Machine learning in neuroradiology: Recent developments and applications

Abstract

Radiology, particularly neuroradiology, has become a major focus of research and industrial investment in artificial intelligence and machine learning (ML). These technologies may help address increasing imaging volumes, workforce shortages, and the need for faster and more consistent interpretation. This article summarizes recent developments in ML applications across neuroradiology. In acute ischemic stroke, ML supports early lesion detection, automated Alberta Stroke Program Early Computed Tomography Score assessment, large-vessel-occlusion detection, infarct core and penumbra estimation, collateral evaluation, workflow prioritization, and outcome prediction. Further applications include cerebral aneurysm detection and prediction of intracerebral hemorrhage expansion and prognosis. In neuro-oncology, current uses include tumor segmentation and classification, molecular-marker prediction, treatment-response assessment, surgical and radiotherapy planning, differentiation of recurrence from pseudoprogression, and prognostication. Additional advances involve image reconstruction, automated quantification, diagnostic classification, and outcome prediction in spine imaging; lesion detection and segmentation in demyelinating disease; and identification and characterization of neurodegenerative disorders. Despite this progress, limited generalizability, insufficient external validation, and poor interpretability remain major barriers to clinical adoption. Explainable artificial intelligence, federated learning, and robust multicenter validation are likely to be central to future clinical implementation.

Key Words: Machine learning; Neuroradiology; Acute ischemic stroke; Neuro-oncology; Explainable artificial intelligence

Core Tip: In this article, we focus on the most recent (post 2023) developments and applications of machine learning (ML) in neuroradiology. The ML models presented here attempt to assist radiologists in stroke imaging, cerebral aneurysm and hemorrhage detection, neuro-oncology, spine imaging, and demyelinating and degenerative diseases. Despite the potential of ML to become an essential player in modern neuroradiology departments, the complexity of ML, especially deep learning models, remains an obstacle to their broad clinical application. The framework of explainable artificial intelligence attempts to address this issue and increase transparency in the model structure and decision-making process.

INTRODUCTION

On the day this introduction was written, 1430 total entries were found in the growing list of artificial intelligence (AI)-medical devices authorized by the Food and Drug Administration for marketing in the United States. Of these, 1094 devices, that is 76.5%, belonged to the radiology panel. The corresponding list of radiology software in the Health AI Register of the European Union contained 308 authorized products, most of which were in the subspecialties of chest radiology (87, 28.2%) and neuroradiology (86, 27.9%). The results of a PubMed database search for “artificial intelligence” and “neuroradiology” showed a steady yearly increase from 160 publications in 2020 to 534 in 2025.

The rapid proliferation of AI-related research in neuroradiology, coupled with the steady rise in commercially available AI-powered medical devices and software solutions tailored for neuroimaging applications, underscores the growing convergence of academia and industry in the advancement of AI-driven innovations in neuroradiology. AI can transform the field of medical imaging by enhancing diagnostic precision, boosting operational efficiency, and improving patient outcomes[1]. Unlike earlier predictions that AI would replace radiologists, AI has emerged as a supportive tool designed to complement radiologists by effectively triaging urgent cases, minimizing diagnostic errors, and ultimately leading to better quality care[1,2].

Contrasting early AI, which required explicit programming for every diagnostic scenario, machine learning (ML) models learn directly from large neuroimaging datasets to identify patterns and adapt according to experience. By autonomously identifying complex patterns, ML continually improves diagnostic accuracy without the need for manual reprogramming[2,3]. Based on the availability of labeled data, ML is broadly categorized into supervised, unsupervised, and semi-supervised learning[2]. Supervised learning relies on expert-annotated datasets to train algorithms to predict predefined outcomes and represents the predominant ML paradigm in radiologic imaging[2]. In contrast, unsupervised learning identifies intrinsic patterns or relationships within unlabeled data without predefined outputs. Semi-supervised learning combines labeled and unlabeled datasets, enabling algorithms to exploit the advantages of both approaches while reducing the dependence on extensive manual annotation[2].

Traditional ML algorithms typically require manual feature extraction, in which experts predefine specific imaging characteristics, such as tumor shape or texture, before the model processes the data. In contrast, deep learning (DL), which is increasingly applied in neuroradiology, utilizes multilayered neural networks to automatically discover intricate patterns directly from raw neuroimaging pixels[3]. This eliminates human-engineered features, enabling DL to achieve unprecedented accuracy in complex diagnostic tasks[3]. However, DL requires advanced computational power, and its models and output procedures are often hidden and inconceivable by healthcare professionals (the so-called “black boxes”), restricting their potential for clinical application[3,4].

This review focuses on some of the most recent advancements in ML in neuroradiology published in the last couple of years. Our review covers some of the most recent advances in ML research on stroke imaging, large vessel obstruction, cerebral aneurysms, intracerebral hemorrhage, neuro-oncology, spine imaging, and demyelinating and neurodegenerative diseases (Table 1). Detection, segmentation, classification, prognostication, and workflow optimization are among the most common tasks performed by ML models. In the last part of this review, concerns regarding generalizability and development, and the rise of explainable or interpretable AI (XAI) are briefly discussed.

Table 1 Summary of the most recent advances in machine learning research on stroke imaging, large vessel obstruction, cerebral aneurysms, intracerebral hemorrhage, neuro-oncology, spine imaging, and demyelinating and neurodegenerative diseases.

