MRI Patterns Distinguish AQP4 Antibody Positive Neuromyelitis Optica Spectrum Disorder From Multiple Sclerosis

Neuromyelitis optica spectrum disorder (NMOSD) and multiple sclerosis (MS) are inflammatory diseases of the CNS. Overlap in the clinical and MRI features of NMOSD and MS means that distinguishing these conditions can be difficult. With the aim of evaluating the diagnostic utility of MRI features in distinguishing NMOSD from MS, we have conducted a cross-sectional analysis of imaging data and developed predictive models to distinguish the two conditions. NMOSD and MS MRI lesions were identified and defined through a literature search. Aquaporin-4 (AQP4) antibody positive NMOSD cases and age- and sex-matched MS cases were collected. MRI of orbits, brain and spine were reported by at least two blinded reviewers. MRI brain or spine was available for 166/168 (99%) of cases. Longitudinally extensive (OR = 203), “bright spotty” (OR = 93.8), whole (axial; OR = 57.8) or gadolinium (Gd) enhancing (OR = 28.6) spinal cord lesions, bilateral (OR = 31.3) or Gd-enhancing (OR = 15.4) optic nerve lesions, and nucleus tractus solitarius (OR = 19.2), periaqueductal (OR = 16.8) or hypothalamic (OR = 7.2) brain lesions were associated with NMOSD. Ovoid (OR = 0.029), Dawson's fingers (OR = 0.031), pyramidal corpus callosum (OR = 0.058), periventricular (OR = 0.136), temporal lobe (OR = 0.137) and T1 black holes (OR = 0.154) brain lesions were associated with MS. A score-based algorithm and a decision tree determined by machine learning accurately predicted more than 85% of both diagnoses using first available imaging alone. We have confirmed NMOSD and MS specific MRI features and combined these in predictive models that can accurately identify more than 85% of cases as either AQP4 seropositive NMOSD or MS.


INTRODUCTION
Neuromyelitis optica spectrum disorder (NMOSD) is an antibody-mediated autoimmune disease of the CNS in which the primary target for inflammation is aquaporin 4 (AQP4), a water channel found on the foot processes of astrocytes (1). NMOSD is frequently a disabling condition when untreated, with high relapse rates and a high risk of permanent neurological deficits following an attack (2). The treatment response in NMOSD is quite distinct to that seen in multiple sclerosis (MS) with NMOSD being highly steroid responsive and having a higher risk of disease recurrence on steroid withdrawal (3). Acute treatment with plasma exchange appears to be particularly effective in NMOSD (4). The risk of NMOSD relapses is diminished by immunosuppressive therapy, anti-B-cell therapies (e.g., rituximab, inebilizumab), anti-IL6 receptor and anticomplement component 5 monoclonal antibodies (5)(6)(7)(8). Some immunomodulatory treatments for MS (β-interferon, fingolimod and natalizumab) may worsen NMOSD and alemtuzumab has shown no discernible benefit in small case series (9)(10)(11). Consequently, it is important to recognise NMOSD early and distinguish it from MS prior to commencing disease modifying therapy.
NMOSD and MS are inflammatory diseases of the CNS and lead to demyelination (12, 13). A considerable degree of overlap in MRI features has been noted (14). However, a number of groups have identified MRI features that are more common in NMOSD than in MS and may be useful in increasing the clinical suspicion for this diagnosis. Long lesions of the optic nerve (at least half the length of the optic nerve) and the spinal cord (at least 3 vertebral segments), and lesions of the area postrema have been incorporated into the current diagnostic criteria for seronegative NMOSD (1). Other MRI features, such as hypothalamic lesions, linear periventricular and brainstem periependymal lesions, and "bridging" lesions of the splenium have also been identified as having specificity for NMOSD when compared with MS (15)(16)(17). Some lesions typical for MS (e.g., perpendicular, ovoid, periventricular lesions, cerebellar peduncle lesions, juxtacortical lesions) appear to be less common in NMOSD (14,18).
We report an observational, cross-sectional study of MRI characteristics in a cohort of AQP4 seropositive NMOSD cases collected across Australia and New Zealand compared with age and sex-matched MS cases from the same region. The aims of the study were to determine the clinical utility of specific MRI lesions and features in distinguishing NMOSD and MS, compare frequencies of specific lesions within cases and map the commonest locations of lesions in the spinal cord. We have gone on to develop an algorithm of MRI lesions and features to assist in distinguishing NMOSD from MS and applied this to the first sets of imaging available in this cohort. The hypotheses being tested were that (1) certain MRI features are associated with either NMOSD or MS (2) numbers of some lesions will be higher in MS than NMOSD, (3) spinal cord lesions will have predilection for different regions, and (4) that these lesions and features can be used to discriminate between these two conditions early in their clinical course, thereby expediting appropriate investigation and therapy.

