Before the widespread adoption of deep learning, a critical conceptual advance was the recognition of image preprocessing as a core analytic infrastructure rather than merely a preliminary step. Bias-field correction and intensity normalization were used to stabilize tissue contrast [46], while skull stripping and nonlinear registration to a shared atlas or reference space enabled voxel-wise and region-wise comparisons across subjects and imaging sites [34,47]. In practice, these steps turned collections of scans into harmonized datasets suitable for population-level analysis and downstream feature engineering.

Pediatric imaging quickly exposed the limitation of adult-centric reference spaces : brain size, myelination, and tissue contrast shift substantially from infancy through adolescence, and routine clinical imaging often includes thicker slices or inconsistent coverage [25]. As a result, adult templates and priors can introduce systematic anatomical mismatch. These challenges motivated the development of age-specific pediatric and infant atlases, which reduce template bias and improved anatomical correspondence [45]. In neonatal imaging, dedicated approaches further showed that automated labeling remained feasible when models reflected newborn contrast and rapid maturation [54]. With scans made comparable, attention naturally shifted to extracting anatomy as explicit measurements through segmentation and morphometric analysis.

Once data could be mapped into common coordinate systems, segmentation became the primary interface between imaging data and quantitative anatomical descriptors. Classical probabilistic formulations combined intensity-based models with spatial regularization; a representative example is the hidden Markov random field-expectation maximization framework, which enforces local spatial consistency under noise and partial volume effects [62]. These approaches enabled scalable tissue classification and structure labeling without the need for exhaustive manual annotation.

In parallel, atlas-based and surface-based pipelines operationalized morphometric analysis at scale. Surface reconstruction frameworks enable estimation of cortical thickness and surface area [15], while automated labeling provided standardized regional volumes across large cohorts [24]. However, these approaches relied on explicit priors and registration assumptions that could become unreliable or less robust in the presence of motion, atypical anatomy, or mass effect—features commonly encountered in pediatric neurosurgical and developmental populations. In response, pediatric-specific model-based frameworks began incorporating stronger anatomical constraints directly into the inference process. For instance, Makropoulos et al. [33] proposed a neonatal whole-brain segmentation approach spanning very preterm to term-equivalent ages by introducing a structural hierarchy and anatomical constraints, improving overlaps against standard atlas-based techniques in a setting where both shape and appearance change rapidly. Importantly, these model-based measurement engines produced derived regional and voxel-wise descriptors that enabled whole-brain statistical analysis and classical predictive modeling on engineered features.

A defining methodological thread of this era was hypothesisdriven statistical mapping. Voxel-based morphometry and the general linear model formalized group comparisons as mass-univariate tests across the brain, enabling reproducible detection of developmental or disease-related differences when coupled to careful multiple-comparisons control and covariate modeling [1].

Beyond structural imaging, diffusion tensor imaging and resting-state functional MRI (fMRI) transformed images into interpretable indices and graphs. Diffusion-derived microstructural measures (e.g., fractional anisotropy) and connectivity matrices capturing correlated spontaneous activity [3,4]. These representations made it possible to ask developmental questions in terms of tracts and networks rather than slices and signals.

Classical machine learning then framed diagnosis and prognosis as feature-to-label mapping, using morphometry, diffusion metrics, connectivity features, and radiomics-style descriptors. Large multi-site benchmarks such as attention deficit hyperactivity disorder (ADHD)-200 highlighted how site effects and non-imaging covariates can dominate performance, underscoring the limits of handcrafted representations in heterogeneous pediatric datasets [6]. In parallel, studies on high-risk infants suggested that early functional connectivity patterns could anticipate later autism spectrum disorder diagnosis, illustrating both the ambition and the constraints of early predictive modeling [18]. In pediatric neuro-oncology, texture-based features were explored to discriminate posterior fossa tumor subtypes from routine imaging [37]. Nevertheless, the earliest clinical traction often arose not from broad predictive modeling, but from automating specific decision-support measurements with immediate bedside relevance.

In conditions like hydrocephalus, the field prioritized fast, interpretable quantification that could be trended over time. Linear indices, including the Evans index and the frontal-occipital horn ratio offered pragmatic surrogates for ventricular size across modalities [21,36]. These measures supported longitudinal communication of disease burden and were routinely used to monitor response to cerebrospinal fluid diversion and to inform decisions regarding shunt placement or revision. At the same time, they inevitably compressed complex 3D ventricular geometry into a small set of 2D ratios, limiting sensitivity to regional shape changes.

