Journal of Neurology (2023) 270:1286–1299 https://doi.org/10.1007/s00415-022-11488-y REVIEW Present and future of the diagnostic work‑up of multiple sclerosis: the imaging perspective Massimo Filippi1,2,3,4,5 · Paolo Preziosa1,2,5 · Douglas L. Arnold6,7 · Frederik Barkhof8,9 · Daniel M. Harrison10,11 · Pietro Maggi12,13 · Caterina Mainero14,15 · Xavier Montalban16 · Elia Sechi17 · Brian G. Weinshenker18 · Maria A. Rocca1,2,5 Received: 25 October 2022 / Revised: 9 November 2022 / Accepted: 10 November 2022 / Published online: 24 November 2022 © The Author(s) 2022 Abstract In recent years, the use of magnetic resonance imaging (MRI) for the diagnostic work-up of multiple sclerosis (MS) has evolved considerably. The 2017 McDonald criteria show high sensitivity and accuracy in predicting a second clinical attack in patients with a typical clinically isolated syndrome and allow an earlier diagnosis of MS. They have been validated, are evidence-based, simplify the clinical use of MRI criteria and improve MS patients’ management. However, to limit the risk of misdiagnosis, they should be applied by expert clinicians only after the careful exclusion of alternative diagnoses. Recently, new MRI markers have been proposed to improve diagnostic specificity for MS and reduce the risk of misdiagnosis. The cen- tral vein sign and chronic active lesions (i.e., paramagnetic rim lesions) may increase the specificity of MS diagnostic criteria, but further effort is necessary to validate and standardize their assessment before implementing them in the clinical setting. The feasibility of subpial demyelination assessment and the clinical relevance of leptomeningeal enhancement evaluation in the diagnostic work-up of MS appear more limited. Artificial intelligence tools may capture MRI attributes that are beyond the human perception, and, in the future, artificial intelligence may complement human assessment to further ameliorate the diagnostic work-up and patients’ classification. However, guidelines that ensure reliability, interpretability, and validity of findings obtained from artificial intelligence approaches are still needed to implement them in the clinical scenario. This review provides a summary of the most recent updates regarding the application of MRI for the diagnosis of MS. Keywords Multiple sclerosis · Magnetic resonance imaging · Diagnosis 10 * Massimo Filippi Department of Neurology, University of Maryland School filippi.massimo@hsr.it of Medicine, Baltimore, MD, USA 11 1 Department of Neurology, Baltimore VA Medical Center, Neuroimaging Research Unit, Division of Neuroscience, Baltimore, MD, USA IRCCS San Raffaele Scientific Institute, Via Olgettina, 60, 12 20132 Milan, Italy Department of Neurology, Cliniques Universitaires Saint 2 Luc, Université Catholique de Louvain, Brussels, Belgium Neurology Unit, IRCCS San Raffaele Scientific Institute, 13 Milan, Italy Department of Neurology, Centre Hospitalier Universitaire 3 Vaudois (CHUV), Lausanne, Switzerland Neurorehabilitation Unit, IRCCS San Raffaele Scientific 14 Institute, Milan, Italy Department of Radiology, Athinoula A. Martinos Center 4 for Biomedical Imaging, Massachusetts General Hospital, Neurophysiology Service, IRCCS San Raffaele Scientific Boston, MA, USA Institute, Milan, Italy 15 5 Harvard Medical School, Boston, MA, USA Vita-Salute San Raffaele University, Milan, Italy 16 6 Department of Neurology, Cemcat, Hospital Vall d’Hebron, NeuroRx Research, Montreal, QC, Canada Autonomous University of Barcelona, Barcelona, Spain 7 McGill University, Montreal, QC, Canada 17 Department of Medical, Surgical and Experimental Sciences, 8 Department of Radiology and Nuclear Medicine, Amsterdam University of Sassari, Sassari, Italy UMC, Vrije Universiteit, Amsterdam, Netherlands 18 Department of Neurology, University of Virginia, 9 Queen Square Institute of Neurology and Centre for Medical Charlottesville, VA, USA Image Computing, University College London, London, UK 13 Vol:.(1234567890) Journal of Neurology (2023) 270:1286–1299 1287 Introduction Table 1  Search strategy and selection criteria Sources Pubmed (https://​www.​ncbi.​nlm.​nih.​gov/​ A diagnosis of multiple sclerosis (MS) requires a sympto- pubmed) matic demyelinating syndrome with objective neurologic Period of time covered From January 1979 until October 2022 findings, the demonstration of a pathological process dis- Search terms “Artificial intelligence”, “Chronic active seminated in space (DIS) and time (DIT) and the exclusion lesions”, “Cortical lesions”, Deep of alternative conditions [1]. learning”, “Diagnostic Criteria”, “Dif- ferential Diagnosis”, “Inflammation”, Recently, the enhanced characterization of clinical and “Iron Rim Lesions”, “Leptomeningeal radiologic features associated with different inflammatory Inflammation”, “Lesion/s”, “MRI”, demyelinating disorders of the CNS [2, 3] and improve- “Machine Learning”, “McDonald cri- ments in neuroimaging and laboratory technologies have teria”, “MOGAD”, “Multiple Sclero- sis”, “NMOSD”, “Paramagnetic Rim contributed to the refinement of the diagnostic work-up of Lesions”, “Primary Progressive”, “Spinal patients with a suspicion of MS [4]. Subsequent iterations Cord”, “Subpial Demyelination”, “White of the McDonald criteria have defined evidence-based Matter” imaging features typical of MS, facilitating earlier fulfill- Selection criteria and 1. Only papers published in English ment of the diagnostic criteria for MS [1]. review preparation 2. The final reference list was generated with the consensus of all co-authors of Since a specific MS biomarker is not available, simpli- this review on the basis of originality fications and easier fulfilment of the diagnostic criteria and relevance to the broad scope of this (e.g., requiring fewer MRI lesions and substituting detec- review, with a focus on the most recent tion of intrathecal immunoglobulin G [IgG] synthesis for articles published in the last three years 3. Experts provided a summary during the DIT) may increase the risk of MS misdiagnosis [5, 6]. meeting of the main findings related to These considerations have prompted extensive research in specific topics of the review. For each the field of neuroimaging to identify novel MRI features topic, a group consensus was reached more specific to MS. Moreover, the use of artificial intelli- and summarized in a first draft, which was circulated among the co-authors for gence (AI) has been recently suggested as a new promising further critical discussion and revision. tool for MS clinical practice [7]. The review represents the final conclu- An international meeting was held on the 3rd of Novem- sions reached by co-authors ber 2021, which involved neurologists and (neuro)radiolo- gists with expertise in MS and its mimics (see Acknowledg- ments for details) to summarize the most recent applications The 2017 McDonald criteria: of MRI in the MS diagnostic work-up but also possible from statements to clinical use future innovations. The key aspects discussed in the meeting included the current evidence regarding the clinical applica- The diagnosis of MS is primarily based on clinical criteria. tion of the 2017 McDonald criteria, promising novel markers Since 2001, MRI has been included in MS diagnostic work- to improve accuracy of diagnosis, and the potential contribu- up to support, supplement, or even replace some clinical tion of AI for MS diagnostic work-up. criteria in excluding differential diagnosis and demonstrating Experts provided a summary related to each topic DIS and DIT [1]. In patients with a typical clinically isolated (see Table 1 for search strategy and selection criteria). syndrome (CIS) suggestive of MS, from their introduction, A group consensus was reached during the meeting and subsequent iterations of the McDonald criteria have sim- summarized in a first draft, which was circulated among plified MS diagnosis, improving sensitivity and preserving the speakers and additional experts in the field for critical accuracy. The last revision of the McDonald criteria (i.e., discussion and revision. the 2017 McDonald criteria) [1] included the removal of any This review summarizes the current state-of-the-art and distinction between symptomatic and asymptomatic lesions, possible future applications of MRI technologies for the and the combination of cortical lesions and juxtacortical diagnostic work-up of MS. lesions to expand the concept of juxtacortical involvement. Increasing attention is also given to the diagnosis and Furthermore, in patients with a typical CIS suggestive of prognostication of subjects with brain MRI abnormali- MS, the presence of CSF-specific oligoclonal bands (OCBs) ties suggestive of MS, but lacking historical accounts of supplants demonstration of DIT (Table 2). prior demyelinating events (i.e., “radiologically isolated Several validation studies in different countries [11–16] syndrome” [RIS] or “prodromic phase of MS”). This is showed that the 2017 McDonald criteria have higher sensi- out-of-scope of the present review and it is described else- tivity, lower specificity and similar accuracy compared with where [8–10]. 