Биомаркеры мультимодального МРТ и машинное обучение по расстройству спектра нейромиелита: обзор.

Introduction

Neuromyelitis optica spectrum disorder (NMOSD) is an inflammatory demyelinating disease of the central nervous system, most commonly associated with pathogenic aquaporin-4 immunoglobulin G antibodies (AQP4-IgG) ( 1 ). It is characterized by a relapsing disease course and severe attacks involving the optic nerves, spinal cord, area postrema, brainstem, and other central nervous system regions ( 2 ). In some cohorts, approximately 60% of patients have been reported to experience relapse within one year ( 3 ), which may lead to cumulative neurological disability, including visual impairment, motor deficits, and other irreversible sequelae. Therefore, early diagnosis, accurate risk stratification, and timely intervention remain critical challenges in NMOSD management.

In recent years, imaging and laboratory markers related to the diagnosis, disease monitoring, and prognostic assessment of NMOSD have continued to evolve. Magnetic resonance imaging (MRI), as a non-invasive and widely available tool, plays an indispensable role in lesion detection and longitudinal assessment ( 4 ). Nevertheless, MRI findings in NMOSD may overlap with those of MS, MOGAD, acute disseminated encephalomyelitis, and several non-inflammatory mimics such as vascular myelopathies. For example, dural arteriovenous fistula at the craniocervical junction may present with longitudinally extensive spinal cord lesions and has been misdiagnosed as NMOSD ( 5 ). Identifying specific and reproducible MRI biomarkers remains important for improving differential diagnosis and individualized disease assessment ( Table 1 ).

Advanced MRI analysis techniques, including voxel-based morphometry (VBM), surface-based morphometry (SBM), diffusion tensor imaging (DTI), and resting-state functional MRI (rs-fMRI) have developed rapidly and can detect subtle structural and functional abnormalities that may be invisible on conventional MRI ( 6 ). By quantifying gray matter atrophy, white matter (WM) microstructural damage, myelin-related changes, and functional connectivity alterations, multimodal MRI may provide further insight into the imaging characteristics and pathophysiological mechanisms of NMOSD.

Machine learning provides a data-driven approach for identifying complex patterns from high-dimensional clinical and imaging datasets and for developing predictive models. In neuroimaging research, machine learning has been increasingly applied to diagnostic classification, prognosis prediction, and treatment-response assessment ( 7 ). Imaging-based machine learning models may therefore offer additional value for NMOSD by integrating conventional MRI features, quantitative imaging biomarkers, and clinical information. However, heterogeneity exists across different studies such as differences in patient populations, imaging protocols, preprocessing strategies, and machine learning methodologies. The clinical applicability is still unclear ( Table 2 ).

This review aims to summarize current evidence on conventional MRI features, advanced quantitative MRI biomarkers, and imaging-based machine learning applications in NMOSD, with a focus on differential diagnosis, relapse risk assessment, treatment-response evaluation, and outcome prediction, so as to provide an integrated perspective for clinical application ( Figure 1 ).

Conventional MRI features in NMOSD

Examination methods based on conventional MRI can provide preliminary qualitative analysis, such as the location, size, morphology and contrast-enhanced pattern of lesion but cannot yield a definitive diagnosis. Additional examinations, such as medical history, electrophysiological tests and cerebrospinal fluid analysis are still required. Even so, conventional MRI remains clinically valuable.

In AQP4-IgG-positive NMOSD, lesions frequently involve regions with high aquaporin-4 expression, including the periependymal regions, area postrema, brainstem, optic nerves, and spinal cord. Therefore, the aforementioned characteristic changes can also be observed on conventional MRI.

Brain MRI may demonstrate lesions in the dorsal medulla and area postrema, as well as periependymal regions surrounding the third and fourth ventricles. Acute lesions may show patchy, cloud-like, or linear periependymal enhancement, whereas partial or complete lesion resolution may be observed during follow-up. On spinal cord MRI, acute lesions are often centrally located on axial images and may involve the central gray matter. On sagittal images, the lesions frequently extend over three or more vertebral segments, although shorter lesions may also occur. Bright spotty lesions, spinal cord swelling, and patchy or ring-like enhancement may be observed. During the chronic stage, lesions may shorten or resolve, while severe attacks may result in cavitation and focal or longitudinal spinal cord atrophy. Optic nerve lesions commonly involve the posterior portion of the optic nerve and may extend to the optic chiasm. Acute lesions may show optic nerve swelling and enhancement, whereas persistent signal abnormalities or optic nerve atrophy may remain during the chronic stage ( 8 , 9 ).

Currently, the widespread clinical application lies in the differential diagnosis between relapsing–remitting multiple sclerosis (RRMS) and NMOSD, which remains a critical clinical challenge due to substantial phenotypic overlap in disease manifestations. MS lesions predominantly affect periventricular regions (exhibiting Dawson’s finger configuration), juxtacortical zones, infratentorial structures, corpus callosum, and cortical gray matter. In contrast, NMOSD typically involves AQP4 antibody enriched regions, including periependymal regions surrounding the third ventricle and cerebral aqueduct, dorsal brainstem adjacent to the fourth ventricle, corticospinal tracts, and hemispheric WM ( 10 ). Longitudinally extensive spinal cord lesions involving at least three vertebral segments and optic nerve lesions involving more than 50% of the nerve length are important characteristic or supportive MRI findings of NMOSD, but they are not mandatory in all patients. Importantly, characteristic MS features—such as ovoid periventricular lesions, cerebellar peduncle involvement, and juxtacortical U-fiber lesions—are infrequent in NMOSD ( 11–13 ).

Although conventional MRI has improved diagnostic discrimination, the interpretation remains subjective and reliant on clinical expertise. Recent neuroradiology-focused reviews further emphasize that longitudinally extensive myelopathy and brainstem or middle cerebellar peduncle lesions are not disease-specific and should be interpreted within a broad differential diagnosis that includes MS, MOGAD, ADEM, inflammatory myelopathies, and non-inflammatory mimics ( 14 , 15 ). Therefore, MRI findings should be interpreted together with clinical phenotype, AQP4-IgG and MOG-IgG status, cerebrospinal fluid findings, enhancement pattern, and lesion evolution, particularly in seronegative or clinically ambiguous NMOSD cases ( 16 ).

Novel MRI technologies and alterations in NMOSD

Conventional MRI has limited sensitivity for subtle abnormalities in normal-appearing tissue and provides little quantitative information on tissue integrity. Advanced quantitative techniques can complement conventional imaging by characterizing structural, microstructural, myelin-related, and functional alterations, offering further insight into tissue injury, though their clinical utility remains under investigation and methodological heterogeneity currently limits routine application.

White matter injuries in NMOSD

Diffusion tensor imaging (DTI), a non-invasive MRI technique based on diffusion-weighted imaging (DWI), evaluates WM microstructure by measuring water molecule diffusion patterns, revealing axonal damage and demyelination ( 17 ). The main parameters of DTI include fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD). FA reflects directional water diffusion coherence, with reduced values indicating impaired WM integrity and correlating with disease severity. MD measures average diffusion magnitude, increasing with edema or cellular necrosis. AD and RD capture diffusion parallel and perpendicular to axons, respectively, elevated AD suggests acute axonal injury and higher RD implies myelin disruption ( 18 ).

