Abstract
BACKGROUND AND PURPOSE: Limbic encephalitis is an autoimmune disease. A variety of autoantibodies have been associated with different subtypes of limbic encephalitis, whereas its MR imaging signature is uniformly characterized by mesiotemporal abnormalities across subtypes. Here, we hypothesized that patients with limbic encephalitis would show subtype-specific mesiotemporal structural correlates, which could be classified by supervised machine learning on an individual level. MATERIALS AND METHODS: T1WI MPRAGE scans from 46 patients with antibodies against glutamic acid decarboxylase and 34 patients with antibodies against the voltage-gated potassium channel complex (including 10 patients with leucine-rich glioma-inactivated 1 autoantibodies) and 48 healthy controls were retrospectively ascertained. Parcellation of the amygdala, hippocampus, and hippocampal subfields was performed using FreeSurfer. Volumes were extracted and compared between groups using unpaired, 2-tailed t tests. The volumes of hippocampal subfields were analyzed using a multivariate linear model and a binary decision tree classifier. RESULTS: Temporomesial volume alterations were most pronounced in an early stage and in the affected hemispheric side of patients. Statistical analysis revealed antibody-specific hippocampal fingerprints with a higher volume of CA1 in patients with glutamic acid decarboxylase-associated limbic encephalitis (P = .02), compared with controls, whereas CA1 did not differ from that in controls in patients with voltage-gated potassium channel complex autoantibodies. The classifier could successfully distinguish between patients with autoantibodies against leucine-rich glioma-inactivated 1 and glutamic acid decarboxylase with a specificity of 87% and a sensitivity of 80%. CONCLUSIONS: Our results suggest stage-, side- and antibody-specific structural correlates of limbic encephalitis; thus, they create a perspective toward an MR imaging-based diagnosis.