To improve objectivity and follow-up, researchers began developing automated ventricle segmentation and volumetric pipelines based on rule-based and model-based image processing. Representative CT work combined anatomical heuristics (e.g., midline detection), intensity modeling, and post-processing constraints to identify ventricular systems and approximate volumetric surrogates [11]. While these pipelines promised more direct quantification for clinical monitoring, they could fail under hemorrhage, severe deformation, or postoperative change, exactly where robust decision support is most needed.

Taken together, the pre-deep learning era made pediatric neuroimaging analyzable by defining measurement spaces, building model-based anatomical engines, and engineering biomarkers for classical prediction. Its persistent bottlenecks—manual feature design, sensitivity to protocol and development, and brittleness in atypical anatomy—set the agenda for deep representation learning in the current stage [6,20].

A fundamental limitation of conventional neuroimaging analysis is its reliance on handcrafted descriptors and human-defined visual cues, which are often insufficient to capture subtle and spatially distributed abnormalities. Such approaches are inherently constrained by prior assumptions about what constitutes a relevant feature, limiting their ability to generalize across heterogeneous imaging protocols, age groups, and disease phenotypes. Deep feature learning fundamentally departs from this paradigm by enabling models to automatically learn hierarchical representations directly from imaging data, bypassing manual feature engineering.

CNNs were the earliest and most widely adopted architectures for deep feature extraction in neuroimaging. By leveraging spatially constrained receptive fields and shared convolutional kernels, CNNs efficiently model local anatomical patterns while preserving spatial coherence. This design allows the encoding of textural, morphological, and intensity-based variations across multiple scales. Across both adult and pediatric studies, CNN-derived representations have consistently outperformed handcrafted features in tasks such as lesion detection, tissue segmentation, and structural abnormality characterization.

In neurosurgical imaging, CNN-based architectures have been widely applied to a broad range of clinically relevant problems, such as neonatal brain MRI segmentation [28], tumor segmentation [30], hemorrhage detection [12], tumor subtype classification, and ventricular morphology analysis [39]. In epilepsy surgery, surface-based CNNs have demonstrated improved sensitivity to subtle cortical thickness and folding abnormalities, facilitating localization of epileptogenic substrates that may be inconspicuous on routine visual assessment [29,48]. Collectively, these studies established CNN-based models as a robust framework for static neuroimaging analysis across age groups, imaging protocols, and neurosurgical disease entities [10].

ViTs extend deep feature learning by addressing a fundamental limitation of convolutional architectures : the implicit bias toward locality. By replacing convolution with self-attention mechanisms, ViTs explicitly model long-range dependencies and global contextual relationships within imaging data. In neuroimaging applications, ViT-based models have been applied to brain tumor analysis and classification in MRI data, often achieving high accuracy by learning global contextual relationships that complement local texture and intensity features [50]. Hybrid architectures that integrate convolutional encoders with transformer blocks, such as transformer-augmented U-Net variants, have demonstrated improved segmentation performance of brain tumors by combining the strengths of local feature extraction with long-range dependency modeling [42]. Collectively, transformer-based approaches represent a powerful extension of deep representation learning for static neuroimaging analysis.

Anomaly detection provides a fundamentally different paradigm for neuroimaging analysis by shifting the focus from disease-specific classification to the identification of deviations from learned normative patterns. Rather than relying on predefined diagnostic labels, anomaly detection models are typically trained on large datasets of normal or near-normal brain images to learn the statistical distribution of typical anatomy and development. Regions or patterns that deviate from this learned distribution are subsequently flagged as potential abnormalities.

This framework is particularly well suited to neurosurgical imaging, where pathological entities may be rare, heterogeneous, or insufficiently characterized to support supervised learning. Early anomaly detection approaches relied primarily on autoencoder-based architectures, in which abnormalities were inferred from elevated reconstruction error. However, conventional autoencoders often overgeneralize normal anatomical variability, limiting sensitivity to subtle or spatially diffuse abnormalities. To address these limitations, more recent methods have introduced variational autoencoders and memory-augmented networks. Variational autoencoders impose probabilistic structure on latent spaces to model uncertainty and distributional shifts [55], while memory-augmented models explicitly store representative patterns of normal anatomy, improving sensitivity to atypical deviations [49]. Representative studies highlighted the clinical relevance of anomaly detection in neurosurgical and neurodevelopmental contexts. Pinaya et al. [38] developed an unsupervised 3D anomaly detection and segmentation framework combining variational autoencoder representations with transformer models trained on healthy brain MRI, showing superior detection of heterogenous pathological appearances defined by deviation from normal anatomy across multiple lesion types and dataset. In pediatric neuroimaging, studies examining neonatal MRI data have shown that unsupervised deep models trained on normative developmental scans can distinguish subtle morphological deviations associated with early brain injury or developmental anomalies, with models identifying abnormalities that were initially missed in radiological readings [40].