13 1288 Table 2  The 2017 McDonald criteria for diagnosis of multiple sclerosis Clinical presentation Clinical presentation Additional data needed for MS diagnosis 13 Relapse-onset (CIS) ≥ 2 clinical relapses and objective clinical evidence of ≥ 2 lesions; None OR ≥ 2 clinical relapses and objective clinical evidence of 1 lesion and clear-cut historical evidence of a prior relapse involving a lesion in a distinct anatomic location ≥ 2 clinical relapses and objective clinical evidence of 1 lesion DIS, demonstrated by: A second clinical relapse implicating a different CNS site OR demonstration of DIS by MRI (≥ 1 lesion in ≥ 2 of the following regions: periventricular, cortical/juxtacortical, posterior fossa, spinal cord) 1 clinical relapse and objective clinical evidence of 2 or more lesions DIT, demonstrated by: A second clinical relapse OR demonstration of DIT by MRI Simultaneous presence of Gd-enhancing and non-enhancing lesions at any time (including symptomatic lesions) or a new T2 and/or Gd-enhancing lesion on FU MRI irrespective of timing of baseline scan OR demonstration of CSF-specific OCBs* 1 clinical relapse and objective clinical evidence of 1 lesion DIS and DIT, demonstrated by: For DIS: A second clinical relapse implicating a different CNS site OR demonstration of DIS by MRI (≥ 1 lesion in ≥ 2 of the following regions: periventricular, cortical/juxtacortical, posterior fossa, spinal cord) For DIT: A second clinical relapse OR demonstration of DIT by MRI Simultaneous presence of Gd-enhancing and non-enhancing lesions at any time (including symptomatic lesions) or a new T2 and/or Gd-enhancing lesion on FU MRI irrespective of timing of baseline scan OR demonstration of CSF-specific OCBs* Progressive-onset (PPMS) 1 year of disability progression (retrospectively or prospectively determined) independent ≥ 2 out of 3 of the following criteria: of clinical relapse • ≥ 1 T2-hyperintense lesions in ≥ 1 areas in the brain characteristic of MS (periventricular, cortical/juxtacortical or infratentorial) • ≥ 2 T2-hyperintense lesions in the spinal cord, with no distinction between symptomatic or asymptomatic lesions • Presence of CSF-specific OCBs CIS clinically isolated syndrome, CNS central nervous system, DIS dissemination in time, DIT dissemination in time, Gd gadolinium, MRI magnetic resonance imaging, MS multiple sclerosis, OCB oligoclonal band, PPMS primary progressive multiple sclerosis *In patients with a typical CIS suggestive of MS fulfilling DIS criteria and with no better explanation for the clinical presentation, the demonstration of CSF-specific OCBs substitutes for the requirement of fulfilling DIT, thus allowing a diagnosis of MS, even if the clinical and MRI findings do not meet the criteria for DIT Journal of Neurology (2023) 270:1286–1299 Journal of Neurology (2023) 270:1286–1299 1289 the 2010 criteria in predicting the second clinical attack not accuracy (91 vs 87%) of the 2017 and 2010 McDonald cri- only in adults, but also in pediatric patients. teria for progressive- and relapse-onset MS were similar, In a recent large multicenter study with 785 CIS patients suggesting to apply a single set of MRI criteria for both suggestive of MS from 9 European centers, the 2017 vs 2010 relapse-onset and progressive-onset patients [22]. McDonald criteria had higher sensitivity (0.83 vs 0.66), Simplification and liberalization of MS diagnostic cri- lower specificity (0.39 vs 0.60), but similar area under the teria raised concerns for an increased risk of misdiagnosis curve (AUC) values (0.61 vs 0.63) [11]. and inappropriate use of treatments [5, 6]. MS misdiagno- The inclusion of lesions in the symptomatic region in ses mostly occur due to inappropriate application of the patients with CIS with a brainstem or spinal cord onset is diagnostic criteria in patients with other inflammatory CNS likely to increase the sensitivity and to decrease specificity, disorders that can meet the DIS and DIT requirements, or without affecting the accuracy of diagnostic criteria [17]. erroneous interpretation of MRI abnormalities [5, 6]. To Although cortical lesion assessment in the diagnostic algo- minimize misdiagnosis a careful exclusion of alternative rithm of CIS patients has been found to increase specific- diagnoses is necessary before applying the 2017 McDonald ity [18, 19], the combination of cortical and juxtacortical criteria since they should be applied primarily in patients lesions to define juxtacortical involvement does not substan- with a typical CIS [1]. tially influence the performance of diagnostic criteria [17]. To this aim, standardized MRI protocols [23], careful In addition to MRI modifications, CSF-derived data also determination of which imaging patterns constitute ‘typi- influence the performance of diagnostic criteria since the cal’ or ‘atypical’ MS features and guidelines for a proper evaluation of the 2017 McDonald criteria without CSF-spe- interpretation of imaging findings [4] are crucial. cific OCB assessment decreased sensitivity (0.74), increased In particular, small-vessel disease caused by chronic specificity (0.54), and preserved AUC values (0.64) [11]. hypoxia represent the most common differential diagno- Moreover, the 2017 McDonald criteria substantially sis for white matter (WM) lesions on brain MRI. This may shorten the time to MS diagnosis, with more CIS patients occur as an age-related phenomenon and is found more fulfilling a diagnosis of MS already at the time of the first commonly in smokers, patients with hypertension, diabetes clinical manifestation and with a single MRI scan [11–13]. and migraine and in various other vascular disorders [24]. The 2017 McDonald criteria shortened the median time Although often recognizable as small, rounded lesions in the to MS diagnosis by 4.6 years compared with the clinical deep WM, when becoming more numerous, they may also criterion alone and by 10 months compared with the 2010 be periventricular. For this reason, the magnetic resonance McDonald criteria (median time to MS diagnosis: 2017 imaging in MS (MAGNIMS) group recommended requir- McDonald criteria = 3.2 months) [11]. This earlier diagnosis ing at least 3 periventricular lesions, especially in elderly is possible not only thanks to modifications of MRI criteria, patients or those with cardiovascular risk factors [1]. This but also to the relevant contribution of CSF-specific OCB criterion has been supported also by a recent multicenter evaluation (median time to MS diagnosis: 2017 McDonald study showing that three periventricular lesions improved criteria without OCBs = 11.4 months) [11]. the specificity and accuracy of the 2017 McDonald DIS cri- An earlier MS diagnosis may facilitate earlier treatment. teria, especially in patients with CIS aged ≥ 45 years [11]. In a study of 1174 patients with CIS suggestive of MS [20], Recent studies explored whether the inclusion of optic the median times from CIS to MS diagnosis and from CIS nerve involvement as a fifth region for DIS fulfilment in to treatment initiation were reduced by 77% and 82%, from CIS patients who had visual-evoked potentials or optic nerve the Poser [21] to the 2017 McDonald criteria [1] periods. A MRI evaluations influenced the performance of the diagnos- significantly lower risk of reaching an Expanded Disability tic criteria [17, 25, 26]. Such an addition slightly improved Status Scale (EDSS) score ≥ 3.0 was also found for patients the diagnostic performance by increasing sensitivity without diagnosed with the most recent diagnostic criteria [20]. substantially losing specificity, both in CIS patients with Up to 15% of MS patients experience a gradual