Machine Learning and Artificial Intelligence

1348639 - Deep learning-based acceleration of Compressed SENSE Cardiac MR imaging – accelerating total scan-times and reducing the number of breath holds.

Friday, January 27, 2023

12:20 PM – 1:20 PM

Location: Pacific Jewel

Presenting Author(s)

HL

Huangling Lu, MD, PhD

Radiology Resident
Leiden University Medical Center
Leiden, Zuid-Holland, Netherlands

Primary Author(s)

HL

Huangling Lu, MD, PhD

Radiology Resident
Leiden University Medical Center
Leiden, Zuid-Holland, Netherlands

Co-Author(s)

Joe f. Juffermans, MSc

PhD Candidate
Leiden University Medical Center
Leiden, Zuid-Holland, Netherlands
NP

Nicola Pezzotti, PhD

Senior Scientist
Philips Research, Netherlands
MS

Marius Staring, PhD

Professor of Machine Learning for Medical Imaging
Leiden University Medical Center, Netherlands
Hildo Lamb, MD, PhD

Radiology Professor
Leiden University Medical Center
Leiden, Zuid-Holland, Netherlands

Background: With compressed sensing (CS) undersampled data points are used for MR image reconstruction to reduce acquisition times with preservation of SNR, but CS tends to simplify image content with higher levels of acceleration. Deep learning (DL) reconstruction methods could accelerate the acquisition process with preservation of high image quality by learning from high complexity images. In cardiac MR imaging high levels of acceleration allows multi-slice imaging during one breath-hold (BH) which could reduce scan-times significantly. We investigate the feasibility of a prospectively assessed DL-based reconstruction technique combined with different levels of acceleration using Compressed Sensing artificial intelligence framework in cardiac MR imaging.

Methods:

A total of 25 healthy subjects are prospectively included. Cardiac MR examinations were performed on a 3T scanner. Whole-heart 14 slices bTFE-SA cine images were acquired using CS only and CS artificial intelligence framework with prospectively assessed DL-based reconstructions (CSAI), both with acceleration factors 1/2/4/6/8/10. BH-time was kept under 15s. Number of BHs, total scan-times including 15s pause between BHs and image quality were assessed. Quantitative and qualitative analyses including biventricular function and visual expert scoring were compared using the two-sided paired T-test and Wilcoxon signed-rank test respectively. Visual expert scoring was performed with focus on blood-to-myocardium contrast, endocardial edge delineation and presence of artifacts.

Results:

Number of BHs with CS-1 as reference (no acceleration and no DL-based reconstruction) decreased from 14 (total scan-time 405s, 1 slice/BH) to 7 BHs with CSAI-2 (total scan-time 195s, 2 slices/BH), to 5 BHs with CSAI-4 (total scan-time 125s, 3 slices/BH) and to 4 BHs with CSAI-6 (total scan time 85s, 4 slices/BH). With CSAI-4 there is a 69% reduction of total scan time. As compared to CS-1 as reference, preliminary results of 4 healthy subjects revealed no significant differences in biventricular end-diastolic (EDV), end-systolic (ESV), stroke volumes, ejection fractions (EF), cardiac outputs and LVmass for CSAI-2/4/6. Compared to CS-1, blood-to-myocardium contrast and endocardial edge delineation were similar for CSAI-2 and CSAI-4. Artefact scoring were similar for CSAI-2 and slightly inferior for CSAI-4. CSAI-6/8/10 were inferior for blood-to-myocardium contrast, endocardial edge delineation and artefact scoring. CSAI-4 were considered excellent to diagnostically adequate with CSAI-6 as adequate to suboptimal.

Conclusion:

Using prospectively Deep Learning-based reconstructions to accelerate whole-heart SA cine imaging is feasible with a reduction of 69% in total scan-time with Compressed Sensing artificial intelligence framework acceleration factor 4 (CSAI-4). CSAI-4 showed unaffected quantitative and qualitative performance considering blood-to-myocardium contrast and endocardial edge delineation and slightly outperformance in artefact scoring.