Ref.	Title	Field of interest
Tan et al[7], 2025	A machine learning model reveals invisible microscopic variation in acute ischemic stroke (≤ 6 hours) with non-contrast computed tomography	Acute ischemic stroke
Chen et al[9], 2025	Automated estimation of ischemic core volume on noncontrast-enhanced CT via machine learning	Acute ischemic stroke
van Poppel et al[12], 2026	Machine learning models for CT-based classification of ischemic stroke onset time within or beyond 4.5 hours: A comparison of approaches	Acute ischemic stroke
Ekingen et al[13], 2025	StrokeNeXt: An automated stroke classification model using computed tomography and magnetic resonance images	Acute ischemic stroke
Ayobi et al[14], 2025	Deep learning-based ASPECTS algorithm enhances reader performance and reduces interpretation time	Acute ischemic stroke
Peng et al[15], 2025	Using deep learning to shorten the acquisition time of brain MRI in acute ischemic stroke: Synthetic T2W images generated from b0 images	Acute ischemic stroke
Xia et al[16], 2025	Machine learning prediction model for functional prognosis of acute ischemic stroke based on MRI radiomics of white matter hyperintensities	Acute ischemic stroke
Jiang et al[17], 2025	A deep learning system for detecting silent brain infarction and predicting stroke risk	Acute ischemic stroke
Liu et al[18], 2023	Prediction of recurrence of ischemic stroke within 1 year of discharge based on machine learning MRI radiomics	Acute ischemic stroke
de la Rosa et al[98], 2025	DeepISLES: A clinically validated ischemic stroke segmentation model from the ISLES’22 challenge	Acute ischemic stroke
Kim et al[22], 2024	Automated prediction of proximal middle cerebral artery occlusions in noncontrast brain computed tomography	Large vessel obstruction
Rai et al[23], 2025	Artificial intelligence-driven detection of large vessel occlusions on NCCT: A multi-institutional study	Large vessel obstruction
Kim et al[24], 2025	Automated detection of large vessel occlusion using deep learning: A pivotal multicenter study and reader performance study	Large vessel obstruction
Steinmetz et al[25], 2024	Impact of deep learning-enhanced contrast on diagnostic accuracy in stroke CT angiography	Acute ischemic stroke/CT angiography
Kim et al[26], 2023	A deep learning-based automatic collateral assessment in patients with acute ischemic stroke	Acute ischemic stroke/collateral network
Jabal et al[29], 2025	Automated CT angiography collateral scoring in anterior large vessel occlusion stroke: A multireader study	Large vessel obstruction/collateral network
Jeon et al[30], 2025	Clinical feasibility of deep learning-driven magnetic resonance angiography collateral map in acute anterior circulation ischemic stroke	Acute ischemic stroke/collateral network
Agarwal et al[19], 2025	Development and internal validation of multimodal machine learning models for predicting eligibility for mechanical thrombectomy in suspected stroke patients using routinely collected clinical and imaging data	Acute ischemic stroke/mechanical thrombectomy
von Braun et al[31], 2025	Prediction of tissue and clinical thrombectomy outcome in acute ischemic stroke using deep learning	Acute ischemic stroke/thrombectomy response prediction
Zhang et al[32], 2024	A deep learning approach to predict recanalization first-pass effect following mechanical thrombectomy in patients with acute ischemic stroke	Acute ischemic stroke/mechanical thrombectomy
González et al[33], 2025	Multimodal deep learning for predicting unsuccessful recanalization in refractory large vessel occlusion	Acute ischemic stroke/mechanical thrombectomy
Li et al[34], 2025	Development of a diagnostic prediction model for post-stroke cognitive impairment in acute large vessel occlusion stroke using multimodal MRI and PET/CT: A study protocol	Acute ischemic stroke/mechanical thrombectomy
Jeevarajan et al[35], 2025	Can CTA-based machine learning identify patients for whom successful endovascular stroke therapy is insufficient?	Acute ischemic stroke/CT angiography
Lehnen et al[37], 2024	Impact of an AI software on the diagnostic performance and reading time for the detection of cerebral aneurysms on time of flight MR-angiography	Cerebral aneurysms
López-Rueda et al[39], 2024	Enhancing mortality prediction in patients with spontaneous intracerebral hemorrhage: Radiomics and supervised machine learning on non-contrast computed tomography	Intracerebral hemorrhage
Liang et al[40], 2025	Prediction of etiology and prognosis based on hematoma location of spontaneous intracerebral hemorrhage: A multicenter diagnostic study	Intracerebral hemorrhage
Ning et al[41], 2025	Noncontrast CT-based deep learning for predicting intracerebral hemorrhage expansion incorporating growth of intraventricular hemorrhage	Intracerebral hemorrhage
Zeng et al[42], 2025	Clinical, radiological, and radiomics feature-based explainable machine learning models for prediction of neurological deterioration and 90-day outcomes in mild intracerebral hemorrhage	Intracerebral hemorrhage
Wang et al[43], 2024	Predicting postoperative rehemorrhage in hypertensive intracerebral hemorrhage using noncontrast CT radiomics and clinical data with an interpretable machine learning approach	Intracerebral hemorrhage
Elsheikh et al[45], 2024	Machine learning-based pipeline for automated intracerebral hemorrhage and drain detection, quantification, and classification in non-enhanced CT images	Intracerebral hemorrhage
Akbari et al[48], 2025	Machine learning-based prognostic subgrouping of glioblastoma: A multicenter study	Neuro-oncology
Leone et al[51], 2025	Virtual biopsy for the prediction of MGMT promoter methylation in gliomas: A comprehensive review of radiomics and deep learning approaches applied to MRI	Neuro-oncology
Su et al[54], 2026	Machine learning-based preoperative predicting TERT promoter mutation and EGFR gene amplification phenotype in IDH wild-type glioblastoma using advanced MR habitat imaging	Neuro-oncology
Topff et al[55], 2025	A data-centric approach to deep learning for brain metastasis analysis at MRI	Neuro-oncology
Holtkamp et al[56], 2025	AI-guided virtual biopsy: Automated differentiation of cerebral gliomas from other benign and malignant MRI findings using deep learning	Neuro-oncology
Chen et al[57], 2025	A multi-modal deep learning model for prediction of Ki-67 for meningiomas using pretreatment MR images	Neuro-oncology
Cheng et al[59], 2026	Attention-based deep learning network for predicting World Health Organization meningioma grade and Ki-67 expression based on magnetic resonance imaging	Neuro-oncology
von Reppert et al[60], 2025	Image-based search in radiology: Identification of brain tumor subtypes within databases using MRI-based radiomic features	Neuro-oncology
Nakaura et al[61], 2026	Intra-axial primary brain tumor differentiation: Comparing large language models on structured MRI reports vs radiologists on images	Neuro-oncology
Song et al[63], 2024	Clinical efficacy of motion-insensitive imaging technique with deep learning reconstruction to improve image quality in cervical spine MR imaging	Spine imaging
Yasaka et al[64], 2024	Super-resolution deep learning reconstruction cervical spine 15T MRI: Improved interobserver agreement in evaluations of neuroforaminal stenosis compared to conventional deep learning reconstruction	Spine imaging
Ramos et al[65], 2024	Fast and accurate 3D spine MRI segmentation using FastCleverSeg	Spine imaging
Bogdanovic et al[66], 2024	AI-based measurement of lumbar spinal stenosis on MRI: External evaluation of a fully automated model	Spine imaging
Nikpasand et al[67], 2024	Automated magnetic resonance imaging-based grading of the lumbar intervertebral disc and facet joints	Spine imaging
van der Graaf et al[68], 2025	AI-based lumbar central canal stenosis classification on sagittal MR images is comparable to experienced radiologists using axial images	Spine imaging
Zhang et al[69], 2024	Deep learning model for the automated detection and classification of central canal and neural foraminal stenosis upon cervical spine magnetic resonance imaging	Spine imaging
Lee et al[70], 2025	Deep learning model for automated diagnosis of degenerative cervical spondylosis and altered spinal cord signal on MRI	Spine imaging
van den Wittenboer et al[71], 2024	Diagnostic accuracy of an artificial intelligence algorithm vs radiologists for fracture detection on cervical spine CT	Spine imaging
Maki et al[72], 2024	Multimodal deep learning-based radiomics approach for predicting surgical outcomes in patients with cervical ossification of the posterior longitudinal ligament	Spine imaging
Schlaeger et al[73], 2023	AI-based detection of contrast-enhancing MRI lesions in patients with multiple sclerosis	Demyelinating diseases
Greselin et al[74], 2024	Contrast-enhancing lesion segmentation in multiple sclerosis: A deep learning approach validated in a multicentric cohort	Demyelinating diseases
Wiltgen et al[75], 2024	A deep learning ensemble for accurate MS lesion segmentation	Demyelinating diseases
Bouman et al[76], 2023	Multicenter evaluation of AI-generated DIR and PSIR for cortical and juxtacortical multiple sclerosis lesion detection	Demyelinating diseases
Rostami et al[77], 2024	Enhancing classification of active and non-active lesions in multiple sclerosis: Machine learning models and feature selection techniques	Demyelinating diseases
Peters et al[78], 2025	AI-based assessment of longitudinal multiple sclerosis MRI: Strengths and weaknesses in clinical practice	Demyelinating diseases
Huang et al[80], 2024	A joint model for lesion segmentation and classification of MS and NMOSD	Demyelinating diseases
Ding et al[81], 2024	Classification of myelin oligodendrocyte glycoprotein antibody-related disease and its mimicking acute demyelinating syndromes in children using MRI-based radiomics: From lesion to subject	Demyelinating diseases
Zhou et al[82], 2024	MRI based demyelinative diseases classification with U-Net segmentation and convolutional network	Demyelinating diseases
Rudolph et al[83], 2024	Artificial intelligence-based rapid brain volumetry substantially improves differential diagnosis in dementia	Neurodegeneration
Pérez-Millan et al[84], 2024	Probabilistic differential diagnosis of frontotemporal dementia and Alzheimer's disease with MRI and CSF biomarkers	Neurodegeneration
Rogeau et al[85], 2024	A 3D convolutional neural network to classify subjects as Alzheimer’s disease, frontotemporal dementia or healthy controls using brain 18F-FDG PET	Neurodegeneration
Sadeghi et al[86], 2024	Detecting Alzheimer’s disease stages and frontotemporal dementia in time courses of resting-state fMRI data using a machine learning approach	Neurodegeneration
Chen et al[87], 2024	An automated hybrid approach via deep learning and radiomics focused on the midbrain and substantia nigra to detect early-stage Parkinson’s disease	Neurodegeneration
Wang et al[88], 2024	Automatic substantia nigra segmentation with Swin-Unet in susceptibility- and T2W imaging: Application to Parkinson disease diagnosis	Neurodegeneration
Vaillancourt et al[89], 2025	Automated imaging differentiation for Parkinsonism	Neurodegeneration
Suh et al[90], 2025	Deep learning-based algorithm for automatic quantification of Nigrosome-1 and parkinsonism classification using susceptibility map-weighted MRI	Neurodegeneration
Mastoi et al[99], 2025	Explainable AI in medical imaging: An interpretable and collaborative federated learning model for brain tumor classification	Explainable AI