DTI studies have identified microstructural abnormalities in both visibly affected T2 lesions and normal-appearing white matter (NAWM) in NMOSD. Reduced FA and increased diffusivity have frequently been reported in pathways particularly vulnerable to NMOSD-related injury, including the optic radiations and corticospinal tracts. Abnormal diffusion measures have also been detected in periependymal regions and other areas anatomically associated with high AQP4 expression, including perichiasmal, periaqueductal, and periventricular regions, as well as thalamic pathways, even in the absence of corresponding abnormalities on conventional MRI ( 16 ). These findings suggest that conventional lesion assessment may underestimate the spatial extent of tissue injury in NMOSD.

Most studies confirm widespread FA reductions in NMOSD white matter and cervical spinal cord ( 19–25 ) in NMOSD. However, MD, AD, and RD changes show less consistency. Mechanistically, AQP4 antibody-mediated attacks increase membrane permeability, disrupting water molecule alignment and reducing anisotropy. Concurrent myelin damage and axonal loss further degrade fiber integrity, typically lowering FA while elevating MD, AD, and RD. However, early-stage patients without significant axonal injury may exhibit normal AD values.

Additionally, DTI has certain value in differentiating NMOSD from MS. Although substantial overlap of imaging features remains, patients with MS and NMOSD show different patterns of microstructural damage in T2 lesions and NAWM areas ( 26 ). The T2-visible lesions are substantially more abundant in MS and can dominate whole-brain WM comparisons ( 27 ). Chen et al. reported higher T2 lesion volumes in 14 of 20 evaluated WM tracts in MS than in NMOSD. Microstructural damage within lesions was also more severe in 13 tracts in MS ( 28 ).

Not only is total lesion volume larger in MS than in NMOSD, but lesion distribution also differs between the two disorders. In comparison with MS, tract-specific studies have reported DTI abnormalities in the fornix, corona radiata, inferior longitudinal fasciculus, optic radiations, and corpus callosum in MS, whereas the corticospinal tracts and optic radiations are frequently involved in NMOSD ( 24 , 25 , 29 ).

Finally, although the abnormal NAWM metrics could be found in both disorders across the evaluated tracts, the damage in MS is more extensive. Kim et al. previously demonstrated widespread occult NAWM abnormalities in NMOSD compared with healthy controls, but the abnormalities were more severe in MS, particularly for AD changes and involvement of association fibers such as the inferior cerebellar peduncle, external capsule, cingulum, superior fronto-occipital fasciculus, and uncinate fasciculus ( 26 ).

Taken together, DTI shows extensive microstructural damage in both visibly affected T2 lesions and NAWM in NMOSD. However, individual DTI metrics are unlikely to provide sufficient diagnostic specificity when used alone. Their greatest potential may lie in regional or multivariable imaging patterns combined with conventional MRI findings, clinical characteristics, and antibody status.

Gray matter structure abnormalities in NMOSD

NMOSD has long been considered an autoimmune disease that primarily affects the WM. However, because WM demyelination by itself cannot explain the full extent of clinical disability such as cognitive decline, renewed interest in gray matter (GM) pathology has been emerging ( 30 ). GM structural abnormalities in NMOSD have been investigated using double inversion recovery (DIR) imaging and high-resolution T1-weighted MRI combined with region-of-interest analyses, voxel-based morphometry (VBM), and surface-based morphometry (SBM). DIR imaging may improve the detection of cortical and juxtacortical lesions, whereas T1-weighted morphometric approaches permit quantitative assessment of regional gray matter volume, cortical thickness, and surface-based measures.

Previous histopathological analyses have revealed meningeal inflammatory infiltrates without concomitant cortical demyelination in NMOSD patients ( 31 ). A finding corroborated by 7 Tesla MRI studies also showed the absence of cortical lesions ( 32 ), suggesting that cortical lesions are usually not seen in patients with NMOSD. However, despite the lack of focal cortical lesions, patients with NMOSD exhibit structural alterations of GM in the cerebrum and cerebellum ( 33 ). Early studies identified focal GM abnormalities in NMOSD patients with unremarkable conventional MRI findings ( 34 ), while subsequent studies have demonstrated focal rather than diffuse cortical atrophy in NMOSD patients compared with healthy controls (HCs), particularly affecting the right middle frontal gyrus, postcentral gyrus cortex ( 35 ), thalamus, prefrontal GM ( 36 ), occipital cortex ( 37 ) and bilateral cerebellar cortices ( 38 ). Emerging evidence further suggests that GM volume loss is predominantly localized to sensorimotor, visual, and associated WM pathways ( 39–41 ). By contrast, some neuroimaging investigations failed to detect significant GM structural alterations ( 42 ). Differences in acquisition protocols, segmentation methods, patient characteristics, and statistical thresholds may contribute to considerable variability across studies.

Notably, recent research has revealed a subset of NMOSD patients manifest cognitive impairment ( 43 ), which is strongly associated with GM pathology. Current research has identified selective thalamic atrophy as a characteristic finding, particularly accentuated in cognitively impaired NMOSD patients relative to cognitively preserved counterparts ( 44 ). Intriguingly, cortical atrophy progression appears to be correlated with age and disease chronicity rather than acute inflammatory burden or lesion load ( 45 ), suggesting that neurodegenerative mechanisms may drive disability accumulation independently of inflammatory activity. These observations suggest that cortical neuroimaging abnormalities and cognitive dysfunction may arise from secondary neurodegenerative processes rather than primary cortical demyelination. The focal GM atrophy has been postulated to represent trans-synaptic degeneration secondary to optic nerve and spinal cord lesions ( 16 ).

Quantitative gray matter measures may also aid differentiation between NMOSD and MS. Compared with NMOSD, MS generally shows more extensive cortical and deep gray matter atrophy, particularly involving the thalamus, caudate nucleus, hippocampus and parahippocampal regions, and other limbic structures. These alterations have been associated with disability, cognitive impairment, and lesion burden in MS ( 46 , 47 ). Conversely, NMOSD demonstrates limited deep gray matter involvement with preserved cortical architecture, potentially explaining its absence of progressive subtypes ( 48 ). However, most of these cohorts were predominantly or exclusively AQP4-IgG-positive; whether these findings extend to seronegative NMOSD or MOGAD remains largely unexplored.

Abnormal neuron activity in NMOSD

Neuronal and synaptic activity spontaneously generate low-frequency blood oxygen level-dependent (BOLD) oscillations (<0.1 Hz) during non-task conditions ( 49 ). Resting-state fMRI (rs-fMRI) detects BOLD oscillations and provides insights into functional neuropathology through three metrics: amplitude of low-frequency fluctuation (ALFF, neuronal activity intensity), regional homogeneity (ReHo, local synchronization), and functional connectivity (FC, interregional network correlations) ( 50 ). These biomarkers collectively characterize resting state cerebral functional architecture and interregional communication, providing critical insights into NMOSD pathophysiology ( 51 ).

NMOSD patients exhibit reduced ReHo and ALFF in cortical regions (precuneus, cingulate gyri, medial frontal gyrus) despite normal structural MRI ( 52 ). Independent component analysis (ICA) of network connectivity demonstrates diminished FC correlating with cognitive dysfunction ( 53 ), indicating widespread neuronal hypoactivity affecting perceptual, motor, and cognitive processing.

Paradoxically, concurrent ALFF elevations emerge in subcortical and cerebellar regions, including the caudate nucleus, parahippocampal gyrus, superior temporal gyrus, thalamus, and middle frontal gyrus ( 54 , 55 ), while enhanced sensorimotor network FC inversely correlates with disability scores ( 56 ). This hyperactivation pattern may reflect compensatory neuroplasticity during motor or cognitive decline ( 51 ). This bidirectional pattern—regional hypoactivity alongside compensatory hyperactivation—suggests dynamic neural resource redistribution, where sustained hyperactivity may ultimately exacerbate neurodegeneration.

