Characterizing progressive cortical thinning in multiple system atrophy: a longitudinal MRI study

Abstract

Longitudinal cortical involvement in Multiple System Atrophy (MSA) remains poorly investigated, as the disease is classically characterized by infratentorial and striatal atrophy. In this study, we characterized the longitudinal progression of cortical thinning in MSA, together with subcortical atrophy, using MRI. We studied 36 clinically diagnosed patients (25 cerebellar, 11 Parkinsonian variant) who underwent T1-weighted MRI at baseline and follow-up (median 0.96 years; interquartile range 0.20). Images were processed with state-of-the-art neuroimaging tools. Longitudinal changes in cortical thickness were assessed using linear mixed-effects models within anatomical (Desikan-Killiany) and functional (Schaefer) cortical parcellations, while volumetric changes were evaluated in subcortical and infratentorial regions using the same statistical framework. Patients exhibited progressive cortical thinning affecting motor, frontal, cingulate and insular regions, with involvement of functional networks including the somatomotor, default mode, frontoparietal and salience/ventral attentional networks (all Bonferroni-corrected, p ≤ 0.001). Over the 1-year time span, subcortical and infratentorial regions, including the pons, putamen, cerebellar gray and white matter, showed significant volume loss (Bonferroni-corrected, p ≤ 0.001) and these changes were more marked than the observed cortical thinning. No significant associations were observed between regions showing thinning over time and disease severity or duration. Our results provide a comprehensive longitudinal characterization of distributed cortical thinning, consistent with the view of MSA as a multisystem disorder extending beyond the infratentorial and subcortical degeneration.

Introduction

Multiple system atrophy (MSA) is a rare, rapidly progressing, adult-onset neurodegenerative disorder characterized by widespread neuronal loss and gliosis associated with glial cytoplasmic inclusions enriched with α-synuclein in oligodendrocytes and neurons^1,2. Average survival rate is approximately 7.5 years after symptom onset³. Patients show a diverse spectrum of motor and non-motor symptoms⁴ affecting—in varying combinations and severities—the extrapyramidal, cerebellar, pyramidal, and autonomic systems^5,6. The disease indeed places a significant burden on patients, their caregivers and healthcare professionals. Magnetic resonance imaging (MRI) is widely adopted in clinical practice and is instrumental for diagnosis⁷. The main structural MRI feature is prominent nigrostriatal degeneration in the Parkinsonian variant (MSA-P), and olivopontocerebellar atrophy in the cerebellar variant (MSA-C). These atrophic patterns can be identified on T1-weighted MRI scans^8–10, indicating their potential use as both endpoints in clinical trials of disease-modifying therapies and as digital biomarkers for monitoring disease progression^2,6,7. Regarding the cerebral cortex, regional, voxel-based morphometry (VBM) and cortical thickness (CT) studies have reported cortical changes^10–17 in MSA, although longitudinal CT has been assessed in a limited number of studies^18–20. The brain is a complex network of functional and structural connections that can be impacted by major subcortical and infratentorial atrophy occurring in MSA. Striatal-thalamocortical loops may be affected by putaminal gray matter loss, reducing input to frontal^21,22 and motor cortices²³, and eventually leading to mild cognitive impairment^24,25 (MCI) which has been reported in up to 44% of patients with MSA^26,27, whereas dementia is quite rare. Disruption of the cerebellum-cortex circuit due to prominent cerebellar loss can lead to reduced functional activity in upstream motor cortices²⁸, as well as altered connectivity with parietal and temporal areas, as measured with resting-state fMRI²⁹. Furthermore, pathological aggregates of α-synuclein have been reported to accumulate in sensorimotor cortices, cingulate cortex, and temporal areas in post-mortem studies^23,30,31. Within this framework, we hypothesize that cortical thinning may reflect both reduced striatal/infratentorial input and MSA-related pathobiological mechanisms directly affecting the cerebral cortex. Moreover, the longitudinal progression, spatial distribution and magnitude of cortical thinning relative to baseline remain poorly understood. In our study, we aimed to: (i) characterize progressive cortical thinning over time alongside subcortical/infratentorial atrophy in a cohort of patients with MSA; (ii) examine whether cortical thinning progression differs by motor phenotype (MSA-C vs MSA-P) or by cognitive status (MSA MCI vs MSA non-MCI); (iii) investigate associations between longitudinal cortical and volumetric imaging changes in regions showing significant decline and clinical measures, including disease severity assessed by UMSARS-I and disease duration.

Results

Longitudinal cortical thickness change

To test our main hypothesis, we applied Linear Mixed Effects (LME) models to CT measures across the 31 Desikan-Killiany (DK) atlas regions of interest (ROIs). A general pattern of progressive cortical thinning was observed across most regions (see Table 1 and Fig. 1A), with multiple ROIs exhibiting thinning at p < 0.05 after Benjamini-Hochberg false discovery rate correction (FDR-BH). However, following the more stringent Bonferroni correction, significant progressive thinning was restricted to six regions: the precentral gyrus (primary motor cortex; M1, β = −0.051 [mm/year], p = 0.001, Bonferroni corrected), the posterior cingulate cortex (PCC; β = −0.033 [mm/year], p = 0.001, Bonferroni corrected), the isthmus cingulate cortex (ICC; β = −0.046 [mm/year], p < 0.001, Bonferroni corrected), the superior frontal gyrus (SF; β = −0.044 [mm/year], p < 0.001, Bonferroni corrected), the caudal middle frontal gyrus (CMF; β = −0.040 [mm/year], p < 0.001, Bonferroni corrected), and the insular cortex (INS; β = −0.041 [mm/year], p < 0.001, Bonferroni corrected). Relative to baseline, these regions showed the following percent decreases: M1 (−2.11 [%]), PCC (−1.43 [%]), ICC (−2.02 [%]), SF (−1.73 [%]), CMF (−1.63 [%]), and INS (−1.43 [%]). Table 1 summarizes all LME results across DK ROIs.