ASPECTS: Alberta stroke programme early computed tomography score; MRI: Magnetic resonance imaging; T2W: T2-weighted; NCCT: Non-contrast computed tomography; PET: Positron emission tomography; CTA: Computed Tomography angiography; AI: Artificial intelligence; MR: Magnetic resonance; MGMT: Methylated-DNA-protein-cysteine methyltransferase; TERT: Telomerase reverse transcriptase; IDH: Isocitrate dehydrogenase; MS: Multiple sclerosis; DIR: Double inversion recovery; PSIR: Phase sensitive inversion recovery; NMOSD: Neuromyelitis optica spectrum disorder; CSF: Cerebrospinal fluid; 18F-FDG PET: 18F-fluoro-deoxy-glucose positron emission tomography; fMRI: Functional magnetic resonance imaging.

LATEST RESEARCH OF ML IN DETECTION, TIMING AND RECURRENCE OF ISCHEMIC STROKE

Key ML applications in ischemic stroke imaging include automated infarct core and penumbra segmentation on computed tomography (CT) and magnetic resonance imaging (MRI), early ischemic change detection, ASPECTS scoring, and prediction of tissue fate[1,2,5,6]. In primary care, where advanced imaging is not available, and in hospitals without expert opinion on stroke imaging, research on non-contrast CT (NCCT)-based ML models shows their potential as triage or diagnostic supportive tools[7-9]. However, reporting quality, heavy computation demands, and interpretability issues of more complex ML models, such as convolutional neural networks (CNNs), have been obstacles to widespread clinical adoption[10,11].

In a systematic review of 16 studies between 2010 and 2025 and a meta-analysis of 13 of them, Gete and Ayele investigated the ability of DL algorithms applied to NCCT to detect acute ischemic stroke (AIS) at the patient level[8]. The pooled sensitivity and specificity were 91% and 90%, respectively. Sensitivity was lower during external validation (82%) than that during internal validation (95%). The Quality Assessment of Diagnostic Accuracy Studies-2 risk of bias assessment showed problems with patient selection and index test domains. The reporting quality was moderate, and explicit external validation was rare. The authors concluded that DL-NCCT has the potential as a tool for triage and decision-supporting, but the gap in generalization hinders its clinical application.

NCCT is the primary imaging examination performed in most hospitals and primary care facilities for stroke evaluation. Tan et al[7] proposed an NCCT-based radiomics model combined with a random forest (RF) model to detect AIS with onset < 6 hours, when density differences between normal and ischemic tissue are often still nuanced. The RF model showed a sensitivity, specificity, and receiver operating characteristic of 77%, 78%, and 0.829 in the validation set and 68%, 81%, and 0.789 in the independent test set, respectively.

Chen et al[9] constructed an automatic segmentation ML algorithm based on Brainomix to estimate the ischemic core on NCCT. Pre-treatment MRI-diffusion weighted imaging (DWI) was used as a reference standard. Core volumes were similar between ML-NCCT and MRI in patients who presented within 1 h of stroke onset, with a significant correlation (r = 0.56, P < 0.001). In patients who presented less than an hour from stroke onset, ML-NCCT struggled to detect early changes in ischemia and underestimated the core volume.

van Poppel et al[12] compared different ML models in their capacity to classify ischemic stroke patients as having onset-to-imaging time (OTI) within or beyond 4,5 hours. Clinical and imaging data, net water uptake levels, and radiomics features were extracted from 993 patients with known OTI from the MR CLEAN and LATE trials. The radiomics-based models outperformed the other models, with an area under the curve (AUC) of 0.81. The texture, shape, and intensity patterns of ischemic areas were stronger in defining OTI than single metrics, such as the net water uptake.

Ekingen et al[13] introduced StrokeNeXt, a lightweight CNN based on a multimodal dataset of 5.336 CT and MRI images, which achieved 93.67% accuracy in stroke detection. When StrokeNeXt was coupled with deep feature engineering, the accuracy increased to 97%, and the classification time was reduced. With a small size of 7.3 million parameters, this CNN model does not require heavy computation, making its integration into the clinical workflow easier.

Could NCCT assessment in stroke be accelerated? Ayobi et al[14] retrospectively studied 200 NCCT scans from five clinical centers with and without the assistance of CINA-ASPECTS (Avicenna.AI), a DL-based algorithm that automatically estimates the ASPECTS score. CINA-ASPECTS reduced the time required for analysis, increased accuracy, and achieved less inter-observer variability than the conventional method did.

Can we gain time during MRI in stroke patients? Peng et al[15] used a DL model, the “Hybrid-Fusion Network”, to generate high-resolution T2-weighted (T2W) images from DWI b0 images. Two radiologists assessed the acquired data. The quality of the brain parenchyma, skull, and middle ear was acceptable or excellent, whereas low-quality images were acquired in the orbit and paranasal sinus.

White matter hyperintensities on MRI-fluid attenuated inversion recovery (FLAIR) images have been linked to a higher risk of stroke and worse prognosis after stroke[16]. Xia et al[16] created a combined model of total white matter hyperintensities radiomics using a support vector machines (SVM) classifier and clinical factors that outperformed the clinical-only and radiomics-only models in predicting AIS prognosis. The sensitivity, specificity, accuracy, and AUC were 89.3%, 84.8%, 86.9%, and 0.939, respectively. A nomogram was devised as a risk assessment tool to assist clinicians in better evaluating AIS prognosis and providing personalized treatment.

Detection of silent brain infarctions (SBIs) in asymptomatic patients with cerebrovascular disease could offer better stroke risk stratification[17]. MRI scanning for general screening is impractical and cost-effective. Jiang et al[17] used 895.640 retinal photographs to pre-train the DeepRETStroke system, a domain-specific foundation model to represent eye-brain connections. Afterwards, 213.762 retinal photographs from an international dataset of six countries on three continents were used to detect SBI and predict AIS. In the internal validation test set, the AUC was 0.901 and 0.769 for predicting stroke and recurrent stroke within five years, respectively. Consistent results were obtained in external validation cohorts. The identification of SBIs without the need for brain imaging could provide an easily accessible, rapid, and cost-effective stroke prevention plan.