clinical and without optic neuritis as the first manifestation of the progression from disease onset (i.e., primary progressive disease [17, 25, 26]. [PP] MS) [1]. According to the 2017 McDonald criteria, Moreover, more distinctive MRI features of MS need to PPMS can be diagnosed in patients with ≥ 1 year of disabil- be identified and validated. Among these, the central vein ity progression independent of clinical relapses who also sign (CVS), leptomeningeal enhancement, subpial demyeli- fulfill at least two of the following three criteria: (1) ≥ 1 nation, and chronic active lesions have been focus of several lesion(s) in 1 or more topography including periventricu- recent lines of research and discussion. lar, cortical/juxtacortical, or infratentorial brain regions; (2) ≥ 2 lesions in the spinal cord; and (3) CSF-specific OCBs (Table 2). A recent study with 117 PPMS patients showed that sensitivity (89 vs 85%), specificity (100 vs 100%) and 13 1290 Journal of Neurology (2023) 270:1286–1299 Fig. 1  The central vein sign. Representative 3D-EPI T2*-magnitude ter lesions. In the magnified views, a central vein running through the images in A axial, B sagittal, and C coronal planes acquired at 3T lesion (red arrows) is visible as a hypointense line (axial and coronal during the injection of gadolinium-based intravenous contrast agent planes) or a hypointense dot (sagittal plane). Abbreviations: 3D-EPI in a 24-years-old relapsing–remitting multiple sclerosis patient. A three-dimensional echo planar imaging conspicuous central vein sign is present in the majority of white mat- The central vein sign Existing evidence from the literature (mainly based on cross-sectional studies) has shown that a 40% CVS posi- Several recent studies have shown that the presence of a vein tive lesion cut-off can accurately discriminate between at the center of WM lesions, the so-called ‘CVS’ [27], can MS and other WM diseases [28]. However, the number of improve the differentiation between MS and its mimics [28]. CVS-positive lesions detected on a patient’s scan signifi- The perivenular nature of MS lesions on ex-vivo histopathol- cantly depends on the type of susceptibility sequence used, ogy is well known since many years. Today, this MS specific the MRI field strength and the injection of intravenous Gd histopathological feature can be detected in-vivo using sus- (which itself is also paramagnetic, increasing the visibility ceptibility-weighted MRI [27, 28]. The percentage of WM of veins) [28]. For all these reasons, the percentage of MS lesions featuring a central vein on susceptibility-weighted CVS lesions and the optimal diagnostic cut-off threshold is MRI is substantially higher in MS compared to other MS higher when using optimized susceptibility-weighted MRI mimics [28–30] such as migraine, small vessel disease, sequences at high and ultra-high field strength (3T and 7T) inflammatory CNS vasculopathies, Susac syndrome, neu- after Gd injection [28]. Among the different susceptibility- romyelitis optica spectrum disorders (NMOSD), and myelin weighted MRI sequences tested so far, the T2*-weighted oligodendrocyte glycoprotein antibody disease (MOGAD). three-dimensional echo-planar imaging (3D-EPI) has shown In MS, most newly developing T2-hyperintense WM the highest capability to detect central veins in MS (Fig. 1) lesions or gadolinium (Gd)-enhancing lesions show the [28, 29]. This is probably due to the high isotropic spa- CVS. In a 2.8 years longitudinal study, 62/153 (40.5%) tial resolution of the 3D-EPI sequence allowing to detect MS patients developed new WM lesions and 50 of them small intralesional veins in an acceptable scan time. Key (80.6%) showed new lesions with the CVS [31]. Moreover, challenges preventing the introduction of CVS assessment 159/233 (68.2%) new T2-hyperintense lesions and 92/114 in clinical practice include: (1) the lack of a standardized Gd-enhancing lesions (80.7%) had the CVS, especially in imaging protocol (mainly because of the limited availability MS patients with a younger age and a higher percentage of of optimized MRI sequences from scanner manufacturers), CVS-positive lesions at baseline [31]. (2) the need of validated CVS-based diagnostic criteria and (3) the need of automated methods to detect the CVS. To 13 Journal of Neurology (2023) 270:1286–1299 1291 Fig. 2  Chronic active lesions. Example of chronic active lesion vis- T2-hyperintense white matter lesion (white arrows among red-coded ualization using susceptibility-weighted MRI. A On 3D axial fluid- lesion mask) show a hypointense rim on phase image derived from a attenuated inversion recovery sequence in a 48-years-old secondary multi echo gradient-echo T2* sequence (C), thus they represent ‘par- progressive multiple sclerosis patient (A–B), two confluent periven- amagnetic rim lesions’ (PRLs) tricular T2-hyperintense white matter lesions and another subcortical overcome these limitations, different CVS-based diagnos- increase in size, at least in the first years after their forma- tic algorithms, including simplified lesion-based diagnostic tion [37, 38]. algorithms (not requiring the analysis of every single lesion) A higher prevalence and/or number of PRLs have been [32], are being tested on a large prospective multicenter set- found to be quite specific and clinically relevant for MS. ting [33]. They have been described in patients with radiologically Concomitantly, existing statistical and deep-learning isolated syndrome (RIS) (12% of all WM lesions, with 61% based methods for automated CVS detection [34, 35] are of RIS patients having ≥ 1 PRLs) [39] and CIS (19.9% of all being tested on a large scale [33] and efforts to improve WM lesions) [40], but not in MS-mimics [40–42], including their performance are ongoing. Finally, although CVS NMOSD, Susac syndrome, and small-vessel disease. could potentially be assessed in all scanner manufactures, In a large cohort of subjects with CIS/MS (n = 254), the different scanner vendors should make optimized and MS mimickers (n = 91) and old healthy controls (n = 271), standardized MRI sequences (like the 3D-EPI) available in the identification of ≥ 1 PRLs was the optimal cut-off to all MS specialized centers. Given its promising diagnostic distinguish CIS/MS patients from MS mimickers and old performance, the CVS is likely to be incorporated in the MS healthy controls (specificity = 99.7%, sensitivity = 24.0%, diagnostic criteria in the near future. AUC = 0.71, 95% confidence interval [CI] = 0.64–0.78). Of note, the fulfilment of ≥ 1 PRLs or ≥ 4 lesions with CVS improved specificity (90.6%), sensitivity (57.9%) and AUC Chronic active lesions (0.83, 95% CI = 0.79–0.87) [42]. In CIS patients, the presence of ≥ 1 PRLs and/or the Pathological studies have revealed that up to 57% of all fulfilment of ‘CVS’ criteria (≥ 3 lesions or 40% threshold chronic WM lesions show a peripheral ‘rim’ of iron-laden of lesions with the CVS) predicted MS conversion after activated microglia/macrophages associated with ongoing 3 years with good sensitivity (70.2–90.4%) and specific- demyelination and axonal loss, around an inactive core with- ity (35.7–85.7%) [40]. Of note, none of the patients who out blood–brain barrier damage [36]. remained CIS after 3 years had any PRLs [40]. Chronic active lesions have been evaluated by looking at Recently, reliable methods have been proposed to auto- susceptibility-weighted MRI scans at high- and ultra-high matically detect PRLs [43, 44]. A fully automated method, field [37] (Fig. 2). On these sequences, chronic active lesions which applied lesion-level radiomic feature extraction and show a paramagnetic hypointense rim (i.e., ‘paramagnetic machine learning on 3D T1-weighted, 3D T2-FLAIR and rim lesions’ [PRLs]), which corresponds to peripheral 3D T2*-phase MRI sequences, showed a strong correlation iron-laden microglia/macrophages [37]. Compared to iron (r = 0.91) with manual rating and an AUC of 0.80 in cor- negative lesions, PRLs have more severe myelin and axon rectly classifying PRLs in MS patients [44]. pathology, a more limited lesional repair and a slow rate of 13 1292 Journal of Neurology (2023) 270:1286–1299 By applying a multimodal patch-based convolutional However, the finding of LME on post-contrast FLAIR neural network (CNN) (RimNet) on 3D T2*-EPI and 3D is not specific to MS. A 7T study described two patterns of T2-FLAIR sequences, a