Deep Learning-Accelerated 3D FLAIR Enables Reliable MS Lesion Detection.
Authors
Affiliations (6)
Affiliations (6)
- From the Imaging Department (S.V., G.B., V.C.), Neurethic Lab, ETHICS EA 7446, Lille Catholic Hospitals, Lille Catholic University, Lille, France [email protected].
- From the Imaging Department (S.V., G.B., V.C.), Neurethic Lab, ETHICS EA 7446, Lille Catholic Hospitals, Lille Catholic University, Lille, France.
- Research and Clinical Translation (D.N.), Magnetic Resonance, Siemens.
- Pixyl, Research and Development Laboratory (V.M.-R.), Grenoble, France.
- Biostatistics Department (L.N., M.B.), Delegation for Clinical Research and Innovation, Lille Catholic Hospitals, Lille Catholic University, Lille, France.
- Department of Neurology (A.K.), Neurethic Lab, ETHICS EA 7446, Lille Catholic Hospitals, Lille Catholic University, Lille, France.
Abstract
Deep learning-based reconstruction has the potential to shorten MRI acquisition while preserving diagnostic image quality, but its reliability for MS lesion detection on 3D FLAIR requires validation. Our aim was to evaluate the diagnostic performance and image quality of a deep learning-reconstructed 3D FLAIR sequence in detecting demyelinating lesions in patients with MS, compared with the conventional reference FLAIR sequence, and to assess the impact of different head coil configurations. In this prospective study, 76 patients with MS underwent 3T MRI using both reference and deep learning-reconstructed FLAIR sequences, with identical spatial and contrast parameters. Imaging was alternately performed using either a 20- or a 64-channel head coil. Two blinded radiologists independently assessed the image quality using a 5-point Likert scale and evaluated lesion detection. The SNR and contrast-to-noise ratios were measured, and automated lesion detection was performed using a certified artificial intelligence device. All clinically relevant lesions (≥3 mm) were detected on the deep learning-reconstructed FLAIR sequence with complete agreement between readers. Six subthreshold lesions (<3 mm) were missed in 3 patients (3.95%; 95% CI, 1.03%-11.88%), all scanned with the 20-channel coil. The reference FLAIR sequence had slightly higher image quality scores overall (mean, 4.86 [SD, 0.35] versus 4.72 [SD, 0.52] <i>P</i> = .01). However, no difference in quality or lesion detection was observed between sequences when using the 64-channel coil. The deep learning-reconstructed FLAIR sequence demonstrated significantly higher SNR and contrast-to-noise ratios across all cases (<i>P</i> < .001), with an additional contrast increase using the 64-channel coil (<i>P</i> = .007). Although subjective evaluation favored the standard reconstruction, quantitative metrics indicated improved image quality with deep learning reconstruction. Automated analysis using an artificial intelligence-based tool confirmed complete concordance in lesion detection between the two sequences. The deep learning-reconstructed FLAIR enables reliable detection of demyelinating lesions in MS with significantly reduced scan time without the loss of diagnostic information. Use of a 64-channel coil further enhances image quality, supporting the integration of this accelerated technique into clinical practice.