Functional MRI may also reveal group-level differences between NMOSD and MS. Although widespread network disruption has been observed in both disorders, more pronounced thalamocortical connectivity abnormalities have been reported in MS, potentially contributing to the greater cognitive impairment observed in some MS cohorts ( 57 ). Task-based fMRI during olfactory stimulation reveals differential activation: NMOSD patients exhibit reduced activation in the right superior temporal gyrus and left paracentral lobule, yet hyperactivation in bilateral insulae, hippocampi, parahippocampal gyri, left inferior orbital gyrus, superior temporal gyrus, putamen, and middle frontal gyrus compared with MS ( 58 ). These functional signatures provide potential diagnostic biomarkers. However, current evidence is based predominantly on small, single-center, and mainly AQP4-IgG-positive cohorts. Their reproducibility across centers and analytical pipelines, as well as their generalizability to seronegative NMOSD and MOGAD, remains to be established.

Application of machine learning in NMOSD

Machine learning is a rapidly growing field at present. Machine learning models incorporating diverse disease biomarkers demonstrate diagnostic and prognostic capabilities that can rival expert performance ( 59 , 60 ). MR imaging biomarkers serve as critical tools for constructing models. In NMOSD, machine learning research mainly focuses on differential diagnosis with MS and prognostic prediction ( 53 , 61–63 ).

Early differential diagnosis of NMOSD and MS

As mentioned above, the early identification of NMOSD from MS is of great significance. A model based on conventional radiological features achieved an accuracy of 87.1% and an AUC of 0.874 ( 64 ). Other region-of-interest-based approaches have reported accuracies or AUC values above 0.80 ( 65–67 ). Deep learning methods have also demonstrated potential, achieving accuracies above 70% ( 63 , 68 , 69 ), with convolutional neural network (CNN)-based image recognition attaining a remarkable accuracy of 98.8% in distinguishing MS from NMOSD ( 63 ). However, such exceptionally high performance should be interpreted cautiously: estimates from small, single-center, or internally partitioned datasets may be inflated by overfitting, class imbalance, scanner-specific characteristics, non-independent training and test samples, or information leakage.

Additionally, a transformer-based multiple-instance learning model integrating brain and spinal cord MRI differentiated MS, AQP4-IgG-positive NMOSD, and MOGAD with an overall accuracy of 81.4% ( 70 ). Similarly, a modified ResNet18 model using five selected axial FLAIR slices achieved an accuracy of 76.1% and an AUC of 0.85 for differentiating NMOSD from MS ( 71 ), with Grad-CAM indicating that WM lesions contributed substantially to classification. However, these models were developed retrospectively at a single center and evaluated using an internal train–test split, without independent external validation.

A subsequent systematic review and meta-analysis reported pooled accuracy, sensitivity, and specificity of 0.87, 0.86, and 0.84, respectively, for machine-learning and deep-learning models differentiating neuroimmunological diseases ( 72 ). However, substantial between-study heterogeneity, particularly for accuracy and specificity, limited the certainty and generalizability of these pooled estimates. A 2025 scoping review of deep-learning applications in neuroinflammatory diseases similarly found that computer-aided diagnosis and radiological segmentation currently dominate the field, whereas prognostic, functional, and treatment-related applications remain less developed ( 73 ).

Regarding novel MRI biomarkers, machine learning models using features of volume, thickness, and surface of 50 cortical GM regions and volumes of the deep GM nuclei achieved an accuracy of 80% for MS-NMOSD differentiation, identifying the thalamus as the most discriminative region ( 74 ). Quantitative susceptibility mapping (QSM)-derived deep gray matter susceptibility features yielded an accuracy of 86.7% and an AUC of 0.902 ( 75 ). Integrating rs-fMRI features with machine learning classifiers achieved 90.9% accuracy (AUC: 0.929) for NMOSD-MS discrimination ( 76 ). Multimodal MRI models combining deep gray matter volume, lesion volume, DTI, and functional connectivity metrics achieved a mean accuracy of 88% ( 53 ). Notably, visible WM lesions, fMRI, cognitive scores, and DTI received high model weights, while other modalities were assigned zero weights, underscoring the insufficiency of single-modal features and emphasizing the necessity of advanced multimodal imaging.

However, most models rely predominantly on cranial MRI, with limited incorporation of spinal cord imaging or optical coherence tomography despite their diagnostic relevance. Recent reviews have highlighted this gap, reporting that fusing brain and spinal cord MRI may improve discrimination and that pooled machine learning and deep learning models show favorable diagnostic performance for MS-versus-NMOSD classification, though these estimates should be interpreted cautiously given substantial differences in disease composition, imaging inputs, algorithms, and validation across included studies ( 72 ).

Exploratory imaging associations with relapse burden and disability

As noted above, a substantial proportion of NMOSD patients experience relapse within the first year ( 77 ). Early use of preventive immunotherapy to reduce the risk of additional relapses for high recurrence risk patients may be especially beneficial ( 78 , 79 ). Although advanced neuroimaging has identified prognostic biomarkers in MS—such as paramagnetic rim lesions (PRLs), subacute enhancing lesions (SELs) ( 80 , 81 ), and translocator protein (TSPO)-active lesions ( 82 )—similar predictive biomarkers remain scarce in NMOSD. Clinical data-driven models have also shown limited potential in relapse prediction ( 83 , 84 ).

Conventional imaging studies suggest associations between brainstem (particularly medullary) involvement and long-term disability ( 85–87 ), while reduced nucleus accumbens volume correlates with increased relapse frequency, and caudate nucleus atrophy correlates with disease duration ( 88 ). Regarding novel MRI biomarkers, only one study established connectome-based predictive modeling (CPM) using DTI to predict disability progression, associating frontoparietal WM disconnection with EDSS scores ( 89 ). However, the study estimated current disability rather than a prespecified future clinical outcome, even though a predictive modeling framework was used. At present, no validated imaging-based model has yet predicted future relapse or disability over a clearly defined prediction horizon.

The scarcity of prognostic models in NMOSD likely reflects the rarity of the disease, limited availability of large longitudinal cohorts, heterogeneity in relapse timing and preventive treatments, inconsistent imaging protocols and outcome definitions, and incomplete separation of AQP4-IgG-positive NMOSD, seronegative NMOSD, and MOGAD. A robust prognostic machine-learning pipeline should therefore use multicenter longitudinal cohorts, prespecified clinical outcomes and prediction horizons, harmonized MRI acquisition and preprocessing, explicit adjustment or stratification for antibody status and treatment exposure, and nested model development with temporal and independent external validation. Calibration, interpretability, and incremental clinical utility should also be assessed before clinical implementation. Given the value of global and regional brain atrophy measures in predicting disability progression and cognitive decline in MS, advanced MRI biomarkers may have potential for future prognostic modeling in NMOSD.