MATERIALS AND METHODS
For this study we were only concerned with AQP4 antibody associated NMOSD and not MOG antibody associated demyelination (MOGAD). Firstly, lesions thought to be associated with NMOSD and MS were identified and defined from the prior literature. MRI from a large cohort of NMOSD and MS cases were then systematically reviewed and the presence and number of those lesions and features in both the "first" or "ever" imaging performed was determined. These data were then used to identify MRI lesions and features associated with NMOSD or MS. Lesion location in the spinal cord was also analysed. Finally, this association data was used to develop predictive algorithms to identify NMOSD and MS based purely on MRI findings.

MRI Lesion and Feature Associations and Definitions
A literature review using the search terms NMOSD (and synonyms-neuromyelitis optica spectrum disorder, neuromyelitis optica, NMO, Devic's disease) and MRI (and synonyms-magnetic resonance imaging, MR imaging) for Title and Abstract fields was performed using MEDLINE and EMBASE databases on 17 January 2019. Searches were restricted to articles written in English. Potentially relevant articles looking at MRI features in NMOSD were identified from review of the title and abstract. These potential articles were then reviewed in full. Inclusion criteria were, (1) assessment of routine MRI lesions and features, (2) populations studied included NMOSD or limited forms of NMOSD (e.g., longitudinally extensive transverse myelitis) where relevant criteria for NMOSD or neuromyelitis optica were met (e.g., AQP4 antibody positive), and (3) original research or latest diagnostic criteria. The following exclusion criteria were applied (1) case reports with no relevant MRI findings, (2) no defined MRI findings or not relevant to NMOSD, (3) specialised MRI techniques or analysis, (4) review article or older diagnostic criteria, and (5) study solely focused on MOGAD. Additional articles were identified from reviewing the bibliographies of included articles and review articles. Articles relating to lesions associated with MS were identified from the 2010 update to the McDonald diagnostic criteria (19) and the literature supporting the MRI criteria contained therein. The available literature of the appearances of MRI lesions in NMOSD and MS were reviewed for evidence of specificity to either NMOSD or MS. However, it should be noted that our aim here was to be as inclusive as possible for any MRI lesion or feature that might be associated with NMOSD or MS. Consequently, there was no requirement for a statistically significant association. In addition, the same literature was reviewed to establish a clear definition for each lesion type that could be utilised in the review of MRI. Finally, for completeness the imaging sequence and anatomical counterparts for all lesions identified were added to the list and defined (e.g., corpus callosum Gd-enhancing lesion).

NMOSD and MS Cohort Case Ascertainment
Potential NMOSD cases and MS cases were referred by 35 adult and paediatric neurologists across 23 clinical centres in Australia and New Zealand specialising in the assessment of patients with inflammatory diseases of the central nervous system as part of a prevalence survey as previously described (20,21). For this study, NMOSD cases were defined according to the 2015 International panel for NMO diagnosis (IPND) criteria (1) with the additional requirement of being positive for AQP4 antibodies using either a tissue-based immunofluorescence technique or positivity on at least two cell-based assays (fixed, Euroimmun R or live, Oxford, UK) as previously described (22). Testing for MOG antibodies was completed on a subgroup of patients using either a fixed cell-based assay (Euroimmun R ) or a fluorescence activated cell sorting live cell-based assay (Westmead Immunology, Sydney). Age and sex-matched MS cases were referred from each centre. The 2010 McDonald criteria (19) were used to confirm the diagnosis of MS. All MS cases were negative for AQP4 antibodies and had no clinical features suspicious for NMOSD. Basic demographic and clinical features were recorded for all cases as per a standardised data collection questionnaire as previously described (20). This study was approved by the Human Research Ethics Committees (HREC) of all participating centres (lead site HREC was Griffith University MED2009/646). All participants gave written informed consent to participate in this study.