A high risk of stroke recurrence may suggest more aggressive management, treatment, and follow-up[18]. A Light Gradient Boosting Machine model based on radiomics features extracted from DWI images and clinical data achieved a sensitivity, specificity, and AUC of 0.85, 0.805, and 0.789, respectively, for predicting recurrence within 1 year[18].

LARGE VESSEL OBSTRUCTION IN AIS

The rapid and accurate detection of large vessel obstruction (LVO) remains a clinical challenge, particularly in resource-limited settings, community hospitals, and pre-hospital environments, where access to advanced neuroimaging and neurointerventional expertise may be limited[19]. LVO accounts for up to 46% of all ischemic stroke cases[20].

The impact of Viz.ai on stroke workflow was studied in a systematic review according to the PRISMA guidelines of 12 studies and 15.595 patients[21]. The pooled analysis showed that Viz.ai reduced computed tomography angiography (CTA)-to-endovascular thrombectomy (EVT) time, CTA-to-recanalization time, and door-to-groin puncture time. However, the analysis did not show statistically significant benefits in patient outcomes, such as intracranial hemorrhage (symptomatic or not), mortality rates, or length of hospital stay.

Kim et al[22] developed the ExtraTrees ML classification model based on NCCT and CTA scans to predict LVO in internal and multicenter external validation datasets. They used four modified Alberta Stroke Program Early CT Score regions (striatocapsular, insula, M1-M3, and M4-M6) to divide the brain and the handcrafted features of asymmetry in density, volume, and Hounsfield units. The clot sign was added using the U-net model. The ML model achieved a robust area under the curve of 0.888 [95% confidence interval (CI): 0.850-0.925], with a sensitivity of 80.1% (95%CI: 72.0-88.1), and a specificity of 88.6% (95%CI: 84.7-92.5). Adding the National Institutes of Health Stroke Scale score to the ExtraTrees model further increased sensitivity to 92.1%.

The ability of a DL model to identify anterior LVO on NCCT was evaluated in a multicenter study involving 612 patients with AIS[23]. “Triage Stroke” (Brainomix 360) detected LVO on NCCT with 67% sensitivity and 93% specificity. The model outperformed radiologists in terms of sensitivity (difference = 20.5%; 95%CI: 8.26-32.78; P = 0.001). Specificity was increased by the addition of the National Institutes of Health Stroke Scale (NIHSS) (99%).

Radiology residents are among the first physicians to read the CTA of an AIS patient. A 2D U-net model increased the sensitivity of LVO detection by four residents by 4% in a multicenter dataset of 595 stroke patients[24]. For isolated middle cerebral artery-M2 occlusions, the sensitivity and specificity were 69% and 96%, respectively.

Low-quality CTA may delay or distort stroke management. A DL model trained with dual-energy CT data was created to enhance the contrast of poorly contrasted CTA images (defined as proximal middle cerebral artery < 350 Hounsfield units)[25]. An iterative reconstruction kernel was used for CTA reconstruction and compared with DL-enhanced CTA. DL-enhanced CTA showed a significantly higher sensitivity for detecting vascular occlusion than conventional CTA (94%; 95%CI: 89.0%-96.8% vs 81.6%; 95%CI: 75.0%-87.1%).

Since time is brain in AIS, the collateral arterial network may be the hour hand of the clock. Kim et al[26] created a supervised DL model to grade collateral status using dynamic susceptibility contrast (DSC) magnetic resonance perfusion images of patients with LVO. The grading model of good collateral probability achieved a C-statistic of 0.91 in the internal dataset and 0.85 in the external dataset, respectively. Good consistency was observed between the DL and expert readers in both the cohorts.

The collateral blood supply volume, which is directly linked to clinical outcomes following EVT, can be quantified using the CTA collateral score[27,28]. Twelve readers with varied experience used an automated e-CTA assistance tool (Brainomix) to grade the collateral network on CTAs from 56 patients with anterior LVO[29]. e-CTA assistance increased the mean accuracy (58.6% to 67.5%, P = 0.003) and mean F1 score (0.574 to 0.676, P = 0.002). The mean reading time was reduced for all readers (103.4 seconds to 59.7 seconds, P = 0.001).

Jeon et al[30] developed the 3D-MROD-net model, a 3D multitask regression and ordinal regression Deep Neural Network, to create DL-driven magnetic resonance angiography (MRA) collateral maps. In a dataset of 296 AIS patients, DL-MRA achieved a shorter time for perfusion grading and better image quality, with excellent inter-method agreement for collateral perfusion.

Agarwal et al[19] developed two supervised Gradient Boosting Machine models to predict eligibility for mechanical thrombectomy in AIS. Model 1 was based on age, NIHSS, last known well time, ASPECTS and infarct volume. Model 2 included the input of present/absent LVO. Model 2 outperformed model 1 in terms of AUC and specificity (0.92 and 82% vs 0.76 and 60%, respectively), whereas the sensitivity was similar (82.5% and 85%, respectively).

A pre-EVT predictive tool that quantifies the individual response to EVT in terms of both tissue and clinical outcomes would assist in decision-making and boost personalized treatment. von Braun et al[31] collected data from 405 anterior LVO stroke patients who underwent thrombectomy. A CNN tissue model predicted a voxel-wise map of the final infarct by integrating multimodal imaging (NCCT, CTA, cerebral blood volume, cerebral blood flow, time-to-maximum), and attention mechanisms. A clinical ML model predicted the NIHSS score at discharge using clinical data and features of the tissue model. The tissue model achieved mean DSC values of 0.45 and 0.49 in the internal and external test cohorts, respectively, indicating good performance. The clinical model had median absolute errors of 1.5 NIHSS points on the internal and 3.0 NIHSS points on the external test datasets, outperforming other ML models.

In patients undergoing endovascular thrombectomy (EVT), successful recanalization is the most critical goal. A high risk of unsuccessful recanalization could engage the family in decision-making and possibly adapt the treatment plan[32]. A hybrid transformer DL model performed well in predicting the first-pass effect using pretreatment MRI (DWI, FLAIR, and apparent diffusion coefficient) or CT (NCCT and CTA) without the need for manual segmentation[32]. First-pass effect is the successful recanalization on the first attempt and is correlated with favorable long-term outcomes. The authors highlight the importance of external validation in a larger cohort as the next step before applying this model in a clinical context.

The EVT outcome after anterior LVO was predicted using a multimodal DL model that combined pre-intervention imaging (NCCT, CTA, and CT perfusion) and clinical data (demographics, comorbidities, and stroke severity)[33]. The target population included not only patients with atherosclerotic disease but also all unsuccessful recanalization cases, as required in real-world clinical practice. Gradient-weighted class activation mapping (Grad-CAM), an XAI tool, was used to weigh the impact of the imaging and clinical factors. The multimodal DL model achieved an AUC of 0.70 ± 0.02 and F1 of 0.39 ± 0.02 in testing, outperforming the model based solely on imaging data.

A post-stroke cognitive impairment (PSCI) prediction model was generated by combining imaging [MRI and positron emission tomography (PET)/CT] with demographic and clinical data[34]. PSCI often presents with subtle symptoms that are masked by motor or sensory dysfunction. Data were obtained from 379 patients with anterior LVO who underwent emergency EVT. Montreal Cognitive Assessment scores at 90 days and 180 days and the modified Rankin Scale (mRS) were set as the primary and secondary outcomes. MRI provides structural information, whereas PET/CT provides metabolic input. A PSCI prediction model could provide an early alert and incentive for personalized treatment. A multicenter study with external validation is necessary for the generalizability of the model.

