Publication Details:

Accelerated Cardiac MR Imaging using a Resolution Enhancement Generative Adversarial Inline Neural Network

Yoon S, Nakamori S, Amyar A, Assana S, Cirillo J, Morales M, Chow K, Bi X, Pierce P, Goddu B, Rodriguez J, Long N, Manning WJ, Nezafat R

Abstract

Background: Cardiac cine can benefit from deep learning-based image reconstruction to reduce scan time and/or increase spatial and temporal resolution.

Purpose: To develop and rigorously evaluate a resolution enhancement generative adversarial inline neural network (REGAIN) that can be readily combined with parallel imaging (PI) or compressed sensing (CS).

Materials and Methods: REGAIN is built on the enhanced super-resolution generative adversarial network by removing upsampling layers with an additional L₁ fast Fourier transform loss. REGAIN was trained using cine images from retrospectively identified 1616 patients (920 males; 56 years ± 16). The trained model was implemented inline and evaluated in 181 subjects (92 males; 48 years ± 20), prospectively enrolled from September 2021 to September 2022. REGAIN was evaluated in (a) breath-hold ECG-gated segmented cine and (b) free-breathing real-time cine. Images were collected with a reduced spatial resolution along the phase-encoding direction using PI or CS, and afterward, spatial resolution was restored using REGAIN. For comparison, the standard PI-accelerated cine images were also collected. The left-ventricular function, volume, and myocardial strain were assessed. Quantitative image sharpness, diagnostic quality, and artifacts were also evaluated. Continuous and categorical variables were compared using Wilcoxon signed-rank tests. Measurement agreement and reproducibility were assessed with Bland-Altman analysis and intraclass correlation coefficients.

Results: REGAIN showed excellent agreement in left-ventricular parameters compared to standard PI, and similar diagnostic quality or artifact (all P > .05). REGAIN enabled the acquisition of more slices per breath-hold compared to standard PI (3.1±1.6 slices vs. 1 slice, respectively). Also, left-ventricular function or volume of REGAIN-accelerated free-breathing real-time cine (20 seconds of total scan time) had excellent agreements with standard PI-accelerated cine (6 minutes of total scan time).

Conclusion: REGAIN-accelerated cardiac cine enables accurate quantification of cardiac function, volume, and myocardial strain.

Publication

Yoon S, Nakamori S, Assana S, Cirillo J, Morales M, Chow K, Bi X, Pierce P, Goddu B, Rodriguez J, Long N, Manning WJ, Nezafat R. Accelerated Cardiac MR Imaging using Resolution Enhancement Generative Adversarial Inline Neural Network. Nov 2022

Code(s)

The goal of this project is to develop an accelerated cardiac MRI from under-sampled k-space in the phase-encoding direction using generative adversarial neural network.

Available Data

This dataset presents image data 1) Breath-hold ECG-gated segmented Cine 2) Free-breathing Real-time Cine (2022-11-07)

Multimedia

ECG-gated Segmented Cine (Short-axis Orientation)

ECG-gated Segmented Cine (2-Chamber Orientation)

ECG-gated Segmented Cine (4-Chamber Orientation)

Leading Author

      Siyeop Yoon, PhD