Review of MRI
MRI of the orbits, brain and spine were provided as raw DICOM images and stored on a secure server. Images were reviewed by at least two independent blinded reviewers (LC, SA, EK and SJS) using eFilm Workstation R 4.2.3, IBM Watson Health software on Eizo R RadiForce MX270W 68 cm monitors. In the case of disagreement, a third blinded expert reviewer (SBh) assessed the imaging and was the final arbiter. All blinded reviewers were trained in the identification of "typical" lesion patterns by the expert blinded reviewer.
Lesion numbers for each set of MRI on each subject were recorded using a bespoke Oracle R Database (Oracle Corporation R , Redwood Shores, California, US). The field strength, slice thickness and administration of gadolinium (Gd) contrast were recorded for each set of images. For analysis the first available imaging was counted separately ("first") and then the maximum number of lesions of each type was also recorded across all subsequent sets of MRI for each patient ("ever"). The only exceptions to this were normal brain (not meeting Paty criteria), Paty criteria (23) and Barkhof criteria (24) which were only considered on the first imaging available. Lesions were then determined to be present (yes or no) on the "first" MRI or "ever" for any MRI on a "per patient" basis. Maximum numbers of each lesion type were also recorder per patient. MRI were defined as being related to a relapse if a documented relapse occurred either in the 90 days prior to the MRI or the 30 days following the MRI. All other MRI were deemed to have been undertaken during a period of remission.
Brain and optic nerve lesions were defined as specified in the prior literature. Conventions for distinguishing periventricular, subcortical, juxtacortical, cortical and tumefactive lesions are illustrated in Figure 1A. White matter lesions meeting criteria for more than one type of lesion that were not tumefactive were allocated according to the following preference order as appropriate: periventricular, cortical, juxtacortical then subcortical. Optic nerve lesions were recorded using standard brain MRI as well as dedicated images of the orbits where available.
Spinal cord lesions were defined and recorded as being in either the superior or inferior half of each vertebral level. This represented a total of 44 potential levels from the superior half of C1 down to the inferior half of L3, although no lesions were recorded at the L3 level as all spinal cords reviewed terminated at L2 or higher. Lesions in the spinal cord were determined to be "short" if they were <3 vertebral bodies (equivalent to <6 adjacent hemi-segmental levels) or "long" if they extended over 3 vertebral bodies or more (25). Where axial images were available spinal lesions were labelled as "central" if they involved all four quadrants of the centre of the spinal cord at a minimum of one level, or "partial" if they involved <4 quadrants of either the central or peripheral cord as illustrated in Figures 1B,C. Whole (axial) cord lesions were defined as involving all 8 central and

Supratentorial
Cloud-like Multiple associated Gd-enhancing white matter lesions with patchy, heterogenous, often subtle, parenchymal enhancement with ill-defined/blurred margins on T1 weighted contrast imaging. Can be associated with periventricular periependymal enhancement ("flame" or "smoke coming out of the mouth" appearance).
NMOSD (45,59,60) Supratentorial Tumefactive Extensive and confluent T2 hyperintense, hemispheric white matter lesion which is >3 cm in its longest diameter and may have peripheral DWI restriction and Gd-enhancement on T1 contrast imaging (see Figure 1A).   Figure 1B). Spinal cord Whole (axial) cord T2 T2 hyperintense lesion involving all 8 of the central and peripheral quadrants of the cord for at least one hemi-vertebral level (see Figure 1B).

NMOSD
NMOSD (84,86,98) Spinal cord Cord swelling T2 hyperintense lesion of the spinal cord associated with significant expansion of the spinal cord (>20% increase in midline sagittal diameter as compared with adjacent sections of spinal cord or loss of CSF space).
NMOSD (74,78,87) Spinal cord Cord atrophy Focal thinning of the spinal cord (more than 20% reduction in midline sagittal diameter as compared with adjacent sections of spinal cord) with or without myelomalacia.
Optic pathway Optic nerve Gd Gd-enhancing lesion of the optic nerve.

Supratentorial Cortical Gd
Gd-enhancing lesion of the cortex.

