Deep learning-accelerated 3D flair for white matter lesion detection in multiple sclerosis: a feasibility study

Deep learning (DL)-based image reconstruction (DLBIR) techniques promise accelerated MRI acquisitions with enhanced image quality. Herein, we compare the image quality of a DLBIR-based 3D FLAIR (3D-FLAIR_DL) with conventional 3D FLAIR (3D-FLAIR_SOC) in a cohort of multiple sclerosis (MS) patients.

Our prospective, reader-blinded study, included 26 MS patients who underwent both sequences on a 3T scanner during the same study session over three months. Two neuroradiologists assessed noise, artifacts, sharpness, overall image quality, and diagnostic confidence using 4-point Likert-like scales. Lesion conspicuity was graded for lesions < 3 mm and ≥ 3 mm. Quantitative metrics included lesion count, apparent signal-to-noise ratio (aSNR), and contrast-to-noise ratio (aCNR). A composite gold standard was used to calculate sensitivity and precision.

3D-FLAIR_DL showed significantly better qualitative scores (p < 0.001) with high inter-reader agreement. Lesion conspicuity and diagnostic confidence improved, especially for < 3 mm lesions. DLBIR images detected 29 additional lesions (10 ≥ 3 mm-sized lesions and 19 < 3 mm lesions), yielding higher sensitivity (99.5% vs. 88.8% for < 3 mm; 100% vs. 97.9% for ≥ 3 mm) without false positives. aSNR and aCNR were significantly higher for DLBIR images (p < 0.001). Acquisition time was reduced by 32% (3:54 vs. 5:44 min).

DLBIR 3D FLAIR significantly improves lesion detection and image quality in MS, supporting its potential integration into standard imaging protocols. As DLBIR algorithms evolve, further validation in larger, diverse cohorts will be essential.