recent study showed a sensitivity LME: “nodular” (i.e., discrete, spherical nodules at the pial (70.6%), specificity (94.9%) and AUC (0.943) comparable surface or subarachnoid space) and “spread/fill” (appearance to manual rating for PRL identification [43]. of contrast spread through the local subarachnoid space). By applying a two-branch feature extraction network Nodular foci were present in 15 of 29 (51%) MS patients and a synthetic minority oversampling network (QSMRim- and in 2 out of 3 (67%) of healthy controls, whereas spread/ Net) on quantitative susceptibility mapping (QSM) and fill foci were present in 22 of 29 (76%) MS patients [51]. T2-weighted FLAIR (T2-FLAIR), the proposed methods In another study, LME was found in 5 out of 66 healthy showed better sensitivity (68%), specificity (98.9%) and controls (5.7%) [52]. accuracy (97.6%) compared to other state-of-the-art methods Brain LME, either as a single or multiple nodular foci, is applied to quantify PRLs on QSM [45]. also frequently seen in patients with other inflammatory neu- rologic conditions (18/51 cases [35.3%]), such as NMOSD (6/11 cases [54.5%]) and MOGAD (3/11 cases [27.3%]), Leptomeningeal enhancement non-inflammatory neurologic diseases (i.e., including small vessel disease, migraine, neurodegenerative diseases, Mild, widespread perivascular inflammatory infiltrates in and compressive myelopathy) (3/38 cases [7.9%]), revers- the meninges are a nearly ubiquitous autopsy finding in ible cerebrovascular constriction syndrome (69/182 cases all forms of MS and approximately 40–50% of those with [37.9%]), and infectious conditions such as human T-lym- SPMS have focal areas of leptomeningeal ectopic lymphoid photropic virus (HTLV) infection (17/38 cases [44.7%]) and tissue (“meningeal follicles”) [46]. Gradients of neuronal human immunodeficiency virus (HIV) (13/61 cases [21.3%]) loss and cortical demyelination emanating from these fol- [52–55]. licles have been found at autopsy, suggesting a direct causa- A recent meta-analysis evaluated the presence of LME tive relationship to cortical lesion formation [47]. Further, in neoplastic neurological (n = 2392 cases), neuroinfectious intrathecal production of inflammatory cytokines associated (n = 1890 cases) and primary neuro-inflammatory diseases with lymphoid follicle activity (i.e., CXCL13, IFN-γ) has (n = 4038) [49]. The LME proportions for these disease been suggested to predict cortical damage on MRI [47, 48]. classes were 0.47 (95% confidence interval [CI] = 0.37;0.57), The potential importance of meningeal inflammation as a 0.59 (95% CI = 0.47;0.69), and 0.26 (95% CI = 0.20;0.35), contributor to MS pathology led to the search for surrogate respectively. In a subgroup analysis for MS (n = 1605 cases), imaging biomarkers. Although meningeal pathology in MS LME proportion was 0.30 (95% CI = 0.21;0.42) with lower does not cause post-contrast enhancement on T1-weighted proportions in RRMS (0.19 [95% CI = 0.13;0.27]) compared imaging, it has long been known that this technique is not as to progressive MS (0.39 [95% CI = 0.30;0.49], p = 0.002) sensitive to meningeal disease as post-contrast fluid attenu- [49]. ated inversion recovery (FLAIR) MRI. FLAIR is > 10 times Accordingly, brain LME on post-contrast FLAIR lacks more sensitive to small concentrations of Gd in CSF and is the specificity necessary for use as a diagnostic biomarker more sensitive to the presence of leptomeningeal enhance- in MS. Moreover, although LME could become a surrogate ment (LME) in conditions such as leptomeningeal carcino- of meningeal inflammation in MS, a direct relationship is matosis and infectious meningitis. not yet established. Recent work showed that post-contrast Although, at present, no study directly compared 3T FLAIR MRI may also directly visualize brain lymphatic and 7T images, a recent meta-analysis showed that, in MS vessels and glymphatic clearance [55, 56], suggesting that patients, higher LME-proportions were found in studies focal blood-meningeal barrier breakdown may not be the imaging at 7T (0.79 [95%-CI 0.64–0.89]) compared to lower only source of enhancement. field strengths (0.21 [95% CI 0.15–0.29], p < 0.001) [49]. With this knowledge, post-contrast FLAIR MRI protocols were tested as a potential surrogate biomarker of menin- Subpial demyelination geal inflammation in MS (Fig. 3). A large study of delayed- acquisition, 3D FLAIR on 3T MRI showed Gd deposition in Cortical demyelinating lesions are an established patho- the leptomeningeal space in 25% of MS patients compared logical feature of MS, can develop from the earliest disease to 2.7% of controls [50]. Two patients from this study later stages and represent one of the main substrates of disease went to autopsy and cellular inflammatory infiltrates were progression [4]. Detection of cortical lesions by MRI can found in regions of leptomeninges correlating to locations be used in clinic as evidence of DIS to aid the diagnosis of enhancing foci during life, supporting LME as a potential of MS. Clinical MRI protocols at 1.5 and 3T are, however, surrogate of meningeal inflammation in MS. hampered in their ability to visualize and characterize the 13 Journal of Neurology (2023) 270:1286–1299 1293 Fig. 3  Leptomeningeal enhancement, cortical lesions and subpial (C) or more extensive (D, E) subpial multiple sclerosis lesions (black demyelination. A, B Example of 7T FLAIR MRI of the brain in arrows) with expansion within white matter/confluence with a juxta- a 49-years-old woman with relapsing–remitting multiple sclero- cortical lesion (E) in patients with multiple sclerosis as seen on ultra- sis before (A) and after (B) the administration of gadolinium-based high resolution T2* gradient echo images at 7 Tesla. A white mat- intravenous contrast agent. A focus of post-contrast pial/subarach- ter lesion is also visible in (E) (white arrow). Abbreviations: FLAIR noid enhancement highlighted by white arrow. Examples of focal fluid-attenuated inversion recovery, MRI magnetic resonance imaging full extent of cortical demyelination in vivo and most of the than double cortical lesion detection relative to 3T MRI [57]. knowledge on cortical MS lesions derives from post-mortem The gain in resolution and delineation across cortical layers studies. Neuropathological examinations have identified dif- allows the identification at 7T of same cortical lesion sub- ferent types of cortical lesions in MS based on their loca- types observed in histopathological examinations [58, 59]. tion within the cortical laminae: type I leukocortical lesions Increased cortical lesion detection achieved at 7T is of located at cortical-subcortical junction; type II, intracortical high interest because subpial lesions are a typical finding plaques, very small lesions completely enclosed in the cortex in MS [60] and their visualization could, therefore, improve without reaching its margins; type III–IV subpial lesions, the MS diagnostic sensitivity and specificity. However, corti- extending downwards from the juxtameningeal pial surface cal lesions and subpial demyelination are also not fully spe- through different cortical laminae or across the full cortical cific for MS since they have been also described in patients depth. Subpial lesions represent the most common type of with acute disseminated encephalomyelitis (ADEM) [60] cortical MS lesions and is vastly underdetected by conven- and MOGAD [53]. tional MRI. Different pulse sequences have been optimized at 7T, Ultra-high field 7T MRI provides several advantages including T2*-weighted gradient-echo, T2-weighted, over lower field MR strength in imaging cortical lesions in FLAIR, double inversion recovery (DIR), and T1-weighted MS (Fig. 3). The increase in signal-to-noise ratio (SNR) magnetization-prepared rapid-acquisition gradient-echo achieved at ultra-high field can be translated into a resolu- (MPRAGE/MP2RAGE), to image cortical lesions in differ- tion in the sub-millimeter range and exquisite contrast for ent MS stages [57–59, 61]. As different sequences seem to imaging details within both gray matter (GM) and WM. favor the identification of different cortical lesion subtypes Although we are still far from revealing the true number (e.g., T2*-weighted gradient-echo for subpial lesion detec- of cortical lesions in vivo, correlative histopathological 7T tion [57]; MP2RAGE for small intracortical