and ℜ(.) is a regularizer. Variable splitting is used [4] for solving this, alternating between data-fidelity incorporating MRI physics, and implicit regularization using neural networks. Recently, zero-shot self-supervised PG-DL (ZS-SSDU) [5] was proposed for subject-specific training without a ground truth or training database. ZS-SSDU splits measurements y^SMS_Ω into three disjoint sets θ, Λ, and Γ. θ enforces data-fidelity in the PG-DL network. Λ defines a k-space loss [4], and Γ is used as k-space validation loss for early stopping [5] to avoid overfitting (Fig. 1). Further, a multi-mask approach was used as data augmentation, partitioning Ω\Γ into K disjoint pairs (θ_k,Λ_k).

Experiments: Free-breathing first-pass perfusion was acquired at 3T on 2 subjects, with outer volume suppression [2], SMS=3, uniform in-plane R=4, resolution=1.7×1.7mm², temporal resolution=116ms.

A 3D unrolled PG-DL network was implemented for subject-specific regularization across all time-frames using ZS-SSDU (Fig. 1). 20% of Ω was randomly selected for Γ. K=100 partitions were used for the remainder. Training was stopped once k-space validation loss increased [5]. Comparisons were made to split-slice GRAPPA (SP-SG) and locally-low rank (LLR) regularized reconstruction. Images were assessed by an experienced reader on a 4-point scale (1:best, 4:worst) for aliasing artifacts, SNR, blurring, and overall quality [4].

Results: Figs. 2 and 3 show results from 2 different subjects. SP-SG suffers from noise amplification. LLR reconstruction reduces noise but leaves residual artifacts. Proposed subject-specific PG-DL eliminates noise and aliasing artifacts, yielding high-quality images across time-frames. Image readings align with these observations.

Conclusion: Proposed subject-specific PG-DL leads to improved perfusion CMR, particularly in later time-frames with limited SNR.