Table 1.

Cortical Thickness progressive change using DK atlas

Lobe	ROI	β Time	CI 95%	tStat	f2 Time	Change to baseline [%]	p-Value
Frontal	Superior	−0.044	[−0.067, −0.021]	−3.82	0.036	−1.73	<0.001
	Caudalmiddle	−0.040	[−0.061, −0.019]	−3.85	0.024	−1.63	<0.001
	Rostralmiddle	−0.018	[−0.036, 0.00]	−2.01	0.007	−0.79	0.048
	Parstriangularis	−0.014	[−0.035, 0.006]	−1.39	0.004	−0.62	0.170
	Parsopercularis	−0.025	[−0.046, −0.005]	−2.48	0.01	−1.02	0.016
	Parsorbitalis	−0.025	[−0.047, −0.002]	−2.2	0.008	−0.97	0.031
	Lateralorbital	−0.027	[−0.044, −0.01]	−3.2	0.021	−1.08	0.002
	Medialorbital	−0.024	[−0.045, −0.004]	−2.33	0.017	−1.05	0.023
	Precentral	−0.051	[−0.082, −0.021]	−3.35	0.028	−2.11	0.001
	Paracentral	−0.044	[−0.073, −0.015]	−3.01	0.018	−1.84	0.004
Parietal	Superior	−0.025	[−0.047, −0.004]	−2.34	0.015	−1.15	0.022
	Inferior	−0.027	[−0.05, −0.004]	−2.37	0.016	−1.13	0.021
	Supramarginal	−0.029	[−0.052, −0.007]	−2.59	0.019	−1.2	0.012
	Postcentral	−0.025	[−0.041, −0.01]	−3.21	0.008	−1.19	0.002
	Precuneus	−0.041	[−0.069, −0.013]	−2.94	0.034	−1.74	0.005
Temporal	Superior	−0.025	[−0.048, −0.001]	−2.11	0.013	−0.9	0.039
	Middle	−0.017	[−0.039, 0.005]	−1.51	0.006	−0.63	0.136
	Inferior	−0.024	[−0.051, 0.002]	−1.81	0.016	−0.89	0.075
	Fusiform	−0.032	[−0.054, −0.01]	−2.87	0.022	−1.22	0.006
	Transverse	0.010	[−0.02, 0.04]	0.65	0.001	0.43	0.516
	Parahippocampal	−0.025	[−0.054, 0.004]	−1.72	0.004	−0.96	0.089
	Entorhinal	−0.041	[−0.08, −0.002]	−2.11	0.02	−1.4	0.038
Occipital	Lateral	−0.018	[−0.04, 0.004]	−1.63	0.008	−0.83	0.109
	Lingual	−0.011	[−0.028, 0.005]	−1.35	0.005	−0.56	0.182
	Cuneus	−0.012	[−0.028, 0.005]	−1.41	0.004	−0.61	0.163
	Pericalcarine	−0.015	[−0.031, 0.00]	−2.04	0.007	−0.96	0.045
Cingulate cx	Rostralanterior	−0.032	[−0.056, −0.008]	−2.67	0.018	−1.22	0.010
	Caudalanterior	−0.026	[−0.048, −0.004]	−2.36	0.012	−1.09	0.021
	Posterior (a)	−0.033	[−0.053, −0.013]	−3.3	0.024	−1.43	0.001
	Isthmus	−0.046	[−0.065, −0.027]	−4.85	0.047	−2.02	<0.001
Insular	Insula	−0.041	[−0.065, −0.018]	−3.5	0.024	−1.43	<0.001

Longitudinal effects of time on cortical thickness ROI, expressed as β coefficients, 95% CI, t−statistics, Cohen’s f² effect sizes and percent change relative to baseline. Rows shown in bold indicate results surviving Bonferroni correction for multiple comparisons. p-values in italics show regions where p < 0.05 corrected with Benjamini-Hochberg false discovery rate.

CI confidence interval. cx: cortex

^a Not significant after outlier sensitivity analysis.

Fig. 1. Cortical thinning in MSA patients.

A Boxplots show cortical thickness values in each Desikan-Killiany atlas ROI. B Boxplots show cortical thickness values in each Schaefer atlas network. Black lines connect average CT at BL with average CT at FUP. Blue: values at baseline. Orange: values at follow-up. The between-subject variability is illustrated by the box, which spans from the 25th to the 75th percentile (i.e., the lower and upper bounds of the box, respectively). The whiskers extend to the most extreme data points that are not considered outliers. Outliers are defined as data points that fall below the 25th percentile minus 1.5 times the interquartile range, or above the 75th percentile plus 1.5 times the interquartile range. VisCent Visual Central, VisPeri Visual Peripheral, SomMot Somatomotor, DorsAttn Dorsal Attention, SalVentAttn Salience Ventral Attention, Cont Fronto Parietal Control, Default Default Mode, TempPar Temporo Parietal network.

