Artificial intelligence in brain tumor diagnosis and surgical planning: Recent advances

Abstract

Background:

Artificial intelligence (AI) is rapidly advancing across medical disciplines, with neurosurgery emerging as a key field for technological innovation. In the management of brain tumors, AI-based platforms have demonstrated considerable potential to enhance diagnostic accuracy, support surgical planning, and assist intraoperative decision-making.

Methods:

A literature search was conducted using PubMed and Cochrane Library databases to identify peer-reviewed studies published in English between 2015 and 2025, using keywords including “artificial intelligence,” “neurosurgery,” “brain tumors,” “machine learning,” “deep learning,” “computer vision,” “natural language processing,” “radiomics,” and “surgical planning.”

Results:

Machine learning and deep learning have improved radiologic detection, classification, and segmentation of brain tumors, while radiomics and radiogenomics enable noninvasive molecular prediction and tumor characterization. AI is increasingly integrated into surgical planning, including brain deformation modeling, fiber tractography, intraoperative histologic assessment, hyperspectral imaging, and intelligent navigation systems. Challenges include limited data availability, algorithm transparency, dataset heterogeneity, and regulatory and infrastructural requirements for clinical implementation.

Conclusion:

AI demonstrates considerable potential to advance brain tumor management, though robust prospective validation and evidence-based implementation are essential prerequisites for safe clinical integration. Continuous technological refinement, multidisciplinary collaboration, ongoing research, and equitable access across healthcare systems, including low-resource settings, will be essential for responsible and effective implementation.

INTRODUCTION

Artificial intelligence (AI), one of the technological breakthroughs of the modern world, enables machines to imitate human problem-solving and decision-making capabilities by learning from extensive datasets and patterns of human activities.[27] AI-driven medical technologies are rapidly evolving from experimental research tools into clinically applicable solutions across a wide range of healthcare domains.[8] Advances in AI, particularly in machine learning (ML) and large-scale data analysis, now allow AI systems to assist clinicians in diagnostic interpretation, therapeutic decision-making, and risk assessment across diverse medical specialties.[20]

AI is increasingly being integrated into neurosurgery, with growing potential to support diagnosis, optimize surgical planning, and enhance the management of complex intracranial pathologies.[60] Patients with brain tumors frequently present with a spectrum of neurological and cognitive symptoms throughout the course of their illness, with the nature and severity of these manifestations largely determined by the size and anatomical location of the lesion.[82] Research efforts have increasingly focused on applying AI to improve the understanding of brain tumor biology and facilitate more personalized patient management strategies.[4]

AI encompasses several domains and techniques. Key approaches include ML, artificial neural networks (ANNs), deep learning (DL), convolutional neural networks (CNNs), computer vision (CV), and natural language processing (NLP). One of the most prominent domains within AI is ML, a subset of AI that involves developing computational models capable of learning from data and making predictions or decisions without explicit programming.[40] ML approaches are generally categorized into three main types: supervised learning, unsupervised learning, and reinforcement learning. In supervised learning, the algorithm is trained using datasets in which the desired output is known, allowing the model to learn the relationship between input variables and target outcomes.[18] In unsupervised learning, the objective is to identify inherent patterns, clusters, or relationships within the data, which can be challenging to evaluate and is often assessed by the model’s performance in downstream supervised tasks.[18] Reinforcement learning involves an algorithm attempting to accomplish a task through a process of trial and error while learning from its successes and mistakes.[71] Other ML techniques that will be discussed in this review include support vector machines (SVMs), a supervised learning method widely used for classification and regression tasks, valued for their efficiency and robustness with large datasets.[14] Support vector regression (SVR) is an extension of the SVM framework designed to predict continuous-valued outcomes, making it suitable for regression problems.[13] Dictionary learning is another ML technique that identifies sparse representations of data by learning a set of basis elements, highly effective for feature extraction, data decomposition, and segmentation tasks.[78]