lesions) [62], MRI assessments have shown ultra-high field MRI can more the information provided by different contrasts is frequently 13 1294 Journal of Neurology (2023) 270:1286–1299 complementary and, when used jointly, can only increase discriminated AQP4-positive NMOSD from MS patients the definition of cortical lesions. Cortical lesion identifica- with an accuracy similar to expert neurologists (accu- tion and segmentation, however, strongly rely on readers’ racy = 71.1% vs 65.9–60.7%), but higher reliability (human training and experience, and the attempts to automatize this intra-rater reliability of 0.47–0.50) [78]. In another study, a process are still far from a clinically acceptable performance CNN algorithm trained on brain FLAIR and T1-weighted [63, 64]. Further improvements in the standardization of sequences showed a higher accuracy compared to that of both acquisitions and processing methods across Centers two expert neuroradiologists in correctly discriminate MS are still needed for a translation of these techniques in clinic. (98.8% vs 72.8–81.8%) from other MS mimics, including The increase in accuracy of cortical lesion detection at NMOSD (88.6 vs 4.4%), migraine (92.2% vs 53–64.8%) 7T could also be used to guide selective characterization of and CNS vasculitis (92.1% vs 45.5–54.6%) [79]. Other ML the different pathological components of cortical lesions, and DL algorithms trained on conventional MRI sequences including demyelination and inflammation, in MS through (FLAIR, PD, T2-weighted and T1-weighted) or proton mag- complementary techniques such as PET, which are able to netic resonance spectroscopy were also able to discriminate estimate the distribution of specific molecules or processes between MS, non-inflammatory WM disorders (hereditary of interest. Investigations combining 7T MRI and PET diffuse leukodystrophy with spheroids and cerebral microan- imaging of neuroinflammation have revealed an association giopathy) [80, 81], or low-grade brain tumors [82]. between cortical demyelination and glial activation [65] as ML algorithms applied on baseline demographic (age, well as heterogeneity in levels of inflammation within corti- sex), clinical (EDSS score and type of onset), and brain cal lesions [66]. This approach could be used in the future MRI features (including WM lesion count, radiomic fea- to stage cortical lesion inflammatory activity in vivo and, tures, regional GM atrophy and cortical thickness) predicted potentially, to improve MS diagnosis. also conversion from CIS to MS with an average accuracy between 71.4 and 92.9% at 1 year [83, 84], between 67.6 and 70.4% at 2 years [85] and between 68.0 and 85.0% at 3 years The contribution of AI follow-up [83, 86]. Although AI methods have been showing increasingly Recent improvements in technologies and the availability promising results in MS diagnostic work-up, several limita- of large amount of data have promoted the application of tions should be taken into account. First, it may be chal- AI algorithms for the diagnostic-work up of MS [7]. Using lenging to interpret neural network decisions. For instance, CNN, a model of deep-learning (DL) tool able to automati- AI methods (especially DL algorithms) may follow “short- cally select the best problem-solving features, recent AI stud- cut” strategies, which, while superficially successful (i.e., ies were able to discriminate between MS patients and HC differentiation between MS patients from HC based on the with an accuracy between 70.2 and 98.8% from T2-weighted presence/absence of lesions), typically fail under different [67, 68], FLAIR [69], or susceptibility-weighted MRI circumstances (i.e., differentiation between MS patients sequences [70]. Machine learning (ML) algorithms, used to from others with WM lesions due to cerebral small vessel learn from specific predefined data features and then make disease). Moreover, possible selection bias and overfitting decisions and trained on quantitative [71–73], diffusion- may overestimate the performance of AI algorithms. AI weighted [74], and resting state functional [74, 75] MRI approaches also need large datasets, which may be chal- sequences, were also able to correctly identify MS patients lenging to obtain due to system availability, high costs, and with an accuracy ranging from 83.7 to 90.0%. heterogeneous acquisition methodologies. To overcome such AI algorithms can contribute discriminating MS from limitation and to allow their use in the clinical scenario, AI mimics. Using random forest, a supervised learning algo- algorithms should be cross-validated in multicenter, pro- rithm, on a set of brain GM imaging measures, lower tha- spective and longitudinal real-world cohorts, to overcome lamic volume together with other measures of brain GM vol- challenges due to variability of image acquisition parameters umes and cortical thickness obtained from high-resolution and scanner models and the presence of heterogeneous dis- T1-weighted sequences were found to discriminate NMOSD tribution of data between an algorithm's training dataset and from MS with an accuracy of 74% [76]. A multiparametric validation. Furthermore, AI algorithms should be integrated approach including data from FLAIR, diffusion-tensor imag- into existing information technology infrastructures and the ing, resting state functional MRI plus clinical and neuropsy- access to the required computing power should be guaran- chological information improved the diagnostic performance teed. Finally, guidelines that ensure reliability and validity of to 88% [77]. findings obtained from AI approaches, and standard thresh- A CNN algorithm applied on brain FLAIR sequences and olds for the accuracy of models required for publication, patients’ clinical information (age at disease onset, age at the which are lacking at present. time of MRI, disease duration, time from the last relapse) 13 Journal of Neurology (2023) 270:1286–1299 1295 Table 3  Summary of newly proposed MRI markers for the diagnostic work-up of MS MRI marker MRI Pathological Specificity in Prediction of Overall Prediction Feasibility in Associations sequence(s) substrate differential MS conver- contribution of disability the clinical with brain diagnosis sion for early MS progression setting damage diagnosis Central vein SWI Perivenular +  +  + + +  +  + ? +  +  + ? sign (T2*-weighted inflamma- 3D-EPI) tion Paramagnetic SWI Chronic active +  +  + +  + +  + +  + +  +  + +  + rim lesions (T2*-weighted lesions 3D-EPI) Leptome- Post-Gd Meningeal + ? ± + +  + + ningeal T2-weighted inflamma- enhancement FLAIR tion Cortical DIR; Cortical +  +  + +  + +  + +  +  + +  + +  + lesions 3D demyelina- T1-weighted tion MPRAGE/ MP2RAGE PSIR Subpial demy- T2*-weighted Cortical +  +  + +  + ? +  +  + + +  + elination gradient- demyelina- echo tion 3D-EPI three-dimensional echo-planar imaging, DIR double inversion recovery, FLAIR fluid-attenuated inversion recovery, Gd gadolinium, MP2RAGE magnetization prepared 2 rapid gradient echoes, MPRAGE magnetization-prepared rapid gradient echo, MRI magnetic resonance imaging, MS multiple sclerosis, PSIR phase-sensitive inversion recovery, SWI susceptibility weighted imaging Conclusions Tesla scanners), whereas leptomeningeal enhancement is not MS-specific, since it can be detected in other inflammatory Accurate criteria in the diagnostic work-up of patients with CNS diseases [52] and with ageing [87]. a suspicion of MS are crucial not only to enable an early In the near future, AI approaches may represent a compli- diagnosis, thus allowing treatment to start sooner, but also to mentary tool for neurologists and neuroradiologists. Beside minimize the risk of misdiagnosis and overdiagnosis. visual-pattern recognition performed by experienced cli- The 2017 McDonald criteria are validated and evidence- nicians of lesional features typical for MS, which may be based criteria that show high sensitivity and accuracy in time-consuming and hardly reproducible, AI may allow to predicting the occurrence of a second clinical attack, sim- identify textures and patterns on MRI sequences that are plify the clinical use of MRI criteria, and allow an earlier beyond the human perception and may further improve the diagnosis and treatment of MS. Some concerns have been diagnostic work-up and patients’ classification. raised due to their low specificity. However, their application Acknowledgements