Enhanced Spinal Cord Lesion Detection in MS Using White-Matter-Nulled 3D MPRAGE with Deep Learning Reconstruction

Fanny Munsch, Amaury Ravache, Takayuki Yamamoto, Bei Zhang, Marion Lacoste, Hikaru Fukutomi, Pauline Buissonnière, Aurélie Ruet, Jean-Christophe Ouallet, Thomas Tourdias, Vincent Dousset
American Journal of Neuroradiology. 2025-12-11; 47(1): 225-233
DOI: 10.3174/ajnr.a8950

Munsch F(1), Ravache A(2), Yamamoto T(3), Zhang B(4), Lacoste M(2), Fukutomi H(5), Buissonnière P(6), Ruet A(6)(7), Ouallet JC(6), Tourdias T(8)(2)(7), Dousset V(8)(2)(7).

Author information:
(1)From the Institut de Bio-imagerie IBIO (F.M., T.T., V.D.), Université
Bordeaux, Bordeaux, France .
(2)CHU de Bordeaux, Neuroimagerie Diagnostique et Thérapeutique (A. Ravache,
M.L., T.T., V.D.), Bordeaux, France.
(3)Department of Diagnostic Imaging and Nuclear Medicine, Graduate School of
Medicine (T.Y.), Kyoto University, Kyoto, Japan.
(4)Canon Medical Systems Europe (B.Z.), Amstelveen, the Netherlands.
(5)Kansai Electric Power Hospital (H.F.), Osaka, Japan.
(6)CHU de Bordeaux, Service de Neurologie (P.B., A. Ruet, J.-C.O.), Bordeaux,
France.
(7)Neurocentre Magendie (A. Ruet, T.T., V.D.), INSERM U1215, Université
Bordeaux, Bordeaux, France.
(8)From the Institut de Bio-imagerie IBIO (F.M., T.T., V.D.), Université
Bordeaux, Bordeaux, France.

BACKGROUND AND PURPOSE: Spinal cord (SC) lesions are critical in MS diagnosis
and progression, yet their detection remains challenging. Conventional sequences
such as 2D T2-weighted FSE and STIR have suboptimal sensitivity, both in the
cervical and thoracic spine. This study evaluated the diagnostic performance of
a new imaging technique for SC lesion detection, the 3D white-matter-nulled
(WMn) MPRAGE sequence, compared with conventional MRI sequences.
MATERIALS AND METHODS: Thirty-eight patients with MS or clinically isolated
syndrome were prospectively evaluated with 3T SC MRI, acquiring 2D T2-weighted
FSE, 2D STIR, 3D MPRAGE, and 3D WMn. A deep learning denoising method was
applied to 3D WMn to compensate the inherent low signal. Four blinded
neuroradiologists independently assessed lesion count, confidence in lesion
detection, and image quality (eg, artifacts). Contrast-to-noise ratio (CNR) was
also computed for all lesions detected on all sequences. Statistical comparisons
were performed across sequences.
RESULTS: In the cervicothoracic spine, the 3D WMn sequence detected
significantly more lesions than 2D T2-weighted FSE (+62%; P < .001), STIR (+47%;
P < .05), and 3D MPRAGE (+50%; P < .01). In the thoracolumbar spine, the 3D WMn
sequence detected significantly more lesions than 2D T2-weighted FSE (+53%, P <
.05). The 3D WMn sequence demonstrated a higher CNR and improved lesion
conspicuity while exhibiting fewer artifacts than STIR. These advantages
contributed to greater rater confidence in lesion detection by using the 3D WMn
sequence, with over 75% of strong confidence reported by the 2 most experienced
raters, along with the highest interreader agreement.
CONCLUSIONS: The 3D WMn sequence associated with deep learning-based denoising
significantly improves SC lesion detection, outperforming all conventional
sequences. Three-dimensional WMn represents, then, a promising alternative for
SC imaging in MS.