After EVT, many patients retain a form of disability despite having small final infarct volumes[35]. Jeevarajan et al[35] used pre-treatment CTA to create a DL model to detect this group of patients who could benefit from additional treatment. The included patients had undergone successful EVT with thrombolysis in cerebral infarction 2b-3 reperfusion after LVO with a small final infarct volumes < 30 mL. The DeepsymNet-v3 model predicted the 90-day mRS 0-2. Of these patients, 56.2% had a 90-day mRS 3-6. The DeepsymNet-v3 model model outperformed the logistic regression model with clinical factors (area under the receiver operating characteristic curve 0.81 vs 0.492, P = 0.006).

IDENTIFICATION OF CEREBRAL ANEURYSMS

In the field of aneurysm detection research, DL models have largely replaced supervised AI tools such as SVM and logistic regression[1]. The goal is to integrate imaging findings with clinical and hemodynamic data.

In a systematic review based on 20 studies published until 12/2023 and 20286 aneurysm cases, ML algorithms were evaluated for their performance in predicting aneurysm rupture risk[36]. Despite the high reported performance (accuracy 0.66-0.90) in predicting aneurysm rupture, the authors raised concerns about the studies’ bias risk, limited applicability, and the scarcity of homogenous data available for meta-analysis. They concluded that no superiority over the existing practice was proven, and prospective multicenter studies are needed before ML deployment.

Using any tool that can assist radiologists in accurately and rapidly identifying aneurysms can have a huge impact on the life expectancy of patients by accelerating their medical management. In a single-center study by Lehnen et al[37], 186 ime-of-flight-MRAs were reviewed by six readers with different experience levels, with or without the assistance of the commercially available CNN-based tool “mdbrain”. MdBrain increased specificity for all readers (including medical students) and sensitivity at the lesion and patient levels for physicians. It also reduced the number of false positives for all the readers. They also observed a moderate but statistically significant reduction in reading times for most of the readers.

ML TOOLS FOR ETIOLOGY, PROGNOSIS AND RELAPSE OF INTRACEREBRAL HEMORRHAGE

Spontaneous intracerebral hemorrhage (sICH) accounts for 10%-20% of cerebrovascular disease cases[38]. Its progression and expansion pose serious threats to patient recovery and life expectancy. López-Rueda et al[39] created radiomics-based supervised RF, SVM, and k-nearest neighbors models to detect patients at high mortality risk at admission. The selected radiomics features were associated with hematoma expansion and edema in the surrounding parenchyma of the brain. The sensitivity and AUC were 0.90-0.95 and 0.97-1, respectively. Multicenter external validation is required to generalize these results.

Liang et al[40] employed lesion-symptom mapping to extract the location features of sICH and develop prognostic and etiological brain maps. NCCT radiomics features from 1162 patients with sICH from three medical centers were combined with clinical data using multiple logistic regression. Different lesion-symptom maps were generated for ICH caused by hypertension, aneurysms, and vascular malformations. The classification of etiology based on location-based features was strong, with a mean AUC of 0.825. The probability maps showed an association between the ICH location and prognosis. The combined hematoma and radiomics features showed a mean AUC of 0.873 for prognosis, outperforming the models of clinical data alone.

Ning et al[41] used DL models to identify patients with ICH at a high risk of hematoma expansion and intraventricular hemorrhage. NCCTs from 775 patients with sICH were used to train and test four 2D/3D CNN models, which were then compared with four baseline ML models. The 2D-ResNet-101 model performed best, with an AUC of 0.777 in the external testing cohort, outperforming the baseline clinical radiologic model and BRAIN score. The 2D CNN model also surpassed the combined logistic and SVM models. An early high-risk alert for hematoma expansion and intraventricular hemorrhage could prove life-saving in the management of ICH.

Zeng et al[42] trained five ML models to predict the 90-day prognosis (mRS-90) and neurologic deterioration in patients with mild ICH who were treated conservatively. The models integrated the clinical data, NCCT imaging indicators, and radiomics features. The SVM models showed the highest F1-scores of 0.667 and 0.846 for predicting neurologic deterioration and 90-day prognosis, respectively. The AUC were 0.846 and 0.970, respectively. Shapley additive explanations (SHAP) analysis highlighted the NIHSS score at admission, NCCT island sign, and time from onset to admission as the best SVM predictors. Pinpointing high-risk mild ICH patients would adapt and personalize the follow-up plan.

Re-hemorrhage after surgical management of hypertensive ICH is critical for survival[43,44]. Wang et al[43] tested four ML models based on a combination of NCCT radiomics features and clinical data to predict risk of post-operative re-hemorrhage. The 609 patients were divided into internal and external validation cohorts. The SVM model, which outperformed the other models, was assessed using the SHAP method. Three NCCT radiomics features and two clinical factors (admission SBP and time to surgery) were identified as the most important predictors of postoperative re-hemorrhage in hypertensive ICH patients. The model can assist in decision-making by detecting patients who require closer follow-up and early personalized interventions.

The assessment of ICH drain location is mostly subjective and experience-based[45]. Elsheikh et al[45] developed an end-to-end pipeline to quantify ICH drain coverage and detect malpositioned drains by using a previously published CNN-based segmentation tool. Their model can be used as a CT computer-aided diagnostic tool to alleviate the workload.

LATEST ML RESEARCH IN THE FIELD OF NEURO-ONCOLOGY

In the past few years, the field of neuro-oncology has expanded beyond the classification and segmentation of tumors. ML research currently focuses on molecular marker prediction, treatment response assessment, including differentiation of pseudo-progression, survival prediction models, surgical planning, decentralized federation models for multicenter collaborations respecting data privacy, and large language models (LLMs) for decision-making support[46]. In a scoping review of 229 studies published after 2022, Szmyd et al[47] concluded that DL studies in neuro-oncology primarily focused on CNS tumor differentiation and surgical planning support. Of these, 52% focused on gliomas.

Akbari et al[48] designed a novel personalized Kaplan-Meier model to divide patients with glioblastoma into three prognostic stratification groups with distinct hazard ratios (favorable/intermediate/poor prognosis). A large dataset of non-enhanced and enhanced T1, T2, and FLAIR images was obtained from a large cohort of 2838 patients with glioblastoma from the ReSPOND consortium of 22 centers across seven countries. A unique overall survival map was created for each patient. Computationally, different features with different predictive powers were extracted. Old age, tumor growth inside the ventricles, tumor surface, and level of tumor enhancement were among the strongest predictors of poor overall survival. Patients in the poor-prognosis subgroup could be offered early alternative or additional interventions such as repeat surgery. Notably, the ML subgrouping framework is readily available to clinicians and is freely accessible online.

Methylation of the O6-methylguanine-DNA methyltransferase promoter is a critical epigenetic biomarker for gliomas, associated with better prognosis and lower temozolomide resistance[49,50]. Leone et al[51] analyzed 34 studies on radiomics and DL for methylation of the O6-methylguanine-DNA methyltransferase methylation according to the PRISMA guidelines. Seven studies used ML, mostly XGBoost and SVM models, instead of DL models. Of the 27 studies that reported diagnostic accuracy, 14 reported a diagnostic accuracy of > 80%. In general, the combination of radiomics and DL increases accuracy, sensitivity, and specificity. Heterogeneity in data sources, study populations, and methodologies limits the generalizability of these models.

Isocitrate dehydrogenase (IDH) mutation and 1p/19q co-deletion are genetic markers of gliomas that affect their classification, treatment, and prognosis[52]. In a review following the PRISMA guidelines of 104 studies and meta-analysis of 72 of them, Farahani et al[53] found a sensitivity and specificity of 80% and 85%, respectively, in predicting IDH mutation using MRI DL-based radiomics models. The sensitivity and specificity for predicting 1p/19q co-deletion were 75% and 82%, respectively. CNNs performed tumor segmentation in 43% of the studies, followed by manual and semi-automatic methods (26% and 12%, respectively). The authors observed a shift toward more complex architectures, such as transformers, after 2024. Despite the good performance of DL-based radiomics in predicting the molecular status of gliomas in a non-invasive manner, the clinical application of these models is hindered by the heterogeneity of multicenter MRI data and the segmentation protocols. After overcoming these obstacles, prospective multicenter trials with external validation will be the next step before clinical application.

