Evaluating SynthSR for MRI harmonization and quantitative analysis of hippocampal volume in autoimmune neurologic diseases

Variability in acquisition protocols limit the use of retrospective clinical magnetic resonance imaging (MRI) for quantitative analysis. SynthSR is an AI tool that harmonizes heterogeneous brain scans into synthetic 1-mm isotropic structural images. We evaluated SynthSR performance in autoimmune encephalitis (AE) and mimics. This multicentre study included 102 patients from the Australian Autoimmune Encephalitis Consortium, and 28 controls, comparing high-resolution reference standard T1-weighted structural scans against low-resolution (n = 27) or contrast-enhanced counterparts (n = 76), with one patient having both. Images were processed via SynthSR, with hippocampal volume measured using FastSurfer. SynthSR processing systematically overestimated hippocampal volume, with high inter-individual variability. Nevertheless, relative consistency between SynthSR-processed and original reference standard remained high (ICC = 0.88, 95% CI 0.74–0.94). Contrast enhancement did not substantially alter performance, but inclusion may require caution. While agreement between synthetic images derived from low-resolution scans and reference standard was moderate (ICC = 0.75, 95% CI 0.36–0.90), agreement with SynthSR-processed reference standard was excellent (ICC = 0.90, 95% CI 0.76–0.96). Longitudinally, no statistically significant difference in volumetric change was detected, although wide limits of agreement indicated substantial individual-level variability. SynthSR facilitates harmonization of heterogeneous MRI for group-level volumetric analyses, provided a consistent processing pipeline is maintained. However, it is not reliable for individual-level patient assessments.