Value of Artificial Intelligence in Neuro-Oncology

Summary

Central nervous system cancers are difficult-to-treat, complex malignancies that remain insufficiently understood and mostly incurable, despite decades of research efforts. Artificial intelligence (AI) is poised to reshape neuro-oncological practice and research, bringing advances in medical image analysis, neuro-molecular-genetic characterisation, biomarker discovery, therapeutic target identification, tailored management strategies, and neurorehabilitation. This Review explores key opportunities and challenges of AI applications along the neuro-oncological care trajectory. We highlight emerging trends in foundation models, biophysical modelling, synthetic data, and drug development, and discuss regulatory, operational and ethical hurdles across data, translation, and implementation gaps. Specifically, we find that near-term clinical translation depends on scaling validated AI solutions for narrowly defined clinical tasks. In contrast, experimental approaches offer broader potential but require technical refinement and resolution of data and regulatory challenges. Addressing both general and neuro-oncology-specific issues is essential to unlock AI’s full potential and ensure its responsible, effective, and needs-based integration into neuro-oncology.

Background

Primary central nervous system (CNS) tumours are the second most common cancers in adolescents and young adults, and the eighth most common cancers in older adults.¹ Secondary CNS tumours, ie, brain metastases, are the most common type of brain tumour, affecting 10–26% of patients who die from their cancer.² The associated healthcare system burden is substantial given high morbidity and mortality rates.^1,2 Overall, CNS neoplasms carry a dismal prognosis, as their unique and complex pathomechanisms challenge the development of effective therapies.

Recent advances in precision diagnostics and assessment,^1,3,4 therapeutic options,^5,6 and translational neuroscience, especially the emerging field of cancer neuroscience,^7,8 have been accompanied and partially enabled by striking innovations in machine learning (ML), perhaps most visibly demonstrated by foundation model-based applications.^9,10 A plethora of AI studies in neuro-oncology showcase its potential in assisting neuro-oncologists along the entire care spectrum and providing scientific insight into the complexities of CNS tumours (figure 3). However, clinical applicability hinges on understanding the specific capabilities and limitations of each method, including where and how it can be used (panel 1). Herein, we critically discuss specific AI applications across the neuro-oncological care trajectory from prevention to rehabilitation, associated challenges and limitations, as well as emerging opportunities yet to pervade translational and clinical neuro-oncology. For in-depth discussions of specific topics, we point the reader towards resources listed in supplementary table 1, appendix p. 3.

Figure 3.

Mapping key machine learning-based technologies along the neuro-oncological care trajectory.

Panel 1. A primer on machine learning for neuro-oncologists and allied health professionals.

Abbreviations. DNN: deep neural network; ML: machine learning

Machine learning (ML) is a subdiscipline of artificial intelligence. To-date, both terms are used quasi-synonymously. ML is the data-driven optimisation of (parametric) functions with respect to pre-defined optimisation objectives. A plethora of ML techniques are used across neuro-oncology, with selected techniques becoming part of the standard armamentarium of a neuro-oncological AI researcher or practitioner.

Clinical imaging-based studies are currently dominated by deep neural networks and radiomics. Deep neural networks (DNNs) are so-called universal function approximators, meaning that they can, in principle, approximate arbitrarily complicated piecewise continuous non-linear mappings, given sufficient parameters, data, and compute resources. This capability is independent of data type, resulting in data- and function-agnosticism, leading to DNNs’ extreme versatility. In practice, optimisation of deep neural networks is limited by various factors aside from resource limitations, and deep learning practitioners have only limited control over exactly which dependencies between which covariates in the data can be learned. However, in most cases, this does not result in obvious errors, ie, DNNs tend to fail “silently”. The fundamentally data-driven nature of the optimisation results in difficulties interpreting the outputs of DNNs and estimating how well they can generalise beyond the data they have been optimised (“trained”) on. Even when an algorithm seems to work on a given validation cohort, there is no theoretical guarantee that will perform comparably on another, underlining the importance of extensive, multi-centre validation studies before introducing AI-based solutions into routine clinical practice. Radiomics is fundamentally different from deep learning, as it relies on the extraction of pre-defined statistical features from images, and subsequent mapping of selected features to a desired target variable (sometimes using DNNs after feature extraction). Pre-defining and selecting features can, but does not ubiquitously, improve performance, potentially warranting a balanced side-by-side evaluation of purely deep learning-based and radiomics feature extraction.

In contrast to imaging-based studies, histo-molecular and multi-omics studies tend to integrate various ML techniques at key steps of extensive data analysis pipelines, which are often equally populated by classical statistical and specialised bioinformatics techniques. As opposed to most ML methods, these techniques can have strong biological priors and may only be applicable to specific data or problem types. ML is often utilised when “classical” statistical approaches face difficulties due to high data dimensionality, the presence of a large number of confounders, non-linear interactions between covariates, and unknown functional relations between covariates of interest. In analyses with thousands of covariates, sample sizes needed to control for spurious correlations with statistical techniques are exceedingly large. Clustering and dimensionality reduction techniques have entered the bioinformatician’s standard repertoire to discover low-dimensional structure in high-dimensional data by finding groups of data points that are close to each other according to a user-defined distance metric (clustering), or by mapping high-dimensional data points to a low-dimensional space, such, that the information retained in the latent encodings captures an aspect of interest that is not visible in the high-dimensional representation (dimensionality reduction). Linking molecular layers by predicting potentially non-linear interactions between covariates is often framed as an ML problem. Again, an ML model can use spurious correlations in the data for predictions, which might require explicit modelling of confounders using biological or structural priors to guide optimisation, which is a highly active field of research.⁷⁵

Epidemiology and prevention

Primary CNS tumours have a global age-standardised incidence rate of 4.63 per 100,000;¹¹ brain metastases of common systemic cancers (lung, breast, prostate, melanoma) occur with an estimated incidence of 9–17%.² Prevention of primary brain tumours is exceptionally difficult given low numbers of actionable oncogenic germline mutations.^1,12 Aside from ionising radiation exposure and rare hereditary syndromes, no other risk factors, including smartphone use, have been conclusively identified.¹ Nonetheless, the growing wealth of molecular data from pan-cancer whole-genome sequencing studies,¹² and concomitant progress in modelling higher-order molecular interactions and inferring molecular traits from non-invasive imaging biomarkers (cf. Early detection and diagnosis) suggest future AI use-cases in predicting clinico-molecular risk factors. For instance, AI-based brain metastasis risk scores derived from bulk RNA-seq data have revealed metastasis-associated epithelial cells in primary lung adenocarcinoma as potential metastatic origins.¹³ Given high brain metastasis incidence rates, similar risk stratification approaches may inform screening and CNS prophylaxis strategies.