This paper reports the conclusions of the “Mag- is recommended only after alternative diagnoses have been netic Resonance Techniques in Multiple Sclerosis, Twenty-third carefully excluded. Advanced Course”, held in Milan, Italy, in November 3­ rd 2021. The To further improve the diagnostic process, novel candi- authors had the following roles: Chair, Massimo Filippi Speakers, date imaging biomarkers, such as CVS and chronic active D. L. Arnold, F. Barkhof, D.M. Harrison, P. Maggi, C. Mainero, X. Montalban, E. Sechi, B.G. Weinshenker, M.A. Rocca; Discussants, P. lesions have been proposed to increase the specificity of MS Preziosa. The course was supported by unrestricted education grants diagnostic criteria, thus reducing the risk of misdiagnosis from Biogen and Merck. The Sponsor had no role in the design and (Table 3). conduct of the study; collection, management, analysis, and interpre- However, they should be further validated and standard- tation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. F. Barkhof is ized before being implemented in the clinical setting. In par- supported by the NIHR biomedical research centre at UCLH. ticular, future studies should ascertain their role to exclude alternative diagnoses but also to diagnose MS in patients Funding None. with both relapse- or progressive-onsets. Data availability This is a review article. Publications cited in the Subpial demyelination is highly specific for MS but manuscript are generally available online, as referenced in the refer- hardly visible at standard field strengths (i.e., 1.5 and 3.0 ence section. 13 1296 Journal of Neurology (2023) 270:1286–1299 Declarations the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will Conflicts of interest Massimo Filippi is Editor-in-Chief of the Journal need to obtain permission directly from the copyright holder. To view a of Neurology, Associate Editor of Human Brain Mapping, Associate copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. Editor of Radiology, and Associate Editor of Neurological Sciences, received compensation for consulting services from Alexion, Almi- rall, Biogen, Merck, Novartis, Roche, Sanofi, speaking activities from Bayer, Biogen, Celgene, Chiesi Italia SpA, Eli Lilly, Genzyme, Jans- References sen, Merck-Serono, Neopharmed Gentili, Novartis, Roche, Sanofi, Takeda, and TEVA, participation in Advisory Boards for Alexion, 1. Thompson AJ, Banwell BL, Barkhof F, Carroll WM, Coetzee T, Biogen, Bristol-Myers Squibb, Merck, Novartis, Roche, Sanofi, Sa- Comi G et al (2018) Diagnosis of multiple sclerosis: 2017 revi- nofi-Aventis, Sanofi-Genzyme, Takeda, scientific direction of educa- sions of the McDonald criteria. Lancet Neurol 17(2):162–173 tional evens for Biogen, Bristol-Myers Squibb, Celgene, Lilly, Merck, 2. Kim HJ, Paul F, Lana-Peixoto MA, Tenembaum S, Asgari N, Novartis, Roche, Sanofi-Genzyme, he receives research support from Palace J et al (2015) MRI characteristics of neuromyelitis Biogen Idec, Merck-Serono, Novartis, Roche, Italian Ministry of optica spectrum disorder: an international update. Neurology Health, Fondazione Italiana Sclerosi Multipla, and ARiSLA (Fondazi- 84(11):1165–1173 one Italiana di Ricerca per la SLA); Paolo Preziosa received speaker 3. Marignier R, Hacohen Y, Cobo-Calvo A, Probstel AK, Aktas O, honoraria from Roche, Biogen, Novartis, Merck Serono, Bristol Myers Alexopoulos H et al (2021) Myelin-oligodendrocyte glycoprotein Squibb and Genzyme, he has received research support from Italian antibody-associated disease. Lancet Neurol 20(9):762–772 Ministry of Health and Fondazione Italiana Sclerosi Multipla; Douglas 4. Filippi M, Preziosa P, Banwell BL, Barkhof F, Ciccarelli O, De L. Arnold reports consulting fees from Biogen, Celgene, Frequency Stefano N et al (2019) Assessment of lesions on magnetic reso- Therapeutics, Genentech, Merck, Novartis, Race to Erase MS, Roche, nance imaging in multiple sclerosis: practical guidelines. Brain and Sanofi-Aventis, Shionogi, Xfacto Communications, grants from 142(7):1858–1875 Immunotec and Novartis, and an equity interest in NeuroRx; Frederik 5. Solomon AJ, Naismith RT, Cross AH (2019) Misdiagnosis of Barkhof acts in steering committee or iDMC member for Biogen, Mer- multiple sclerosis: Impact of the 2017 McDonald criteria on clini- ck, Roche, EISAI and Prothena. Consultant for Roche, Biogen, Merck, cal practice. Neurology 92(1):26–33 IXICO, Jansen, Combinostics. Research agreements with Merck, Bi- 6. Solomon AJ, Pettigrew R, Naismith RT, Chahin S, Krieger S, ogen, GE Healthcare, Roche, Co-founder and shareholder of Queen Weinshenker B (2021) Challenges in multiple sclerosis diagnosis: Square Analytics LTD; Daniel M. Harrison has received consulting misunderstanding and misapplication of the McDonald criteria. fees from Genentech, EMD-Serono, Biogen, and Sanofi-Genzyme, has Mult Scler 27(2):250–258 received research support from EMD-Serono and Genentech, has re- 7. Bonacchi R, Filippi M, Rocca MA (2022) Role of artificial intel- ceived royalties and writing fees for UpToDate, Inc. and the American ligence in MS clinical practice. Neuroimage Clin 35:103065 College of Physicians; Pietro Maggi received research support from 8. De Stefano N, Giorgio A, Tintore M, Pia Amato M, Kappos L, Fund for Scientific Research (F.R.S, FNRS), Cliniques universitaires Palace J et al (2018) Radiologically isolated syndrome or subclini- Saint-Luc Fonds de Recherche Clinique and Biogen, speaker fees from cal multiple sclerosis: MAGNIMS consensus recommendations. Sanofi-Genzyme and Biogen; Caterina Mainero has nothing to dis- Mult Scler 24(2):214–221 close; Xavier Montalban has received speaking honoraria and travel 9. Okuda DT, Mowry EM, Beheshtian A, Waubant E, Baranzini SE, expenses for participation in scientific meetings, has been a steering Goodin DS et al (2009) Incidental MRI anomalies suggestive of committee member of clinical trials or participated in advisory boards multiple sclerosis: the radiologically isolated syndrome. Neurol- of clinical trials in the past years with Abbvie, Actelion, Alexion, Bio- ogy 72(9):800–805 gen, Bristol-Myers Squibb/Celgene, EMD Serono, Genzyme, Hoff- 10. Calabrese M, Marastoni D, Crescenzo F, Scalfari A (2021) mann-La Roche, Immunic, Janssen Pharmaceuticals, Medday, Merck, Early multiple sclerosis: diagnostic challenges in clinically and Mylan, Nervgen, Novartis, Sandoz, Sanofi-Genzyme, Teva Pharma- radiologically isolated syndrome patients. Curr Opin Neurol ceutical, TG Therapeutics, Excemed, MSIF and NMSS; Elia Sechi has 34(3):277–285 nothing to disclose; Brian G. Weinshenker reports personal fees from 11. Filippi M, Preziosa P, Meani A, Dalla Costa G, Mesaros S, Novartis, MedImmune/VielaBio, Horizon, Alexion, Chugai, Roche, Drulovic J et al (2022) Performance of the 2017 and 2010 revised Genentech Mitsubishi-Tanabe and UCB Biosciences, has a patent of McDonald criteria in predicting MS diagnosis after a clinically NMO-IgG for diagnosis of neuromyelitis optica with royalties paid to isolated syndrome: a MAGNIMS study. Neurology 98(1):e1–e14 RSR Ltd, Oxford University, Hospices Civil de Lyon, and MVZ La- 12. van der de Vuurst Vries RM, Mescheriakova JY, Wong YYM, bor PD Dr Volkmann und Kollegen GbR; Maria A. Rocca received Runia TF, Jafari N, Samijn JP et al (2018) Application of the 2017 consulting fees from Biogen, Bristol Myers Squibb, Eli Lilly, Jans- revised McDonald Criteria for multiple sclerosis to patients with a sen, Roche and speaker honoraria from Bayer, Biogen, Bristol Myers typical clinically isolated syndrome. JAMA Neurol 75(11):1392–8 Squibb, Bromatech, Celgene, Genzyme, Merck Healthcare Germany, 13. Hyun JW, Kim W, Huh SY, Park MS, Ahn SW, Cho JY et al Merck Serono SpA, Novartis, Roche, and Teva, she receives research (2019) Application of the 2017 McDonald diagnostic criteria for support from the MS Society of Canada and Fondazione Italiana Scle- multiple sclerosis in Korean patients with clinically isolated syn- rosi Multipla. She is an Associate Editor for Multiple Sclerosis and drome. Mult Scler 25(11):1488–1495 Related Disorders. 