Telomerase reverse transcriptase (TERT) gene promoter mutation is a molecular hallmark of poor-prognosis IDH-wild-type glioblastomas[52]. Amplification of the epidermal growth factor receptor (EGFR) is a potential prognostic factor[52]. Su et al[54] divided 149 patients with pre-operative conventional MRI, DWI, and DSC perfusion-weighted imaging into training, validation and independent test sets. Using apparent diffusion coefficient and cerebral blood volume maps, the solid tumor component was subdivided into multiple habitat-based subregions using a k-means clustering method. In the training cohort, these habitat features were used to develop RF models for predicting TERT promoter mutation and EGFR amplification. For TERT promoter mutation, the AUC was 0.87, with an accuracy, sensitivity, and specificity of 82.1%, 75.9%, and 76.3%, respectively. For EGFR amplification, the AUC was 0.85, with corresponding values of 80.7%, 70.4%, and 75.9%, respectively.

Topff et al[55] recognized the low performance of AI models in detecting small brain metastases (BMs) and developed a modified nnU-net model to identify and segment BMs of any size. In their retrospective multicenter study, including 1623 patients with BMs and patients with cancer but not BMs, the sensitivity was 98% and 97.4% in the internal and external (n = 260) cohorts, respectively. The sensitivity for BMs with a size of less than 3 mm remained high (93.3%). A mean of 0.6 false-positive results per patient was noted, with 60.3% of the false-positive results being attributed to small blood vessels.

Differentiating gliomas from other intracranial pathologies, such as metastasis, is crucial for patient management. Holtkamp et al[56] studied a diverse group of 1280 patients with gliomas, BMs, inflammatory lesions, hemorrhage, and meningiomas from a single center. Radiomics features were extracted from FLAIR, non-enhanced and enhanced T1 images. The best-performing XGBoost and RF models achieved AUC of 0.96-1 and 0.94, respectively, in differentiating gliomas from other individual intracranial pathologies and from a combined group of pathologies. This ML tool can directly move gliomas to genetic profiling.

Chen et al[57] developed a DL model to predict Ki-67 status in meningioma patients. High Ki-67 levels, a proliferation biomarker, are associated with high recurrence rates and poor prognosis[58]. A total of 1239 patients from three hospitals were retrospectively analyzed in this study[57]. The input for the multimodal DL model that performed best was 3D MRIs, radiological characteristics (e.g., peritumoral edema and intratumoral necrosis), and radiomics. The AUC were 0.797 and 0.808 in the internal and external datasets, respectively. Interestingly, the model can be integrated into radiology workstations and supports the daily workflow. Moreover, the model used radiological characteristics to draw the attention of DL to the relevant MRI regions so that the data drawn would be more informative. The predictive model also estimated tumor growth at 3 years and 5 years, with AUC of 0.756 and 0.727, respectively.

Cheng et al[59] generated fully automated attention-based DL models to predict the World Health Organization meningioma grade and Ki-67 expression levels. The 952 patients from the two medical centers were divided into training, internal, and external validation groups, with the latter being large and consisting of 314 participants. The 2D ResNet50 model achieved the highest AUC in predicting meningioma grade (0.90), and the 2.5D Swin Transformer performed best in predicting Ki67 expression (AUC 0.89). Both models outperformed the clinical and radiomics models. Noninvasive prediction of meningioma grade and Ki-67 expression could assist clinicians in creating an early personalized treatment plan.

ML could also assist the daily workflow of neuroradiologists by expanding the limited information that a text-based search term can provide to the enriched information of an image-based search approach[60]. Von Reppert et al[60] studied retrospectively 295 patients who underwent surgery or radiotherapy. This study aimed to create a model that identifies reference MRIs that are highly similar to the query cases. A CNN model was used for semi-automated tumor segmentation, and radiomics features were extracted from the segmented tumors. A development and held-out test cohort were created. A nearest-neighbor algorithm was applied to retrieve the most similar cases. The retrieval performance was assessed using the mean average precision at k, and the neuroradiologists independently scored the visual similarity. The best-performing dimensionality reduction technique showed a mean average precision at 5 ranging from 78% to 100% by tumor type.

Nakaura et al[61] tested multiple LLMs to provide the top five differential diagnoses based on preoperative structured MRI reports of 137 surgically confirmed intra-axial brain tumors. Generative pre-trained transformer 4 performed best, achieving accuracies of 65.7%, 84.7%, and 90.5% in the top 1, 3, and diagnoses, respectively, compared with the accuracies of 85.4%, 94.9%, and 94.9% by board-certified neuroradiologists. LLMs were effective in diagnosing glioblastoma but struggled with CNS lymphoma. Characteristically, generative pre-trained transformer 4 had 0% top 1 accuracy for CNS lymphoma. The fact that LLMs showed a relatively low accuracy in suggesting the most probable diagnosis but accurately enlisted the five most probable diagnoses renders them differential diagnosis generators rather than definitive diagnostic tools[62]. Moreover, the tendency of LLMs to miss rare diagnoses could lead to anchoring bias by guiding clinicians toward a restricted panel of common pathologies[62]. Aligning AI models with updated neuro-oncology guidelines and databases is necessary to increase their diagnostic accuracy and reduce safety concerns[62].

APPLICATIONS OF ML IN SPINE IMAGING

Recent work in spine imaging suggests that ML, particularly DL, is moving from proof-of-concept to clinically applicable tasks in the acquisition, reconstruction, automated quantification, diagnostic classification, and prognostication. In cervical MRI, DL-based reconstruction has been used to improve the image quality and reduce noise. Song et al[63] showed that combining a motion-insensitive sequence (PROPELLER) with DL reconstruction improved signal-to-noise and contrast-to-noise ratios and reduced noise in cervical spine MRI, supporting a role for ML in protocol optimization and not only post hoc interpretation. Likewise, super-resolution DL reconstruction improved interobserver agreement for grading cervical neuroforaminal stenosis on 1.5-T MRI, showing that AI may improve image quality and reporting consistency[64].

Another major area is automated segmentation and quantitative measurements. Ramos et al[65] developed FastCleverSeg, which is a semi-automatic 3D spine MRI segmentation method for vertebrae, vertebral discs, and paraspinal soft tissues, showing how ML can accelerate extraction of morphologic biomarkers from volumetric spine MRI datasets. Regarding lumbar stenosis, Bogdanovic et al[66] externally validated a fully automated AI model for MRI-based quantification of lumbar spinal canal stenosis and found close agreement with fellowship-trained musculoskeletal radiologists for thecal sac measurements across heterogeneous 1.5T and 3T exams.

DL is also directly applied to common degenerative conditions. Nikpasand et al[67] reported automated MRI-based grading of lumbar intervertebral disc and facet joint degeneration, targeting two of the most frequently used semiquantitative scales in spine imaging. For lumbar central canal stenosis, van der Graaf et al[68] showed that an AI model using only sagittal MR images performed comparably to experienced radiologists who had access to axial images, suggesting a pathway for more efficient triage or preliminary grading workflows. In the cervical spine, Zhang et al[69] developed a DL pipeline for the automated detection and classification of central canal and neural foraminal stenosis on MRI, with performance analogous to subspecialist radiologists but faster. Lee et al[70]. took this a step further by developing a DL model that could automatically identify both degenerative cervical spondylosis and abnormal spinal cord signal on MRI. This is important because it shows that AI systems can provide more than a single diagnosis, which is closer to real radiology reports and the way radiologists interpret and report several relevant findings from the same examination.