Early detection and diagnosis

Imaging-based diagnosis

Magnetic resonance imaging (MRI) is routinely used to detect and assess disease type, location, and extent via macrostructural anatomical information provided by pre- and post-contrast-enhanced T1- and T2-weighted MRI.¹ Structural MRI-trained AI models have been utilised across all standard preprocessing and diagnostic tasks, encompassing denoising, registration, artefact correction, tumour detection, segmentation, classification, and grading (cf. supplementary table 1, appendix p. 3, for reviews).^14–17 AI-assisted diagnosis may reduce difficulties associated with small lesion sizes, heterogeneity, and ambiguous clinical-radiographic presentation, mitigating inter-rater variability and risks from invasive diagnostic procedures.¹⁷ Indeed, multicentric studies recently demonstrated that deep learning-informed CNS tumour detection and classification improves neuroradiologists’ CNS tumour classification accuracy and volumetric assessment.^14,17 Nonetheless, performance is mostly only demonstrated for small, homogeneous cohorts, while extensive prospective, multicentric validation and consecutive translation into clinical practice remain largely unmet needs.¹ Although most applications are developed for well-delineated tasks on a single imaging modality, models simultaneously trained on multiple modalities or for multiple tasks increasingly emerge, with initial results suggesting non-inferiority to unimodal models on selected neuro-oncological tasks.¹⁶

Aside from tumour type, grade, and extent visible on clinical imaging, the molecular-genetic and metabolic profiles of primary CNS tumours, brain metastases, and treatment-related adverse effects profoundly impact management strategy and prognosis,^1,18 as reflected in the revised WHO Classification of Tumours of the Central Nervous System 2021.¹⁸ Hence, AI-based translational inference of molecular characteristics from structural neuroimaging data has garnered considerable interest as a cost-effective, non-invasive alternative to conventional tissue-based diagnosis.^15,19,20 Indeed, radiomics (extraction of pre-defined quantitative medical image descriptors), radiogenomics (radiomics-based prediction of molecular traits), and, more recently, deep learning (panel 1), have shown promise in predicting clinically relevant genetic alterations (eg, isocitrate dehydrogenase (IDH)-mutational status and 1p/19q co-deletion) and epigenetic profiles (eg, O⁶-methylguanine-DNA methyltransferase [MGMT] promoter methylation status) from structural and diffusion-weighted MRI^16,19 and 6-¹⁸F-fluoro-_L-DOPA positron emission tomography.²⁰ Although radiomics/learned features used for translational inference sometimes have morphological/clinical correlates, the data-driven nature of AI necessitates tissue-based corroboration and biological annotation to validate imaging-based studies for improved clinical utility.²¹ This can entail molecular annotation of AI-extracted features via correlation with histological or immunohistochemical traits, eg, immune tumour microenvironment (TME) macrophage infiltration,²² or tumour subtype-specific differentially modulated pathways potentially implicated in pathogenesis.^23,24

Beyond neuroimaging-informed prediction of histo-molecular properties via translational inference, AI may catalyse utilisation of non-standard or non-human-interpretable imaging modalities, eg, ultrasound radio frequency signals for intraoperative molecular glioma diagnosis²⁵ or glioma grade-specific cerebrovascular dysregulation biomarkers extracted from blood-oxygen-level-dependent functional MRI.²⁶

If data are noisy, scarce, or unavailable, generative models can be trained to produce synthetic images, such as super-resolved magnetic resonance spectroscopy for improved metabolic characterization of IDH-mutant gliomas,¹⁵ or postcontrast structural MRI generated de novo from pre-contrast images for CNS tumour grading.²⁷ Yet, a lack of underlying mechanistic/causal models present a fundamental upper bound to the explanatory power of association-based approaches. Instead, biophysical or physics-constrained models of medical image generation processes may not only produce realistic synthetic data, but confer causally grounded explainability to extracted imaging features; corresponding inverse models can then be endowed with physically plausible structural constraints, as recently demonstrated for chemical exchange saturation transfer (CEST)-MRI (table 1).²⁸

Table 1.

A selection of ten high-impact publications that showcase use cases of machine learning in different areas of neuro-oncology. Consensus was reached across all co-authors to select papers for their emphasis on common pathologies, clinical relevance, methodological rigour, novelty, and potential impact on neuro-oncological practice and research.

