Multiple sclerosis (MS), a chronic inflammatory demyelinating and degenerative disorder of the central nervous system (CNS), affects nearly three million people worldwide. It is the most common cause of non-traumatic neurological disability in young to middle-aged adults (1). Clinical deterioration is essentially driven by neuronal loss, which may be the consequence of (i) acute damage in newly-formed demyelinating lesions, resulting from recurrent autoimmune responses mediated by the peripheral immune system, and/or (ii) a more gradually installing neurodegeneration. The latter is believed to arise from a (non-exclusive) combination of mitochondrial dysfunction in chronically demyelinated axons, submeningeal lymphocytic clustering with damage to the underlying cortex, and pathogenic microglial activity around slowly expanding focal lesions and/or diffusely throughout the white matter (2, 3). MS is traditionally categorized into relapsing–remitting (RR), secondary progressive (SP) and primary progressive (PP) subtypes based on clinical presentation. In RR MS, tissue injury is primarily attributed to acute demyelinating lesions, whereas chronic neurodegeneration predominates in the progressive phenotypes. Recent literature endorses the stance that those phenotypes should not be seen as strictly separated entities but rather as a spectrum in which both key processes often occur together, albeit in varying proportions (3–6).

T1 contrast-enhancing and T2 hyperintense lesions are well-established magnetic resonance imaging (MRI) biomarkers of acute inflammation in MS (7). These markers are widely used in clinical practice to aid diagnosis and monitor disease activity; the latter refers to the occurrence of relapses or new focal lesions on MRI. However, their correlation with clinical outcomes is modest and often inconsistent, a phenomenon commonly named as the clinico-radiological paradox (7).

Brain volume loss (BVL) has emerged as a complementary MRI biomarkers, reflecting neurodegeneration in MS. BVL correlates with concurrent and future disability, both physical and cognitive, even in early stage of the disease (8, 9). Histological studies have demonstrated a strong association between cortical thickness measured on MRI and post-mortem brain samples, supporting the validity of BVL as a reliable indicator of actual brain atrophy (10).

BVL has increasingly been incorporated as primary or secondary endpoint in immunomodulating disease-modifying treatment (DMT) trials (11–14). A large meta-analysis evaluated over 13,000 patients across 13 pivotal studies. Brain volume changes were measured starting 6–12 months after treatment initiation to account for potential pseudo-atrophy bias, which is an apparent reduction in brain volume that can occur early after anti-inflammatory treatment due to resolution of oedema rather than neurodegeneration (15). The results showed that the effect of a therapy on BVL significantly correlates with its impact on disability outcomes, independent of its anti-inflammatory properties (16). To distinguish disease-related from age-related brain changes, a cut-off of −0.4% annual percentage volume change (PVC) has been proposed to define “pathological” BVL, offering 80% specificity and 65% sensitivity in categorizing patients (17).

Despite its research relevance, BVL has not yet been widely implemented in routine clinical practice (18). Technical factors, such as the use of different scanners, can introduce variability that often exceeds the magnitude of actual brain volume changes in uncontrolled settings (19). Only a few real-world studies in MS have assessed the value of longitudinal BVL using different scanners and primarily relied on statistical adjustments (20–22), which may limit the clinical applicability of their findings.

Our study addresses this gap by investigating the clinical relevance of longitudinal BVL over 4–5 years in a real-world MS cohort, explicitly accounting for scanner variability using a similarity index as a post-acquisition harmonization method. We employed the registration-based icobrain ms algorithm (icometrix, Leuven, Belgium), a CE-marked and FDA-cleared automated method to quantify BVL in MS (23). By combining robust real-world data with reliable volumetric quantification while also controlling for scanner variation, our study advances previous works and provides actionable insights for the potential integration of BVL into clinical monitoring.

BVL has recently gained attention as an MRI-derived proxy for neurodegeneration in patients with MS. The implementation of brain volumetry into routine clinical practice, though, has been hindered, mainly due to the risk of measurements becoming less precise once they are conducted outside strictly standardized research settings. In this longitudinal study, we aimed to validate the clinical relevance of BVL in a real-world cohort of patients with MS while explicitly taking into scanner differences using a similarity index as post-acquisition harmonization approach. We did not observe significant differences in annualized BVL between MS and HC essentially failing the primary objective of this study. However, worsening functional ability in patients with MS, as measured by 9HPT scores, was linked to increased atrophy in both TGM and CGM. Additionally, clinical decline on the composite MSFC_SDMT outcome measure was associated with WB, TGM, and CGM volume loss. We could not predict “pathological” BVL based on baseline demographics or clinical status.