Supratentorial Periventricular Gd
Gd-enhancing lesion of the periventricular region.

Supratentorial
Subcortical T2 T2 hyperintense white matter lesion which is >3 mm but <3 cm in diameter and is not periventricular or juxtacortical (see Figure 1A).

Supratentorial
Juxtacortical Gd Gd-enhancing lesion of the juxtacortical region.

Supratentorial
Rounded corpus callosum T2 Round, soft edged T2 lesion contained within wholly within the body of the corpus callosum (similar to the "snowball" lesion of Susac's syndrome).

Supratentorial
Other corpus callosum T2 Any other corpus callosum lesion not described elsewhere.

(Continued)
Frontiers in Neurology | www.frontiersin.org  Figure 1A). MS (14,33,34) Supratentorial Juxtacortical T2 T2 hyperintense lesion adjacent to and following the cortex and involving the cortical U-fibre layer (see Figure 1A).  Spinal cord Partial T2 spinal cord T2 hyperintense lesion of the spinal cord involving fewer than 4 central or peripheral quadrants of the cord (see Figure 1C). (81,85,103,(106)(107)(108) NMOSD, neuromyelitis optica spectrum disorder; MS, multiple sclerosis. peripheral quadrants on any one axial section. The presence of Gd-enhancement was noted for all spinal and brain lesions where contrast had been administered. The distribution of spinal cord lesions has been illustrated using a "heat map" analogy (26). Long (≥3 vertebral segments) lesions, short lesions and Gd-enhancing lesions were colour coded and represented in columns for each subject with horizontal rows representing a vertebral half-segment. A summative approach using all spine MRI per patient was used. Summary frequencies of lesions at each hemi-vertebral level across all subjects were calculated on a proportional frequency basis.

Statistical Analysis
Frequencies are expressed as n/N (%) and continuous data are presented as median (range) if not normally distributed or mean (standard deviation) if normally distributed. All analyses have been conducted on a per patient basis, thus "N" relates to the number of cases and not the number of scans. Comparisons of frequencies of different types of lesion between NMOSD and MS have been made using the "ever" data from all scans performed. Odds ratios and 95% confidence intervals were calculated using Haldane-Anscombe correction where frequencies for either NMOSD or MS were zero (27,28). Odds ratios >1 favoured NMOSD and values <1 favoured MS. Results where the 95% confidence intervals did not cross 1 were taken as significant. No correction for multiple testing was applied for the following reasons: (1) there was prior evidence for association for many of the lesions and features assessed; (2) the NMOSD and MS cases were matched for age and sex, making adjustment for these baseline characteristics unnecessary; and (3) these were observational data (29). The relative value in distinguishing NMOSD from MS was assessed using sensitivity and specificity. Where imaging was not available the denominator was reduced. Lesion counts were assessed using appropriate non-parametric methods (Mann-Whitney Utest) (30). All analyses were performed on a per subject basis to not artificially inflate the "N" as might occur in a per MRI analysis. Statistical comparisons were performed using Statistical Package for Social Science (SPSS R ) v25 (IBM R ; Chicago, US).
To create a predictive model for NMOSD and MS two approaches were undertaken utilising the "ever" MRI dataset. The first was an iterative approach in which MRI lesions and features associated with each condition were considered as present (1) or absent (0) and combined in either a summative (overall score) or regional (optic nerve, brain or spinal cord) subgrouping with variable weighting for each factor and region and variable cut-offs for each summative or regional score. Cut-offs were progressively increased using Excel R , Microsoft R (Seattle, US) until either a specificity of 1.00 was achieved or sensitivity fell below 0.50. Models with the highest specificity, aiming for a specificity >0.90, and highest sensitivity were then tested by sequentially removing features and increasing weightings and/or cut-offs for sub-scores until optimal values were achieved. The second approach used machine learning to create a predictive algorithm based on the same data using STATA R statistical software v14 (StataCorp R 2015, College Station, TX, USA). The utility of each model was assessed using weighted means (for both NMOSD and MS prediction) for the true positive rate (TPR), false positive rate (FPR), precision (positive predictive value), F-measure and receiver operator characteristic (ROC) area under the curve, with NMOSD as the positive state and MS as the negative state in the same "ever" data. Precision, F-measure and ROC area are all measures of accuracy and have a value from zero to one. The best performing models were then tested and compared with existing predictive models for MS (23,31), NMOSD (14,18) and NMOSD/MOGAD (32-34) (in the "first" MRI dataset). This was felt to most accurately reflect the clinical situation of a patient undergoing diagnostic evaluation. Again, weighted means for TPR, FPR, precision, Fmeasure and ROC area were used to asses diagnostic utility of the derived algorithms and prior methods. A sensitivity analysis including only MRI performed within 5 years of symptom onset was performed.