Treatment-response assessment and monitoring

Quantitative measures of tissue injury, including diffusion, myelin-sensitive, functional, and metabolic imaging markers may provide complementary information beyond conventional lesion assessment ( 90 , 91 ). However, evidence supporting imaging-based machine-learning models for treatment-response assessment in NMOSD remains extremely limited. To date, no externally validated model has been established for monitoring therapeutic response or predicting treatment-related outcomes specifically in NMOSD. Nevertheless, a recent longitudinal study using machine-learning-assisted lesion segmentation found that inflammatory lesion volume increased significantly during remission in MS (3,493 → 4,430 mm 3 , p < 0.001), but remained relatively stable in AQP4-IgG-positive NMOSD (640 → 930 mm 3 , p = 0.129), suggesting distinct patterns of subclinical disease activity and potentially different MRI-monitoring strategies ( 92 ).

Additionally, a recent study in MS has suggested that deep-learning methods may identify contrast-enhancing lesions from non-contrast MRI and that myelin water fraction may serve as a marker of remyelination ( 93 ), particularly relevant given the association between NMOSD contrast-enhancing lesions and poor recovery ( 94 ). Thus, advanced MRI-based prognostic and therapeutic evaluation studies may emerge as future priorities in NMOSD. However, these findings represent indirect evidence and cannot be extrapolated directly to NMOSD because of differences in disease mechanisms, lesion characteristics, and treatment response. Prospective NMOSD-specific studies with standardized longitudinal imaging and predefined treatment-response outcomes are therefore required.

Methodological considerations

Advanced MRI biomarkers and machine learning approaches have shown potential for improving early differential diagnosis, disease assessment, and treatment monitoring in NMOSD. However, clinical implementation of advanced imaging biomarkers and machine learning models remains limited.

Firstly, the clinical applicability depends on robust validation, model interpretability, and generalizability across different scanners, protocols, and patient populations. However, most available studies are based on relatively small or single-center cohorts, with heterogeneous MRI scanners, acquisition protocols, feature extraction methods, and patient populations. Direct comparison across these studies remains difficult given differences in diagnostic criteria, antibody status, acquisition and preprocessing, feature selection, algorithm choice, and validation strategy.

Additionally, most available studies were retrospective cohorts, and relied on internal cross-validation rather than independent external validation, with limited assessment of calibration, interpretability, and generalizability to seronegative NMOSD, MOGAD, and atypical presentations, which may lead to overestimated performance and limited generalizability.

In addition, several models were developed in geographically and ethnically homogeneous cohorts, and their performance may not generalize to populations with different demographic, genetic, disease-subtype, and healthcare characteristics. External validation across ethnically and geographically diverse cohorts is therefore essential before clinical implementation.

Moreover, feature selection, preprocessing, and hyperparameter tuning should be performed exclusively within the training data or cross-validation folds to minimize selection bias and information leakage. In addition to discrimination metrics, future studies should assess model calibration and incorporate interpretability approaches, such as feature-importance analysis, SHAP values, or saliency maps, to determine whether predictions are based on clinically plausible imaging patterns.

Future studies should prioritize multicenter prospective validation, standardized MRI acquisition and post-processing pipelines, transparent handling of class imbalance, independent external validation with decision-curve analyses, and greater incorporation of spinal cord and optic nerve imaging, compared directly against established clinical, serological, and radiological assessment to determine whether machine learning provides reproducible, clinically meaningful incremental value.

Conclusion and limitation

Advanced MRI techniques have revealed structural, microstructural, and functional abnormalities beyond lesions visible on conventional MRI in NMOSD. Current imaging-based machine-learning evidence is predominantly focused on differential diagnosis, particularly the distinction between NMOSD and MS, whereas applications in relapse prediction, disability progression, treatment-response assessment, and therapeutic monitoring remain exploratory and lack adequate independent external validation. Accordingly, these tools should remain adjunctive and should not replace clinical assessment or antibody testing. Their clinical applicability depends on robust validation, model interpretability, and generalizability across different scanners, protocols, and patient populations.

The present review has two additional limitations. First, its technical coverage is restricted, as detailed discussion focuses mainly on three relatively established advanced MRI approaches, namely DTI, structural morphometric analysis, and resting-state functional MRI. Other potentially relevant techniques, including magnetization transfer imaging and magnetization transfer ratio, myelin water fraction imaging, diffusion kurtosis imaging, neurite orientation dispersion and density imaging, free-water imaging and other advanced diffusion models, quantitative susceptibility mapping and susceptibility-weighted imaging, arterial spin labeling and other perfusion MRI techniques, magnetic resonance spectroscopy, quantitative spinal cord MRI, and quantitative imaging of the optic nerves and optic chiasm, were not systematically reviewed. Future reviews could provide a more comprehensive synthesis of these modalities.

Second, the available literature does not sufficiently distinguish among AQP4-IgG-positive NMOSD, seronegative NMOSD, and MOGAD. These populations differ in pathophysiology, imaging patterns, clinical course, and diagnostic uncertainty, and many studies included predominantly AQP4-IgG-positive patients or did not clearly report serostatus. Therefore, the summarized findings may not be fully generalizable to seronegative NMOSD or MOGAD. Future reviews and imaging studies should pay greater attention to diagnostically challenging cases, including AQP4-IgG-negative or equivocal presentations, atypical optic neuritis or myelitis, and vascular, infectious, inflammatory, or autoimmune mimics.

Edited by: Valentina Tomassini , University of Studies G. d'Annunzio Chieti and Pescara, Italy

Reviewed by: João Vitor Gerdulli Tamanini , Federal University of São Paulo, Brazil

Xiaoya Chen , First Affiliated Hospital of Chongqing Medical University, China

Author contributions

JZ: Writing – review & editing, Writing – original draft. HX: Writing – review & editing, Data curation. MZ: Writing – review & editing. ZZ: Writing – original draft, Methodology. YZ: Writing – review & editing, Visualization, Writing – original draft.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

1. Tan CT Mao Z Qiu W Hu X Wingerchuk DM Weinshenker BG . International consensus diagnostic criteria for neuromyelitis optica spectrum disorders . Neurology . ( 2016 ) 86 : 491 – 2 . doi: 10.1212/WNL.0000000000002366 , 26833940

2. Bichuetti DB Oliveira EM Souza NA Tintor M Gabbai AA . Patients with neuromyelitis optica have a more severe disease than patients with relapsingremitting multiple sclerosis, including higher risk of dying of a demyelinating disease . Arq Neuropsiquiatr . ( 2013 ) 71 : 275 – 9 . doi: 10.1590/0004-282x20130020 23689404

3. Wingerchuk DM Lennon VA Pittock SJ Lucchinetti C Weinshenker B . Revised diagnostic criteria for neuromyelitis optica . Neurology . ( 2006 ) 66 : 1485 – 9 . doi: 10.1212/01.wnl.0000216139.44259.74 16717206

4. Nakashima I Fujihara K Miyazawa I Misu T Narikawa K Nakamura M . Clinical and MRI features of Japanese patients with multiple sclerosis positive for NMO-IgG . J Neurol Neurosurg Psychiatry . ( 2006 ) 77 : 1073 – 5 . doi: 10.1136/jnnp.2005.080390 , 16505005 PMC2077753

5. Iampreechakul P Sopchokchai I Wangtanaphat K Chuntaroj S Wattanasen Y Hangsapruek S . Holocord myelopathy misdiagnosed as neuromyelitis optica spectrum disorder (NMOSD): a unique case of dural arteriovenous fistula at the craniocervical junction along first spinal nerve . Surg Neurol Int . ( 2024 ) 15 : 209 . doi: 10.25259/SNI_138_2024 , 38974535 PMC11225421