ANNs represent another key approach within AI. ANN models are designed to emulate the information-processing mechanisms of the human brain. Their architectures are inspired by biological nervous systems and consist of interconnected neurons organized in complex, nonlinear networks. Key components of ANN modeling include data input, network architecture design, determination of hidden layers, simulation of network activity, and adjustment of weights and biases through learning algorithms. ANNs have facilitated significant advancements in AI, with applications in fields such as image and speech recognition, medical data analysis, and robotics.[51] Building upon ANNs, DL employs computational models composed of multiple processing layers that can automatically learn multi-level representations of data. By extracting increasingly abstract features at each layer, DL models have significantly advanced performance in domains such as speech recognition, visual object recognition, and medical applications including genomics and drug discovery. These models use the backpropagation algorithm to repeatedly adjust internal parameters, allowing each layer to transform inputs into increasingly abstract representations.[44] DL is increasingly used in clinical neuroscience for magnetic resonance imaging (MRI) analysis, tumor segmentation, and subtype prediction, showing promise as a decision-support tool.[17] Among DL architectures, CNNs are widely used for image analysis. CNNs are composed of convolutional, pooling, and fully connected layers and automatically learn spatial hierarchies of features from imaging data using the backpropagation algorithm. They have become dominant in CV and are increasingly applied in medical imaging, including radiology. CNNs have been widely used for tasks such as tumor classification and segmentation, enabling automated and accurate analysis of lesions on computed tomography (CT) and MRI scans.[79]

CV is an AI application for image interpretation, enabling software to recognize and characterize objects. It often uses ML to develop algorithms capable of detecting patterns, classifying structures, and analyzing images.[65] CV-based approaches enable accurate and efficient segmentation and classification of brain tumors on CT images, supporting clinical diagnosis and treatment planning.[2] NLP is another domain of AI that involves computational methods for analyzing and modeling human language.[32] NLP is being applied in the screening of electronic health records (EHRs), speech-to-text extraction of clinical data at the point of care,[29] and the extraction of brain tumor diagnoses from outpatient clinic letters.[7]

This narrative review aims to explore recent advances and emerging applications of AI in the diagnosis and surgical planning of brain tumors, as well as the challenges and future prospects of AI in precision neuro-oncology. A literature search was conducted using the PubMed and Cochrane Library databases to identify relevant peer-reviewed studies published in English between 2015 and 2025, focusing on the application of AI in brain tumor diagnosis, surgical planning, and neurosurgical management. The following keywords were used individually and in combination: “artificial intelligence,” “neurosurgery,” “brain tumors,” “machine learning,” “deep learning,” “computer vision,” “natural language processing,” “radiomics,” and “surgical planning.”

DISCUSSION

AI applications in brain tumor diagnosis

Brain tumor diagnosis traditionally relies on neuroimaging and histomolecular analysis of resected or biopsied tissue.[48] Contrast-enhanced CT and MRI are widely used noninvasive modalities for this purpose. Conventional MRI sequences (T1, T2, and T2 fluid-attenuated inversion recovery [FLAIR]) provide detailed information on tumor volume and morphology, forming the basis for quantitative imaging analyses.[12] Nowadays, AI has emerged as a promising approach for rapid brain tumor diagnosis, with growing evidence supporting its role in neuro-oncologic imaging.[12]

In this section, we review AI applications in brain tumor diagnosis across three domains: radiologic detection, classification, and segmentation; radiomics and molecular prediction; and tumor characterization for preoperative risk assessment.

Brain tumor detection, classification, and segmentation

Several AI frameworks have recently been developed for automated brain tumor detection, classification, and segmentation. A DL-based you only look once version 7 model combined with CNNs was applied to MRI for automated detection of gliomas, meningiomas, and pituitary tumors. Enhanced with attention mechanisms and multi-scale feature fusion, the framework achieved high performance (99.5% precision, 99.3% recall), supporting brain tumor diagnosis.[1] CV techniques have been effectively applied to classify brain tumors using CT imaging, with multiple classifiers demonstrating high diagnostic performance. In one framework encompassing six tumor types, including meningioma, glioma, schwannoma, neurofibromatosis, chondrosarcoma, and chordoma, an overall accuracy of 97.83% was achieved.[2] Shrot et al. applied an ML-based approach incorporating magnetic resonance (MR) morphologic, diffusion tensor, and perfusion imaging to develop an automatic tumor classification algorithm. A binary SVM achieved classification accuracies of 95.7% for glioblastoma, 92.7% for metastases, 97% for meningiomas, and 91.5% for primary central nervous system (CNS) lymphoma.[69] However, these models were predominantly evaluated on small, single-institution datasets, and prospective multicenter validation is necessary to confirm generalizability across diverse clinical settings.