NMOSD and MS MRI Lesion Identification and Definitions From Prior Literature
The initial literature search identified 426 potential articles relating to MRI features in NMOSD, of which 118 were selected for full review. These in turn yielded 71 included articles as summarised in Figure 2. A further five articles of relevance to NMOSD were identified from review of bibliographies of both the included articles and 10 review articles on MRI in NMOSD. Reasons for exclusion of articles are summarised in    specific lesions together with their definitions and references are listed in Table 1. Thirty-four lesion types or MRI characteristics were identified as being potentially associated with NMOSD and 19 were potentially associated with MS. Some lesion types have been described in both MS and NMOSD. In this situation, lesions were either assigned the disease with greatest evidence for specificity in diagnosis (e.g., large supratentorial T2 lesions in MS) or were left neutral (e.g., cortical T2 lesions). Finally, natural anatomical counterparts to all lesion types or features were added and defined with no disease association as listed in Table 1. Most lesion types and definitions will be familiar to all, but two perhaps require further explanation. The first is whole (axial) cord lesion, which in this context refers to a cord lesion that involves the full axial section of the cord (i.e., all 8 central and peripheral quadrants of a single axial section as defined in Figures 1B,C). The second is "pyramidal" corpus callosum lesions which is a pyramidal shaped lesion arising from the inferior margin of the body of the corpus callosum, which we realised were the most commonly pictured callosal lesion in MS, although never really defined (18,109

NMOSD and MS Cohorts and MRI Availability
There were 67 NMOSD cases who met 2015 IPND criteria, had full clinical data and were positive for AQP4 antibodies. There were 101 MS cases who met the 2010 McDonald criteria, had full clinical data, had no features suspicious for NMOSD, and were negative for AQP4 antibodies. MRI of orbits, brain or spine was available for 66/67 (99%) of NMOSD cases and 100/101 (99%) of MS cases (99% MRI availability overall). Consequently, one NMOSD case and one MS case were excluded from further analysis. Basic demographic and clinical features of the included cases are given in Table 2. These data indicate that NMOSD and MS cases were well-matched for age and sex but differed with regard to clinical features in a pattern that has been previously identified for these two disorders (104,122,123). All 66 included NMOSD cases were positive for AQP4 antibodies [38/66 (58%) using a cell-based assay and 28/66 (42%) using tissue immunofluorescence]. All 100 included MS cases were negative for AQP4 antibodies with 52/100 (52%) being tested on a cell-based assay. In our previously published analysis, the specificity of our tissue-based indirect immunofluorescence assays was 99.7% (95% CI 98.4-99.9%) (22). In addition, 48/66 (73%) of NMOSD cases and 52/100 (52%) of MS cases were tested for MOG antibodies and all were negative (22).
The availability of MRI is summarised in Table 3. In total, 733 sets of MRI were reviewed. MRI brain was available for 62/66 (94%) and MRI spine was available for 61/66 (92%) of NMOSD cases. MRI brain was available for 100/100 (100%) and MRI spine was available for 86/100 (86%) of MS cases. Dedicated MRI of the orbits was more frequently available in NMOSD than in MS cases [18/66 (27%) of NMOSD cases vs. 6/100 (6%) of MS cases, P < 0.001]. MRI in NMOSD were more likely to have been obtained during a relapse (50% for brain and 49% for spine MRI in NMOSD vs. 13% for brain and 16% for spine MRI in MS; p < 0.0001). The availability of MRI per patient in NMOSD and MS is shown in Figure 3. The median time to first imaging from first symptoms was 9 months for brain and 12 months for spine MRI in the NMOSD cohort. The equivalent times for MS were both 10 years reflecting the greater disease duration of these cases from a historical cohort. Many MS cases had onset prior to 2000 and it was not possible to obtain DICOM files for MRI performed prior to the early 2000's as these were generally not centrally stored prior to then.