6. Pache F Zimmermann H Finke C Lacheta A Papazoglou S Kuchling J . Brain parenchymal damage in neuromyelitis optica spectrum disorder - a multimodal MRI study . Eur Radiol . ( 2016 ) 26 : 4413 – 22 . doi: 10.1007/s00330-016-4282-x , 27012555

7. Bratic B Kurbalija V Ivanovic M Oder I Bosnic Z . Machine learning for predicting cognitive diseases: methods, data sources and risk factors . J Med Syst . ( 2018 ) 42 : 243 . doi: 10.1007/s10916-018-1071-x , 30368611

8. Jarius S Aktas O Ayzenberg I Bellmann-Strobl J Berthele A Giglhuber K . Update on the diagnosis and treatment of neuromyelits optica spectrum disorders (NMOSD) – revised recommendations of the Neuromyelitis Optica study group (NEMOS). Part I: diagnosis and differential diagnosis . J Neurol . ( 2023 ) 270 : 3341 – 68 . doi: 10.1007/s00415-023-11634-0 , 37022481 PMC10267280

9. Cortese R Prados Carrasco F Tur C Bianchi A Brownlee W de Angelis F . Differentiating multiple sclerosis from AQP4-Neuromyelitis Optica Spectrum disorder and MOG-antibody disease with imaging . Neurology . ( 2023 ) 100 : e308 – 23 . doi: 10.1212/WNL.0000000000201465 , 36192175 PMC9869760

10. Carnero Contentti E Okuda DT Rojas JI Chien C Paul F Alonso R . MRI to differentiate multiple sclerosis, neuromyelitis optica, and myelin oligodendrocyte glycoprotein antibody disease . J Neuroimaging . ( 2023 ) 33 : 688 – 702 . doi: 10.1111/jon.13137 , 37322542

11. Kim W Park MS Lee SH . Characteristic brain magnetic resonance imaging abnormalities in central nervous system aquaporin-4 autoimmunity . Mult Scler J . ( 2010 ) 16 : 1229 – 36 . doi: 10.1177/1352458510376640 20685766

12. Liao MF Chang KH Lyu RK Huang CC Chang HS Wu YR . Comparison between the cranial magnetic resonance imaging features of neuromyelitis optica spectrum disorder versus multiple sclerosis in Taiwanese patients . BMC Neurol . ( 2014 ) 14 : 218 . doi: 10.1186/s12883-014-0218-8 , 25433369 PMC4264553

13. Matthews L Marasco R Jenkinson M Küker W Luppe S Leite MI . Distinction of seropositive NMO spectrum disorder and MS brain lesion distribution . Neurology . ( 2013 ) 80 : 1330 – 7 . doi: 10.1212/WNL.0b013e3182887957 , 23486868 PMC3656462

14. Tamanini JVG Sabino JV Cordeiro RA Mizubuti V Villarinho LL Duarte JÁ . The role of MRI in differentiating demyelinating and inflammatory (not infectious) myelopathies . Semin Ultrasound CT MR . ( 2023 ) 44 : 469 – 88 . doi: 10.1053/j.sult.2023.03.017 , 37555683

15. Tamanini JVG Ribeiro GAS Kimura AT Borella LF Freddi TA Reis F . Hyperintense lesions of the middle cerebellar peduncle and beyond: a pictorial essay . Radiol Bras . ( 2024 ) 57 : e20240001 . doi: 10.1590/0100-3984.2024.0001 39830058 PMC11740878

16. Wang J Cui C Lu Y Chang Y Wang Y Li R . Therapeutic response and possible biomarkers in acute attacks of neuromyelitis optica spectrum disorders: a prospective observational study . Front Immunol . ( 2021 ) 12 : 720907 . doi: 10.3389/fimmu.2021.720907 34421925 PMC8372759

17. Alexander DC Dyrby TB Nilsson M Zhang H . Imaging brain microstructure with diffusion MRI: practicality and applications . NMR Biomed . ( 2019 ) 32 : e3841 . doi: 10.1002/nbm.3841 , 29193413

18. Chen YJ Nabavizadeh SA Vossough A Kumar S Loevner LA Mohan S . Wallerian degeneration beyond the corticospinal tracts: conventional and advanced MRI findings . J Neuroimaging . ( 2017 ) 27 : 272 – 80 . doi: 10.1111/jon.12404 , 28072502

19. Qian W Chan Q Mak H Zhang Z Anthony MP Yau KKW . Quantitative assessment of the cervical spinal cord damage in neuromyelitis optica using diffusion tensor imaging at 3 tesla . J Magn Reson Imaging . ( 2011 ) 33 : 1312 – 20 . doi: 10.1002/jmri.22575 , 21590999

20. Benedetti B Valsasina P Judica E Martinelli V Ghezzi A Capra R . Grading cervical cord damage in neuromyelitis optica and MS by diffusion tensor MRI . Neurology . ( 2006 ) 67 : 161 – 3 . doi: 10.1212/01.wnl.0000223637.65208.7c , 16832101

21. Rivero RL Oliveira EM Bichuetti DB Gabbai AA Nogueira RG Abdala N . Diffusion tensor imaging of the cervical spinal cord of patients with neuromyelitis optica . Magn Reson Imaging . ( 2014 ) 32 : 457 – 63 . doi: 10.1016/j.mri.2014.01.023 , 24636569

22. Cacciaguerra L Carotenuto A Pagani E Mistri D Radaelli M Martinelli V . Magnetic resonance imaging evaluation of perivascular space abnormalities in Neuromyelitis Optica . Ann Neurol . ( 2022 ) 92 : 173 – 83 . doi: 10.1002/ana.26419 , 35596582 PMC9544484

23. Chen Z Lou X Liu M Huang D Wei S Yu S . Assessment of optic nerve impairment in patients with neuromyelitis optica by MR diffusion tensor imaging . PLoS One . ( 2015 ) 10 : e0126574 . doi: 10.1371/journal.pone.0126574 , 25962183 PMC4427268

24. Yu C Lin F Li K Jiang T Qin W Sun H . Pathogenesis of normal-appearing white matter damage in neuromyelitis optica: diffusion-tensor MR imaging . Radiology . ( 2008 ) 246 : 222 – 8 . doi: 10.1148/radiol.2461062075 , 18033757

25. Rueda Lopes FC Doring T Martins C Cabral FC Malfetano FR Pereira VCSR . The role of demyelination in neuromyelitis optica damage: diffusion-tensor MR imaging study . Radiology . ( 2012 ) 263 : 235 – 42 . doi: 10.1148/radiol.12111241 , 22438446

26. Kim SH Kwak K Hyun JW Joung A Lee SH Choi YH . Diffusion tensor imaging of normal-appearing white matter in patients with neuromyelitis optica spectrum disorder and multiple sclerosis . Eur J Neurol . ( 2017 ) 24 : 966 – 73 . doi: 10.1111/ene.13321 , 28643955

27. Zhao DD Zhou HY Wu QZ Liu J Chen XY He D . Diffusion tensor imaging characterization of occult brain damage in relapsing neuromyelitis optica using 3.0T magnetic resonance imaging techniques . NeuroImage . ( 2012 ) 59 : 3173 – 7 . doi: 10.1016/j.neuroimage.2011.11.022 , 22108642

28. Chen X Roberts N Zheng Q Peng Y Han Y Luo Q . Comparison of diffusion tensor imaging (DTI) tissue characterization parameters in white matter tracts of patients with multiple sclerosis (MS) and neuromyelitis optica spectrum disorder (NMOSD) . Eur Radiol . ( 2024 ) 34 : 5263 – 75 . doi: 10.1007/s00330-023-10550-1 , 38175221