Study	Objective	Resources	Methods and Results
Hollon et al (2023). Artificial-intelligence-based molecular classification of diffuse gliomas using rapid, label-free optical imaging34	Intra-operative diagnosis via SRH-based prediction of molecular in diffuse gliomas	Endpoints: Molecular classification accuracy, F1 score Data: Whole-slide SRH images of fresh, unprocessed gliomas of 373 adult-type glioma patients, public glioma genetic data of 2777 patient from 6 resources (training); 153 diffuse glioma patients (testing) Study type: Retrospective multicentre study	Image and genetic encoder training: CNN vision encoder pre-trained on multi-label supervised contrastive loss Genetic alteration prediction: Transformer pre-trained on vision and genetic embeddings for prediction of masked genetic alterations Molecular classification: Accuracy 0.947 (IDH), 0.941 (1p19q co-deletion), 0.910 (ATRX); F1 scores 0.963 (IDH), 0.966 (1p19q co-deletion), 0.947 (ATRX). Direct classification of SRH images into 3 WHO CNS5 classes: subgroup classification accuracy overall 0.915; in patient 55 years or younger 0.944
Sanchez-Aguilera et al (2023). Machine learning identifies experimental brain metastasis subtypes based on their influence on neural circuits39	Characterization of brain metastasis subtypes via differential neural activity modulation signatures	Endpoints: Metastasis type prediction accuracy Data: Electrophysiologically recorded neural activity (LFPs from all groups recorded at 7 and 10 days post implantation) and categorical features from 3 organotropic mouse brain cancer cell lines (482N1-BrM, E0771-BrM, B16/F10-BrM) and 1 control Study type: Prospective single-centre study	Effect of brain metastasis on neural activity: Reduced cortical and hippocampal activity ipsilateral to metastasis across all frequency bands; contralateral effects only in 482N1-BrM; activity differences do not reflect locomotor activity or volume conduction effects, LFP differences not explained by mass effect or peritumoral inflammatory milieu; metastasis type and transcriptomic program correlate Decision Tree predicts metastasis type: PCA for projection of features on linear subspace, Decision Tree for metastasis type prediction 10 days after injection with accuracy 0.77 (0.75 – 0.79), and before 9 days of injection with accuracy 0.73 (0.64 – 0.82)
Vollmuth et al (2023). Artificial intelligence (AI)-based decision support improves reproducibility of tumour response assessment in neuro-oncology: An international multi–reader study17	Assessment of AI–based volumetric tumour burden quantification to improve standardisation and reliability of response assessment between evaluators compared to standard RANO criteria	Endpoints: Concordance correlation coefficient for time-to-progression Data: 3D T1w, 3D cT1w, 2D axial FLAIR, 2D axial T2w, DWI with ADC maps from two imaging timepoints from 30 diffuse glioma patients Study type: Retrospective multicentre study	Performance improvement due to AI-based decision-support: CNN for computational skull stripping and tumour segmentation; concordance between raters increased from 0.77 (0.69 – 0.88) (RANO-based) to 0.91 (0.82 – 0.95) (AI-assisted). The effect was more pronounced for lower-grade glioma (CCC 0.70 to 0.90) than for glioblastoma (CCC 0.83 to 0.86). The inter-rater agreement on time-to-progression increased in terms of standard deviation, evaluators with less years of experience rated the AI-based decision support as more helpful
Perlman et al (2022). Quantitative imaging of apoptosis following oncolytic virotherapy by magnetic resonance fingerprinting aided by deep learning28	Utilisation of simulated chemical-exchange-saturation-transfer MR fingerprinting for training a deep neural network to detect and quantify apoptotic response to oncolytic virotherapy	Endpoints: Not applicable Data: CEST MRI 8 – 11 days post tumour implantation, 48 and 72 hours post oncolytic virotherapy treatment in U87ΔEGFR human glioblastoma orthotopic mouse model and 1 healthy human control Study type: Prospective single-centre study	Simulated CEST MRF and model training: Numerical simulation of expected signals for 70 million tissue parameter combinations (T1, T2, B0 inhomogeneities, and 4 semi-solid, amide chemical-exchange parameters) for 2 acquisition protocols. Tissue parameter inference from simulated fingerprints: Two small densely connected deep neural networks were trained to consecutively extract 4 chemical exchange parameters from given simulated MRI. Treatment response monitoring: Quantitative response maps from mouse models were compared at the 3 different measurement time points: all signals significantly altered in tumour after OV inoculation (the first three decreased, only the tumour amide proton exchange rate increased), which indicates lower pH and lower protein concentration, suggesting apoptosis Tissue-based validation of molecular findings: Histological stains (H&E, Coomassie) and immunohistochemistry images (HSV, Caspase-3) were used to verify response maps Validation in healthy human control: After adaptation to clinical 3T MRI scanner, all maps were in agreement with previous studies
Migliozzi et al (2023). Integrative multi-omics networks identify PKCδ and DNA-PK as master kinases of glioblastoma subtypes and guide targeted cancer therapy54	Therapeutic target identification via multi-omic characterisation of glioblastoma; discovery of master kinases orchestrating phenotypic hallmark acquisition in glioblastoma functional niches	Endpoints: Glioblastoma subtype prediction AUC Data: Genetic, epigenetic, transcriptomic, proteogenomic, metabolomic, lipidomic, and phospho-proteomic data from 92 IDH wild-type glioblastoma patients from CPTAC and 282 IDH wild-type glioblastoma patients from TCGA Study type: Retrospective multicentre study	Functional glioblastoma subtype discovery: k-nearest neighbours classifier trained on gene expression profiles uncovers 4 functional glioblastoma subtypes Kinase–substrate phosphosite interactome inference: Ensemble of support vector machine classifiers trained on known substrate–kinase pairs for interactome construction and model-based kinase activity prediction identified PKCδ and DNA-PKcs master kinases in the functional glycolytic/plurimetabolic and proliferative/progenitor glioblastoma subtypes, respectively. BJE6–106 (PKCδ inhibitor) and nedisertib (DNA activated protein kinase inhibitor) were suggested as potential targeted agents Probabilistic IDH-wildtype glioblastoma prediction: Multinomial regression model with lasso penalty trained on RNA-seq to predict functional glioblastoma subtype, AUC 0.83
Wang et al (2021). Machine learning revealed stemness features and a novel stemness-based classification with appealing implications in discriminating the prognosis, immunotherapy and temozolomide responses of 906 glioblastoma patients51	Introduction and clinical validation of a novel stemness-based classification as prognostic predictors for glioblastoma; utilisation of multi-omic analysis to reveal targetable pathways; patient stratification to find glioblastoma cohort potentially responsive to ICIs	Endpoints: Stemness-subtype predictor accuracy, sensitivity, specificity, AUC Data: Predicted mRNA stemness index-associated differentially expressed genes for 2 stemness subtypes that were found via clustering and annotation (literature) and clinical patient status as label from 868 glioblastoma patients from 2 databases (training and validation); 388 glioblastoma patients from 1 database and 1 institute (testing) Study type: Retrospective multicentre study	mRNA stemness index (mRNAsi) annotations: mRNAsi negatively correlated with immune infiltration levels, uncorrelated with tumour purity, positively correlated with T cell subsets, negative with macrophages, high mRNAsi better OS (HR 0.80) poorer PFS (HR 1.28), mRNAsi independent of other clinical variables; differential expression analysis reveals 130 DEGs Stemness-based subtyping and annotation: k-means-based consensus clustering revealed two distinct subtypes with different tumour immune microenvironments, immunogenomic patterns (subtype 1 has higher CNA and mutation burden, more sensitive to immunotherapy, but resistant to TMZ), and clinical variables (eg, subtype 1 younger patients) Identification of subtype-specific compounds: DEGs between subtypes and pathway annotation revealed 34 and 76 potential compounds for types 1 and 2, respectively ML-based stemness predictor: (1) Train 4 different classifiers on stemness-associated DEGs to predict status of stemness subtype; (2) find intersecting highly predictive genes; (3) train final multivariate logistic regression classifier on these genes to predict status, resulting in AUC 0.96, accuracy 0.93, sensitivity 0.91, specificity 0.94
Wang et al (2023). Neuropathologist-level integrated classification of adult-type diffuse gliomas using deep learning from whole-slide pathological images31	Digital pathology-based molecular glioma diagnosis from annotation-free whole-slide images	Endpoints: Diagnostic AUC Data: H&E WSIs and diagnostic class label of 1702 diffuse gliomas from 1 institute, and 922 diffuse glioma patients from 3 institutes Study type: Retrospective multicentre study	Pre-training and phenotype-based clustering: CNN pre-training on patch-level labels of 644,896 patches; subsequent k-means clustering of output feature vectors (k = 9). Training of k separate CNN classifiers on patches from each cluster separately for prediction of 6 diagnostic classes. Patient-level classifier training: Three single-cluster classifiers were better than all-cluster baseline; training of final patient-level classifier on all patches from three clusters for patient-level classifier training (275,741 patches), and aggregation of final patient-level result; classifier achieved AUCs of 0.932 – 0.994 on internal testing cohort, 0.923 – 0.987 on external testing cohort 1, 0.904 – 0.952 on external testing cohort 2 Interpretation of class activation maps: Regions of interest computed via class activation maps were evaluated by pathologist, well-aligned with pathological morphology (eg, highlighting necrosis and microvascular proliferation)
Zheng et al (2023). Spatial cellular architecture predicts prognosis in glioblastoma33	Utilization of spatial transcriptomics and deep learning to map WSIs to glioblastoma transcriptional subtypes; linking transcriptional subtype to clinical outcome; uncovering OS-related regional gene expression programs	Endpoints: Transcriptomic subtype prediction AUC Data: 75,625 transcriptomic spots and 69,647 H&E WSI patches from 22 glioblastoma patients from 3 public sources, and additional 8 and 312 glioblastoma patients from 2 public sources (training and validation); 98 glioblastoma patients from 1 public source (testing) Study type: Retrospective multicentre study	Identification of spatially resolved transcriptional subtypes: NMF to find five meta-gene-modules associated with known transcriptional glioblastoma subtypes. Subsequent training of CNN on H&E WSI patches to predict dominant transcriptional tumour cell type and immune cell type per patch, achieving AUC 0.93 (tumour), 0.80 (T cells), 0.84 (macrophages) Prognosis prediction from histology images and biological annotation: CNN training on H&E patches to predict a composite score of C-index and brier scores (aggressiveness), which significantly differs between transcriptional subtypes (mesenchymal-hypoxia type most aggressive, oligodendrocyte-progenitor type least aggressive). Gene set analysis on transcriptomics dataset reveals differentially modulated genes related to high aggressiveness (eg, injury response) and lower aggressiveness (eg, neuronal development)
Kondepudi et al (2024). Foundation models for fast, label-free detection of glioma infiltration29	Fast and accurate detection of glioma infiltration in fresh, unprocessed surgical tissue using a Vision Foundation Model	Endpoints: Degree of infiltration (mean AUC, MAE) Data: 11,462 whole-slide SRH images of fresh, unprocessed surgical specimens of 2,799 adult-type glioma patients (pre-training); 3,560 whole-slide SRH images (896 patients) (testing); 767 IDH-mutant and 659 IDH-wild-type diffuse gliomas (1,130 total specimens) (prospective clinical trial for testing) Study type: Prospective international multicentre single-arm, non-inferiority, diagnostic clinical trial	Vision FM training: Patch feature extraction with hierarchical self-supervised learning, Vision Transformer pre-training with contrastive loss. Subsequent fine-tuning with ordinal representation learning on 100 times downscaled dataset, expert-labelled for tumour infiltration degree Prospective trial results: Test mean AUC between 0.922 and 0.886 in three medical centres (US and Europe) Simulated interventional trial for FM as surgical adjunct FM-based approach (0.981) outperforms standard-of-care intraoperative surgical adjuncts FLAIR (0.763), 5-ALA fluorescence (0.890)
Drexler et al (2024). A prognostic neural epigenetic signature in high-grade glioma40	Definition of epigenetically defined neural signature from transcriptomic profile; linking signature to OS and PFS; characterisation and potential biomarker identification	Endpoints: OS, PFS, functional connectivity in glioblastoma; stability of epigenetic neural classification Data: Multi-cohort, multi-omics profiles (DNA methylation, transcriptomics, proteomics, spatial transcriptomics) of 5,047 CNS tumours (including 1,058 glioblastoma patients from combined cohorts and 187 from TCGA), additional single-cell and patient-derived xenograft datasets, and a validation set of 72 diffuse midline gliomas Study type: Retrospective multicentre study	Epigenetic neural subgrouping and prediction: Stratification of IDH-wild-type glioblastomas into high- and low-neural groups derived from DNA methylation; high-neural tumours exhibit synaptic, stem-like states and worse survival outcomes; training of graph neural network on spatial transcriptomic microenvironments to predict neural signature (F1-score 0.98); weighted correlation network analysis correlates gene expression modules with neural signature Functional connectivity and invasion: High-neural tumours integrated into neuron-to-glioma networks, displayed greater peritumoral connectivity in MEG/fMRI, and showed higher proliferation/migration in co-culture and xenograft models Spatiotemporal stability and surgical benefit: The neural signature remains stable across spatially distinct biopsies and upon recurrence; high-neural glioblastomas benefited less from near-complete resection compared to low-neural; elevated serum BDNF correlated with the high-neural subgroup and higher seizure incidence; high-neural signature in H3 K27-altered diffuse midline gliomas associated with worse outcomes