It has been widely accepted that MS patients present with accelerated BVL compared to healthy individuals (9, 20, 21). The majority of studies on BVL, both in standardized and real-world settings, use the structural image evaluation with normalization of atrophy, or SIENA, method for PVC quantification, whereas the effect of change in MRI scanner strength (1.5 or 3.0 T) over time is usually taken into account with mixed-effect regression models (20–22). We have used the icobrain ms software to quantify BVL and applied a similarity index as post-acquisition harmonization method to consider scanner switches. We found PVC rates below the typical values of brain atrophy in MS (36), with only 15 of our 72 participants (21%) actually demonstrating annualized WB volume loss below 0.4%. However, this discrepancy cannot be due to the use of another tool for PVC quantification, as consistency between both methods for real-world BVL analysis has already been demonstrated (37). The negative outcome of our first objective (primary endpoint) does not seem to be influenced by a treatment effect, but may have been due to the unequal distribution of age between the MS and HC cohorts. Recent research suggest that age-related BVL accelerates significantly around 60 years of age (38). As shown in Supplementary Figure 1, the greatest mean BVL was observed in the oldest age category (55–70 years) for both HC and MS cohorts, where there is an important imbalance between the number of people with MS and HC. This imbalance may contribute to the apparent lack of difference in brain atrophy between the groups at the overall group level. Limited statistical power may have also contributed to our negative finding to some extent. Our sample sizes were sufficient to detect medium-to-large effects, but smaller differences in BVL could have gone undetected, increasing the risk of a type II error. Finally, survivor bias may have played a role: patients with more aggressive disease courses may have been less likely to remain in long-term follow-up, resulting in an overrepresentation of more stable individuals in the MS cohort and attenuating observed differences compared to HC. Nonetheless, some other studies have also reported low brain atrophy rates in MS (39, 40), further supporting the reliability of our BVL measures. Despite having a cohort with less global BVL than expected, these patients with MS still exhibited signs of disability worsening, which we were able to connect to GM volume loss.

MS was classically considered a disease of white matter (WM), but the intensity and relevance of GM involvement has become increasingly evident over the past two decades. GM volume loss, particularly in the deep nuclei, can occur early in the disease course and independent of focal WM lesions (41, 42). Moreover, several longitudinal studies have demonstrated that the volumetric decline of GM structures is not only more pronounced as compared to WM, but also more strongly associated with clinical outcomes (43, 44). Interestingly, the increased rate of WB volume loss observed in MS patients with progression independent of relapse activity, as compared to those that remained stable, could mainly be attributed to changes in the cerebral cortex (45). Our result are in line with previous findings, reporting significant associations between (C)GM atrophy and disease worsening, according to the 9HPT score and MSFC_SDMT (46, 47). BVL in GM can manifest according to different spatial patterns, which is relevant to specific clinical manifestations and even phenotypes (48–50). Regional GM atrophy may more accurately reflect the status of certain individual clinical measures than global estimates. For example, thalamic atrophy seems to be associated with cognitive impairment reflected by the SDMT in patients with MS, while overall GM atrophy does not (51). While we did not find statistically significant DMT effects on BVL, possibly due to a lack of statistical power, others have recently reported an increase of thalamic volume with natalizumab, suggesting a potential neuroprotective role.⁵² So, even though we found an association between (C)GM atrophy and overall disease progression, more locally defined anatomic regions, such as the thalamus, may even be more informative about certain clinical as well as therapeutic aspects of MS. Unfortunately, parcellation of the CGM and segmentation of the DGM is not part of the default icobrain ms pipeline (and thus not validated for use in data from clinical routine) but may still be an interesting objective for future studies. Next to the importance of spatial patterns in brain atrophy, an important temporal factor might be at play when considering the intricate nature of the relationship between WM and GM pathology. A recent systematic review concluded that global GM atrophy appears to be secondary to focal WM lesions in early stages of the disease, and only later in the disease course will adopt a more independent character due to other neurodegenerative processes that are still not fully understood (52). This may correspond to the gradual change in inflammatory pathways throughout the disease course, with a predominant role of invading T lymphocytes in the formation of WM lesions in relapsing MS (1), and meningeal inflammation, amongst others, likely acting as a driving force for cortical atrophy in the progressive phase, which seemingly develops according to a clear gradient of neuronal loss that turns inwards from the boundary between GM and cerebrospinal fluid (53, 54). Such shift may even happen very early in the disease, as CIS patients that progress toward RR MS already show a 3.4 fold increase in GM atrophy, but no change in WM when compared to HC (46).