29. Roosendaal SD Geurts JJ Vrenken H Hulst HE Cover KS Castelijns JA . Regional DTI differences in multiple sclerosis patients . NeuroImage . ( 2009 ) 44 : 1397 – 403 . doi: 10.1016/j.neuroimage.2008.10.026 , 19027076

30. Kawachi I Nishizawa M . Significance of gray matter brain lesions in multiple sclerosis and neuromyelitis optica . Neuropathology . ( 2015 ) 35 : 481 – 6 . doi: 10.1111/neup.12216 , 26079808

31. Saji E Arakawa M Yanagawa K Toyoshima Y Yokoseki A Okamoto K . Cognitive impairment and cortical degeneration in neuromyelitis optica . Ann Neurol . ( 2013 ) 73 : 65 – 76 . doi: 10.1002/ana.23721 , 23378324

32. Kister I Herbert J Zhou Y Ge Y . Ultrahigh-field MR (7 T) imaging of brain lesions in Neuromyelitis Optica . Mult Scler Int . ( 2013 ) 2013 : 398259 . doi: 10.1155/2013/398259 , 23431447 PMC3568863

33. Liu Y Jiang X Butzkueven H Duan Y Huang J Ren Z . Multimodal characterization of gray matter alterations in neuromyelitis optica . Mult Scler (Houndmills, Basingstoke, England) . ( 2017 ) 24 : 1308 – 16 . doi: 10.1177/1352458517721053 , 28741987

34. Rocca MA Agosta F Mezzapesa DM Martinelli V Salvi F Ghezzi A . Magnetization transfer and diffusion tensor MRI show gray matter damage in neuromyelitis optica . Neurology . ( 2004 ) 62 : 476 – 8 . doi: 10.1212/01.WNL.0000106946.08741.41 , 14872036

35. Kim SH Kwak K Hyun JW Jeong IH Jo HJ Joung A . Widespread cortical thinning in patients with neuromyelitis optica spectrum disorder . Eur J Neurol . ( 2016 ) 23 : 1165 – 73 . doi: 10.1111/ene.13011 , 27108769

36. Chanson JB Lamy J Rousseau F Blanc F Collongues N Fleury M . White matter volume is decreased in the brain of patients with neuromyelitis optica . Eur J Neurol . ( 2013 ) 20 : 361 – 7 . doi: 10.1111/j.1468-1331.2012.03867.x , 22978568

37. Cortese R Battaglini M Prados F Gentile G Luchetti L Bianchi A . Grey matter atrophy and its relationship with white matter lesions in patients with myelin oligodendrocyte glycoprotein antibody-associated disease, Aquaporin-4 antibody-positive neuromyelitis optica spectrum disorder, and multiple sclerosis . Ann Neurol . ( 2024 ) 96 : 276 – 88 . doi: 10.1002/ana.26951 , 38780377

38. Duan Y Zhuo Z Li H Tian DC Li Y Yang L . Brain structural alterations in MOG antibody diseases: a comparative study with AQP4 seropositive NMOSD and MS . J Neurol Neurosurg Psychiatry . ( 2021 ) 92 : 709 – 16 . doi: 10.1136/jnnp-2020-324826 , 33687975 PMC8223649

39. Schiavi S Petracca M Battocchio M el Mendili MM Paduri S Fleysher L . Sensory-motor network topology in multiple sclerosis: structural connectivity analysis accounting for intrinsic density discrepancy . Hum Brain Mapp . ( 2020 ) 41 : 2951 – 63 . doi: 10.1002/hbm.24989 , 32412678 PMC7336144

40. von Glehn F Jarius S Cavalcanti Lira RP Alves Ferreira MC von Glehn FHR Costa e Castro SM . Structural brain abnormalities are related to retinal nerve fiber layer thinning and disease duration in neuromyelitis optica spectrum disorders . Mult Scler . ( 2014 ) 20 : 1189 – 97 . doi: 10.1177/1352458513519838 , 24477120

41. Duan Y Liu Y Liang P Jia X Ye J Dong H . White matter atrophy in brain of neuromyelitis optica: a voxel-based morphometry study . Acta Radiol . ( 2014 ) 55 : 589 – 93 . doi: 10.1177/0284185113501815 , 23966367

42. Blanc F Noblet V Jung B Rousseau F Renard F Bourre B . White matter atrophy and cognitive dysfunctions in neuromyelitis optica . PLoS One . ( 2012 ) 7 : e33878 . doi: 10.1371/journal.pone.0033878 , 22509264 PMC3317931

43. Calabrese M Oh MS Favaretto A Rinaldi F Poretto V Alessio S . No MRI evidence of cortical lesions in neuromyelitis optica . Neurology . ( 2012 ) 79 : 1671 – 6 . doi: 10.1212/WNL.0b013e31826e9a96 , 22993282

44. Hyun JW Park G Kwak K Jo HJ Joung A Kim JH . Deep gray matter atrophy in neuromyelitis optica spectrum disorder and multiple sclerosis . Eur J Neurol . ( 2017 ) 24 : 437 – 45 . doi: 10.1111/ene.13224 , 28009104

45. Tian DC Xiu Y Wang X Shi K Fan M Li T . Cortical thinning and ventricle enlargement in neuromyelitis optica spectrum disorders . Front Neurol . ( 2020 ) 11 : 872 . doi: 10.3389/fneur.2020.00872 , 32973658 PMC7481470

46. Weier K Eshaghi A Magon S Andelova M Radue EW Kappos L . The role of cerebellar abnormalities in neuromyelitis optica--a comparison with multiple sclerosis and healthy controls . Mult Scler (Houndmills, Basingstoke, England) . ( 2015 ) 21 : 757 – 66 . doi: 10.1177/1352458514554051 , 25480863

47. Duan Y Liu Y Liang P Jia X Yu C Qin W . Comparison of grey matter atrophy between patients with neuromyelitis optica and multiple sclerosis: a voxel-based morphometry study . Eur J Radiol . ( 2012 ) 81 : e110 – 4 . doi: 10.1016/j.ejrad.2011.01.065 , 21316170

48. Liu Y Fu Y Schoonheim MM Zhang N Fan M Su L . Structural MRI substrates of cognitive impairment in neuromyelitis optica . Neurology . ( 2015 ) 85 : 1491 – 9 . doi: 10.1212/wnl.0000000000002067 , 26423432

49. Savoldi F Rocca MA Valsasina P Riccitelli GC Mesaros S Drulovic J . Functional brain connectivity abnormalities and cognitive deficits in neuromyelitis optica spectrum disorder . Mult Scler J . ( 2020 ) 26 : 795 – 805 . doi: 10.1177/1352458519845109 , 31079538

50. Barkhof F Haller S Rombouts SA . Resting-state functional MR imaging: a new window to the brain . Radiology . ( 2014 ) 272 : 29 – 49 . doi: 10.1148/radiol.14132388 , 24956047

51. Yang Y Rui Q Wu X Chen X Han S Yang Y . Altered functional connectivity associated with cognitive impairment in neuromyelitis optica spectrum disorder . Mult Scler Relat Disord . ( 2022 ) 68 : 104113 . doi: 10.1016/j.msard.2022.104113 35987110

52. Liang P Liu Y Jia X Duan Y Yu C Qin W . Regional homogeneity changes in patients with neuromyelitis optica revealed by resting-state functional MRI . Clin Neurophysiol . ( 2011 ) 122 : 121 – 7 . doi: 10.1016/j.clinph.2010.05.026 , 20621555

