Application of deep-learning to the seronegative side of the NMO spectrum

• Published in: Journal of neurology 2022-03, Vol.269 (3), p.1546-1556
• Main Authors: Cacciaguerra, Laura, Storelli, Loredana, Radaelli, Marta, Mesaros, Sarlota, Moiola, Lucia, Drulovic, Jelena, Filippi, Massimo, Rocca, Maria A.
• Format: Article
• Language: English
• Subjects: Algorithms; Aquaporin 4; Aquaporins; Autoantibodies; Deep Learning; Humans; Learning; Magnetic resonance imaging; Medicine; Medicine & Public Health; Multiple sclerosis; Myelin; Myelin-Oligodendrocyte Glycoprotein; Myelitis; Neural networks; Neuritis; Neurology; Neuromyelitis; Neuromyelitis Optica - diagnostic imaging; Neuroradiology; Neurosciences; Oligodendrocyte-myelin glycoprotein; Optic neuritis; Original Communication; Patients; Phenotypes
• ISSN: 0340-5354; 1432-1459
• DOI: 10.1007/s00415-021-10727-y

Objectives To apply a deep-learning algorithm to brain MRIs of seronegative patients with neuromyelitis optica spectrum disorders (NMOSD) and NMOSD-like manifestations and assess whether their structural features are similar to aquaporin-4-seropositive NMOSD or multiple sclerosis (MS) patients. Patients and methods We analyzed 228 T2- and T1-weighted brain MRIs acquired from aquaporin-4-seropositive NMOSD ( n  = 85), MS ( n  = 95), aquaporin-4-seronegative NMOSD [ n  = 11, three with anti-myelin oligodendrocyte glycoprotein antibodies (MOG)], and aquaporin-4-seronegative patients with NMOSD-like manifestations (idiopathic recurrent optic neuritis and myelitis, n  = 37), who were recruited from February 2010 to December 2019. Seventy-three percent of aquaporin-4-seronegative patients with NMOSD-like manifestations also had a clinical follow-up (median duration of 4 years). The deep-learning neural network architecture was based on four 3D convolutional layers. It was trained and validated on MRI scans of aquaporin-4-seropositive NMOSD and MS patients and was then applied to aquaporin-4-seronegative NMOSD and NMOSD-like manifestations. Assignment of unclassified aquaporin-4-seronegative patients was compared with their clinical follow-up. Results The final algorithm differentiated aquaporin-4-seropositive NMOSD and MS patients with an accuracy of 0.95. All aquaporin-4-seronegative NMOSD and 36/37 aquaporin-4-seronegative patients with NMOSD-like manifestations were classified as NMOSD. Anti-MOG patients had a similar probability of being NMOSD or MS. At clinical follow-up, one unclassified aquaporin-4-seronegative patient evolved to MS, three developed NMOSD, and the others did not change phenotype. Conclusions Our findings support the inclusion of aquaporin4-seronegative patients into NMOSD and suggest a possible expansion to aquaporin-4-seronegative unclassified patients with NMOSD-like manifestations. Anti-MOG patients are likely to have intermediate brain features between NMOSD and MS.