From a clinical perspective, anomaly detection is best understood as a complementary screening and prioritization tool rather than a replacement for disease-specific diagnostic models. By highlighting cases or regions that deviate from normative patterns, anomaly detection systems guide expert attention toward potentially abnormal findings that may otherwise be overlooked, especially in rare, novel, or poorly characterized conditions, expanding the scope of static neuroimaging analysis beyond known disease entities.

Generative AI introduces a complementary paradigm to discriminative deep learning by modeling the underlying data distribution rather than solely optimizing task-specific predictions. In static neuroimaging, generative models, such as generative adversarial networks (GANs) and diffusion-based frameworks have emerged as powerful tools for addressing fundamental constraints including data scarcity, imaging burden, and modality limitations.

Data scarcity remains a major obstacle in pediatric and rare neurosurgical conditions, where the prevalence of specific pathologies is inherently low and the acquisition of large, well-annotated datasets is often impractical. GANs have been shown to synthesize anatomically plausible brain images and lesion-specific variations, effectively augmenting limited datasets and mitigating class imbalance. For example, Bowles et al. [5] demonstrated that GAN-based augmentation of brain MRI lesions can introduce realistic structural variability while preserving clinically meaningful imaging characteristics, leading to improved generalization of downstream segmentation models under low-data conditions. Similar challenges are observed in neonatal brain MRI, where synthetic learning approaches using contrast-independent, label-driven image generation have been shown to enable robust brain segmentation despite limited annotated data and strong domain variability [52]. Such generative augmentation extends beyond conventional transformation-based techniques by expanding the diversity of neuroanatomical and pathological patterns available for model training.

Beyond data augmentation, generative AI has been increasingly applied to cross-modality image synthesis to compensate for missing or degraded imaging sequences and to harmonize heterogeneous datasets. MRI-to-MRI synthesis has been explored to reconstruct missing data and improve robustness under heterogeneous acquisition protocols. Cross-modality synthesis, MRI-to-PET generation has enabled approximation of metabolic information directly from structural MRI (sMRI) [8]. Similarly, MRI-to-CT synthesis using GAN has been applied to generate CT-equivalent images from T2-weighted MRI without additional radiation exposure [41]. These approaches are especially relevant in pediatric neurosurgery, where minimizing radiation dose and invasive imaging procedures is a critical clinical priority. By providing surrogate imaging contrasts that preserve diagnostically and procedurally relevant information, generative models offer a practical pathway to reduce imaging burden while maintaining the information required for neurosurgical decision-making.

AI in neurosurgery has evolved from task-specific, offline models toward systems that are increasingly integrated across the clinical workflow. Despite this progress, most current systems remain limited in their ability to adapt to intraoperative variability and continuous data streams. These limitations are particularly pronounced in pediatric neurosurgery, where rapid neurodevelopment and anatomical variability challenge static, adult-centric models. Accordingly, the field is transitioning toward dynamic and interactive AI systems that operate across surgical phases and support real-time clinical decision-making. Here, static AI refers to offline or single-time-point analytic systems, whereas dynamic AI denotes temporally adaptive models integrated within the surgical workflow.

The applications of AI in neuroimaging is undergoing a fundamental transition from task-specific algorithms to brain foundation models. Unlike traditional deep learning models trained for a single purpose (e.g., tumor segmentation), brain foundation models leverage large-scale pre-training on diverse neural signals, spanning fMRI, electroencephalogram (EEG), and sMRI, to learn generalized brain representations that can be adapted to various downstream clinical tasks [27,64]. This shift is particularly transformative for pediatric neurosurgery, where labeled data is scarce, and pathologies are highly heterogeneous.

Recent studies demonstrate the capacity of such models to capture complex brain structure-function relationships [53]. Furthermore, for conditions like epilepsy which require high temporal resolution, Neuro-GPT applies a masked autoencoder approach to EEG data. In parallel, masked autoencoder-based approaches applied to EEG data have shown robust performance in seizure classification under low-data conditions, offering a potential strategy for monitoring rare pediatric epilepsy syndromes [14].