For tumor segmentation, a DL framework combining fully CNNs with conditional random fields achieved accurate tumor segmentation with both spatial and appearance consistency. Evaluated on multimodal brain tumor image segmentation challenge (BRATS) 2013–2016 datasets, the model demonstrated competitive performance across MRI sequences (T1c, T2, and FLAIR) while enabling faster slice-by-slice segmentation.[83] AI-based brain tumor image analysis (referred to as BT) software provides rapid and reproducible tumor segmentation and volumetric assessment of glioblastomas, distinguishing tumor subregions (edema, necrosis, nonenhancing, and enhancing) from healthy tissue using routine MRI sequences. Its performance is comparable to expert manual segmentation and standard clinical tools, supporting preoperative imaging assessment and precise surgical planning.[62] In addition, Laukamp et al. demonstrated that a multiparametric deep-learning model, DeepMedic, can accurately detect and segment meningiomas on routine MRI. The automated segmentations strongly correlated with manual readings (Dice coefficients 0.81 for total tumor and 0.78 for contrast-enhancing tumor), supporting improved diagnosis, therapy planning, and monitoring.[42] Nevertheless, most segmentation models were benchmarked on standardized datasets such as BRATS, and performance may vary considerably when applied to real-world clinical imaging with heterogeneous acquisition protocols.

For brain metastases, the DL-based brain metastasis segmentation system significantly improved the segmentation accuracy of brain metastasis lesions and reduced clinician workload by approximately 42% in an evaluation of over 10,000 metastatic lesions, supporting its integration into neuro-oncologic imaging and radiotherapy planning.[50] Senders et al. used NLP to automatically extract and quantify brain metastases from MRI reports, achieving 83% accuracy and enabling faster, standardized clinical data analysis.[68] In a comparative analysis, ChatGPT-4o was benchmarked against experienced radiologists for brain tumor diagnosis. The model detected 95.7% of lesions and characterized MRI features effectively, although limitations were observed in lesion localization and differentiation of extra-axial from intra-axial tumors.[57] However, current studies remain predominantly retrospective, and prospective clinical validation is required before these tools can be reliably integrated into routine neuro-oncologic practice.

Radiomics, radiogenomics, and molecular prediction

Radiomics is a subfield of AI that focuses on computing, identifying, and extracting quantitative imaging features, and using them to develop predictive or prognostic mathematical models.[48] It is increasingly being explored for applications in neuro-oncology, with analyses of brain tumors primarily performed using conventional MRI sequences.[48]

Radiogenomics is a novel research field that establishes associations between radiological imaging features and underlying genomic or molecular expression, with the goal of enhancing diagnostic precision and enabling personalized oncological management.[73] By integrating radiomics with genomic data in a noninvasive fashion, radiogenomics holds potential for molecular characterization, survival prediction, and illumination of oncogenic mechanisms without the need for surgical tissue sampling.[73] DL algorithms have demonstrated excellent diagnostic performance in this domain, achieving a 96.0% positive posttest probability for isocitrate dehydrogenase (IDH) mutation prediction in gliomas on validation datasets[34] and accurately determining O6-Methylguanine-DNA methyltransferase promoter methylation status in glioblastoma using large-scale, multi-sequence MRI datasets with 3D context-preserved DL architectures.[39] Furthermore, DL-based models incorporating data augmentation and feature selection methods have shown effective prediction of 1p/19q codeletion status in low-grade glioma patients,[46] while endto-end noninvasive imaging tools have been developed for preoperative prediction of epidermal growth factor receptor mutation status from brain MRI, potentially facilitating individualized treatment planning.[9]