To assess robustness with respect to atlas choice and to describe thinning over functional networks, the analysis was repeated using the 17 Networks Schaefer parcellation (see Fig. 1B and Table 2). Widespread cortical thinning was again observed at the FDR-BH corrected level (p < 0.05). Under Bonferroni correction, thinning was observed across all DMN subnetworks (DMNA: β = −0.032 [mm/year]; DMNB: β = −0.029 [mm/year]; DMNC: β = −0.041 [mm/year], all p < 0.001), Fronto Parietal Control subnetworks (CONTA: β = −0.026 [mm/year]; CONTB: β = −0.025 [mm/year]; CONTC: β = −0.039 [mm/year]; p < 0.001), Salience/Ventral Attentional subnetworks (SalVentAttnA: β = −0.042 [mm/year]; SalVentAttnB: β = -0.032 [mm/year], both p < 0.001) and somatomotor network (SomMotA: β = −0.044 [mm/year], p < 0.001). See Table 2 for all LME results associated with Schaefer parcellation. No significant Phenotype × Time or MCIstatus × Time interaction effects were observed (p > 0.05).

Table 2.

Cortical thickness progressive change using Schaefer atlas

Network	β Time	CI 95%	tStat	f2 Time	Change to baseline [%]	p-Value
VisCent	−0.019	[−0.039, 0.00]	−1.97	0.011	−0.88	0.053
VisPeri	−0.016	[−0.03, −0.001]	−2.16	0.009	−0.79	0.035
SomMotA	−0.044	[−0.066, −0.022]	−4.01	0.026	−2.00	<0.001
SomMotB	−0.027	[−0.045, −0.008]	−2.92	0.014	−1.08	0.005
DorsAttnA	−0.024	[−0.047, −0.002]	−2.15	0.019	−1.04	0.035
DorsAttnB	−0.031	[−0.053, −0.008]	−2.69	0.017	−1.31	0.009
SalVentAttnA	−0.042	[−0.063, −0.021]	−4.06	0.04	−1.63	<0.001
SalVentAttnB	−0.032	[−0.05, −0.015]	−3.62	0.025	−1.29	<0.001
LimbicA	−0.033	[−0.065, −0.001]	−2.07	0.021	−1.16	0.042
LimbicB	−0.023	[−0.039, −0.007]	−2.89	0.02	−0.93	0.005
ContA	−0.026	[−0.041, −0.011]	−3.39	0.017	−1.1	0.001
ContB	−0.025	[−0.04, −0.011]	−3.46	0.016	−1.02	<0.001
ContC	−0.039	[−0.061, −0.016]	−3.43	0.042	−1.8	0.001
DefaultA	−0.032	[−0.047, −0.017]	−4.13	0.029	−1.27	<0.001
DefaultB	−0.029	[−0.045, −0.013]	−3.63	0.02	−1.1	<0.001
DefaultC	−0.041	[−0.064, −0.019]	−3.65	0.046	−1.59	<0.001
TempPar	−0.024	[−0.05, 0.003]	−1.77	0.012	−0.9	0.081

Longitudinal effects of time on cortical thickness ROI, expressed as β coefficients, 95% CI, t-statistics, Cohen’s f² effect sizes and percent change relative to baseline. Rows shown in bold indicate results surviving Bonferroni correction for multiple comparisons. p-values in italics show regions where p < 0.05 corrected with Benjamini-Hochberg false discovery rate. CI confidence interval. VisCent Visual Central; VisPeri Visual Peripheral, SomMot Somatomotor, DorsAttn Dorsal Attention, SalVentAttn Salience Ventral Attention, Cont Fronto Parietal Control, Default Default Mode, TempPar Temporo Parietal network.

For the bilateral analysis, we repeated the LME analysis after removing outliers (see Fig. S4 and Tables S2, S3 in the Supplementary Materials). As for the DK atlas, results were unchanged, except for the posterior cingulate cortex, which was no longer significant after Bonferroni correction following outlier removal. As for the Schaefer parcellation scheme, all networks remained significant after outlier exclusion. Additional hemisphere-specific longitudinal analyses are reported in the Supplementary Materials (Tables S5–S8). These analyses were overall consistent with the bilateral results, with most significant effects showing concordance between hemispheric and bilateral approaches, while a limited number of regions showed significance in only one hemisphere (see Table S11). The baseline cross-sectional comparison between patients with MSA and HC is reported in the Supplementary Materials (see Figs. S1, S2).

Longitudinal volumetric changes

We explored volumetric longitudinal changes to confirm the macroscopic patterns of atrophy from T1-weighted images. We observed significant volume reductions over time in both putamen and infratentorial regions (see Table 3A and Fig. 2A). A significant fixed effect of time was found in the putamen (β = −0.024 [% ICV] per year, Bonferroni corrected p < 0.001), pons (β = −0.046 [% ICV] per year, Bonferroni-corrected p < 0.001), cerebellar white matter (β = −0.10 [% ICV] per year, Bonferroni-corrected p < 0.001) and cerebellar gray matter (β = −0.21 [% ICV] per year, Bonferroni-corrected p < 0.001). The interaction terms Phenotype × Time and MCIstatus × Time were not statistically significant (p > 0.05). There was an overall reduction in volume over time, as also confirmed by the relative change with respect to baseline (Putamen: −4.18 [%], cerebellar WM −5.50 [%], cerebellar gray matter −3.54 [%], pons −6.21 [%], see Table 3A). Furthermore, a significant effect of motor phenotype was observed (see Table 3B and Fig. 2B). Compared with MSA-C, patients with MSA-P exhibited greater putaminal atrophy (β = −0.12 [% ICV], p < 0.001 Bonferroni corrected), while MSA-C showed a greater extent of pontine atrophy than MSA-P (β = 0.17 [% ICV], p = 0.01 Bonferroni corrected). As we did for ROI_CT, we removed them and repeated the LME analysis to assess potential differences in the results (see Suppl. Fig. S5 and Table S4). The overall finding remained consistent: the fixed effect of time and motor phenotype retained their statistical significance (with unchanged significance levels) in the same regions. See Tables S9 and Suppl. Fig. S10 for the hemisphere-specific volumetric analyses. For putamen, cerebellar WM and GM the hemisphere-specific analyses were fully concordant with the bilateral findings (see Table S11). Baseline cross-sectional comparison between patients with MSA and HC is reported in the Supplementary Materials (see Suppl. Fig. S3).