53. Eshaghi A Riyahi-Alam S Saeedi R Roostaei T Nazeri A Aghsaei A . Classification algorithms with multi-modal data fusion could accurately distinguish neuromyelitis optica from multiple sclerosis . Neuroimage Clin . ( 2015 ) 7 : 306 – 14 . doi: 10.1016/j.nicl.2015.01.001 , 25610795 PMC4297886

54. Wang F Liu Y Li J Sondag M Law M Zee CS . Abnormal brain function in neuromyelitis optica: a fMRI investigation of mPASAT . Eur J Radiol . ( 2017 ) 95 : 197 – 201 . doi: 10.1016/j.ejrad.2017.08.012 , 28987668

55. Liu Y Liang P Duan Y Jia X Wang F Yu C . Abnormal baseline brain activity in patients with neuromyelitis optica: a resting-state fMRI study . Eur J Radiol . ( 2011 ) 80 : 407 – 11 . doi: 10.1016/j.ejrad.2010.05.002 , 20605694

56. Liu Y Xiong H Li X Zhang D Yang C Yu J . Abnormal baseline brain activity in Neuromyelitis Optica patients without brain lesion detected by resting-state functional magnetic resonance imaging . Neuropsychiatr Dis Treat . ( 2020 ) 16 : 71 – 9 . doi: 10.2147/NDT.S232924 , 32021200 PMC6955618

57. Liu Y Duan Y Huang J Ren Z Ye J Dong H . Multimodal quantitative MR imaging of the thalamus in multiple sclerosis and neuromyelitis optica . Radiology . ( 2015 ) 277 : 784 – 92 . doi: 10.1148/radiol.2015142786 , 26043265

58. He S Peng T He W Gou C Hou C Tan J . Comparative study of brain fMRI of olfactory stimulation in neuromyelitis optica spectrum disease and multiple sclerosis . Front Neurosci . ( 2021 ) 15 : 813157 . doi: 10.3389/fnins.2021.813157 , 35082598 PMC8785660

59. Betrouni N Delval A Chaton L Defebvre L Duits A Moonen A . Electroencephalography-based machine learning for cognitive profiling in Parkinson's disease: preliminary results . Mov Disord . ( 2019 ) 34 : 210 – 7 . doi: 10.1002/mds.27528 , 30345602

60. Morales DA Vives-Gilabert Y Gómez-Ansón B Bengoetxea E Larrañaga P Bielza C . Predicting dementia development in Parkinson's disease using Bayesian network classifiers . Psychiatry Res Neuroimaging . ( 2013 ) 213 : 92 – 8 . doi: 10.1016/j.pscychresns.2012.06.001 , 23149030

61. He T Zhao W Mao Y Wang Y Wang L Kuang Q . MS or not MS: T2-weighted imaging (T2WI)-based radiomic findings distinguish MS from its mimics . Mult Scler Relat Disord . ( 2022 ) 61 : 103756 . doi: 10.1016/j.msard.2022.103756 35397290

62. Hagiwara A Otsuka Y Andica C Kato S Yokoyama K Hori M . Differentiation between multiple sclerosis and neuromyelitis optica spectrum disorders by multiparametric quantitative MRI using convolutional neural network . J Clin Neurosci . ( 2021 ) 87 : 55 – 8 . doi: 10.1016/j.jocn.2021.02.018 , 33863534

63. Rocca MA Anzalone N Storelli L del Poggio A Cacciaguerra L Manfredi AA . Deep learning on conventional magnetic resonance imaging improves the diagnosis of multiple sclerosis mimics . Investig Radiol . ( 2021 ) 56 : 252 – 60 . doi: 10.1097/RLI.0000000000000735 , 33109920

64. Clarke L Arnett S Bukhari W Khalilidehkordi E Jimenez Sanchez S O'Gorman C . MRI patterns distinguish AQP4 antibody positive neuromyelitis optica spectrum disorder from multiple sclerosis . Front Neurol . ( 2021 ) 12 : 722237 . doi: 10.3389/fneur.2021.722237 , 34566866 PMC8458658

65. Etemadifar M Norouzi M Alaei SA Karimi R Salari M . The diagnostic performance of AI-based algorithms to discriminate between NMOSD and MS using MRI features: a systematic review and meta-analysis . Mult Scler Relat Disord . ( 2024 ) 87 : 105682 . doi: 10.1016/j.msard.2024.105682 , 38781885

66. Huang J Xin B Wang X Qi Z Dong H Li K . Multi-parametric MRI phenotype with trustworthy machine learning for differentiating CNS demyelinating diseases . J Transl Med . ( 2021 ) 19 : 377 . doi: 10.1186/s12967-021-03015-w , 34488799 PMC8419989

67. Ma X Zhang L Huang D Lyu J Fang M Hu J . Quantitative radiomic biomarkers for discrimination between neuromyelitis optica spectrum disorder and multiple sclerosis . J Magn Reson Imaging . ( 2018 ) 49 : 1113 – 21 . doi: 10.1002/jmri.26287 , 30408268

68. Wang Z Yu Z Wang Y Zhang H Luo Y Shi L . 3D compressed convolutional neural network differentiates neuromyelitis optical spectrum disorders from multiple sclerosis using automated white matter Hyperintensities segmentations . Front Physiol . ( 2020 ) 11 : 612928 . doi: 10.3389/fphys.2020.612928 , 33424635 PMC7786373

69. Kim H Lee Y Kim YH Lim YM Lee JS Woo J . Deep learning-based method to differentiate neuromyelitis optica spectrum disorder from multiple sclerosis . Front Neurol . ( 2020 ) 11 : 599042 . doi: 10.3389/fneur.2020.599042 , 33329357 PMC7734316

70. Huang C Chen W Liu B Yu R Chen X Tang F . Transformer-based deep-learning algorithm for discriminating demyelinating diseases of the central nervous system with neuroimaging . Front Immunol . ( 2022 ) 13 : 897959 . doi: 10.3389/fimmu.2022.897959 , 35774780 PMC9238435

71. Seok JM Cho W Chung YH Ju H Kim ST Seong JK . Differentiation between multiple sclerosis and neuromyelitis optica spectrum disorder using a deep learning model . Sci Rep . ( 2023 ) 13 : 13 . doi: 10.1038/s41598-023-38271-x , 37468553 PMC10356911

72. Petrosian D Giedraitiene N Kizlaitiene R Jatuzis D Kaubrys G Vaisvilas M . Assessing the effectiveness of machine learning and deep learning in differentiating neuroimmunological diseases: a systematic review and meta-analysis . Front Neurol . ( 2026 ) 16 : 579206 . doi: 10.3389/fneur.2025.1579206 , 41602983 PMC12832442

73. Demuth S Paris J Faddeenkov I De Sèze J Gourraud PA . Clinical applications of deep learning in neuroinflammatory diseases: a scoping review . Rev Neurol . ( 2025 ) 181 : 135 – 55 . doi: 10.1016/j.neurol.2024.04.004 , 38772806

74. Eshaghi A Wottschel V Cortese R Calabrese M Sahraian MA Thompson AJ . Gray matter MRI differentiates neuromyelitis optica from multiple sclerosis using random forest . Neurology . ( 2016 ) 87 : 2463 – 70 . doi: 10.1212/WNL.0000000000003395 , 27807185 PMC5177679