In addition to degeneration, trauma is an important application of this technique. In a retrospective diagnostic accuracy study, van den Wittenboer et al[71] found that an AI algorithm for cervical spine fracture detection on CT had lower sensitivity than attending radiologists, although the specificity was similar, and the model found many fractures that radiologists had missed. This serves as an important reminder that current ML tools should be considered adjunct tools for triage, prioritization, or second reading, instead of autonomous systems in emergency imaging.

Finally, ML is beginning to contribute to prognostication. Maki et al[72] used a multimodal DL-based radiomics approach to predict postoperative outcomes in patients with cervical ossification of the posterior longitudinal ligament and their findings suggested that spine imaging biomarkers can be integrated with clinical data for individualized risk stratification. Overall, the most promising near-term uses of AI in spine neuroradiology are image reconstruction, automated measurements, and support for grading degenerative disease, while independent use for trauma assessment or more complex decisions still requires caution.

ML IN DETECTION AND DIFFERENTIAL DIAGNOSIS OF DEMYELINATING DISEASES

Among demyelinating disorders, ML applications in neuroradiology have mostly been established in multiple sclerosis (MS), with the literature being smaller and growing in neuromyelitis optica spectrum disorder (NMOSD) and myelin oligodendrocyte glycoprotein antibody-associated disease (MOGAD). Since 2023, ML studies have mostly focused on: (1) Automated lesion detection/segmentation; (2) Identifying active or contrast-enhancing lesions; (3) Synthesizing diagnostically useful sequences from conventional MRI; and (4) Differential diagnosis between overlapping inflammatory demyelinating syndromes. Despite technical progress, most studies remain retrospective and single-center; therefore, clinical translation still depends on external validation and workflow integration[73].

DL has demonstrated the clearest practical value in MS lesion analysis. In 2023, Schlaeger et al[73] reported that AI-based detection of contrast-enhancing MS lesions on brain MRI came close to human reader performance, although errors persisted for small, infratentorial, and low-contrast lesions, supporting a “second-reader” role instead of a fully autonomous one. In 2024, Greselin et al[74] validated a 3D U-Net for the automatic detection and segmentation of contrast-enhancing lesions using multicenter clinical MRI from the Swiss MS Cohort (a dataset including 372 scans from 280 MS patients), addressing a particularly relevant neuroradiologic task because enhancing lesions are central to diagnosis and treatment monitoring. In 2024, Wiltgen et al[75] developed Lesion Segmentation Tool (LST)-AI, an open-source DL extension of the earlier LST tool for the automated segmentation of MS white matter lesions on MRI. This extension uses a combination of three 3D U-Nets trained on 491 paired T1-weighted and FLAIR scans with expert manual lesion annotations and was designed to better handle the imbalance between white matter lesions and non-lesioned white matter. In addition, they added a lesion location annotation module that classifies lesions according to anatomical category (according to the 2017 McDonald criteria). In testing 103 public cases, LST-AI outperformed existing public methods, including LST, Sequence Adaptive Multimodal SEGmentation, and nnUNet, with excellent performance in the multiple sclerosis lesion segmentation-1 challenge. The authors therefore recommend replacing LST with LST-AI, which they provide as an open-source model. These results illustrate how lesion segmentation is becoming both more accurate and more reproducible[75].

Another important direction is sequence enhancement and synthesis. Bouman et al[76] showed that AI-generated double inversion recovery and phase-sensitive inversion recovery images derived from conventional MRI performed well for cortical and juxtacortical MS lesion detection in a seven-center dataset. Because cortical lesions remain difficult to visualize on routine protocols, this approach can be even more useful for centers where dedicated double inversion recovery/phase-sensitive inversion recovery acquisitions are unavailable, inconsistent, or not acquired due to a heavy clinical workload. Moreover, this type of application is important because it not only classifies disease but also attempts to improve the visibility of diagnostically meaningful lesion patterns on standard clinical examinations[76].

ML has also been applied to disease activity and longitudinal monitoring. In 2024, Rostami et al[77] evaluated 16 ML models and one sequential DL model for distinguishing active from non-active MS plaques on T2W MRI. The sequential DL model achieved the highest AUC (95.6%), implying that in selected settings, non-contrast MRI-based activity assessment could eventually reduce gadolinium reliance[77]. In 2025, Peters et al[78] found that AI-based longitudinal assessment had a very high negative predictive value for new or enlarging MS lesions in routine follow-up MRI; however, the positive predictive value was limited due to false positives, especially across different scanners. Such results are a useful reminder that longitudinal ML tools may be best suited for triage and prioritization instead of final reporting[78].

Regarding the differential diagnosis of demyelinating diseases, recent studies have expanded beyond MS. A 2024 systematic review and meta-analysis found that AI models discriminating MS from NMOSD using MRI features achieved a pooled accuracy of 82%, with substantial heterogeneity in acquisition, features, and validation methods[79]. Individual studies published since 2024 also support this trend. Specifically, Huang et al[80] developed a joint model for lesion segmentation and classification of MS vs NMOSD, reporting 74.87% Dice for segmentation and 92.36% classification accuracy. In pediatrics, Ding et al[81] used multisequence MRI radiomics to distinguish MOGAD from non-MOGAD acute demyelinating syndromes, with subject-level AUCs of approximately 0.84-0.85 on internal and external testing, thus showing the potential of differentiating these two diseases. Similarly, Zhou et al[82] proposed a joint segmentation-classification framework for differentiating pediatric NMOSD from acute disseminated encephalomyelitis, emphasizing how ML may help in clinically unclear inflammatory white-matter disorders.

Generally, the most recent literature (2023-2025) proposes that ML in demyelinating diseases is moving beyond proof-of-concept and toward more practical neuroradiology applications, especially for lesion segmentation, detection of enhancing lesions, longitudinal follow-up, and differential diagnosis. However, there are still limitations, such as modest sample sizes, class imbalance, limited external validation, and scanner/protocol variability among studies. Understandably, the most realistic scenario in the near future would not be the replacement of neuroradiologists but supporting them in daily tasks such as reporting accuracy, efficiency, and standardization in MS, with emerging extensions to NMOSD and MOGAD.

THE EXPANDING FIELD OF ML IN NEURODEGENERATION

ML is increasingly being used in neurodegenerative disorders for disease state identification and to improve our understanding of structural and metabolic imaging of clinically relevant processes. A 2024 study assessed a DL-based fast brain volumetry technique in 55 individuals with Alzheimer’s disease (AD), frontotemporal dementia (FTD), and healthy controls. This study found that AI-assisted volumetry improved the differentiation between AD and FTD and significantly reduced the post-processing time (less than five minutes)[83].

Instead of simply classifying patients into strict diagnostic groups, more recent studies have shifted towards providing decision support that is closer to real clinical practice by combining multiple data sources to estimate the likelihood of different diagnoses. To further illustrate this, Pérez-Millan et al[84] developed an ML architecture that differentiates between AD and FTD using MRI and cerebrospinal fluid biomarkers. In a cohort of 215 patients with AD, 103 with FTD, and 173 controls, MRI demonstrated an 82% accuracy rate in distinguishing AD from FTD. By adding cerebrospinal fluid neurofilament light and 14-3-3, the diagnosis was more certain and increased the number of high-probability cases. This type of probabilistic approach can be very helpful in neurodegenerative diseases, where overlapping atrophy patterns can make binary labels more error-prone[84].