Researchers have developed integrated AI frameworks combining DL, radiomics, and habitat analysis for preoperative glioma grading using multiparametric MRI. Evaluated in 847 patients, with external validation in 213, the dual-stream model achieved high accuracy (area under the curve [AUC] 94.6%) in distinguishing high- from low-grade gliomas (LGG), highlighting its potential as a noninvasive diagnostic and surgical planning tool.[81] In a further multicenter validation, integrating habitat radiomics with multiparametric MRI enabled accurate prediction of glioma grade and molecular characteristics, including IDH1 mutation, Ki-67 expression, and P53 status.[84] AI frameworks have been explored for other primary CNS tumors, expanding their diagnostic and predictive capabilities. A diffusion radiomics model combining recursive feature elimination and a random forest classifier accurately identified atypical primary CNS lymphoma (AUC 0.944–0.984) and outperformed conventional radiomics approaches and standard MRI metrics. Its performance was comparable to experienced neuroimaging readers (AUC 0.825–0.930), demonstrating its utility for preoperative diagnostic support.[33]

ML-integrated radiomics combining multiparametric MRI features with clinical parameters enabled noninvasive prediction of molecular subgroups in medulloblastoma. The model demonstrated strong diagnostic performance, particularly for Wingless-type and Sonic Hedgehog subgroups, supporting preoperative molecular characterization.[80]

Similarly, MRI-based radiomics models have been developed for craniopharyngiomas and skull base tumors. Chen et al. developed an MRI-based radiomics model for the noninvasive prediction of pathological subtypes and genetic mutation status in craniopharyngiomas, while Li et al. developed a multiparametric MRI-based radiomics signature to differentiate skull base chordomas from chondrosarcomas.[10,43] Despite these advances, most radiomics and radiogenomics models remain limited by small sample sizes, single-institution development, and variability in imaging protocols, highlighting the need for large-scale, multicenter prospective validation before routine clinical implementation.

Tumor characterization and preoperative risk assessment

AI can aid tumor characterization and preoperative risk assessment by predicting molecular alterations, invasion, and progression. A SVM algorithm was developed to noninvasively predict 1p/19q codeletion status in presumed LGG using preoperative T1 and T2 MRI sequences, along with patient age and sex. In external validation on 129 patients, the model achieved an AUC of 0.72, outperforming the average of neurosurgeons (AUC 0.52) but slightly below neuroradiologists (AUC 0.81), demonstrating its utility for preoperative molecular characterization.[74] Hale et al. demonstrated that ML algorithms including ANNs, SVMs, and k-nearest neighbors can accurately predict meningioma grade from preoperative MRI features (AUC 0.89), outperforming conventional statistical methods.[23] A deep CNN was evaluated for predicting histopathological grading of meningiomas using MRI-derived imaging parameters. The model demonstrated high accuracy in distinguishing benign from atypical or anaplastic meningiomas on apparent diffusion coefficient maps, highlighting the potential of diffusion-based MRI for AI-assisted preoperative tumor grading and risk stratification.[5]

Furthermore, a multiparametric MRI-based radiomics model using brain-tumor interface features compared several machine-learning algorithms, including LightGBM, logistic regression, multilayer perceptron, random forest, SVM, and XGBoost. The XGBoost model achieved the highest performance (AUC 0.913 for training, 0.897 for testing), supporting its potential as a noninvasive tool for preoperative prediction of meningioma brain invasion.[11] Finally, metabolic profiling using AI-assisted whole-brain MR spectroscopy enabled early prediction of progression in high-grade gliomas. By analyzing metabolic signatures across brain regions, the model achieved strong predictive performance (AUC ≈ 0.86) and identified the choline-to-creatine ratio as a key marker of early disease advancement.[66] Nevertheless, these models were largely developed on limited single-center datasets, and independent external validation is needed to establish their reliability across varied clinical and imaging environments.

In summary, AI-based approaches demonstrate significant potential in brain tumor diagnosis by improving imaging-based detection, classification, and segmentation. Radiomics and radiogenomics enable noninvasive assessment of molecular features, tumor behavior, and preoperative risk stratification. Collectively, these tools may assist in surgical planning and support more individualized patient management strategies. However, their clinical application remains limited by lack of external validation, dataset heterogeneity, and absence of large-scale prospective studies.