We acknowledge several limitations in the present study. First, despite our efforts to minimize the potential confounding impact of acute inflammatory activity or edema on BVL quantification, a pseudo-atrophy effect may not be limited to the presence of active lesions only. Recent literature has shown (i) that up to 25% of acute clinical events identified as relapse do not appear to be associated with lesional changes on MRI (55), (ii) contrast-enhancement within acute focal lesions typically lasts for approximately 4 weeks, whereas the pseudo-atrophy effect seems to reach a plateau only after 16 to 20 weeks (56, 57), (iii) inflammation in MS is not restricted to focal lesions and may occur diffusely throughout the normal-appearing white matter as well (58), and (iv) the exact pathophysiological mechanism behind pseudoatrophy is still not fully understood and may involve other processes (besides accelerated water loss/fluid shifts), such as changes in glial cells (15, 56). Second, during our selection process, we experienced important loss of participants that drastically reduced the number of patients in our final sample. The subgroup with pathological BVL was particularly small (n = 12), which may have further limited our statistical power for the final objective. Previous studies have reported drop-out rates ranging from 37 to 82% (20–22, 59), and notably, one of these indicated a higher failure rate when focusing exclusively on 3D (82%) versus 2D (56%) imaging (59). Our drop-out rate of 56% aligns with the existing literature and we included 3D acquisition only. Notably, data loss in our study more frequently stemmed from applying the similarity index with a cut-off of 0.15 (33%), than from failing to meet key technical standards (16%). Even though we decided to continue with a lower similarity index cut-off (0.15 rather than 0.20) to retain as much patients as possible, we still experienced significant data loss due to this measure. Third, the similarity index does not account for various biological factors that can influence longitudinal volume changes, such as timing of scans, hydration state, lifestyle factors, and comorbidities. Nonetheless, a recent study suggests that technical causes contribute more significantly to variations in brain volume measures than physiological factors (60), reinforcing the value of using a similarity index as a post-acquisition harmonization approach in our analysis. Fourth, we used retrospective data for the MS cohort and relied on clinical data from two timepoints only (baseline and follow-up), limiting our ability to track confirmed disability worsening. On the other hand, we did employ a range of disability measures. This comprehensive approach allows us to capture various dimensions of disability, which is crucial in a heterogeneous condition like MS. Fifth, we used the SDMT for calculating MSFC scores, whereas this metric normally involves the Paced Auditory Serial Addition Test (PASAT) (61). This does not allow us to accurately compare our findings with others from the field. Still, both scores are representatives of cognition (particularly information processing speed) and using the SDMT score instead of the PASAT may even be an improvement due to its slightly better predictive validity, particularly when considering longitudinal data (62–64). Finally, we have only included the total EDSS score without its individual functional systems (i.e., the visual, brainstem, pyramidal, cerebellar, sensory, bowel and bladder, and cerebral functions) (28). In this real-world study, collecting such detailed information during routine medical visits is often challenging due to time constraints. To overcome this limitation, we have tried to take a multifaceted approach, by incorporating various measures of disability in MS.

Using the automated icobrain ms algorithm with similarity index-based post-acquisition harmonization to address scanner variability, we did not detect accelerated BVL in this real-world MS cohort compared to healthy individuals. Nonetheless, GM volume loss remains clinically relevant in MS, as it was associated with disability worsening according to the 9HPT and the MSFC_SDMT. To enable routine clinical use of BVL measurements, future research should prioritize robust post-acquisition correction methods that explicitly account for scanner differences. In the meantime, a practical recommendation for clinical settings is to perform follow-up scans on the same scanner whenever feasible, to enhance the reliability of longitudinal BVL assessments.