Table 3.

LME results for longitudinal regional volume

A	β Time	CI 95%	T Stat	f2 Time	Change to baseline [%]	p-Value
Putamen	−0.024	[−0.032, −0.017]	−6.22	0.025	−4.18	<0.001
WM	−0.103	[−0.15, −0.056]	−4.42	0.021	−5.50	<0.001
GM	−0.206	[−0.315, −0.097]	−3.78	0.02	−3.54	<0.001
Pons	−0.046	[−0.056, −0.035]	−8.61	0.018	−6.21	<0.001

B	β Phenotype			f2 Phenotype	MSA-P vs MSA-C [%]
Putamen	−0.12	[−0.18, −0.07]	−4.37	0.53	−19.18	<0.001
WM	0.28	[0.02, 0.53]	2.16	0.13	16.49	0.030
GM	0.48	[1, 1.85]	1.85	0.09	8.82	0.070
Pons	0.17	[0.05, 0.3]	2.78	0.22	27.33	0.010

A Longitudinal effects of time on ROI volumes, expressed as β coefficients, 95% CI, t-statistics, Cohen’s f² effect sizes and percent change relative to baseline normalized by ICV. B Cross-sectional effects of motor phenotype (MSA-P vs MSA-C) on ROI volumes, reported as β coefficients, 95% CI, t-statistics, Cohen’s f² and group differences expressed as percent ICV. Rows shown in bold indicate results surviving Bonferroni correction for multiple comparisons.

CI confidence intervals, ICV intracranial volume, MSA-P Parkinson phenotype, MSA-C cerebellar phenotype, GM cerebellar gray matter, WM cerebellar white matter.

Fig. 2. Longitudinal volume changes and between-phenotype differences.

A Longitudinal volumetric change. Blue: values at baseline (BL). Orange: values at follow-up (FUP). Black lines connect the group means at BL and FUP. B Phenotype difference in volumes. Orange: values for MSA-P. Dark gray: values for MSA-C. Boxplots show volumes as a fraction of ICV. The between-subject variability is illustrated by the box, which spans from the 25th to the 75th percentile (i.e., the lower and upper bounds of the box, respectively). The whiskers extend to the most extreme data points that are not considered outliers. Outliers are defined as data points that fall below the 25th percentile minus 1.5 times the interquartile range, or above the 75th percentile plus 1.5 times the interquartile range. WM cerebellar white matter; GM cerebellar gray matter.

Associations between clinical measures and cortical thinning or volume loss

We hypothesized that statistically significant longitudinal changes in cortical thickness or regional volume loss (expressed via the ΔROI_SIG fixed effect) would be associated with clinical worsening and/or longer disease duration. However, no significant associations were observed (p > 0.05), suggesting that there is no evidence for a linear relationship between volume or CT change and disease progression within the follow-up interval of this study.

Discussion

In this longitudinal study, we characterized structural brain changes over a 1-year follow-up in patients with MSA, with a particular focus on progressive cortical thinning occurring alongside subcortical and infratentorial volumetric atrophy. In regions showing significant longitudinal change, cortical thickness declined by approximately 1.4–2.1% across anatomically defined regions and by 1.0–2.0% in functionally defined networks, whereas volumes declined more markedly, by approximately 3.5–6.2%. Cortical thinning was observed across frontal, motor, cingulate and insular cortices using anatomical parcellation. Decline in CT was also examined using functional parcellation and thinning involved multiple networks, including the somatomotor, default mode, frontoparietal, and salience/ventral attention networks. Overall, our results extend prior work and provide new information about the spatially distributed cortical loss in MSA.

Cortical thinning in MSA remains relatively underexplored. In our study, we aimed to address this knowledge gap by describing spatially distributed thinning over a 1-year follow-up period. Our study included a relatively large longitudinal sample size and employed state-of-the-art surface-based methods, which are more reliable than voxel-based³², or Laplacian approaches that have been previously employed^19,20. Unlike these methods, FreeSurfer provides histologically validated³³, anatomically constrained, and reproducible CT estimates³², particularly in longitudinal and clinical studies³⁴.

Thinning is reported in both anatomically-defined ROIs and functionally-defined motor and non-motor systems, consistent with the recent multifaceted characterization of MSA²⁷ that encompasses executive, attentional and visuospatial dysfunctions. Notably, although the effect sizes for progressive thinning were low, thinning was detected over a short 1-year follow-up. Indeed, the pathological process in MSA primarily affects nigrostriatal or olivopontocerebellar regions, while cortical involvement occurs in later stages of the disease^35,36. Therefore, the lower magnitude of cortical thinning relative to subcortical changes may be compatible with the known predominance of deep-region involvement in MSA³⁷. Somatosensory areas showed significant longitudinal thinning at both the anatomical and network levels, involving M1 and SomMotA functional networks. This finding is consistent with a previous longitudinal VBM study in MSA-P¹⁹. In addition, we observed thinning in SF and CMF cortices and within the frontoparietal network. The involvement of this network, which is implicated in executive control³⁸, is compatible with the executive dysfunction reported in MSA^39,40. Bilateral frontal and parietal areas are interconnected via the corpus callosum, which has been reported to be affected in MSA⁴¹, raising the possibility that disconnection may contribute to the observed cortical changes.