MRI Features of NMOSD and MS
The odds ratios for occurrence in NMOSD vs. MS at any time for MRI lesion types and features are shown in Figure 4 and Supplementary Table 2. Longitudinally extensive spinal cord T2, bright spotty cord T2, whole (axial) cord T2, bilateral optic nerve T2/Gd-enhancing, Gd-enhancing spinal cord T1, nucleus tractus solitarius T2, periaqueductal T2, Gd-enhancing optic nerve T1, third ventricular periependymal T2, spinal cord swelling, central spinal cord T2, optic nerve T2, hypothalamic T2, initial brain MRI normal and spinal cord atrophy lesions and features were found to be associated with NMOSD. In addition, leptomeningeal Gd-enhancing T1, central medullary T2, optic chiasm T2 or Gd-enhancing T1, cloud-like Gd-enhancing T1, area postrema T2, longitudinal T2 optic nerve and longitudinal corticospinal tract T2 lesions had odds ratios that crossed one, but were only seen in NMOSD cases and not in MS. Examples of each of these NMOSD lesions is illustrated in Figure 5. NMOSD spinal cord lesions and features were seen at some time point ("ever") in 45/62 (73%) of NMOSD cases. NMOSD brain lesions were seen in 23/62 (37%) and optic nerve lesions in 25/62 (40%) of NMOSD cases at some point in time.
Examples of the above MS associated lesions are illustrated in Figure 6.
The commonest types of T2 lesion seen (periventricular, subcortical, juxtacortical and large T2) and total T2 lesion counts were more numerous in MS than NMOSD as shown in Figure 7. The median (range) total number of T2 lesions seen on any MRI brain were 4 (0-47) for NMOSD and 14 (0-54) for MS (p < 0.001; Mann-Whitney U-test). There were no differences in the frequencies of patch and punctate white matter lesions in NMOSD and MS (Figure 7). The following lesions previously noted in NMOSD: linear periventricular periependymal, bridging splenium, brainstem periependymal, cystic, heterogeneous corpus callosum, cerebral peduncle, punctate and patch lesions, were all found with   Table 2). Examples of these lesions from NMOSD and MS cases are illustrated in Figure 8.
No Gd-enhancing T1 lesions of the cortex, corpus callosum, basal ganglia, hypothalamic region or brainstem were seen in NMOSD or MS cases. In addition, no anterior midbrain, posterior reversible encephalopathy syndrome-like, Balo-like, floor of fourth ventricle T2 or ring-enhancing Gd-enhancement of the spinal cord lesions were seen.

Spinal Cord Lesion Location
The distribution of spinal cord lesions at any stage of disease for each subject are summarised in Figure 9 as a "heat map." A longitudinally extensive spinal cord lesion was seen in one MS case in the cervical region (Figure 10). It was only through review of prior imaging that it was clear that this lesion had arisen over time due to a coalescence of smaller lesions rather than a single focus of inflammation but remained in our count of longitudinally extensive spinal cord lesions as the appearance met our definition. Gd-enhancement was only evident in the spinal cords of 1/85 (1%) MS cases. The predominant site for cord lesions in MS was the cervical region, with lesions between C2 and C7 being seen in 10-30% of cases. MRI spine was persistently normal in 26/86 (30%) of MS cases. In contrast, longitudinally extensive spinal cord lesions were seen in 43/61 (70%) and Gdenhancement was evident in 19/61 (31%) of NMOSD cases. No spinal cord lesions were evident in 7/61 (11%) of NMOSD cases. The most common location for spinal cord lesions in NMOSD were the cervical (C2-C6) and mid-thoracic regions (T2-T8) where lesions at each hemi-vertebral level were present in more than 30% of cases.

