Differentiation of Anti-NMDAR Encephalitis and Autoimmune Limbic Encephalitis Using Histogram Analysis based on Multiparametric MRI

Objective: To evaluate the combined diagnostic value of clinical features, conventional MRI findings, and histogram metrics derived from multiparametric MRI (mp-MRI) in differentiating anti-N-methyl-D-aspartate receptor (NMDAR) encephalitis from autoimmune limbic encephalitis (ALE). Methods: This retrospective study analyzed baseline clinical and brain MRI data from 76 anti-NMDAR encephalitis and 59 ALE patients. Bilateral hippocampi were manually delineated as regions of interest, and histogram metrics were extracted from fluid-attenuated inversion recovery (FLAIR), T2-weighted imaging (T2WI), and apparent diffusion coefficient (ADC) maps. Clinical characteristics, conventional MRI features, and histogram metrics were compared between the two groups. Independent predictors were identified using logistic regression. Diagnostic models based on clinical variables, conventional MRI, and single-modality or multimodal histogram metrics were constructed and assessed using area under the curve (AUC). Results: Independent clinical predictors included age, prodromal symptoms, psychiatric or behavioral abnormalities, memory impairment, cerebrospinal fluid nucleated cell count, and elevated IgM levels (all P < 0.05). For conventional MRI, bilateral hippocampal volume and abnormal hippocampal signal served as independent predictors (all P < 0.05). The clinical model (AUC = 0.858), conventional MRI model (AUC = 0.770), and mp-MRI histogram model (AUC = 0.737) each outperformed single-modality histogram models. Integrating histogram metrics with clinical and conventional MRI features significantly enhanced diagnostic performance, with the combined mp-MRI model achieving the highest accuracy (AUC = 0.947). Conclusion: Histogram metrics derived from mp-MRI are promising biomarkers for differentiating anti-NMDAR encephalitis from ALE, and the integration with clinical and conventional MRI indicators in combined models significantly improves diagnostic accuracy.