A total of 94 participants were included in this cross-sectional case-control study and assigned to one of three groups.

All long-COVID patients were recruited from the long-COVID outpatient clinics of the Department of Internal Medicine IV (Infectiology) and the Department of Neurology at the University Hospital of Jena. Participants were included based solely on a confirmed long-COVID diagnosis to ensure a broad symptom spectrum; without requiring specific symptom profiles. A positive polymerase chain reaction (PCR) test was used to verify SARS-CoV-2 infection at both clinics. The diagnosis of long-COVID was based on the 2021 NICE and AWMF guidelines, which defined long-COVID as symptoms newly occurring after SARS-CoV-2 infection, not explained by other medical conditions, and persisting for > 4 weeks after infection onset (27, 28). The presence and duration of long-COVID symptoms were systematically recorded in a descriptive way via self-report, the symptom spectrum was as multi-facetted as in representative population-based studies with pronounced fatigue and cognitive impairment (29). Accordingly, 61.1% of the PCn cohort and 73.1% of the PCcog cohort reported cognitive impairment, while 75% of the PCn group and 96.2% of the PCcog group reported fatigue. Additional symptoms are listed in Table 1 (see also Supplementary Table 2 ). On average, the PCn group experienced 5.14 long-COVID symptoms, whereas the PCcog group reported 6.9 (see Table 1 ). Medical history, including information on the timing of previous SARS-CoV-2 infection(s), COVID-vaccination status and the severity of the acute COVID-19 infection as defined by the World Health Organization (WHO), was collected by a certified physician. Severity of COVID-19 was categorized into five levels based on the clinical manifestations and disease progression: uninfected (score 0), ambulatory mild disease (scores 1-3), hospitalized moderate disease (scores 4-5), hospitalized severe disease (scores 6-9), and deceased (score 10) (30). The mean WHO severity among long-COVID participants (both groups) was 2.27 (range 1 to 5, SD 0.89). Patients were enrolled and assessed in the study on average 9.9 months after infection (range 1 to 24.5 months, SD 4).

Long-COVID individuals were stratified based on their cognitive performance assessed using the MoCA score (31): those with a MoCA score ≥ 26 were assigned to the cognitively unimpaired long-COVID group (PCn; n=36; mean age 41.8; SD 11.4), while participants with a MoCA score < 26 were assigned to the cognitively impaired long-COVID group (PCcog; n= 28; mean age 50.04; SD 15). A healthy control group (HC n= 30; mean age yrs 42.0; SD 10.8) was included for comparison. All participants self-identified as either male or female, and subsequent analyses were stratified accordingly by sex. To enable sex-specific analysis, each group was further subdivided by sex. Female and male participants were labeled f-PCn/m-PCn, f-PCcog/m-PCcog and f-HC/m-HC, respectively. Group distributions were as follows: the cognitively unimpaired long-COVID group (PCn) included 23 female (f-PCn) and 13 male participants (m-PCn); the cognitively impaired long-COVID group (PCcog) comprised 15 female (f-PCcog) and 13 male participants (m-PCcog); and the healthy control group (HC) consisted of 13 female (f-HC) and 17 male participants (m-HC)

The HC group was recruited via announcements of the study in the local newspaper and on social media accounts of the clinic. To ensure that participants in the HC group had not previously been infected with SARS-CoV-2 at the time of assessment, the absence of SARS-CoV-2–specific antibodies was confirmed by serological testing. Serology results were further validated using a Western blot to distinguish between antibodies arising from natural infection and those induced by vaccination.

All participants were screened via telephone to exclude those with past or current psychiatric disorders and current addiction. Additional exclusion criteria included contraindications for magnetic resonance imaging (MRI), diseases of the nervous system, a history of traumatic brain injury or loss of consciousness, unmedicated internal medical conditions and severe cognitive impairment (IQ < 80). To exclude the latter, IQ was estimated using the German Multiple Choice Vocabulary Test B (MWT-B) (32).

The study protocol was approved by the local ethics committee of the Jena University Hospital. All participants gave written informed consent. Tables 1 , 2 summarize the demographic and psychometric data.

Each participant underwent high-resolution T1-weighted MRI scans using a standard quadrature head coil and an axial 3-dimensional magnetization-prepared rapid gradient echo (MP-RAGE) sequence (TR 2400 ms, TE 2.22 ms, α 8°, 208 contiguous sagittal slices, FoV 256 mm, voxel resolution 0.8 x 0.8 x 0.8 mm, acquisition time 6:38 min) on a 3 Tesla Siemens Prisma fit (Siemens, Erlangen, Germany). All scans were checked for imaging artefacts.

VBM analysis was performed using the CAT12 (Computational Anatomy Toolbox 12) developed by the Structural Brain Mapping group at University Hospital Jena, Germany, and implemented in SPM12 (Statistical Parametric Mapping, Institute of Neurology, London, UK). The T1-weighted images underwent bias-field correction to account for field homogeneity, followed by spatial normalization using the DARTEL algorithm (33). The images were segmented into white matter, gray matter and cerebrospinal fluid (34). To improve the accuracy of the segmentation process, it was extended to account for partial volume effects (35). Adaptive maximum a posteriori estimation was applied, and a hidden Markov random field model was used. To exclude artefacts at the grey-white matter boundary, an internal grey matter threshold of 0.1 was applied. After pre-processing, all scans were subjected to an automated quality control protocol. As CAT12 does not apply a fixed motion-exclusion threshold; instead, we used its image quality ratings and visual inspection to identify and exclude scans affected by motion artifacts. 2 participants from the long- COVID patient group had to be excluded from further analysis due to poor image quality at that point. The remaining images were smoothed with an 8 mm FWHM Gaussian kernel, which represents a widely used compromise between sensitivity and anatomical specificity, satisfies assumptions of Gaussian random field theory for voxel-based inference, and facilitates comparability with prior morphometry studies (36, 37).