We observed progressive thinning also in the isthmus cingulate cortex and within the DMN. Previous studies reported a relationship between cingulate thinning and cognitive score in MSA-P¹⁰ and reported altered DMN connectivity^42,43. These regions have previously been implicated in executive and memory-related processes^42–44. However, the cognitive implications of these cortical changes cannot be directly inferred from the present data, as specific cognitive assessments were not performed. The insular cortex exhibited progressive thinning over time, consistent with prior structural imaging studies across both MSA phenotypes^13–15,19. Insular loss has been associated with executive and attentional deficts²² and is anatomically consistent with our observation of salience/ventral attentional network involvement. Because autonomic dysfunction is a common clinical symptom in MSA⁴⁵, this spatial pattern may be of interest for future studies investigating this potential association.

Overall, the observation that longitudinal thinning was distributed across functionally defined networks motivates future studies investigating within- and between-network functional connectivity, as well as their specific cognitive correlates. Such work may help clarify whether the distributed cortical changes observed here are meaningfully organized at the level of functional systems and may further inform the conceptualization of MSA as a multisystem disorder²⁷. Taken together, these findings suggest that cortical involvement in MSA is detectable but modest in magnitude and may benefit from complementary approaches, such as structural morphological similarity analyses of cortico-cortical relationships^46,47, to be more fully characterized. This may be relevant for future clinical research, where combining measures of cortical organization with established subcortical and infratentorial markers could improve the characterization of disease progression.

In parallel with cortical thinning, all subcortical and infratentorial regions showed significant volume reductions between BL and FUP, extending previous longitudinal insights^8,48. In particular, the putamen and pons are well-established as pivotal regions of neurodegeneration in MSA-P (putamen) and MSA-C (pons)⁹. Consistent with this evidence, we reported greater putaminal atrophy in MSA-P and greater pontine atrophy in MSA-C at the cross-sectional level. Although differential volumetric involvement across motor phenotype has been reported^17,49, there is concurrent evidence that putaminal volume loss also occurs in MSA-C, albeit to a lesser extent⁸, and that pontine volume loss in MSA-P can occur at a rate comparable to that in MSA-C⁹. Indeed, together with the middle cerebellar peduncle, the pons has been proposed as an MRI biomarker for both phenotypes⁵⁰, highlighting the complexity of the disease, which is further characterized by differences in onset, severity and the dynamic evolution of volumetric atrophy patterns across phenotypes²⁷. Furthermore, cerebellar WM showed a higher rate of volume reduction than cerebellar GM (i.e., -5.50% vs -3.54% relative to baseline), suggesting more extensive WM impairment within the cerebellum⁶. This aligns with the known pathology of MSA, where α-synuclein fibrils aggregate first in oligodendrocytes, leading to axonal degeneration^1,2. The volumetric reduction rates observed in our study (in the range −3.54% to −6.21%) are in line with those reported in a previous study⁸ that used an earlier version of the FreeSurfer pipeline and a multicenter dataset with a comparable sample size. Taken together, these volumetric findings corroborate prior evidence supporting pons, putamen and cerebellar WM volumes as promising MRI markers of structural involvement in MSA^4,50.

In our study, the absence of differential longitudinal trajectories across motor phenotype and cognitive classification at baseline does not provide strong support for subgroup-specific progression over the current follow-up interval. However, the unbalanced and limited sample size may in part explain these negative results, rather than biological equivalence among the subgroups. Nonetheless, as reported by Jellinger et al²⁷, neuroimaging correlates of MCI in MSA are not consistent, with some studies not identifying cortical abnormalities despite MCI diagnosis^50–52 and some other cross-sectional studies^15,17 suggesting that cortical thinning may contribute to cognitive decline²².

In our cohort, no significant associations emerged between imaging changes and clinical measures, including disease duration and severity as measured with UMSARS-I. Similar negative findings have been reported in previous MSA studies, including both longitudinal^8,19 and cross-sectional^14,53 investigations of structural imaging measures and clinical variables. This may reflect differences in the temporal dynamics⁹ and sensitivity of imaging versus clinical measures⁸, as well as the limited sensitivity of clinical scales such as UMSARS to subtle cortical changes¹⁹. It is also possible that more detailed neuropsychological evaluations targeting specific cognitive domains would have been more sensitive to these relationships. These findings may have implications for future clinical research and study designs, as broad clinical measures such as UMSARS and baseline cognitive classification may not fully capture the same timescale of change reflected by longitudinal structural MRI over short follow-up intervals. Therefore, future longitudinal studies and clinical trials of disease-modifying therapies may benefit from including more sensitive and granular multidomain assessments⁸.