AI-assisted surgical planning and intraoperative support for brain tumors

The integration of AI into surgical planning has demonstrated substantial promise for improving surgical outcomes.[3] AI platforms have the potential to enhance both the safety and effectiveness of tumor resections in brain tumor surgery.[76]

This section reviews the role of AI in surgical planning and intraoperative support for brain tumors, focusing on applications such as preoperative functional mapping, surgical trajectory planning, fiber tractography, brain deformation modeling, intraoperative histologic assessment, hyperspectral imaging, and other emerging AI-driven operative technologies.

Preoperative functional mapping and surgical planning

AI techniques are increasingly being integrated into preoperative brain tumor surgical planning to optimize functional preservation and operative strategy. DL-based resting-state functional MRI mapping has been developed to support presurgical evaluation of the eloquent cortex. A three-dimensional CNN generated accurate language and motor network maps with reduced scanning time while maintaining high fidelity to standard acquisitions, reliably identifying eloquent regions in patients with brain tumors to facilitate precise surgical planning and functional preservation.[49] Dundar et al. developed an AI-based surgical planning system that incorporates a heuristic optimization algorithm and a Q-learning reinforcement learning model to identify optimal cranial entry points and generate safe cortico-tumoral surgical trajectories. [18] This approach highlights the potential of AI to improve surgical precision and minimize injury to critical neuroanatomical structures during tumor resection.[19] However, these models require validation across larger, diverse patient cohorts to confirm their clinical reliability and transferability beyond controlled research environment.

Ishankulov et al. demonstrated that traditional ML models applied to intraoperative cortico-cortical evoked potential data can predict postoperative speech dysfunction in eloquent brain areas. This approach enables early identification of functional risk during tumor resection, and integration of such AI-driven predictions into surgical workflows may enhance safety and help preserve critical language function.[27] Staartjes et al. developed a deep neural network to preoperatively predict gross-total resection in 140 patients undergoing endoscopic transsphenoidal surgery for pituitary adenomas. [70] Incorporating 16 preoperative radiological and procedural variables, the model achieved excellent performance (AUC 0.96, accuracy 91%, sensitivity 94%, specificity 89%), outperforming both logistic regression and the Knosp classification. These results highlight the potential of AI to support individualized risk stratification and enhance the likelihood of complete tumor removal.[70] Nevertheless, the relatively small sample sizes and single-institution designs of these studies limit generalizability, and large-scale prospective studies are warranted before routine clinical integration.

An emerging application of NLP in surgical planning involves the use of large language models (LLMs) for tumors involving eloquent cortical areas. Given the demonstrated alignment between LLM architectures and human language processing networks, these models have been proposed as potential individualized virtual simulation tools for language-eloquent brain tumor surgery. If clinically validated, they may enable preoperative prediction of postoperative language deficits, recovery trajectories, and optimization of individualized treatment strategies, complementing established modalities such as direct electrocortical stimulation and functional MRI.[54] While LLMs demonstrate advanced capabilities across complex tasks and hold broader potential in neurosurgery, current research predominantly addresses basic applications, with limited focus on performance optimization and reproducibility. Standardized reporting frameworks, acknowledgment of LLM stochasticity, and methodologically rigorous validation approaches will be essential to advance their clinical utility.[61]

AI-enhanced fiber tractography

Fiber tractography using diffusion MRI has become an essential tool in neurosurgery, enabling noninvasive visualization and delineation of critical white matter pathways for preoperative planning.[30] Building on this foundation, Kumar et al. developed BrainTract, a metaheuristic optimization-based CNN, to segment white matter fiber tracts and analyze structural connectivity from diffusion MRI, achieving an accuracy of 97.1% for fiber tract segmentation.[38] Similarly, Korycinski et al. developed HyTract, a hybrid framework integrating ANNs with a path search algorithm to accurately reconstruct neural fiber pathways near the surgical field.[36] These approaches collectively highlight the potential of AI-enhanced tractography to improve preoperative mapping of critical neural pathways and support safer surgical planning during tumor resection. However, most current models are validated on limited datasets and controlled imaging conditions, and their robustness in clinically challenging scenarios, such as peritumoral edema, mass effect, and white matter tract distortion commonly seen in brain tumors, remains insufficiently studied and requires prospective evaluation.