Diagnostic Utility of MRI Features
The relative frequencies, sensitivities, specificities and odds ratios for associated MRI features in the diagnosis of NMOSD and MS are given in Supplementary Table 2. It is notable that most of the MRI features associated with NMOSD have high specificity but low sensitivity, except for longitudinally extensive spinal cord lesions. In diagnosing MS, ovoid, periventricular, pyramidal corpus callosum, other corpus callosum, splenium,

Predictive Modelling Based on MRI Features
The outcome of the predictive modelling for NMOSD and MS based on MRI features seen on "ever" imaging is summarised in Supplementary Table 3. A total of 134 models were tested for NMOSD and 55 for MS but only the best models (models 1-4) are shown. For both MS and NMOSD predictive models, summative scores of features (see Supplementary Table 3) with weightings (for longitudinally extensive spinal cord and bilateral optic nerve lesions in NMOSD and ovoid lesions in MS) gave the best results. The NMOSD score gave a weighted mean precision of 0.921. Exclusion of spinal MRI data (model 5) reduced this to 0.802. A combined approach using both NMOSD/MS scores according to the rule, if the NMOSD score × 3.5 > MS score then the diagnosis is NMOSD and if not, the diagnosis is MS, gave an overall precision of 0.935. The machine learning decision tree is shown in Figure 11 and performed slightly better with precision of 0.939. The performance of these latter two models is shown in comparison to previous methods of predicting MS and NMOSD using the "first" MRI dataset in Table 4.
In MS the Paty criteria (23) had a high sensitivity (0.910) but low specificity (0.468) for MS and was poor in predicting NMOSD. The Barkhof dissemination in space criteria (124) had higher specificity (0.774) in predicting MS with reasonable sensitivity (0.570). The previously published MRI criteria (14,18,(32)(33)(34) for distinguishing NMOSD from MS performed well in identifying MS, but performed no better than the Barkhof criteria in identifying NMOSD. The NMOSD/MS score gave overall precision of 0.876 with the machine learning decision tree giving precision of 0.871. These models were able to accurately distinguish NMOSD from MS in 144/166 (87%) and 143/166 (86%) of cases, respectively, based on first available imaging alone. A sensitivity analysis restricted to first scans performed in the 5 years since onset of symptoms showed very similar results (precision for NMOSD/MS score = 0.900 and machine learning algorithm = 0.864).