Some limitations of the current work warrant more discussion. First, we did not have a longitudinal control group of healthy older adults for direct comparison, which prevents a full disentanglement of MSA-related structural changes from age-related effects. Nevertheless, analyzing within-subject longitudinal changes within the same MSA cohort remains informative, as it allows modeling of inter-individual variability and repeated measures. This approach has been previously adopted in previous research on cortical atrophy^54–56, although it does not by itself establish disease specificity. Importantly, prior longitudinal studies, in healthy older individuals using freesurfer-based processing pipelines, have reported cortical thinning and subcortical/infratentorial atrophy rates over one-year time span lower than those observed in our MSA population^57–59. Although comparisons with prior literature should be interpreted cautiously and cannot substitute for longitudinal HC data, this observation is compatible with the possibility that part of the observed thinning reflects disease-related mechanisms beyond normative aging. Nonetheless, future longitudinal studies including healthy older control groups will be essential to better determine the extent to which progressive cortical thinning is specific to MSA rather than to normative aging. Second, the unbalanced distribution between phenotypes, particularly the limited number of MSA-P patients, may have reduced the statistical power to detect longitudinal motor phenotype effects and limited the generalizability of the findings. In future studies, more balanced cohorts will be required to identify potential differences in progression patterns between MSA-P and MSA-C patients. With larger and more balanced sample sizes—although challenging to achieve in MSA—alternative approaches, such as vertex-based surface analyses, may also be considered. Third, the short follow-up could have been insufficient to detect more robust longitudinal changes in cortical structures. However, we know that MSA is a rapidly progressive disorder, thus affecting the possibility to comply with longer MRI follow-up. Indeed, previous longitudinal MRI studies in this population mostly presented similar short intervals^{8,20,41,48,60,61}. Fourth, we chose not to harmonize the dataset for acquisition protocol differences, as minor protocol variations did not affect structural estimates (see Methods and Supplementary Materials). Moreover, this decision is supported by several factors: (i) each patient was scanned with the same acquisition protocol across time points; (ii) our dataset was collected at a single center using the same scanner, with only minor modifications to the MPRAGE protocol, which are expected to introduce less variability than that typically observed in multi-scanner and/or cross-manufacturers datasets^62,63; and (iii) harmonization methods may obscure meaningful biological variability⁶⁴, particularly when this effect is subtle⁶⁵.

In summary, this study provides longitudinal evidence of progressive cortical thinning in anatomically and functionally defined regions and networks. While subcortical and infratentorial volumetric atrophy remains the key structural attribute of the disease, cortical thinning emerges as an additional and previously understudied feature of structural change that is detectable over a 1-year follow-up interval. While larger multicenter studies are needed to confirm these findings, we suggest that our study contributes to the current understanding of MSA as a multisystem disorder and highlights the importance of incorporating cortical measures in future longitudinal and multimodal MRI studies.

Methods

Patients

Thirty-six patients (25 with MSA-C, 11 with MSA-P variant) were enrolled at the center for neurodegenerative diseases at the University of Salerno. The population was consecutively recruited as patients referred to our tertiary care center for MSA. All patients were diagnosed as probable MSA according to the Second consensus criteria⁴⁵, since the MRI acquisitions were performed before the new Movement Disorder Society (MDS) criteria were available. When we retrospectively reviewed diagnoses according to the MDS MSA criteria, all patients met criteria for clinically established MSA. Features not supporting a diagnosis of MSA according to the Second consensus criteria were excluded at diagnosis. By retrospective review, no exclusion features from the MDS MSA criteria were present in the whole sample. MSA-C patients generally present a milder motor progression as compared to MSA-P, allowing them to be more compliant with the MRI follow-up, thus explaining the unbalanced sample. Participants were excluded in the presence of other neurological or psychiatric disorders or significant abnormalities on clinical MRI, including cerebrovascular lesions such as stroke, lacunar infarcts, or severe white matter hyperintensities.

All patients underwent clinical evaluation including UMSARS scale and MRI at BL and FUP (see Table 4). All patients underwent an II level neuropsychological evaluation. Due to the lack of MSA-specific criteria for the diagnosis of MCI, it was diagnosed according to the Movement Disorders Society criteria for MCI in Parkinson’s disease⁶⁶ (PD). MCI was diagnosed according to PD-MCI Level II (comprehensive assessment) criteria, requiring formal neuropsychological testing with at least two tests in each of the five cognitive domains: Attention/Working memory (evaluated by the brief version of the Stroop Color and Word Test (i.e., Time interference score) and Trail Making Test-A), Executive function (evaluated by Clock Drawing Test, immediate copying test of Rey-Osterrieth Complex Figure, phonological fluency test, Trail making test:B-A), Language (evaluated by semantic fluency task, auditory comprehension task, words, non-words and sentences Repetition tasks), Memory (evaluated by immediate and delayed recall of the Rey’s auditory 15-word learning test and Brief Story test), Visuospatial function (evaluated by Benton Judgment of Line Orientation test and Constructional Apraxia Test). MCI was diagnosed when the impairment was shown on at least two neuropsychological tests, represented by either two impaired tests in one cognitive domain or one impaired test in two different cognitive domains. Clinical variables that were assessed in relation with MRI variables were disease duration and severity, as assessed by UMSARS part I score, that evaluates activities of daily living in the past 2 weeks and is less affected by the effects of symptomatic therapy as compared to UMSARS part II, that is the motor examination scale. The study was performed in accordance with the Declaration of Helsinki and was approved by the local ethics committee (Comitato Etico Campania Sud; approval no. 77, September 7, 2016). All participants provided written informed consent prior to participation in the study. For further information about the population, see Table 4.

Table 4.

Demographic information. Normally distributed variables are reported as mean ± standard deviation

	MSA all	MSA-C	MSA-P	P-value MSA-C vs MSA-P
Sample size	36	25	11
Age at BL	60.6 ± 6.4	60.9 ± 5.9	59.7 ± 7.7	0.6a
Sex	18 F/18 M	12 F/13 M	6 F/5 M	>0.9b
FUP time (years)	0.96 (0.20)	0.97 (0.19)	0.97 ± 0.14	0.9a
MCI status at BL	20 no/16 yes	13 no/12 yes	7 no/4 yes	0.7b
Disease duration (years)	3 (2)	3.0 ± 1.4	2.9 ± 1.3	0.8c
UMSARS-I at BL	19.1 ± 5.2	18.2 ± 4.3	21.2 ± 6.6	0.1c
UMSARS-I at FUP	25.0 ± 5.9	24.6 ± 5.1	26.0 ± 7.4	0.5c
	HC	P-value HC vs MSA
Sample Size	50
Age	63.9 ± 9.2	0.07a
Sex	26 F/24 M	>0.9b

Non-normally distributed variables are reported as median (interquartile range).