Brain deformation modeling

Intraoperative brain shift remains a major limitation of preoperative imaging-based guidance, as deformation of neural structures can compromise anatomical accuracy and potentially result in unintended resection of healthy cortex. Biomechanical models, such as finite element methods (FEMs) and thin plate splines, provide valuable insight into brain shift during surgery, with FEM demonstrating superior accuracy compared to simpler interpolation techniques.[21] Building on these physics-based frameworks, real-time AI-assisted brain shift modeling has been developed using ANNs and SVR. The system predicts intraoperative tumor displacement with submillimeter accuracy (<0.2 mm), enabling precise soft-tissue visualization and enhancing augmented reality-assisted surgical guidance.[72] However, these models have been validated in limited institutional settings, and their real-time performance across varied surgical scenarios, tumor types, and intraoperative imaging modalities requires broader prospective evaluation.

AI-based intraoperative histologic assessment

AI-driven approaches are increasingly being applied to intraoperative histologic assessment, enabling rapid and automated classification of CNS tumors to support neurosurgeons, especially in settings with limited access to neuropathologists.[36] Stimulated Raman histology (SRH) is a label-free optical imaging technique that provides rapid, high-resolution histologic information from unprocessed brain tissue.[55] Hollon et al. developed a near real-time DL-based tumor diagnosis system using SRH and CNNs. Trained on over 2.5 million SRH images, the model delivered automated tumor classification within 150 s and achieved diagnostic accuracy comparable to conventional histopathology (94.6% vs. 93.9%), offering a rapid, label-free approach to support intraoperative decision-making and reduce reliance on traditional frozen section workflows.[24] Further developments have enhanced SRH-guided intraoperative analysis. An integrated CV platform combining SRH with a CNN-based AI model enables rapid intraoperative evaluation of skull base tumor specimens. The system supports real-time assessment of tumor margins in benign and malignant skull base tumors including meningiomas, pituitary adenomas, and schwannomas, facilitating informed surgical decisions, reducing residual tumor burden, and potentially improving oncologic outcomes.[31] DeepGlioma is a rapid AI-based system (<90 s) for intraoperative molecular classification of diffuse gliomas by integrating SRH with genomic data. It achieved a mean molecular classification accuracy of 93.3 ± 1.6%, enabling real-time identification of IDH mutation, 1p/19q co-deletion, and Alpha Thalassemia/ Mental Retardation Syndrome X-Linked (ATRX) mutation to guide surgical strategy and optimize the extent of resection.[25] RapidLymphoma is a DL-based intraoperative diagnostic pipeline designed to differentiate primary diffuse large B-cell lymphoma of the CNS (primary central nervous system lymphoma) from other CNS tumors using SRH of fresh surgical tissue. The system delivers automated diagnostic feedback within approximately 3 min, achieving a balanced accuracy of 97.8% ± 0.9% in a prospective cohort of 160 cases, outperforming conventional frozen section analysis (77.8%), thereby supporting rapid intraoperative decision-making and surgical planning.[64] Despite these promising findings, most SRH-based AI systems have been developed and validated in high-resource academic centers, and their implementation in resource-limited settings remains challenging due to requirements for specialized imaging platforms, computational infrastructure, and trained personnel.