Abbreviations. ADC: Apparent diffusion coefficient; AI: Artificial intelligence; AUC: Area under the receiver-operator characteristic curve; BDNF: Brain-derived neurotrophic factor; CCC: Concordance correlation coefficient; CNA: Copy number alteration; CNN: Convolutional neural network; CNS: Central nervous system; CPTAC: Clinical Proteomic Tumor Analysis Consortium; DEG: Differentially expressed gene; DWI: Diffusion-weighted imaging; FLAIR: Fluid-attenuated inversion recovery; GAN: Generative adversarial network; GLM: Generalised linear model; H&E: Haematoxylin and eosin; HSV: Herpes simplex virus; LFP: Local field potential; ICI: Immune checkpoint inhibitor; IDH: Isocitrate dehydrogenase; MAE: Mean absolute error; ML: Machine learning; MRI: Magnetic resonance imaging; NMF: Non-negative matrix factorization; OS: Overall survival; OV: Oncolytic virotherapy; PCA: Principal component analysis; RANO: Response assessment in neuro-oncology; SRH: Stimulated Raman histology; SVM: Support vector machine; TCGA: The Cancer Genome Atlas; TME: Tumour microenvironment; TTP: Time to progression; WHO: World Health Organisation; WSI: Whole-slide image

Digital neuropathology

Digital neuropathology aims to semi-automate manual histological or immunohistochemical assessment, including precision diagnosis, molecular characterisation, and neuropathological workflow optimisation. Deep learning-based digital pathology trained on whole-slide images (WSIs) has demonstrated utility across various standard diagnostic tasks including tumour localisation, segmentation, grading, and molecular classification,^29–32 with AUCs often exceeding 0.9 (table 1). As neuropathological assessment cannot per se exhaustively inform molecular diagnosis, translational inference can enable spatially resolved profiling via inference of histo-molecular tissue properties from standard WSI, such as patch-wise prediction of glioblastoma transcriptional subtypes, revealing survival-related regional gene expression programs.³³

Besides biopsy-based precision diagnosis, rapid molecular diagnosis is paramount for guiding the resective risk-benefit-strategy. Here, stimulated Raman histology (SRH)^29,34 (table 1) or nanopore methylation sequencing data³⁵ enable deep learning-based rapid molecular classification, assisting neurosurgeons with intraoperative tumour margin assessment and epigenetic subtyping.

Neuropathological evaluation is often complicated by ambiguous presentation of treatment-related adverse effects, eg, pseudoprogression and tissue necrosis. While AI-assisted tools may augment diagnostic certainty in ambivalent cases, lack of consensus definitions and reference standards, sampling bias, and frequent presence of “mixed lesions” (containing both tumour cells/foci and treatment-related pathology) remain principal barriers.³⁶

Despite significant potential for automating arduous histopathological tasks or enhancing diagnostic consistency and granularity, neuropathological datasets are typically too large for cost-effective digitisation, storage, and manual labelling; even if digitised and annotated, general applicability remains limited by differences in acquisition protocols across institutions and high inter-rater variability. Hence, neuropathology may broadly benefit from the development of foundation models (FMs) – deep neural networks pre-trained task-agnostically on vast datasets in a weakly or self-supervised manner and consecutively adaptable to specific downstream tasks (panel 2). Although not yet validated in the context of auto-labelling pipelines, a recent prospective multicentric clinical trial demonstrated superiority of an FM trained on annotation-free SRH images compared to standard-of-care surgical adjuncts in identifying glioma infiltration during surgery.²⁹ Despite encouraging preliminary results, pervasive FM adoption is decelerated by prohibitive resource requirements and ethical/regulatory hurdles (cf. Opportunities, Pitfalls, and Future Directions).

Panel 2. Generalist foundation models in neuro-oncology.

Abbreviations. AI: artificial intelligence; FM: Foundation model

Foundation models (FMs) are deep neural networks that undergo weakly or self-supervised pre-training on large-scale, possibly multi-modal, datasets (containing up to trillions of samples); they can flexibly adapt to a diverse set of downstream tasks, either by supervised training on much smaller annotated datasets (“fine-tuning”), or novel training schemes, such as “instruction fine-tuning” via natural language instructions. After (instruction) fine-tuning, they can adapt to novel tasks from only examples or instructions (“in-context learning”).^9,10 This suggests a future role of FM-based medical AI systems with truly generalist capabilities, assisting clinicals across varied clinically relevant tasks, ranging from medical question answering, image classification, radiology report generation and summarisation, and genomic variant calling. Such generalist FM-based assistants could, eg, suggest a diagnosis, integrating molecular data into its considerations, give recommendations informed by the most recent literature and clinical trial data, assist in the operating room by delineating tumour boundaries, recommend treatments, and give a prognosis based on the data gathered at all prior steps, enabling human-AI collaboration.^9,10 Beyond their role as conversational assistants, it is thought that neuro-oncological FMs can integrate non-human interpretable data, such as molecular or neural activity data directly into their training without encoding the data as natural language first, augmenting human capabilities rather than simply acting as assistant technology. FMs are yet to be validated in real-world clinical contexts, facing challenges with reliability, alignment with human intent and values, unquantifiable extrapolation capabilities, privacy concerns due to potential reproduction of sensitive patient information, and their requirement for extremely large training datasets paired with vast compute infrastructure.