The statistical analysis was performed using the general linear model approach, implemented in SPM12. Groups were compared using two-sample t-tests. To account for associated variance, we included total intracranial volume (TIV), age and sex as confounding variables in the VBM analysis. As a non-parametric statistic, we applied threshold-free cluster enhancement (TFCE) with 5000 permutations (38) to all analyses and corrected for multiple comparisons via the family wise error method (FWE) at p<0.05. For atlas labelling of significant clusters we used the Neuromorphometrics Atlas (http://www.neuromorphometrics.com). We first performed this analysis for the overall groups of healthy controls (HC), long-COVID patients without cognitive symptoms according to MoCA (PCn) and long-COVID patients with cognitive impairment (PCcog). In a second step we performed these analyses separately for female and male participants, adding TIV and age as confounding variables. Complementary statistical analysis of clinical and demographic data was conducted in IBM SPSS Statistics (version 29 0.2.0.). A significance level of p≤ 0.05 was applied. The chi-square test assessed sex distribution among groups. Kruskal-Wallis tests were used to compare age, years of education and TIV. Mann-Whitney U tests examined differences in WHO severity of acute COVID-19 infection and time since infection between PCn, f-PCn, m-PCn and PCcog. f-PCcog, m-PCcog. ANCOVA was performed to assess the influence of covariates on MoCA scores.

Overall analyses and sex-separated analyses of the HC, PCn and PCcog groups revealed several significant clusters (p<0.05, FWE-corrected) showing GMV alterations between the HC and both PC groups.

In the HC>PCcog comparison, several brain regions showed overlapping GMV reductions in both sexes, indicating shared structural alterations. Both m-PCcog and f-PCcog participants exhibited GMV reductions in the left ventral diencephalon, thalamus, hippocampus, and parahippocampal gyrus. Among these, the thalamus and hippocampus are particularly critical for cognitive function (39, 40). Men additionally showed alterations in right occipital areas. These findings are consistent with previous findings on GMV alterations associated with long-COVID. Diez-Cirarda et al. reported GMV reductions in limbic areas, among others, associated with cognitive dysfunction (41). Similarly, a UK Biobank study investigating brain changes in 401 participants over a long-term follow-up period found reductions in gray matter thickness and tissue contrast in the parahippocampal gyrus and orbitofrontal cortex, along with an overall greater progression of cognitive decline in COVID-19 patients (20).

The hippocampus has likewise emerged as a region of concern in individuals recovering from SARS-CoV-2 infection. As a key structure for cognition and particularly episodic memory (42), it is critically implicated in the pathophysiology of neurodegenerative disorders, such as Alzheimer`s disease, and psychiatric conditions including major depressive disorder (43–45). In the context of long-COVID, hippocampal structural and functional alterations, potentially affecting adult neurogenesis, have been linked to memory loss and an accelerated progression of neurodegenerative processes (46, 47).

In the present study, long-COVID patients exhibited consistently reduced hippocampal GMV compared to HC, in line with findings from Capelli et al., Kamasak et al. and Invernizzi et al. (48–50). However, contrasting results have been reported in two large-scale studies that found increased hippocampal GMV (51), with Lu et al. additionally describing a positive correlation with memory impairment (Lu et al., 2020). These divergent findings underscore the complexity of structural brain alterations associated with long-COVID.

The coexistence of both increases and decreases in GMV of different brain regions may reflect a dynamic interplay between neurodegenerative processes and compensatory mechanisms, such as neuroplasticity (52, 53). Neuroinflammation, a key focus of long-COVID research, is thought to contribute to structural brain changes through cytokine-mediated disruption of the blood-brain barrier, neurovascular damage, and impaired neurogenesis (54–62). Evidence from previous coronavirus outbreaks (SARS, MERS) supports this mechanism (63, 64). These pro-inflammatory responses resemble those implicated in cancer therapy-related cognitive impairment, suggesting shared pathophysiological pathways (65, 66). Under certain conditions, however, microglial activation may promote neurogenesis, depending on cytokine profiles and concentrations (67). This dual role could help explain the heterogeneous pattern of GMV alterations observed in long-COVID.

Beyond hippocampal and cortical regions, the amygdala has been implicated in COVID-19 related neurocognitive changes (50, 68). Invernizzi et al. reported structural and functional alterations not only in the hippocampus but also the left amygdala, where reduced connectivity was shown to specifically mediate spatial working memory deficits (50). A cross-sectional study involving 75 individuals, including COVID-19 survivors with and without brain fog and healthy controls, found that both COVID-19 groups showed reduced gray matter concentrations in the left inferior temporal gyrus, left fusiform gyrus, and right orbital gyri compared to healthy controls. In addition, participants with brain fog exhibited further reductions in the bilateral caudate nuclei, right putamen/pallidum, and amygdala (68).

This highlights the role of limbic circuitry in cognitive sequalae following SARS-CoV-2 infection. Complementary longitudinal work in healthy individuals without prior SARS-CoV-2 infection linked transient volumetric increases in the amygdalae to stress- and anxiety-related processes following the COVID-19 outbreak and lockdown, with GMV gradually decreasing over time after lockdown relief (69). In our study, however, the amygdala did not emerge as a region of interest. Nevertheless, other limbic structures showed relevant alterations.

Although the overall analysis was secondary to the sex-stratified results, it revealed several regions of interest relevant to cognitive functioning, including the hippocampus, entorhinal area, posterior cingulate gyrus, angular gyrus, and planum temporale (see Supplementary Table 1 and Supplementary Figure 1 ).