In addition, ML appears promising in molecular and functional neuroimaging. In 2024, Rogeau et al[85] released a 3D CNN using brain 18F-fluorodeoxyglucose-PET images from 199 AD, 192 FTD, and 200 cognitively normal individuals. This demonstrated that DL can categorize AD and FTD and accurately detect normal controls. Sadeghi et al[86] also applied ML to resting-state functional (fMRI) time-course data to determine the stages of AD and FTD and concluded that this multimodal model based on resting-state fMRI and clinical data accurately distinguishes AD with mild cognitive impairment from FTD and healthy controls. These findings support the hypothesis that fMRI can also provide helpful information about a disease besides simple structural atrophy evaluation; however, uniformity is still an important issue regarding clinical translation.

Regarding parkinsonian neurodegeneration, the most dynamic field seems to be the automated examination of the substantia nigra and other midbrain indicators. Chen et al[87] integrated radiomic and DL features using neuromelanin-sensitive/quantitative susceptibility mapping-derived MRI, concentrating on the midbrain and substantia nigra in individuals with early Parkinson’s disease (PD). They found that their hybrid support vector machine model achieved an accuracy of 96.3% on an independent test set and 95.8% on an external validation set, which was better than models using only radiomics, and it highlighted the helpfulness of mixing learned and hand-crafted picture features[87]. Wang et al[88] further demonstrated that DL can automate substantia nigra segmentation on susceptibility-weighted and T2W MRI with their Swin-Unet methodology achieving a Dice coefficient of 0.832 on susceptibility weighted imaging, and subsequent logistic regression yielding a test AUC of 0.894 using automatically segmented susceptibility weighted imaging features, which indicated a performance comparable to manually labeled regions[88]. Instead of depending on subjective visual assessment, the above studies used quantitative methods that can be repeated[87,88].

Moreover, newer studies have shifted their focus from identifying idiopathic PD to distinguishing between various syndromes. The multicenter Automated Imaging Differentiation for Parkinsonism project employed 3T diffusion MRI in combination with support vector ML to differentiate between Parkinsonian disorders. This shows that ML can surpass case-control PD classification by differentiating PD, multiple system atrophy-parkinsonian type, and progressive supranuclear palsy in a way that is more beneficial with regard to therapy[89]. Recent work on parkinsonian neurodegeneration by Suh et al[90] aimed to make nigrosome-1 imaging a more objective and clinically useful biomarker. They applied a DL model to susceptibility map-weighted imaging, which was able to automatically detect and quantify nigral hyperintensity with high diagnostic accuracy for idiopathic PD. In the validation cohort, the automated quantification achieved AUCs of 0.915 and 0.928 for the left and right nigrosome-1, respectively, while the classification of nigrosome-1 abnormality reached AUCs of 0.967 and 0.976. Nigral hyperintensity volume was also lower in patients with more advanced disease, indicating a potential value for severity assessment and patient stratification, and not only for diagnosis[90]. Overall, this study emphasizes how DL can convert a visually assessed MRI feature into a rapid and reproducible imaging marker with possible clinical utility in PD[90].

Considering the above, the most recent literature (from 2024 onwards) suggests that ML applications in neurodegenerative diseases do not simply focus on proof-of-concept but rather explore clinically applicable methods. However, most studies employ selected cohorts, retrospective designs, or carefully curated imaging pipelines, making it essential to validate the results in different scanners, institutions, and individuals with ambiguous diagnoses before confidently applying them routinely.

REACHING OUT TO THE CLINICIAN

The gradual transition from other ML models to DL, especially CNNs, may increase robustness and accuracy, depending on the application domain and data used[91]. However, the complexity and difficulty in interpretability also increase[11]. DL architectures are based on hidden layers and millions of parameters, rendering them incomprehensible to healthcare providers[92]. Clinicians often stand between a patient with a clinical question and a model output without understanding the pathway between them. The application of AI models in clinical decision-making has a direct effect on patients, but there are also safety, ethical, and legal implications in play. The lack of interpretability, the so-called “black box” issue, is among the many reasons for the delayed widespread clinical adoption of AI. Interpretability, often interchanged with explainability (explainable AI, XAI), enables clinicians to trace how a DL model reaches its suggestions for diagnosis, treatment, and prognosis. In addition to the problem of lack of interpretability in ML, Carriero et al[93] highlighted the relativity of the transparency information needed depending on the stakeholder, that is, clinicians, patients, and guideline developers.

The main questions asked are what the logic behind the model is and how it shapes its output[11]. In a systematic review of 133 studies following the PRISMA guidelines between 2014 and 2023, Rodríguez Mallma et al[94] identified 11 ML models and 12 explainability methods in the study of 20 brain diseases. The SHAP was the most common model-agnostic explanation technique used in these studies. SHAP assigns weights to all input features and observes the output of their combinations[94,95]. The local interpretable model-agnostic explanation is another popular explainability tool[92]. Local interpretable model-agnostic explanation, which is typically faster than SHAP, highlights the significant features of the input data that lead to a particular decision[92]. Regarding the creation of the model output, uniform manifold approximation and projection for dimension reduction and Grad-CAM are among the most popular methods for generating transparency[92,96]. Grad-CAM, which is frequently applied in neuroradiology AI studies, assists in interpreting DL models.

CONCLUSION

ML has demonstrated growing potential across the full spectrum of neuroradiology, from the rapid detection of subtle ischemic changes in the brain parenchyma, large vessel occlusions, and intracranial hemorrhage to nuanced characterization of neuro-oncological lesions, neurodegenerative processes, and demyelinating diseases. Multicenter prospective validation studies are essential to confirm that models generalize across diverse patient populations, imaging protocols, and scanner vendors. The contribution of the Ischemic Stroke Lesion Segmentation (ISLES) challenges (https://www.isles-challenge.org/) since 2015 in promoting stroke imaging by boosting collaboration between experts in neurointervention, radiology, and computer science is well-established. The latest ISLES 2024 challenge predicts the post-treatment stroke infarct from pre-interventional acute stroke imaging and clinical data[97]. All ISLES data are freely accessible to the research community. DeepISLES, a robust and freely available segmentation tool for ischemic lesions, was developed by combining the best-performing models of the 2022 ISLES challenge and validated on a large external dataset (n = 1685)[98].

The integration of XAI frameworks is fundamental to building trust, enabling practitioners and patients to understand model outputs rather than accept them as black boxes. Greater emphasis should be placed on federated learning approaches, which allow institutions to collaborate without compromising patient privacy. Federated learning is a decentralized ML method for training AI models by bringing the model to the data, instead of the opposite[99]. In this way, the raw data do not exit the local server, and their privacy is inherently protected. Mastoi et al[99] integrated a pretrained GoogLeNet architecture for classification with an explainable federated learning framework utilizing Grad-CAM and Saliency map visualizations. The proposed model identified gliomas, meningiomas, no tumors, and pituitary tumors with 94% classification accuracy on a dataset of 7.042 MRI images.

Flanders and Geis[5] wrote that “we have only begun to scratch the surface in terms of understanding the optimum fit of imaging AI tools in the current complex medical landscape”. Particularly in neuroradiology, this could be rephrased as “we have only begun to scratch the surface in terms of understanding the optimum fit of imaging AI tools, and the surface keeps expanding”. Considering the vast quantity of ML research on neuroradiology, it is essential to underline that even a flawless AI model does not necessarily guarantee a strong clinical impact[1]. Over- or underestimating the role of an AI model in a clinical context could lead to medical errors by neglecting the expertise of clinicians or discarding valuable information it can provide.