# A pilot study assessing the clinical utility of deep learning-reconstructed 3D-echo-planar-imaging-based quantitative susceptibility mapping in multiple sclerosis

## TL;DR

This study shows that deep learning improves the quality of MRI scans for multiple sclerosis, making it easier to detect disease markers.

## Contribution

The study demonstrates that deep learning reconstruction enhances 3DEPI-based QSM for multiple sclerosis, improving biomarker detection and image quality.

## Key findings

- DLR-3DEPI-based QSM significantly improves confidence in identifying MS biomarkers compared to conventional reconstruction.
- DLR-3DEPI-based QSM achieves excellent consistency with GRE-based QSM despite reduced acquisition time.
- High inter-method agreement suggests DLR improves image quality without altering clinical perception.

## Abstract

Quantitative susceptibility mapping (QSM) has emerged as a promising paraclinical tool in multiple sclerosis (MS). This retrospective pilot study aims to evaluate whether a recently proposed deep learning-assisted, k-space-operating reconstruction, denoising and super-resolution technique (DLR) applied on 3D-echo-planar-imaging (3DEPI) protocols, has the potential to improve the quality and clinical utility of QSM in MS, at 3T. Secondarily, we assess whether applying DLR vs. a conventional reconstruction (CR) can improve the quality of QSM based on noise-susceptible, fast 3DEPI protocols.

3T MRI 3DEPI-data were acquired on seven MS patients and offline-reconstructed using CR and DLR. A sample size of 433 lesions was identified, based on FLAIR segmentation. Two experts, independently and method-blinded, rated lesion-wise the CR- and DLR-3DEPI-derived QSM, assessing the confidence in identifying paramagnetic rim lesions (PRLs), central vein sign (CVS), QSM hyper/isointense lesions and image quality. Gradient-recalled-echo (GRE), 2- and 1-average 3DEPI (acquisition time: 7:02, 3:44, and 1:56 min, respectively) from a healthy individual were offline-reconstructed using CR and DLR. Derived QSM maps were compared visually and quantitatively.

Deep learning reconstruction-3DEPI-based QSM was rated significantly higher for the confidence in identification of the MS-specific biomarkers (hyper/isointense lesions: P < 0.001, CVS: P = 0.01) and overall image quality (P < 0.001), compared to CR-3DEPI-based. Inter-method agreement was high for both raters (Cohen’s κ = 0.98/0.92), suggesting that DLR improves the quality without changing the rater’s perception of the individual QSM-related clinical findings. Additionally, QSM derived from fast DLR-3DEPI with a fourfold acquisition-time reduction compared to GRE, exhibited excellent visual and quantitative consistency with GRE-based QSM.

Our results constitute a first demonstration of the enhanced quality and clinical utility of the DLR-3DEPI-based QSM in MS.