BL baseline, FUP follow-up, MCI Mild Cognitive Impairment, UMSARS-I Unified Multiple System Atrophy Rating Scale Part I, HC healthy controls.

^aT-test.

^bexact Fisher test for categorical variables.

^cMann-Whitney U test.

MRI acquisition

All subjects underwent MRI using the same 3T MAGNETOM Skyra (Siemens Healthcare GmbH, Erlangen, Germany), equipped with 16-channels phase array head and neck coil. To map anatomical brain structures, we acquired high-resolution whole-brain 3D T1-weighted images using a Magnetization Prepared Rapid Gradient Echo (MPRAGE) sequence. Over the course of the study, three slightly different protocols were used, varying in repetition time (TR), Echo Time (TE), flip angle and voxels resolution:

• Protocol 1: TR = 2.4 s, TE = 2.3 ms, resolution = 1 × 1 × 1 mm, flip angle = 8°,  N = 20

• Protocol 2: TR = 2.3 s, TE = 2.9 ms, resolution = 1 × 1 × 1.2 mm, flip angle = 9°,  N = 13

• Protocol 3: TR = 2.0 s, TE = 2.9 ms, resolution = 1 × 1 × 1 mm, flip angle = 8°,  N = 3.

Generalized auto calibrating partially parallel acquisitions (GRAPPA) acceleration factor was 2 for all the protocols. Importantly, each subject was scanned using the same protocol at both BL and FUP. In addition, we performed a statistical analysis to assess the potential impact of protocol differences on CT estimates (see Supplementary Materials), which did not reveal a significant effect of acquisition protocol on CT estimates. Therefore, given the within-subject consistency in acquisition parameters in our longitudinal design and the absence of a systematic effect of protocol parameters, no additional image harmonization was performed.

T1-weighted images processing

All T1-weighted images (36 subjects × 2 timepoints) were analyzed with the longitudinal Freesurfer stream (version 8.0.0)⁶⁷, that is widely adopted and well-established for obtaining unbiased estimates of longitudinal CT and volumetric measures. The pipeline follows a three-step procedure whose details are described in prior works^8,9,67,68:

• Each timepoint is processed independently through the standard cross-sectional pipeline (recon-all with option -3T) that performs intensity normalization, skull stripping, Talaraich registration, subcortical segmentation, cortical surface reconstruction, parcellation.

• An unbiased within-subject template was then created using recon-all -base, which combines information from both timepoints. This step overall reduces processing bias towards a timepoint and variability.

• Each timepoint is further processed with recon-all-long, using the within subject template as a reference. The timepoints are aligned to the same template, thus enhancing sensitivity to longitudinal changes.

As for volumetric analysis, we extracted the segmentation obtained with Synthseg⁶⁹, that is provided as segmentation results by the version 8.0.0 of Freesurfer. We selected cerebellar WM and GM as well as putamen. For pons volume estimation we used a post processing python script (segment_subregions with options –long-base). Hence, we estimated regional volumes in a total of four ROIs (ROI_VOLS, putamen, cerebellar GM and WM, pons) that were normalized over total Synthseg-estimated intracranial volume (ICV), thus expressing volumes as percentage of ICV. We focused on these four regions as they are well-established in recent MSA literature^8,9. As for CT analysis, we used parcellation from the Desikan-Killiany atlas (DK), composed of 31 ROIs per hemisphere. In the main analysis, CT values were averaged across hemispheres, resulting in 31 bilateral ROIs per subject (ROI_CT). Complementary analyses were also performed separately for the left and right hemisphere. In addition, cortical thickness was analyzed using the Schaefer functional atlas⁷⁰ (400 parcels, 17 networks), with mean CT values aggregated bilaterally within each network, as previously described⁷¹. Furthermore, additional hemisphere-specific analyses were performed separately for the left and right hemisphere. The supplementary Schaefer atlas analysis was performed to assess whether progressive thinning in MSA was detectable irrespective of the parcellation scheme and to investigate CT longitudinal change at the level of a different organizational scheme. Finally, we performed a sensitivity analysis excluding ROI-specific outliers to assess the robustness of the LME results on bilateral ROIs. All the details of this analysis are reported in the Supplementary Materials.

Quality control procedure

At the time of acquisition, the radiologist and MR engineer re-acquired the 3D T1-weighted image if quality was deemed insufficient. Subsequently, after the recon-all commands, we visually inspected Freesurfer outputs to identify gross reconstruction errors, which were not observed. We then applied fully automated methods to detect any potential poor-quality surface reconstructions. First, we used the fsqc toolbox (https://github.com/Deep-MI/fsqc), which provides quantitative metrics for automated quality assessments of pre-processed structural MRI scans. Second, we computed the Euler number, which has been shown to correlate with manual quality ratings and to strongly agree with subjects labeled as unusable by expert raters^72–74. The Euler number is defined as $E=2-2\times (\mathit{total\; number\; of\; defects}$), where the number of topological defects is calculated separately per hemisphere in the previous step. More negative Euler numbers indicate a higher number of surface reconstruction defects. Third, we applied a threshold of −120, i.e., subjects with an Euler number lower than −120 were excluded from the analysis. This threshold has been used in a recent study that employed freesurfer surfaces reconstructions in a large cohort across the lifespan, including participants in our age range⁷⁵. According to this criterion, no subjects in our sample fell below the threshold at either baseline or follow-up. Therefore, no participants were excluded from the analysis.