In structural imaging, emerging foundation models address domain-shift challenges that commonly degrade the performance of adult-trained algorithms when applied to pediatric brains. Large-scale three-dimensional segmentation models enable zero-shot or few-shot adaptation across imaging modalities, while contrastive learning approaches emphasize anatomical consistency across developmental stages [13]. Together, these methods provide a framework for longitudinal tracking of morphological change in the developing brain and for improving generalization across age groups and acquisition protocols [2].

Spatial computing addresses a longstanding challenge in neurosurgery : translating two-dimensional imaging data into three-dimensional surgical understanding. This challenge is amplified in pediatric neurosurgery, where anatomical structures are smaller and developmental variations are substantial.

Clinical studies have shown that immersive visualization can influence surgical strategy in complex pediatric cases, including craniosynostosis, vascular malformations, and central nervous system tumors [7]. Virtual reality-based rehearsal allows surgeons to explore anatomical relationships preoperatively and refine operative trajectories, while mixed reality systems can overlay three-dimensional models onto the operative field to support intraoperative orientation. Compared with conventional navigation systems that require attention shifts to external monitors, head-mounted displays allow hands-free interaction while maintaining visual focus on the surgical field, which may improve ergonomic efficiency during prolonged procedures.

Beyond operative guidance, immersive technologies have demonstrated additional value in pediatric care by reducing perioperative anxiety and procedural pain. Augmented and virtual reality interventions have been associated with reduced stress markers and subjective pain scores in children undergoing invasive procedures, supporting their role as adjunctive, non-pharmacological tools in pediatric neurosurgical workflows [43,56]. Emerging applications also include remote collaboration and tele-mentoring, enabling expert guidance during complex procedures in settings with limited local expertise [35].

Physical AI represents the integration of computer vision and robotic systems to extend AI beyond image interpretation into physical interaction with the surgical environment. In pediatric neurosurgery, where operative corridors are narrow and tissue tolerance is limited, this integration is particularly relevant for minimizing iatrogenic injury.

Recent computer vision approaches enable real-time recognition of surgical instruments and tissue types within the operative field. These systems support workflow analysis, safety monitoring, and intraoperative visualization by identifying instruments and segmenting anatomical structures in real time [61]. In parallel, robotic platforms provide motion scaling and tremor filtration, which are advantageous for high-precision pediatric procedures performed in confined spaces. However, hardware constraints, including instrument size and workspace limitations, remain significant barriers to widespread pediatric adoption [58].

Current robotic systems largely operate under direct surgeon control, but ongoing research explores semi-autonomous assistance through AI-driven feedback mechanisms [9]. Vision-based estimation of tissue properties has been proposed to compensate for the lack of tactile feedback, while preoperative imaging-derived constraints may be used to define virtual safety boundaries that restrict robotic motion near critical structures [44]. Although these concepts remain under active development, they illustrate a trajectory toward AI-assisted environments that enhance precision and safety without displacing surgical authority.

Generative AI and multimodal agent systems represent a shift from perceptual automation toward cognitive support within the neurosurgical workflow. Early applications of large language models focused on text-based tasks, including clinical documentation and knowledge retrieval, where performance comparable to trainee-level responses has been reported. More recent multimodal models extend these capabilities to the interpretation of medical images, enabling joint reasoning across visual and textual inputs.

Although current image interpretation accuracy remains limited, early studies demonstrate the feasibility of generating preliminary differential diagnoses and contextual explanations from radiological data. Domain-specific vision-language models trained on neurosurgical images and captions further improve task relevance by embedding specialty-specific visual semantics. Additional work has proposed treating volumetric medical images as temporal sequences, enabling generative models to analyze three-dimensional data in a manner analogous to video processing.

Beyond interpretation, agent-based systems are being explored to coordinate information flow and support clinical decision-making. Multi-agent frameworks capable of planning, execution, and optimization have demonstrated efficiency gains in treatment planning tasks, highlighting the potential for reducing cognitive burden in complex workflows [57,63]. Nevertheless, challenges related to hallucination, bias, and reliability remain substantial. Current consensus emphasizes the necessity of human oversight and rigorous validation, particularly in high-stakes pediatric neurosurgical environments, to ensure that generative systems function as supportive tools rather than autonomous decision-makers [60].