SynthSR can harmonize autoimmune brain MRI, but hippocampal volumes drift high
Retrospective MRI analysis has long run into the same obstacle: scans acquired on different protocols, at different resolutions, and sometimes after contrast injection do not behave like a single dataset. In autoimmune neurologic disease, where hippocampal injury can influence diagnosis and follow-up, that inconsistency becomes more than a technical nuisance. It can blur volumetric readouts that clinicians would otherwise want to trust.
This multicenter study tested whether SynthSR, an AI-based harmonization tool, could convert heterogeneous clinical scans into synthetic 1-mm isotropic structural images suitable for quantitative analysis. The cohort came from the Australian Autoimmune Encephalitis Consortium and included 102 patients plus 28 controls. Investigators compared high-resolution reference standard T1-weighted structural scans with lower-quality source images, including 27 low-resolution scans and 76 contrast-enhanced counterparts; one patient contributed both types. Hippocampal volume was then measured with FastSurfer after image processing.
The headline result was hard to ignore: SynthSR systematically overestimated hippocampal volume and did so with substantial inter-individual variability.
Agreement stayed strong at the group level, even as individual estimates wandered
Despite that upward volume bias, the overall relationship between SynthSR-processed scans and the original reference standard remained strong. The intraclass correlation coefficient was 0.88, with a 95% confidence interval of 0.74 to 0.94. In practical terms, the method tracked the reference standard closely enough to support group-level comparisons when the same processing pipeline was used consistently. The size of the coefficient matters because it suggests the signal is still detectable across a mixed real-world dataset, even when the input images are imperfect.
The picture changed when the analysis moved toward more granular comparisons. Agreement between synthetic images derived from low-resolution scans and the reference standard was only moderate, with an intraclass correlation coefficient of 0.75 and a wide 95% confidence interval of 0.36 to 0.90. When those low-resolution scans were compared with SynthSR-processed reference standard images, agreement improved to excellent, with an intraclass correlation coefficient of 0.90 and a 95% confidence interval of 0.76 to 0.96. That pattern suggests the harmonization step can rescue some of the variability introduced by degraded acquisition, but only within the constraints of the processing framework itself.
Contrast enhancement did not substantially alter performance. The inclusion of contrast-enhanced scans still requires caution, but the tool did not appear to collapse when gadolinium was part of the imaging mix.
Longitudinal change was not significantly different, yet the patient-level spread stayed wide
Serial analysis did not identify a statistically significant difference in volumetric change over time. Even so, the limits of agreement were wide enough to show substantial individual-level variability. In a disease such as autoimmune encephalitis, where hippocampal damage can evolve over time and where subtle structural loss may be clinically meaningful, wide dispersion at the single-patient level is not a minor technical footnote. It limits confidence in using SynthSR as a bedside decision tool for an individual person’s scan-to-scan change.
The biology helps explain why this problem is difficult. Hippocampal measurements sit at the intersection of acquisition quality, tissue contrast, and segmentation algorithms. When source scans vary in resolution or include post-contrast signal changes, an algorithm tasked with generating a standardized structural image may impose a form of synthetic regularity that smooths over true anatomy as well as noise. That tradeoff can improve consistency for cohort studies while still distorting absolute volume estimates in a specific patient.
A useful research tool, not a standalone clinical ruler
The most defensible interpretation is narrow and clinically important. SynthSR appears able to facilitate harmonization of heterogeneous MRI datasets for group-level volumetric analysis, but only if the processing pipeline remains consistent from case to case. The tool is not reliable for individual-level patient assessment.
For neuroimmunology clinics, that distinction affects how retrospective imaging repositories are used. It supports exploratory analysis across legacy scans collected under different protocols, which is exactly where many autoimmune encephalitis cohorts live. It does not justify treating a single synthetic hippocampal volume as a substitute for expert radiologic judgment, especially when decisions hinge on whether an individual patient is losing tissue over time.
The study also reinforces a familiar imaging principle: automation can standardize a dataset, but it cannot fully erase the physics of image acquisition. Low-resolution scans carry structural information differently from high-resolution reference images, and contrast administration introduces a separate layer of variability. A model may compensate for some of that drift, but the correction is not identical to native acquisition.
Why this matters for autoimmune encephalitis research
Autoimmune encephalitis often prompts serial MRI review because hippocampal involvement can be central to the disease process. Quantitative volumetry is attractive in that setting because it promises a more objective readout than visual inspection alone. The present data show that this promise is real at the population level, but fragile when applied to one patient at a time.
In the context of a multicenter retrospective study, an intraclass correlation coefficient of 0.88 is encouraging. It means harmonization can make older, heterogeneous scans usable for analysis when the question is framed across groups. The same dataset also shows why the hippocampus is a demanding target: systematic overestimation, wide limits of agreement, and performance shifts across source image types make it risky to use as a precise clinical biomarker without a stable acquisition and processing chain.
The practical takeaway is restrained and clear. SynthSR may open retrospective autoimmune neurology datasets to more rigorous volumetric work, but the resulting numbers should be treated as research measurements rather than individual patient verdicts. The remaining problem is not whether harmonization helps at all; it is how much error can be tolerated before a synthetic hippocampal volume stops reflecting the patient sitting in front of the scanner.