Hyperspectral Imaging

Another emerging intraoperative technique, Hyperspectral Imaging (HSI), provides real-time label-free tissue characterization during brain tumor surgery. HSI is a hybrid imaging modality that combines a digital photographic camera with a spectrographic unit, enabling contactless and nondestructive biochemical analysis of living tissue. By capturing spectral information across multiple wavelengths, HSI provides both quantitative and qualitative data on tissue composition at the molecular level in a contrast-free manner. This capability allows objective discrimination between different tissue types, including differentiation of healthy and pathological tissue.[6] ML combined with HSI has emerged as a promising intraoperative tool for brain tumor detection and delineation. By analyzing both spectral and spatial features, this approach enables real-time tumor boundary delineation in primary and secondary brain tumors, supporting maximal safe resection while preserving normal tissue.[45] Manni et al. developed a DL-enhanced HSI approach using a 3D-2D hybrid CNN for intraoperative brain tissue classification. The system distinguished tumor tissue, healthy parenchyma, and blood vessels with approximately 80% accuracy, providing real-time support for glioblastoma resection through improved tumor boundary delineation and surgical guidance.[52] However, the relatively modest classification accuracy reported in current studies, combined with the technical complexity of HSI integration into standard operative workflows and its sensitivity to surgical lighting conditions and tissue artifacts, underscores the need for further technological refinement and prospective clinical validation.

AI for intraoperative navigation and workflow optimization

AI-driven technologies are increasingly being used for intraoperative navigation and workflow optimization in brain tumor surgeries to enhance surgical outcomes. A retrospective study of 40 brain tumor patients demonstrated that DL-based reconstruction of accelerated intraoperative MRI was preferred over conventional compressed sensing in 83% and 98% of cases by two neuroradiologists, but only 20% by a neurosurgeon. DL reconstructions also scored higher in image quality metrics in 72%, 72%, and 14% of cases for the three readers, respectively, although artifacts and reduced signal were noted.[56] In another study, Wei et al. applied a segmentation dictionary learning algorithm for intraoperative MRI navigation during glioma resection, reporting improved extent of resection and reduced postoperative neurological deficits compared with conventional surgery.[75] ML has been applied to automatically analyze endoscopic transsphenoidal pituitary surgery videos, enabling workflow optimization by accurately recognizing surgical phases and steps. Using a combined CNN and recurrent neural network model, the system achieved 91% accuracy for phase recognition and 76% for step recognition, with potential applications in surgical education and real-time intraoperative guidance.[37] Nevertheless, the small sample sizes, retrospective designs, and variability in surgeon acceptance across these studies highlight the need for prospective validation and systematic assessment of AI integration within real-world intraoperative workflows.

Robotics in neurosurgery

Surgery is increasingly benefiting from robotic assistance, and in neurosurgery, these technologies hold significant potential to improve patient care and outcomes.[22] Robot-assisted stereotactic brain biopsy, in particular, has become a widely adopted and valuable tool in modern neurosurgical practice.[53] Although robotics in neurosurgery remain in its infancy globally due to high costs and specialized skill requirements, advances in AI, computing power, connectivity, and precision targeting are expected to drive meaningful expansion in the future, enhancing procedural safety, efficiency, and overall patient outcomes.[22]

In summary, AI-driven technologies demonstrate considerable potential to enhance multiple domains of neurosurgical practice, including preoperative planning, intraoperative guidance, diagnostic support, and workflow optimization. However, their clinical translation remains heterogeneous, and widespread adoption will require robust prospective validation, cost-effectiveness evaluation, infrastructure development, and standardization across diverse healthcare settings. Key studies discussed in this review, including AI models, tumor applications, datasets, and performance metrics, are summarized in Table 1.

Table 1:

Summary of key artificial intelligence studies in brain tumor diagnosis and surgical planning.