DISCUSSION
We have undertaken a cross-sectional, comparison study of MRI data in AQP4 antibody positive NMOSD and MS cases analysing for typical lesions that have previously been described for each condition. We have compared how frequently these lesions are seen in the two disorders and have found patterns that are largely consistent with previously published reports (14,18,37,42,45,49,81). We note that NMOSD associated lesions in the brain were generally single lesions but collectively were seen in 40% of cases. Longitudinally extensive spinal cord lesions were seen in a majority of NMOSD cases (70%) at some point during their disease. Initially normal MRI brain (not meeting Paty criteria) was seen in 16% of NMOSD cases. The frequency of supratentorial T2 lesions, excepting the NMOSD specific lesions, was higher in MS compared to NMOSD. Gd-enhancing lesions of the spinal cord and optic nerve were more common in NMOSD and Gd-enhancing lesions in the brain were more common in MS. Analysis of lesion location in the spinal cord indicates that longitudinal, central cord lesions affecting the whole (axial) cord with swelling and Gd-enhancement are more common in NMOSD and short, partial cord lesions are more common in MS. The predictive value in distinguishing NMOSD from MS for these individual MRI features was limited but using scores for NMOSD and MS based on the occurrence of typical lesions for each or a machine learning decision tree proved useful in distinguishing the two conditions. When applied to the current dataset these criteria accurately predicted the correct diagnosis in over 85% of cases based on first available imaging alone and performed better than previously defined criteria (14,18,(32)(33)(34). These findings need to be replicated in an independent cohort, but if proven to be accurate have the potential to be useful in directing appropriate further investigation with AQP4 antibody testing and in situations where AQP4 antibody testing is not available may provide some guidance with regards to treatment. We did not see any anterior midbrain, PRES-like or Balolike lesions in either NMOSD cases or MS cases. This probably reflects the relative rarity of these lesion types and conclusions about their frequency in the two disorders must await analysis in larger datasets. Similarly, leptomeningeal Gd-enhancement, chiasmal, central medullary and area postrema lesions were only seen in NMOSD, but given their low frequency, definitively excluding their occurrence in MS would require a dataset of several hundred if not thousands of cases. We were not able to confirm postulated associations with NMOSD for cystic brain, tumefactive, bridging splenium, brainstem periependymal, linear periventricular periependymal, heterogenous corpus callosum, pencil-like corpus callosum or cerebral peduncle lesions, which were seen with similar frequency in NMOSD and MS. This exemplifies the need to include large numbers of "control" MS comparison cases prior to defining lesions as being associated with NMOSD. These lesions appear to be more generically associated with demyelination of the central nervous system. Punctate and patch lesions have been noted as one of the most commonly seen brain lesions in NMOSD. In the present study these lesions were seen just as commonly in MS and with similar lesion counts within cases. We would note that these lesions are similar to the lesions described in migraine (125), a condition that is likely to be frequently seen in both NMOSD and MS as all three conditions are more common in women (126)(127)(128).
Two recent studies have confirmed the value of the Matthews criteria in distinguishing MS from NMOSD (34,129). These studies found similar sensitivity for detecting MS at 79.8% (34) and 82.9% (129) compared to 75% in the present study and neither of these figures are as higher as the 91% sensitivity for diagnosing MS seen with the NMSOD/MS score method and the machine learning algorithm identified here. Our results accord well with a recent similar study comparing AQP4 antibody associated NMOSD and MS (130). We would suggest that a high MS score on MRI, in the absence of clinical features suggestive of NMOSD, AQP4 antibody testing is not indicated. Another study has looked specifically at corpus callosum lesions in NMOSD and MS, in common with the present study finding that overall corpus callosum lesions are seen equally as commonly in all regions (131). This study found that corpus callosum lesions in NMOSD were more likely to be diffuse, have blurred margins and be heterogeneous which is somewhat at odds with the data presented here. Similarly, "extensive" or bridging lesions of the splenium were seen more commonly in NMOSD but were also seen in MS (131) whilst these lesions seen equally in the two conditions in the present larger dataset. Others have also looked at the value of Brain MRI in distinguishing NMOSD and MOGAD from MS, finding lesion distribution is helpful (32).
Strengths of the present study include the comprehensive literature search for potentially associated MRI features, the large dataset of NMOSD and MS cases with near complete availability of brain and spine imaging, definition of NMOSD cases through highly specific assays for AQP4 antibodies, age and sex matching of MS cases, blinded assessment of scans and an expansive approach to predictive modelling to optimise the classification of the NMOSD and MS cases. Potential weaknesses of the study include the historical cohort nature of the MS cases leading to differences in disease duration and timing of first imaging. We have also not been able to test our predictive models in a truly independent dataset. However, the associated features used in the predictive models had all been previously identified as being associated with NMOSD or MS. The fact that more MRI in NMOSD were undertaken during a relapse may contribute to the greater frequency of Gd-enhancement seen in the optic nerve and spinal cord. Finally, the low frequency and difference between the two groups in dedicated orbital imaging, means that the findings in relation to optic nerve lesions may be underestimated. It should also be noted that these predictive models do not take into account the pre-test probability for a diagnosis of NMOSD, which in some regions of the world (e.g., Japan) may be as high as 50% (132) vs. perhaps 1% in populations of European ancestry (21).
In conclusion, the MRI features of NMOSD and MS are distinct and can be used to distinguish the two conditions with high precision early in the disease course. This makes our final predictive models of potential value in making initial management and treatment decisions, particularly in settings where AQP4 antibody testing is not readily available.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article will be made available to suitably qualified researchers subject to approval by the Griffith University HREC.

ETHICS STATEMENT
This study was approved by Griffith University HREC. Written informed consent to participate in this research was provided by all participants or their legal guardian/next of kin.

AUTHOR CONTRIBUTIONS
This study was undertaken by the Australia and New Zealand Neuromyelitis Optica (ANZ NMO) Collaboration and all authors are members of this group. As such all authors took part in the design of this study and were involved in participant recruitment, data collection, laboratory testing of samples or MRI analysis. The primary data analysis was conducted by LC, SAB, and JSu. The group was coordinated by a group chaired by SAB, and consisted of AGKK, DM, MM, JDEP, MS, and BT. All authors have approved the final manuscript.

FUNDING
This project was undertaken by the ANZ NMO Collaboration and was supported by funding from Multiple Sclerosis Research Australia (11-038), the Brain Foundation, Griffith University and the Gold Coast Hospital Foundation. The work in Oxford was supported by the National Health Service National Specialised Commissioning Group for Neuromyelitis Optica.