Molecular-genetic tumour characterisation

Most CNS neoplasms are characterised by high histo-molecular inter- and intra-tumoral heterogeneity. Analysis of molecular data is fraught with difficulties including high data dimensionality, presence of confounders, and non-linear higher-order interactions between molecular covariates,³⁷ implicating certain AI techniques as viable alternatives to more classical statistical approaches, whenever appropriate. Here, AI’s data-type agnosticism (panel 1) enables prediction of molecular modalities, for instance, predicting transcription factor binding affinity for identification of modulatory single nucleotide polymorphisms³⁷ or inferring major genetic alterations (IDH-, TERT promoter-, and ATRX mutation status, 1p19q co-deletion status) from global DNA methylation levels in gliomas, with prediction accuracies exceeding 0.9.³⁸

Beyond molecular alterations, concomitant systemic effects of CNS tumours on the nervous system (local tissue innervation changes and modulatory effects on tumour phenotype, facilitating or enabling hallmark capability acquisition)⁸ suggest that certain tumour subtype characteristics manifest in differential connectivity and electrophysiological profiles (both ipsi- and contralesionally).^39–41 Indeed, experimental brain metastasis models exhibit atypical calcium-dependent activity in the TME and differentially expressed genetic profiles, suggesting a delicate interplay of molecular programs and neural activity patterns (table 1).³⁹ These insights showcase potential for AI-based “activity/connectivity fingerprinting” in identifying novel biomarkers and associated precision treatments. Indeed, integration of electrophysiological and connectomic characteristics of tumours and affected neural structures derived from intraoperative electrocorticography or magnetoelectrography – shown to be prognostic in glioblastoma^40,42 – may inform future neural activity-based risk stratification.

Neuro-oncological treatment

Treatment administration

Resection followed by radiotherapy and chemotherapy is recommended for the therapeutic management of most malignant as well as several benign central nervous system tumors.³ Here, all approaches fundamentally rely on imaging and digital data processing to delineate the therapeutic target (table 3). This foremost includes radiation treatment planning and delivery.⁴³ Over recent years, AI-based segmentation models have been implemented in routine clinical practice to delineate certain normal tissue structures for avoidance. Ongoing work aims to apply AI for delineation of tumour target structures, including gliomas and brain metastases.⁴⁴ In morphologically simple instances without significant anatomic distortion or alteration from prior treatments, these approaches may improve workflow efficiency, potentially allowing for greater consistency and quality in routine practice or clinical trials. However, standardisation of AI algorithm development and imaging input in defining all biologically relevant regions for treatment, especially for patients with infiltrative malignant gliomas, remains an unmet need.⁴⁵ This also has implications for surgical approaches, as removal of the tumour portion beyond the contrast enhancing core has been associated with increased survival in glioblastoma patients.⁴⁶ However, distinguishing non-enhancing tumour from oedema with only scattered tumour cells remains a salient discussion point.⁴ Advanced imaging with AI-assisted analysis may eventually guide resection extent by predicting tumour extent beyond contrast enhancement on pre-operative imaging.

Table 3.

Selected applications and associated key considerations for medical AI in neuro-oncology stratified by regulatory stage. Listed applications and publications were selected for didactic purposes, emphasising clinical relevance, novelty, proof of principle, and potential impact on neuro-oncological practice and research. Applications with 510(k) clearance were identified by searching the FDA AI/ML-Enabled Medical Devices List for relevant FDA-approved AI-based medical devices (up until January 2025).⁸⁶

Regulatory stage	Application / AI type (selected examples)	Key considerations/barriers to clinical implementation
Available with FDA 510(k) or equivalent regulatory clearance	Automated visualization, segmentation, registration, volumetric quantification, and labelling of brain structures from MR images for diagnosis and radiotherapy planning (TeraRecon Neuro, TeraRecon, Inc., 2022; ClearPoint Meastro Brain Model, ClearPoint Neuro Inc., 2022; NS-HGlio, Neosoma Inc., 2022; VBrain, Vysioneer, Inc., 2021; NeuroQuant, CorTechs Labs, Inc., 2017)) Noise reduction in MR images (SubtleMR, Subtle Medical Inc., 2023) Generation, visualization, and evaluation of pseudo-CT from MR images (ART-Plan, TheraPanacea, 2023)	Disconnect between research and approval of medical devices/utilisation in routine clinical practice: data, validation, and implementation gaps Data gap: lack of high-quality, annotated, centralised, anonymised, diverse, representative data for model development Validation gap: lack of international, multicentric, prospective trials to demonstrate clinical viability of novel methods, accounting for potential transfer penalty or performance deterioration over time Implementation gap: disconnect between approval of medical device and active clinical use; successful reimbursement/financial viability often not guaranteed/achieved despite approval; demonstration of improved outcomes critical to success beyond formal approval
Clear pathway to regulatory approval and commercialisation	Identification, location, characterisation of suspicious area(s) on digitised WSIs (Paige Prostate, Paige.AI Inc., 2020): approved in context of prostate tumour detection, feasibly transferable to neuro-oncological use-cases FM-based application fine-tuned to specific use-cases, eg, based on Health Developer Foundation74	Difficult/unclear recertification for improved models or different application of existing model Licensing issues in the context of open-sourced models developed in research contexts (eg FMs); patentability of novel models developed on top of open-source model/method
Unclear pathway to regulatory approval or commercialisation	Generalist, promptable models, zero-shot generalisation in novel contexts9,29 Generative models for synthetic data generation28	Need for novel regulatory frameworks for novel technologies, eg FMs or biophysical models (ensure safety, equity, and bias in models with strong zero-shot generalisation abilities or trained to learn in-context) Recertification infeasible for each new use-case Need for standardised protocols, prospective validation, and a shift toward federated learning to enable large, decentralised collaborations across institutions

Abbreviations. AI: Artificial intelligence; CT: Computed tomography; EU: European Union; FDA: Food and Drug Administration; FM: Foundation model; MR: Magnetic resonance; WSI: Whole-slide image

AI-assisted imaging guidance is already used in the operation room, with feedback systems analysing the operative video in real-time.⁴⁷ Such analyses may eventually provide haptic or acoustic feedback about tissue areas with a higher likelihood of harbouring tumour, interpret signals from outside the immediate surgical field (like fluorescent signals from tumour remnants or electrophysiological monitoring), or yield early warnings when unnecessarily risky steps are taken. Based upon label-free optical imaging methods or methylation profiles, deep neural networks have also been used to intraoperatively diagnose molecular glioma subtypes, further guiding resective risk-benefit-strategy (table 1).^34,35 Optimal imaging methods preserving the cytoarchitecture can detect tumour cells in morphologically unaffected appearing brain parenchyma in real-time, allowing tailored resection according to the intraoperative (tissue-based) delineation of the expansive tumour. Further reducing the margins for human errors in a stressful clinical environment, robotic systems performing stereotactic sampling and even image-guided microsurgery or placement of spinal hardware were shown non-inferior to conventional techniques.^48,49 AI’s role in radiation treatment planning has also been investigated for stereotactic radiosurgery and conventionally fractionated external beam radiation therapy. To date, these limited studies report variable feasibility and clinical acceptability. Incorporation of clinical outcomes including tumour control, toxicity, and functional neurologic status should be emphasised in prospective studies of AI implementation.