Challenges and future prospects

The growing use of AI in healthcare introduces ethical and legal challenges, particularly regarding informed consent and the ownership of patient-derived data used for algorithm development. The use of large-scale data further complicates established principles of privacy and confidentiality.[63] Many AI models lack transparent and interpretable parameters, making it difficult for clinicians and researchers to understand how decisions are generated. This “black box” nature of AI systems poses challenges in regulated healthcare environments and hinders regulatory evaluation of algorithm reliability and safety.[4] Another important concern relates to the potential risks associated with clinical automation. Faulty or inadequately trained algorithms may generate erroneous outputs, while excessive reliance on automated systems may contribute to reduced clinical vigilance and potential de-skilling of physicians.[59] In addition, the clinical implementation of AI is often limited by the accessibility of healthcare data for ML applications. Clinical information is frequently fragmented across multiple systems, including imaging archives, pathology databases, EHRs, and prescribing platforms, making data integration challenging.[35] The application of AI in neurosurgery is often constrained by relatively small datasets compared with the large-scale data typically used in ML research, which may restrict the scalability and generalizability of AI models despite promising performance in individual studies.[15] These data-related barriers are compounded by practical challenges, such as the substantial technological infrastructure and financial investment required to implement AI in neurosurgery.[76] These practical barriers are further amplified in low-resource hospitals, where limited digital infrastructure presents a significant implementation challenge.[77] Furthermore, many AI models in neuro-oncology have been trained on datasets that overrepresent certain tumor types, imaging protocols, and patient demographics, predominantly from high-income countries, potentially limiting their applicability and equity across diverse global populations. Finally, patient trust in AI technology represents a critical barrier; while many patients and their families accept the use of AI in neurosurgery, fully autonomous systems are less acceptable, with the majority preferring that the neurosurgeon retain ultimate control.[58]

Despite these challenges, the future of AI in neurosurgery holds significant promise. Widespread adoption will require regulatory approval, integration with EHR systems, sufficient standardization, incorporation into clinician training, and sustainable reimbursement mechanisms.[16] Regulatory clearance from national and international healthcare authorities, including the U.S. Food and Drug Administration, the European Medicines Agency, and equivalent bodies in other regions such as China’s National Medical Products Administration, will be essential prerequisites for the safe and lawful clinical deployment of AI-based neurosurgical tools. Progress will depend on close collaboration between surgeons, data scientists, and engineers, allowing clinical insight to guide meaningful data analysis and algorithm development. Expanding participation in clinical data registries at local, national, and international levels may help overcome current data limitations and improve the availability of integrated clinical, imaging, and molecular datasets. Transparency and interpretability in AI algorithms will ensure accountability and maintain high standards of patient care.[24] Multicenter external validation of AI models across diverse institutions, patient populations, and imaging protocols will be equally critical to confirm their generalizability and clinical reliability beyond the research settings in which they were developed. AI should also be progressively incorporated into postgraduate surgical education to enhance learning and build surgeons’ confidence in using these emerging platforms. AI-based platforms can support surgical training by identifying individual strengths and weaknesses and providing targeted modules to refine specific skills, potentially improving both training efficiency and surgical performance.[67] Surveys indicate that many surgeons and members of the wider surgical team are receptive to the integration of AI technologies in neurosurgical practice.[43] Ultimately, the meaningful integration of AI in neurosurgery will depend on clinical engagement, continuous evaluation of ethical, educational, and practical considerations, and ongoing research to implement evidence-based guidelines that can bring real change to the management of brain tumors.

CONCLUSION

AI demonstrates considerable and growing potential to advance brain tumor management, and its careful, evidence-based adoption may meaningfully enhance diagnostic precision, surgical outcomes, and patient care. Diagnostic and intraoperative tools should be continuously updated to maintain accuracy and support safe clinical decision-making. Close collaboration among surgeons, engineers, and data scientists, along with ongoing research, is essential to optimize AI’s impact on patient care and surgical performance. Importantly, privacy and consent principles should be established, together with public education initiatives, to foster trust in these technologies. Finally, ensuring equitable access across healthcare settings, including low-resource hospitals, will be vital to prevent disparities in neurosurgical care.

Ethical approval:

Institutional review board approval is not required.

Declaration of patient consent:

Patient’s consent is not required as there are no patients in this study.

Financial support and sponsorship:

Nil.

Conflicts of interest:

There are no conflicts of interest.

Use of artificial intelligence (AI)-assisted technology for manuscript preparation:

The authors confirm that there was no use of artificial intelligence (AI)-assisted technology for assisting in the writing or editing of the manuscript and no images were manipulated using AI.

Disclaimer

The views and opinions expressed in this article are those of the authors and do not necessarily reflect the official policy or position of the Journal or its management. The information contained in this article should not be considered to be medical advice; patients should consult their own physicians for advice as to their specific medical needs.