75. Yan Z Liu H Chen X Zheng Q Zeng C Zheng Y . Quantitative susceptibility mapping-derived Radiomic features in discriminating multiple sclerosis from neuromyelitis optica spectrum disorder . Front Neurosci . ( 2021 ) 15 : 765634 . doi: 10.3389/fnins.2021.765634 , 34924934 PMC8678528

76. Liang X Zeng Q Zhu Y Wang Y He T Wu L . Differentiation between multiple sclerosis and neuromyelitis optic spectrum disorders with multilevel fMRI features: a machine learning analysis . Sci Rep . ( 2025 ) 15 : 1909 . doi: 10.1038/s41598-024-84508-8 , 39809878 PMC11733006

77. Zhang W Cui L Zhang Y Wang W Wang R Liu Z . Questioning the existence of monophasic neuromyelitis optica spectrum disorder by defining a novel long-term relapse-free form from a large Chinese population . J Neurol . ( 2020 ) 267 : 1197 – 205 . doi: 10.1007/s00415-019-09685-3 , 31932912

78. Sherman E Han MH . Acute and chronic management of neuromyelitis optica spectrum disorder . Curr Treat Options Neurol . ( 2015 ) 17 : 48 . doi: 10.1007/s11940-015-0378-x , 26433388 PMC4592697

79. Kessler RA Mealy MA Levy M . Treatment of neuromyelitis optica spectrum disorder: acute, preventive, and symptomatic . Curr Treat Options Neurol . ( 2016 ) 18 : 2 . doi: 10.1007/s11940-015-0387-9 , 26705758 PMC4807395

80. Preziosa P Pagani E Meani A Moiola L Rodegher M Filippi M . Slowly expanding lesions predict 9-year multiple sclerosis disease progression . Neurol Neuroimmunol Neuroinflamm . ( 2022 ) 9 : null . doi: 10.1212/NXI.0000000000001139 , 35105685 PMC8808355

81. Calvi A Tur C Chard D Stutters J Ciccarelli O Cortese R . Slowly expanding lesions relate to persisting black-holes and clinical outcomes in relapse-onset multiple sclerosis . Neuroimage Clin . ( 2022 ) 35 : 103048 . doi: 10.1016/j.nicl.2022.103048 , 35598462 PMC9130104

82. Polvinen E Matilainen M Nylund M Sucksdorff M Airas LM . TSPO-detectable chronic active lesions predict disease progression in multiple sclerosis . Neurol Neuroimmunol Neuroinflamm . ( 2023 ) 10 : e200133 . doi: 10.1212/NXI.0000000000200133 37349108 PMC10291892

83. Solomon JM Paul F Chien C Oh J Rotstein DL . A window into the future? MRI for evaluation of neuromyelitis optica spectrum disorder throughout the disease course . Ther Adv Neurol Disord . ( 2021 ) 14 : 17562864211014389 10.1177/17562864211014389 PMC8111516 34035837

84. Palace J Lin DY Zeng D Majed M Elsone L Hamid S . Outcome prediction models in AQP4-IgG positive neuromyelitis optica spectrum disorders . Brain . ( 2019 ) 142 : 1310 – 23 . doi: 10.1093/brain/awz054 , 30938427 PMC6487334

85. Mealy MA Mossburg SE Kim SH Messina S Borisow N Lopez-Gonzalez R . Long-term disability in neuromyelitis optica spectrum disorder with a history of myelitis is associated with age at onset, delay in diagnosis/preventive treatment, MRI lesion length and presence of symptomatic brain lesions . Mult Scler Relat Disord . ( 2019 ) 28 : 64 – 8 . doi: 10.1016/j.msard.2018.12.011 , 30554040 PMC6397677

86. Wang Y Zhang L Zhang B Dai Y Kang Z Lu C . Comparative clinical characteristics of neuromyelitis optica spectrum disorders with and without medulla oblongata lesions . J Neurol . ( 2014 ) 261 : 954 – 62 . doi: 10.1007/s00415-014-7298-7 , 24609971

87. Zhou Y Zhong X Shu Y Cui C Wang J Wang Y . Clinical course, treatment responses and outcomes in Chinese paediatric neuromyelitis optica spectrum disorder . Mult Scler Relat Disord . ( 2019 ) 28 : 213 – 20 . doi: 10.1016/j.msard.2018.12.038 , 30623860

88. Chien C Oertel FC Siebert N Zimmermann H Asseyer S Kuchling J . Imaging markers of disability in aquaporin-4 immunoglobulin G seropositive neuromyelitis optica: a graph theory study . Brain Commun . ( 2019 ) 1 : fcz026 . doi: 10.1093/braincomms/fcz026 , 32954267 PMC7425339

89. Kim M Choi KS Hyun RC Hwang I Kwon YN Sung JJ . Structural disconnection is associated with disability in the neuromyelitis optica spectrum disorder . Brain Imaging Behav . ( 2023 ) 17 : 664 – 73 . doi: 10.1007/s11682-023-00792-4 , 37676409

90. Cortese R Giorgio A Severa G De Stefano N . MRI prognostic factors in multiple sclerosis, neuromyelitis optica spectrum disorder, and myelin oligodendrocyte antibody disease . Front Neurol . ( 2021 ) 12 : 679881 . doi: 10.3389/fneur.2021.679881 , 34867701 PMC8636325

91. Caverzasi E Papinutto N Cordano C Kirkish G Gundel TJ Zhu A . MWF of the corpus callosum is a robust measure of remyelination: results from the ReBUILD trial . Proc Natl Acad Sci USA . ( 2023 ) 120 : e2217635120 . doi: 10.1073/pnas.2217635120 , 37155847 PMC10193980

92. Hong SG Kim KH Kang Y-R Hyun J-W Kim S-H Kim HJ . Comparing lesion volume dynamics between multiple sclerosis and neuromyelitis optica spectrum disorder during remission using machine-learning segmentation . J Clin Neurol (Seoul, Korea) . ( 2025 ) 21 : 433 – 8 . doi: 10.3988/jcn.2025.0199 , 40878307 PMC12411297

93. Narayana PA Coronado I Sujit SJ Wolinsky JS Lublin FD Gabr RE . Deep learning for predicting enhancing lesions in multiple sclerosis from noncontrast MRI . Radiology . ( 2020 ) 294 : 398 – 404 . doi: 10.1148/radiol.2019191061 , 31845845 PMC6980901

94. Xu Y Ren Y Li X Xu W Wang X Duan Y . Persistently gadolinium-enhancing lesion is a predictor of poor prognosis in NMOSD attack: a clinical trial . Neurotherapeutics . ( 2021 ) 18 : 868 – 77 . doi: 10.1007/s13311-020-00973-9 , 33469828 PMC8423888

Glossary

Neuromyelitis optica spectrum disorder

Myelin oligodendrocyte glycoprotein antibody-associated disease

Magnetic resonance imaging

voxel-based morphometry

resting-state functional MRI

Diffusion tensor imaging

Diffusion-weighted imaging

Surface-based morphometry

Fractional anisotropy

mean diffusivity

Axial diffusivity

Radial diffusivity

Gray matter

Region of interest

Blood oxygen level-dependent

Amplitude of low-frequency fluctuation

Regional homogeneity

Functional connectivity

Multiple sclerosis

Convolutional neural network

Quantitative susceptibility mapping

Connectome-based predictive modeling

Myelin water fraction

Magnetization transfer ratio

Magnetic resonance spectroscopy