Systemic treatment and therapeutic target identification

While delaying tumour progression is possible, curative treatment options remain elusive for high-grade adult-type diffuse gliomas; glioblastoma and H3-K27M altered diffuse midline gliomas (DMGs) are largely unresponsive to chemotherapy. Despite many trials, no novel agents have been approved in the past decade, and the utility of targeted agents remains limited to small patient subsets.¹

Since few genetic alterations with clinically validated therapeutic implications are known,^18,50 identification of patient strata responsive to certain therapeutic approaches may help improve patient management towards tailored therapy recommendations. Disappointing results in glioblastoma treatment with immune checkpoint inhibitors (ICIs) have prompted attempts to select ICI-responsive patient strata based on glioblastoma stemness, using AI-based stemness predictors from RNA-seq profiles (table 1). It was shown that high-stemness tumours correlate with programmed death-ligand 1 (PD-L1) inhibitor susceptibility and higher overall survival (OS), genomic instability, and differential immune TME profiles.⁵¹ AI-based prediction of single-sample gene enrichment analysis scores of selected signatures along histomorphologic glioblastoma niches and subsequent clustering identified a proteomic MYC-KRAS-hypoxia program promoting intra-tumoral heterogeneity. The axes correlate with aggressiveness, differential drug sensitivities, and relative chemoresistance.⁵² Interestingly, pharmacological drug sensitivities were explained better by these protein-defined axes than by established transcriptional subtypes, implicating immune checkpoint protein CD276 (vasculature) as a potential ICI target rather than PD-L1.⁵² Cross-type proteogenomic studies comparing low-grade glioma and glioblastoma identified IDH1 and EGFR to be mutationally exclusive and highlighted tumour-type specific kinase preference, signalling potential targets in the differentially enriched pathways.⁵³

Despite vast numbers of suggested druggable targets, their respective roles in contributing to phenotypic tumour variations remain unclear. Identification of nexus molecules orchestrating these phenotypic variations is pivotal.⁵⁴ Recent studies have demonstrated AI’s utility to infer kinase–substrate phosphosite interactomes⁵⁴ in glioblastoma, yielding identification of highly active master kinases⁵⁴ in the functional glycolytic/plurimetabolic and proliferative/progenitor glioblastoma subtypes, respectively (table 1).⁵⁴ To inform patient allocation for prospective clinical trials, functional glioblastoma subtypes were predicted using RNA-seq data from frozen tissue samples (AUC above 0.83 for all classes).⁵⁴ Target identification in combination with AI-driven de novo compound design could profoundly speed up development of novel therapies, but as the first artificially designed drugs are only now entering clinical trials, their effectiveness remains to be seen (supplementary table 1, appendix p. 3).

A plethora of other therapeutic avenues is on the horizon,⁵ including, but not limited to, novel immunotherapies, chimeric antigen receptor (CAR) T-cells, oncolytic virotherapy, multimodal combination therapies, optimised therapy sequencing, and drugs targeting “neural-cancer cross-talk” and nervous system-cancer networks.⁷ Clinical trials on tumour network-disconnection strategies targeting gap junctions in glioblastoma and inhibition of glutamatergic neuron-to-glioma synapses are ongoing,⁷ warranting investigations into AI-assisted identification of structural and molecular drivers behind nervous system-cancer network function, novel biological agents, and the utility of electrophysiological/neural signalling biomarkers.

Prognostication, response prediction and assessment

Molecular-genetic risk stratification and prognosis

Besides overall clinical prognostic factors like age and clinical status, tumour-specific risk strata are primarily determined by histo-molecular characteristics (eg, MGMT promoter methylation status, H3-K27M mutation in DMGs). Translational inference from structural MRI and diffusion tensor imaging can improve outcome predictions, such as OS and progression-free survival (PFS), and has been linked to risk score-specific differentially enriched pathways and histo-molecular signatures.^23,24,40

However, current stratification schemes often fail to capture the significant morphometric, epigenetic, and transcriptomic tumoral intra-group heterogeneity, warranting efforts to refine established imaging-derived or molecular-genetic strata.^{33,40,55–59} As an example, AI-derived cytoarchitectonic features from WSIs in lower-grade gliomas correlate with OS, revealing poor-prognosis subtypes marked by high mutational load, frequent copy-number alterations, and considerable tumour-infiltrating lymphocyte presence.⁵⁶ In some tumour entities, such as high-grade astrocytoma with piloid features (HGAP), morphometric subclasses alone offer limited utility because intra-group differences evade standard assessment. However, hierarchical clustering of epigenetic profiles in HGAP reveals clinically salient subtypes, such as a neurofibromatosis type 1-mutated hypermethylated HGAP subtype with reduced tumour purity and poor PFS.⁵⁷ Beyond HGAP, clustering and dimensionality reduction approaches (panel 1) delineate distinct epigenetic subclasses in DMGs and glioblastomas,^57–59 characterised by atypical histo-molecular features and differential clinical outcomes. For instance, H3.3-K27M-mutated DMGs tend to have poor OS and are unresponsive to chemotherapy,¹ whereas H3.3-K27M tumours with BRAF or FGFR1 co-mutations form a distinct epigenetic-transcriptomic cluster with divergent histology and slightly favourable prognosis.⁵⁹ Glioblastoma further demonstrates how transcriptional divergence can complicate prognosis: the transcriptomic high-neural subtype exhibits hypomethylated CpG sites and upregulation of genes associated with synaptic integration, fostering neuron-to-glioma synapse formation negatively associated with OS and PFS (table 1).⁴⁰ This transcriptional heterogeneity is a key driver of therapy resistance in glioblastoma, as the co-existence of transcriptional niches hinders effective targeting of subtype-specific molecular alterations.⁶⁰ Mapping these regional gene expression programs to tumour morphology, connectivity, and signalling in turn opens the door for AI-driven identification of novel biomarkers in magnetoencephalography, digital pathology, or blood serum – enabling a spatially resolved understanding of molecular programs in glioblastoma (table 1).^33,40 Ultimately, synergies between clinical imaging, deep molecular profiling, and AI-powered stratification may foster the discovery of novel prognostic biomarkers and associated tumour subgroups, providing guidance for more personalised prognostication and therapeutic management, including targeted, subgroup-specific interventions.

Response prediction and assessment

Sequential, rigorous methodical review of imaging in neuro-oncology is critical for treatment decisions, response assessment, and distinguishing treatment-related adverse effects (eg, pseudoprogression, tissue necrosis) from tumour progression. AI-assisted tumour volume quantification in structural MRI has yielded substantial reduction in inter-rater variability of time-to-progression estimation (table 1)¹⁷ and increased reliability in detecting pseudoprogression in high-grade gliomas.⁶¹ Similarly, co-registered spherical-projected contrast-enhanced structural MRI integrated with dose maps enabled prediction of post-stereotactic radiosurgery brain metastasis local control outcomes (AUC 0.89).⁶²

An exciting frontier accompanying advances in treatment delivery include novel imaging platforms for iterative evaluation of tumour status throughout a conventional radiotherapy course. Improvements in understanding the biologic relevance of changes observed throughout the course of radiotherapy may allow a departure from rigid approaches towards a paradigm of response-adaptive radiotherapy, ie, responsiveness to emerging treatment resistance even during the course of radiotherapy, allowing improvements in the therapeutic ratio.⁶³ Similar ideas have been applied to imaging-based disease monitoring of glioblastoma patients during post-radiation medical treatment, with the depth and duration of response being predictors for survival in glioblastoma.⁶⁴

However, significant intra-group outcome heterogeneity in large classes of CNS neoplasms is not yet reflected in current management recommendations. AI-based therapy response prediction may enable more granular risk and sensitivity stratification, as demonstrated for various entities and therapies.^{28,50,51,65,66} Recent transcriptomic analyses of the glioblastoma immune TME suggest utility of risk stratification and therapy response prediction based on an AI-predicted stemness index⁵¹ and immune cell-associated long noncoding RNA.⁶⁵ Glioblastoma groups organised along the stemness axis, inferred and predicted from gene expression data (AUC 0.96), are characterised by distinct mutational burden, immune TME, and immunogenomic patterns, and differential sensitivity to ICIs (table 1).⁵¹ Radiomics/radiogenomics was utilised for response prediction to targeted therapies (eg, dabrafenib plus trametinib in BRAF^V600E mutant gliomas)⁵⁰ for low- and high-grade glioma, and ICI therapy (eg, predicting PD-L1 expression levels), as demonstrated in studies on molecular brain metastasis profiling from structural MRI.⁶⁶ Excitingly, simulated CEST-MR fingerprint images have been utilised to monitor apoptosis as an early predictor of treatment response to oncolytic virotherapy, where a deep neural network could infer metabolic changes from quantitative maps of chemical exchange parameters in a glioblastoma orthotopic mouse model (table 1).²⁸ Yet, whether changes in novel response parameters are clinically meaningful warrants rigorous clinical efficacy validation, inter alia to mitigate overinterpretation.