14. Fadda G, Brown RA, Longoni G, Castro DA, O’Mahony J, Ver- hey LH et al (2018) MRI and laboratory features and the per- Open Access This article is licensed under a Creative Commons Attri- formance of international criteria in the diagnosis of multiple bution 4.0 International License, which permits use, sharing, adapta- sclerosis in children and adolescents: a prospective cohort study. tion, distribution and reproduction in any medium or format, as long Lancet Child Adolesc Health 2(3):191–204 as you give appropriate credit to the original author(s) and the source, 15. Hacohen Y, Brownlee W, Mankad K, Chong WK, Thompson A, provide a link to the Creative Commons licence, and indicate if changes Lim M et al (2020) Improved performance of the 2017 McDon- were made. The images or other third party material in this article are ald criteria for diagnosis of multiple sclerosis in children in a included in the article's Creative Commons licence, unless indicated real-life cohort. Mult Scler 26(11):1372–1380 otherwise in a credit line to the material. If material is not included in 13 Journal of Neurology (2023) 270:1286–1299 1297 16. Wong YYM, de Mol CL, van der de Vuurst Vries RM, van Pelt sclerosis with a simplified three-lesion algorithm. Mult Scler ED, Ketelslegers IA, Catsman-Berrevoets CE et al (2019) Real- 24(6):750–757 world validation of the 2017 McDonald criteria for pediatric 33. Ontaneda D, Sati P, Raza P, Kilbane M, Gombos E, Alvarez E MS. Neurol Neuroimmunol Neuroinflamm 6(2):e528 et al (2021) Central vein sign: a diagnostic biomarker in multi- 17. Filippi M, Preziosa P, Meani A, Ciccarelli O, Mesaros S, Rovira ple sclerosis (CAVS-MS) study protocol for a prospective mul- A et al (2018) Prediction of a multiple sclerosis diagnosis in ticenter trial. Neuroimage Clin 32:102834 patients with clinically isolated syndrome using the 2016 MAG- 34. Maggi P, Fartaria MJ, Jorge J, La Rosa F, Absinta M, Sati P NIMS and 2010 McDonald criteria: a retrospective study. Lan- et al (2020) CVSnet: a machine learning approach for automated cet Neurol 17(2):133–142 central vein sign assessment in multiple sclerosis. NMR Biomed 18. Filippi M, Rocca MA, Calabrese M, Sormani MP, Rinaldi F, 33(5):e4283 Perini P et al (2010) Intracortical lesions: relevance for new 35. Dworkin JD, Sati P, Solomon A, Pham DL, Watts R, Martin ML MRI diagnostic criteria for multiple sclerosis. Neurology et al (2018) Automated integration of multimodal MRI for the 75(22):1988–1994 probabilistic detection of the central vein sign in white matter 19. Preziosa P, Rocca MA, Mesaros S, Meani A, Montalban X, lesions. AJNR Am J Neuroradiol 39(10):1806–1813 Drulovic J et al (2018) Diagnosis of multiple sclerosis: a multi- 36. Frischer JM, Weigand SD, Guo Y, Kale N, Parisi JE, Pirko I centre study to compare revised McDonald-2010 and Filippi-2010 et al (2015) Clinical and pathological insights into the dynamic criteria. J Neurol Neurosurg Psychiatry 89(3):316–318 nature of the white matter multiple sclerosis plaque. Ann Neurol 20. Tintore M, Cobo-Calvo A, Carbonell P, Arrambide G, Otero- 78(5):710–721 Romero S, Rio J et al (2021) Effect of changes in MS diag- 37. Dal-Bianco A, Grabner G, Kronnerwetter C, Weber M, Kornek nostic criteria over 25 years on time to treatment and progno- B, Kasprian G et al (2021) Long-term evolution of multiple sis in patients with clinically isolated syndrome. Neurology sclerosis iron rim lesions in 7 T MRI. Brain 144(3):833–847 97(17):e1641–e1652 38. Rahmanzadeh R, Lu PJ, Barakovic M, Weigel M, Maggi P, 21. Poser CM, Paty DW, Scheinberg L, McDonald WI, Davis FA, Nguyen TD et al (2021) Myelin and axon pathology in multi- Ebers GC et al (1983) New diagnostic criteria for multiple sclero- ple sclerosis assessed by myelin water and multi-shell diffusion sis: guidelines for research protocols. Ann Neurol 13(3):227–231 imaging. Brain 144(6):1684–1696 22. Shatila M, Ciccarelli O, Brownlee W (2021) ECTRIMS 2021— 39. Suthiphosuwan S, Sati P, Absinta M, Guenette M, Reich DS, ePoster. Mult Scler J 27(2):134–740 Bharatha A et al (2020) Paramagnetic rim sign in radiologically 23. Wattjes MP, Ciccarelli O, Reich DS, Banwell B, de Stefano N, isolated syndrome. JAMA Neurol 77(5):653–655 Enzinger C et al (2021) 2021 MAGNIMS-CMSC-NAIMS consen- 40. Clarke MA, Pareto D, Pessini-Ferreira L, Arrambide G, Alber- sus recommendations on the use of MRI in patients with multiple ich M, Crescenzo F et al (2020) Value of 3T susceptibility- sclerosis. Lancet Neurol 20(8):653–670 weighted imaging in the diagnosis of multiple sclerosis. AJNR 24. Geraldes R, Jurynczyk M, Dos Passos G, Prichler A, Chung K, Am J Neuroradiol 41(6):1001–1008 Hagens M et al (2020) Distinct influence of different vascular 41. Maggi P, Sati P, Nair G, Cortese ICM, Jacobson S, Smith BR risk factors on white matter brain lesions in multiple sclerosis. J et al (2020) Paramagnetic rim lesions are specific to multiple Neurol Neurosurg Psychiatry 91(4):388–391 sclerosis: an international multicenter 3T MRI study. Ann Neu- 25. Brownlee WJ, Miszkiel KA, Tur C, Barkhof F, Miller DH, Cic- rol 88(5):1034–1042 carelli O (2018) Inclusion of optic nerve involvement in dis- 42. Meaton I, Altokhis A, Allen CM, Clarke MA, Sinnecker T, semination in space criteria for multiple sclerosis. Neurology Meier D et al (2022) Paramagnetic rims are a promising diag- 91(12):e1130–e1134 nostic imaging biomarker in multiple sclerosis. Mult Scler. 26. Vidal-Jordana A, Rovira A, Arrambide G, Otero-Romero S, Rio https://​doi.​org/​10.​1177/​13524​58522​11186​77 J, Comabella M et al (2021) Optic nerve topography in multiple 43. Barquero G, La Rosa F, Kebiri H, Lu PJ, Rahmanzadeh R, sclerosis diagnosis: the utility of visual evoked potentials. Neurol- Weigel M et al (2020) RimNet: a deep 3D multimodal MRI ogy 96(4):e482–e490 architecture for paramagnetic rim lesion assessment in multiple 27. Sati P, Oh J, Constable RT, Evangelou N, Guttmann CR, Henry sclerosis. Neuroimage Clin 28:102412 RG et al (2016) The central vein sign and its clinical evaluation 44. Lou C, Sati P, Absinta M, Clark K, Dworkin JD, Valcarcel AM for the diagnosis of multiple sclerosis: a consensus statement from et al (2021) Fully automated detection of paramagnetic rims in the North American Imaging in Multiple Sclerosis Cooperative. multiple sclerosis lesions on 3T susceptibility-based MR imag- Nat Rev Neurol 12(12):714–722 ing. Neuroimage Clin 32:102796 28. Castellaro M, Tamanti A, Pisani AI, Pizzini FB, Crescenzo F, 45. Zhang H, Nguyen TD, Zhang J, Marcille M, Spincemaille P, Calabrese M (2020) The use of the central vein sign in the diag- Wang Y et al (2022) QSMRim-Net: Imbalance-aware learning nosis of multiple sclerosis: a systematic review and meta-analysis. for identification of chronic active multiple sclerosis lesions on Diagnostics (Basel) 10(12):1025 quantitative susceptibility maps. Neuroimage Clin 34:102979 29. Sinnecker T, Clarke MA, Meier D, Enzinger C, Calabrese M, De 46. Howell OW, Reeves CA, Nicholas R, Carassiti D, Radotra B, Stefano N et al (2019) Evaluation of the central vein sign as a Gentleman SM et al (2011) Meningeal inflammation is wide- diagnostic imaging biomarker in multiple sclerosis. JAMA Neurol spread and linked to cortical pathology in multiple sclerosis. 76(12):1446–1456 Brain 134(Pt 9):2755–2771 30. Maggi P, Absinta M, Grammatico M, Vuolo L, Emmi G, Carlucci 47. Magliozzi R, Howell OW, Reeves C, Roncaroli F, Nicholas G et al (2018) Central vein sign differentiates multiple sclerosis R, Serafini B et al (2010) A Gradient of neuronal loss and from central nervous system inflammatory vasculopathies. Ann meningeal inflammation in multiple sclerosis. Ann Neurol Neurol 83(2):283–294 68(4):477–493 31. Al-Louzi O, Letchuman V, Manukyan S, Beck ES, Roy S, 48. Magliozzi R, Howell OW, Nicholas R, Cruciani C, Castellaro Ohayon J et al (2022) Central vein sign profile of newly devel- M, Romualdi C et al (2018) Inflammatory intrathecal pro- oping lesions in multiple sclerosis: a 3-year longitudinal study. files and cortical damage in multiple sclerosis. Ann Neurol Neurol Neuroimmunol Neuroinflamm 9(2):1120 83(4):739–755 32. Solomon AJ, Watts R, Ontaneda D, Absinta M, Sati P, Reich DS 49. Ineichen BV, Tsagkas C, Absinta M, Reich DS (2022) Lep- (2018) Diagnostic performance of central vein sign for multiple tomeningeal enhancement in multiple sclerosis and other 13 1298 Journal of Neurology (2023) 270:1286–1299 neurological diseases: a systematic review and meta-analysis. 