Statistical analysis

We tested the normality of demographic variables using Shapiro-Wilk test. Variables showing a normal distribution (p > 0.05) are reported as mean ± standard deviation; otherwise, they are reported as median and interquartile range. Group differences between MSA-C and MSA-P groups were evaluated using non-parametric Mann-Whitney U test or t-test, according to data distribution. For categorical variables (i.e., sex and MCI status at baseline), Fisher’s exact test was used.

Longitudinal changes in regional CT and volumetric measures were assessed using a linear-mixed effects model (LME). For each ROI, the following model was fitted

$$ROI\sim 1+Time+Phenotype+MCIstatus+Age+Sex+\left(1|SubjID\right)$$ 1

Where ROI represents either cortical thickness (ROI_CT) or regional volume (ROI_VOL). A random intercept for subject (SubjID) was included to account for repeated measures (BL and FUP) and for subject-specific variability. Fixed effects included Time (modeled as a categorical variable with two levels: BL and FUP), Age, Sex, Phenotype (MSA-C or MSA-P) and MCI status at BL (MSA MCI or MSA non-MCI). Time is the primary effect of interest, as it captures progressive decline. Phenotype and MCI were also included to assess potential cross-sectional differences. We did not include random slopes due to convergence issues for the parameter estimates, likely related to the limited number of timepoints per subject. To account for multiple comparisons across ROIs, statistical significance was assessed using Bonferroni correction within each analysis. Separate corrections were applied for volumetric ROIs (4 ROIs), CT ROIs estimated from the DK atlas (31 ROIs), and network-level CT ROIs estimated from Schaefer atlas (17 ROIs), using Bonferroni thresholds of p = 0.05/nROIs (corresponding to p < 0.0125, p < 0.0016 and p < 0.0029, respectively). For the sake of completeness, we also reported results surviving Benjamini-Hochberg false discovery rate correction (FDR-BH, p < 0.05).

Then, to investigate whether the longitudinal trajectories differed by motor phenotype or MCI subgroup, additional LME models were fitted, including interaction terms between Phenotype and Time and between MCI status and Time. The model retained the same random-effects structure (random intercept for subjects) and was adjusted for age and sex. Assumptions of all the LME models were assessed through visual inspection of residual plots to identify systematic patterns and quantile-quantile plots of the residuals to detect potential deviations from normality. Potential collinearity among fixed effects was assessed with variance inflated factors (VIFs), with values below 5 implying absence of problematic collinearity^76,77. All LME models were fitted using R lmer function of the lme4 package⁷⁸.

To facilitate comparisons across ROIs and with the literature, model parameter estimates (β coefficients) are additionally reported as relative percent change with respect to baseline average value of each ROI^58,59. We calculated effect sizes estimates using Cohen’s f² which is defined as⁷⁹: $f^2=(R_2^2-R_1^2)/(1-R_2^2)$; where the $R_2^2$ represents the variance explained by the full model and $R_1^2$ represents the variance explained by the model excluding the fixed effect of interest (e.g., Time or Phenotype). Effect sizes were estimated using marginal R² as defined by the R function r.rsquaredGLMM() from the MuMIN package in R⁸⁰. In the context of LME models, effect size estimation requires cautious interpretation^81,82. Consistent with previous studies^83–85, effect sizes are reported here for the sake of results interpretation. For each region, we report β estimates with 95% confidence intervals, t-statistics, f², relative change to baseline and p-values.

To assess whether longitudinal changes in cortical thinning (ROI_CT) and infratentorial or subcortical volumes (ROI_VOL) were associated with clinical severity (UMSARS-I) or disease duration, we fitted another linear model. We restricted our analysis to only those ROIs that survived Bonferroni correction for multiple comparisons. The linear model was structured as follows:

$$y\sim 1+{\mathrm{\Delta }\mathit{ROI}}_{\mathit{SIG}}+\mathit{Sex}+\mathit{Age}+\mathit{Phenotype}+\mathit{MCIstatus}$$ 2

Where y represents either the longitudinal change in UMSARS-I clinical scale between FUP and BL or disease duration. ΔROI_SIG indicates longitudinal change in ROIs surviving correction for multiple comparisons with the previous LME modeling. The ΔROI_SIG fixed effect specifically tests whether longitudinal change in CT (or volumes) is associated with changes in the clinical measures, after correcting for motor subtype, MCI status, age and sex. For the main objective of the study, we assessed progressive cortical changes using a within-subject longitudinal design. In addition, we assessed baseline atrophy by comparing patients with MSA with a separate group of 50 healthy controls (HC) acquired on the same scanner. Group differences were tested using an ANCOVA model with group (MSA vs HC) as the between-subject factor and age and sex as covariate. Three separate ANCOVA models were fitted: one for CT in the DK atlas, one for CT in the Schaefer atlas, and one for volumes.

Author contributions

(1) Study design; (2) Data acquisition; (3) Data processing and statistical analysis; (4) first manuscript draft; (5) manuscript revision. R.I. (1,3,4,5); M.A. (1,5), F.T. (2,5); D.M. (2,5); M.P. (2,5); R.M. (2,5); A.C. (2,5); S.P. (2;5); F.E. (2,5); M.S. (5); P.B. (5); S.P. (1,5); M.T.P. (1,2,5).

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to privacy and ethical restrictions, but are available from the corresponding author (mpellecchia@unisa.it) upon reasonable request.

Code availability

Scripts to run the MRI and statistical analyses are available at the following link: https://github.com/Rak-prog/MSA_long_paper.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41531-026-01454-0.