Management and Rehabilitation

Tumour- and treatment-related complications

Management of tumour and treatment-related adverse effects represents a growing neuro-oncological challenge given recent improvements in cancer survivorship, particularly for patients with brain metastases. The nature and burden of potential toxicities has evolved substantially with increased utilisation of stereotactic radiosurgery and adjuvant integration of targeted therapies.⁶⁷ Late CNS toxicities include cognitive dysfunction, leukoencephalopathy and brain atrophy, cerebrovascular complications, endocrinopathies, and brain tissue necrosis.^67,68 Differential spatial radiographic patterns provide an opportunity for AI-assisted monitoring of treatment-induced tissue necrosis to circumvent unnecessary exposure to antineoplastic therapies or delivery of treatments that undermine the efficacy of systemic therapies. To unveil molecular correlates of radiotoxicity, AI can predict radiotoxicity-associated gene expression signatures correlating with OS in glioblastoma.⁶⁹ Efforts to prevent or reduce CNS radiotoxicity are imperative; AI can improve radiation field planning to spare proximal organs-at-risk, and support the introduction of reliable non-invasive surveillance platforms for risk stratification, detection, and adequate management of treatment-related toxicities (supplementary table 1, appendix p. 3).

Neurocognitive and motor rehabilitation

Tumour-related complications or adverse effects can be progressive and irreversible, resulting in cognitive impairment and permanent neurologic disability, diminishing quality of life and even impacting survival.⁶⁸ AI-assisted neurocognitive rehabilitation has demonstrated promising results in paralysed persons with speech and motor deficits via intracortical brain-computer interfaces (BCIs) deciphering attempted handwriting movements from motor cortex activity,⁷⁰ as well as subdural multielectrode arrays or non-invasive interfaces decoding speech from neural activity.^71,72 Conversely, generative models for BCI- or neurostimulation-mediated manipulation of spatiotemporal neural activity can potentially restore motor or impaired higher-order neurocognitive functions. Transposing these ideas to neuro-oncology can aid in optimising brain health through function restoration, improving communication abilities, active societal participation, and quality of life, especially in long-term survivors with cancer-related neurocognitive impairments.

Opportunities, Pitfalls, and Future Directions

AI-based decision support systems hold promise to assist clinicians across the entire neuro-oncological care trajectory (table 2). Selected applications for well-circumscribed tasks can augment specialist care^14,17 or have already attained regulatory approval (table 3). Insufficient generalisability, reproducibility, and data accessibilitiy, lack of consensus definitions and reference standards for validation, ethico-legal issues, and regulatory hurdles remain principal barriers to broad applicability.³⁶

Table 2.

Clinical challenges and future directions for AI in neuro-oncology.

Clinical Challenge	Opportunities and Future Directions
Discovery of predictors for disease onset, effective preventive risk stratification	AI-driven integration of GWAS data12 with other molecular layers, functional annotation for risk group identification based on molecular profiling (eg, enabling identification of novel mutations/interacting gene modules,52 preventive liquid biopsies or structural MRI for identified risk groups, etc.) Prevention of metastatic spread via oncogenic driver identification13 Prediction of likelihood of therapy-related adverse effects
Reduction of human error and inter-rater variability in diagnosis and treatment response assessment	Diagnostic neuroradiological assistance via CNS tumour detection, classification, grading, and segmentation, and synthetics data generation14,16,84 Multi-modal models for longitudinal integration of clinico-molecular variables and care trajectory-wide assistance10
Histo-molecular precision diagnosis	Inference of molecular alterations from non-invasive clinical imaging (MRI, CT, PET, ultrasound) or histological imaging (WSI, cryosections)15,19,20,22,31,32,34 Annotation of known histo-molecular subtypes with differentially enriched pathways Discovery of novel histo-molecular subtypes33,55–59 Prediction of relevance of molecular alterations of unknown clinical significance
Personalised/agile treatment planning and delivery	Radiation therapy planning43 Tumour segmentation, prediction of optimal resection margins, closed-loop surgical guidance systems47 Real-time intra-operative diagnosis based on ultrasound25 or optical methods34 Robotics-aided tissue sampling or microsurgery49 Personalised therapy recommendation based on AI-predicted response51 Iterative status evaluation, response assessment, and monitoring by imaging-based volumetric assessment,17 metabolic imaging,28 discrimination of pseudoprogression from true progression,61 metastatic invasion, and agile treatment administration
Personalised risk stratification	OS, PFS, treatment-related adverse effect prediction23,51,85 Discovery of association between predicted risk score and molecular-genetic traits, such as intra-tumoral heterogeneity33 Introduction of novel risk stratification schemes based on immunological or molecular-genetic tumour traits51 Adverse event prediction and monitoring to prevent treatment-related complications
Restorative neurorehabilitation	Non-invasive neural interfaces to mitigate loss of function due to tumour- or treatment-related neurocognitive sequelae72 Invasive BCIs to mitigate or restore loss of sensory, motor, or cognitive function by recording and ML-based decoding of neural activity, activity-based control of language interface or neuroprosthetic devices70,71
Discovery of novel drug targets and biological agents	Identification of phenotype-associated molecules as potential druggable targets54 Generative de novo design of small molecules or proteins with pre-defined properties,75 computational high-throughput testing based in biophysical model of agent-target interaction
Mechanistic models of disease	Integrative molecular data analysis51,53,54 Biophysical/causal models of disease dynamics75

Abbreviations. AI: Artificial intelligence; BCI: Brain-computer interface; F-DOPA-PET: 6-¹⁸F-fluoro-_L-DOPA positron emission tomography; GWAS: Genome-wide association study; ML: Machine learning; OS: Overall survival; PFS: Progression-free survival; WSI: Whole-slide image

Here, we highlight recent progress in promising emerging applications across FM-based neuro-oncological assistants, biophysical modeling, and synthetic data for drug discovery and inverse problems, and critically discuss associated challenges across current data, validation, and implementation gaps.

Foundation model-based neuro-oncology assistants

A surge of interest in FMs, particularly in digital pathology, signals a shift towards generalist models trained on vast (multi-modal) datasets at scale, adaptable to varied downstream tasks (panel 2). Serving as a backbone for well-delineated applications,^29,31 they may open up pathways to regulatory approval. Yet, reproducible validation studies and rigorous benchmarking are essential to compare FM-based applications against performant narrow models.

In contrast, instruction-fine-tuned FMs may aid practising clinicians across a broad range of tasks as generalist neuro-oncological assistants (panel 2) that communicate through natural language interfaces.⁹ However, uncertain zero-shot generalisation capabilities, potential bias in disfavour of identifiable groups, and unsatisfactory alignment currently preclude pertinent regulatory pathways towards clinical adoption. Alongside development and implementation of FM governance guidelines,⁷³ technical progress is required to guarantee truthfulness and strict adherence to ethical principles, particularly in safety-critical scenarios. Existing access inequities, potentially exacerbated by vast data and compute requirements for pre-training, require preemptive mitigation, prompting attempts to bridge the digital divide by democratising access and local deployment through open-sourcing medical FM checkpoints.⁷⁴

Biomarker discovery, biophysical modelling, synthetic data

AI’s data-type agnosticism (panel 1) is conducive to identification of preventive, diagnostic, and prognostic biomarkers, inference of histo-molecular properties from routinely collected clinical imaging for non-invasive diagnosis/continual monitoring, and risk stratification into known, refined, or novel diagnostic, therapeutic, and prognostic groups.