67. Wang SH, Tang C, Sun J, Yang J, Huang C, Phillips P et al (2018) Neuroimage Clin 33:102939 Multiple sclerosis identification by 14-layer convolutional neural 50. Absinta M, Vuolo L, Rao A, Nair G, Sati P, Cortese IC et al network with batch normalization, dropout, and stochastic pool- (2015) Gadolinium-based MRI characterization of leptomenin- ing. Front Neurosci 12:818 geal inflammation in multiple sclerosis. Neurology 85(1):18–28 68. Zhang Y-D, Pan C, Sun J, Tang C (2018) Multiple sclerosis iden- 51. Harrison DM, Wang KY, Fiol J, Naunton K, Royal W 3rd, Hua tification by convolutional neural network with dropout and para- J et al (2017) Leptomeningeal enhancement at 7T in multiple metric ReLU. J Comput Sci-Neth 28:1–10 sclerosis: frequency, morphology, and relationship to cortical 69. Eitel F, Soehler E, Bellmann-Strobl J, Brandt AU, Ruprecht K, volume. J Neuroimaging 27(5):461–468 Giess RM et al (2019) Uncovering convolutional neural net- 52. Absinta M, Cortese IC, Vuolo L, Nair G, de Alwis MP, Ohayon work decisions for diagnosing multiple sclerosis on conventional J et al (2017) Leptomeningeal gadolinium enhancement across MRI using layer-wise relevance propagation. Neuroimage Clin the spectrum of chronic neuroinflammatory diseases. Neurology 24:102003 88(15):1439–1444 70. Lopatina A, Ropele S, Sibgatulin R, Reichenbach JR, Gullmar D 53. Tzanetakos D, Tzartos JS, Vakrakou AG, Breza M, Velonakis G, (2020) Investigation of deep-learning-driven identification of mul- Stathopoulos P et al (2022) Cortical involvement and leptome- tiple sclerosis patients based on susceptibility-weighted images ningeal inflammation in myelin oligodendrocyte glycoprotein using relevance analysis. Front Neurosci 14:609468 antibody disease: a three-dimensional fluid-attenuated inversion 71. Yoo Y, Tang LYW, Brosch T, Li DKB, Kolind S, Vavasour I recovery MRI study. Mult Scler 28(5):718–729 et al (2018) Deep learning of joint myelin and T1w MRI features 54. Asgari N, Flanagan EP, Fujihara K, Kim HJ, Skejoe HP, Wuerfel in normal-appearing brain tissue to distinguish between multi- J et al (2017) Disruption of the leptomeningeal blood barrier in ple sclerosis patients and healthy controls. Neuroimage Clin neuromyelitis optica spectrum disorder. Neurol Neuroimmunol 17:169–178 Neuroinflamm 4(4):e343 72. Neeb H, Schenk J (2019) Multivariate prediction of multiple scle- 55. Wu CH, Lirng JF, Ling YH, Wang YF, Wu HM, Fuh JL et al rosis using robust quantitative MR-based image metrics. Z Med (2021) Noninvasive characterization of human glymphatics and Phys 29(3):262–271 meningeal lymphatics in an in vivo model of blood-brain barrier 73. Fooladi M, Sharini H, Masjoodi S, Khodamoradi E (2018) A leakage. Ann Neurol 89(1):111–124 novel classification method using effective neural network and 56. Absinta M, Ha SK, Nair G, Sati P, Luciano NJ, Palisoc M et al quantitative magnetization transfer imaging of brain white mat- (2017) Human and nonhuman primate meninges harbor lym- ter in relapsing remitting multiple sclerosis. J Biomed Phys Eng phatic vessels that can be visualized noninvasively by MRI. 8(4):409–422 Elife. https://​doi.​org/​10.​7554/​eLife.​29738 74. Zurita M, Montalba C, Labbe T, Cruz JP, da Dalboni Rocha 57. Kilsdonk ID, Jonkman LE, Klaver R, van Veluw SJ, Zwanen- J, Tejos C et al (2018) Characterization of relapsing-remitting burg JJ, Kuijer JP et al (2016) Increased cortical grey matter multiple sclerosis patients using support vector machine classi- lesion detection in multiple sclerosis with 7 T MRI: a post- fications of functional and diffusion MRI data. Neuroimage Clin mortem verification study. Brain 139(Pt 5):1472–1481 20:724–30 58. Mainero C, Benner T, Radding A, van der Kouwe A, Jensen 75. Sacca V, Sarica A, Novellino F, Barone S, Tallarico T, Filip- R, Rosen BR et al (2009) In vivo imaging of cortical pathol- pelli E et al (2019) Evaluation of machine learning algorithms ogy in multiple sclerosis using ultra-high field MRI. Neurology performance for the prediction of early multiple sclerosis from 73(12):941–948 resting-state FMRI connectivity data. Brain Imaging Behav 59. Pitt D, Boster A, Pei W, Wohleb E, Jasne A, Zachariah CR 13(4):1103–1114 et al (2010) Imaging cortical lesions in multiple sclerosis with 76. Eshaghi A, Wottschel V, Cortese R, Calabrese M, Sahraian MA, ultra-high-field magnetic resonance imaging. Arch Neurol Thompson AJ et al (2016) Gray matter MRI differentiates neu- 67(7):812–818 romyelitis optica from multiple sclerosis using random forest. 60. Junker A, Wozniak J, Voigt D, Scheidt U, Antel J, Wegner C Neurology 87(23):2463–2470 et al (2020) Extensive subpial cortical demyelination is specific 77. Eshaghi A, Riyahi-Alam S, Saeedi R, Roostaei T, Nazeri A, Agh- to multiple sclerosis. Brain Pathol 30(3):641–652 saei A et al (2015) Classification algorithms with multi-modal 61. Jonkman LE, Klaver R, Fleysher L, Inglese M, Geurts JJ (2015) data fusion could accurately distinguish neuromyelitis optica from Ultra-high-field MRI visualization of cortical multiple sclerosis multiple sclerosis. Neuroimage Clin 7:306–314 lesions with T2 and T2*: a postmortem MRI and histopathology 78. Kim H, Lee Y, Kim YH, Lim YM, Lee JS, Woo J et al (2020) study. AJNR Am J Neuroradiol 36(11):2062–2067 Deep learning-based method to differentiate neuromyelitis 62. Beck ES, Reich DS (2018) Brain atrophy in multiple sclerosis: optica spectrum disorder from multiple sclerosis. Front Neurol how deep must we go? Ann Neurol 83(2):208–209 11:599042 63. Fartaria MJ, Sati P, Todea A, Radue EW, Rahmanzadeh R, 79. Rocca MA, Anzalone N, Storelli L, Del Poggio A, Cacciaguerra O’Brien K et al (2019) Automated detection and segmentation L, Manfredi AA et al (2021) Deep learning on conventional mag- of multiple sclerosis lesions using ultra-high-field MP2RAGE. netic resonance imaging improves the diagnosis of multiple scle- Invest Radiol 54(6):356–364 rosis mimics. Invest Radiol 56(4):252–260 64. La Rosa F, Beck ES, Maranzano J, Todea RA, van Gelderen P, 80. Theocharakis P, Glotsos D, Kalatzis I, Kostopoulos S, Georgiadis de Zwart JA et al (2022) Multiple sclerosis cortical lesion detec- P, Sifaki K et al (2009) Pattern recognition system for the dis- tion with deep learning at ultra-high-field MRI. NMR Biomed crimination of multiple sclerosis from cerebral microangiopathy 35(8):e4730 lesions based on texture analysis of magnetic resonance images. 65. Herranz E, Gianni C, Louapre C, Treaba CA, Govindarajan ST, Magn Reson Imaging 27(3):417–422 Ouellette R et al (2016) Neuroinflammatory component of gray 81. Mangeat G, Ouellette R, Wabartha M, De Leener B, Platten M, matter pathology in multiple sclerosis. Ann Neurol 80(5):776–790 Danylaite Karrenbauer V et al (2020) Machine learning and mul- 66. Herranz E, Louapre C, Treaba CA, Govindarajan ST, Ouellette tiparametric brain MRI to differentiate hereditary diffuse leukod- R, Mangeat G et al (2020) Profiles of cortical inflammation in ystrophy with spheroids from multiple sclerosis. J Neuroimaging multiple sclerosis by (11)C-PBR28 MR-PET and 7 Tesla imaging. 30(5):674–682 Mult Scler 26(12):1497–1509 13 Journal of Neurology (2023) 270:1286–1299 1299 82. Eksi Z, Ozcan ME, Cakiroglu M, Oz C, Aralasmak A (2021) 86. Zhang H, Alberts E, Pongratz V, Muhlau M, Zimmer C, Wiestler Differentiation of multiple sclerosis lesions and low-grade brain B et al (2019) Predicting conversion from clinically isolated syn- tumors on MRS data: machine learning approaches. Neurol Sci drome to multiple sclerosis-an imaging-based machine learning 42(8):3389–3395 approach. Neuroimage Clin 21:101593 83. Wottschel V, Alexander DC, Kwok PP, Chard DT, Stromillo ML, 87. Freeze WM, Schnerr RS, Palm WM, Jansen JF, Jacobs HI, Hoff EI De Stefano N et al (2015) Predicting outcome in clinically isolated et al (2017) Pericortical enhancement on delayed postgadolinium syndrome using machine learning. Neuroimage Clin 7:281–287 fluid-attenuated inversion recovery images in normal aging, mild 84. Wottschel V, Chard DT, Enzinger C, Filippi M, Frederiksen JL, cognitive impairment, and alzheimer disease. AJNR Am J Neu- Gasperini C et al (2019) SVM recursive feature elimination analy- roradiol 38(9):1742–1747 ses of structural brain MRI predicts near-term relapses in patients with clinically isolated syndromes suggestive of multiple sclero- sis. Neuroimage Clin 24:102011 85. Bendfeldt K, Taschler B, Gaetano L, Madoerin P, Kuster P, Mueller-Lenke N et al (2019) MRI-based prediction of conversion from clinically isolated syndrome to clinically definite multiple sclerosis using SVM and lesion geometry. Brain Imaging Behav 13(5):1361–1374 13