However, if data are scarce or carry inherent domain-specific structure, this domain agnosticism can impede convergence to satisfactory solutions. To counteract this, deep neural networks can be endowed with inductive biases to encode domain-inherent causal, relational, or semantic structure (eg, symmetries, intrinsic dimensionality, biophysical/geometric/relational priors, probabilistic/causal structure).⁷⁵ Relational, chemico-physical, and causal priors can expedite drug target identification and spatial transcriptomics,⁴⁰ generative compound design,⁷⁵ and disentangle treatment effects from confounders,⁷⁶ respectively. These applications can expand the neuro-oncological therapeutic armamentarium and guide optimal patient allocation into prospective clinical trials based on patient characteristics beyond standard inclusion criteria.⁵⁴ Novel therapeutics demonstrating activity for selected CNS tumours need prospective evaluation alongside standard local therapies (surgery, radiation therapy) to ensure translation into meaningful improvements in clinical outcomes.

If data are entirely unavailable, generative models producing synthetic data may present a resourceful alternative to extensive data collection, but do not inherently guarantee data realism. Extensive clinical validation can mitigate this to an extent, but lack of mechanistic/causal models behind synthetic data generation processes precludes unconditional generalisability. In contrast, models with strict bio-physical inductive biases may allow parametrisation/simulation of some medical imaging modalities from first principles, opening up an avenue for solving associated inverse problems – the process of recovering unknown parameters/states from observed measurements – eg, for histo-molecular parameter estimation.²⁸ Furthermore, concurrent simulation of physiological and measurement processes, eg, neural activity and electroencephalography, may enable in-vitro experimentation with “neural activity fingerprints” in disease models,^7,41 akin to strategies already utilised in de novo compound design and virtual screening.⁷⁵

Mitigating data gaps, balancing innovation and regulation

Adopting and scaling AI in neuro-oncology, inter alia hinges on closing a tripartite data gap: collection (absence, scarcity, inaccessibility); (ii) standardisation (digitisation, storage, harmonisation) ; (iii) annotation.

In most institutions, large, representative, and high-quality datasets are entirely absent, scarce (particularly for rare disease subtypes), or inaccessible to the wider research community due to stringent data protection requirements. Collection and access are critical for model development and validation in multi-centric international trials to ensure generalisiability across institutions, regions, and patient groups. To counteract overutilisation of small, biased samples – often in disfavour of minorities and vulnerable groups – and leverage existing and future AI applications in neuro-oncological care, efforts aimed at broad data collection, representative patient population selection, standardised dataset curation, and processing are paramount.^36,77

Existing data are siloed due to inconsistent digitisation and a lack of standardised recording, storage, centralisation, harmonisation, and deidentification practices. Variation between institutions, inconsistent data quality, and selection bias currently impede swift validation and integration of novel algorithms into clinical settings due to transfer penalties or performance deterioration upon deployment. Neuro-oncological stakeholder groups hence promote the use of standardised data acquisition protocols and implementation of standard storage, digitisation, and processing practices.³⁶

Most AI applications to-date are supervised, ie, they require annotated data for training. Yet, data labelling is often costly, arduous, and prone to inter-rater variability, particularly in neuropathology, where data annotation is not part of standard clinical routine. This annotation gap may be alleviated using AI-based auto-labelling pipelines, eg, leveraging pre-trained FMs adaptive to novel tasks with just a few annotated datapoints (panel 2).

AI-based decision support systems require prospective validation and seamless integration into routine neuro-oncological practise. The wealth of neuro-oncological AI research harshly contrasts with a conspicuous lack of prospective validation studies and current paucity of AI applications obtaining regulatory approval. Efforts are underway to mitigate these validation and implementation gaps, eg, by hosting research community-wide tumour segmentation challenges,⁷⁸ open-sourcing pre-trained medical AI models,⁷⁴ and developing AI-based medical devices for several key neuro-oncological tasks (table 3). Beyond validation studies, developers must align with existing regulatory frameworks (WHO,⁷⁹ UNESCO,⁸⁰ and OECD)⁸¹ to obtain regulatory approval, and, critically, provide evidence that novel technologies indeed outperform existing workflows through meaningful improvements in clinical outcomes. Here, fast-track approval procedures by regulatory agencies (eg, FDA breakthrough device designation) offer opportunities for dynamic yet responsible innovation. Existing medical devices may serve as a roadmap to regulatory approval and financial viability for future applications (table 3); reimbusement strategies need to balance patient outcomes and cost-effectiveness.⁸² Devices without clear pathway to regulatory approval (eg, generalist assistants) require the invention of flexible recertification pathways and novel governance frameworks. Guidelines must cover fairness, reproducibility, robustness, and adherence to pre-defined ethical standards, anchored in principles of human-centred, value-based, equity-driven,⁸³ and balanced regulation without stifling innovation.

Conclusion

Integrating AI into neuro-oncological practise and research holds tremendous potential, as already demonstrated in non-invasive precision diagnosis and response assessment from clinical and non-routine imaging, translational inference, the discovery of novel biomarkers and histo-molecular tumour subclasses, treatment administration and monitoring, foundational research into druggable targets for incurable CNS neoplasms, risk stratification, and BCI-based neurorehabilitative devices. Exploratory future opportunities entail (generalist) neuro-oncology assistants, biophysical/causal modelling, eg, neural-cancer interactions, synthetic data generation, drug and drug target discovery, and patient stratification for administration of tailored therapies. Development and deployment critically hinge on solving core challenges pertaining to data gaps, clinical validation of assistive technologies and AI-based classification schemes, tissue-based corroboration of biomarkers, rooting generative models of data and disease in causal, biophysical insights, validation of druggable targets, and resolving core ethico-legal and regulatory hurdles to ensure responsible, person-centred, equitable, and needs-based integration of AI-based tools into neuro-oncology.

Figure 1.

PRISMA flow diagram.

Figure 2.

The distribution of machine learning methods (x-axis) per neuro-oncological data type (y-axis). The circle size is proportional to the number of abstracts referencing method-data type pairs in the 2675 pre-filtered articles. Our search identified that the current literature is largely dominated by studies on clinical imaging and radiomics/deep learning and classical statistical methods, which are also prominently used for omics studies; fewer studies investigate more niche modalities and advanced techniques. For raw data please see supplementary table 2, appendix p. 4.

Abbreviations. CT: Computed tomography; dMRI: Diffusion MRI; ML: Machine learning; MRI: Magnetic resonance imaging; PET: Positron emission tomography

Colour code. Blue: Clinical imaging (MRI, dMRI, CT, PET); purple: Neuropathology; green: Omics

Search strategy and selection criteria.

References for this review were identified through searches of PubMed/MEDLINE, ArXiv, and Google Scholar with pre-defined MeSH term-based search phases and terms including “glioma”, “glioblastoma”, “machine learning”, “radiomics”, “foundation model”, “omics” from January 2020 until December 2024 (cf. appendix pp. 9 – 10). Articles were screened for relevance based on exclusion and inclusion criteria, defined separately for “clinical” and “experimental” studies (full methodology in appendix pp. 1 – 2). The 275 remaining publications were subsequently evaluated in depth. After exclusion of 234 articles, and manual addition of 11 further articles of high relevance, 52 original articles were included in the final selection (figure 1). The distribution of papers per category is visualised in figure 2.