Abstract
Most neuroimaging applications involve a multi-step pipeline encompassing image acquisition, reconstruction, enhancement, registration, segmentation, diagnosis, and prognosis. These steps are fundamental for diagnosing neurological disorders, tracking disease progression, and guiding treatment. Recent advances in artificial intelligence (AI), particularly deep learning, are transforming each stage by improving efficiency, image quality and resolution, accuracy, and clinical utility. This review surveys recent advances and emerging trends across diverse neuroimaging modalities, including multimodal studies that integrate imaging with clinical and molecular data. New AI approaches address long-standing challenges in neuroimaging: physics-informed models incorporate prior knowledge to improve reconstruction, self-supervised learning mitigates the lack of ground-truth data in incomplete datasets, graph neural networks capture the non-Euclidean nature of connectomics, generative diffusion models predict missing contrasts and enable cross modal synthesis, and data harmonization techniques reduce scanner and site variability to improve generalizability. Despite these advances, key barriers such as heterogeneous and biased datasets, limited benchmarking, and regulatory challenges impede the translation of these methods into clinical workflows. We conclude by highlighting priorities for developing reliable, generalizable, and interpretable that can advance both neuroimaging research and real-world patient care.
Introduction
Neuroimaging plays a central role in diagnosing neurological diseases, 1 monitoring progression, and guiding treatment. 2 Yet, traditional pipelines remain constrained by lengthy scan durations, complex reconstructions, and labor-intensive processing steps such as segmentation and registration. As patient volumes grow and scanners produce ever larger datasets, these challenges increasingly strain clinical workflows.
Artificial intelligence (AI), particularly deep learning (DL), has emerged as a transformative solution to these challenges.3–5 Unlike hand-engineered pipelines, these models learn directly from raw data, capturing complex patterns in high-dimensional datasets that often escape conventional analysis. Over the last decade, such methods have moved from proofs-of-concept to clinically relevant tools, demonstrating their ability to accelerate and stabilize scans, 6 reconstruct or enhance signals under challenging undersampled conditions,7–9 segment anatomy with greater consistency, 10 and robustly register multi-contrast or multi-subject data.11,12 In this review, we discuss applications of AI across magnetic resonance imaging (MRI), positron emission tomography (PET), and computed tomography (CT), which represent key modalities relevant to neuroimaging research. Other modalities such as single-photon emission computed tomography (SPECT) remain important in specific clinical contexts, and recent work has highlighted the growing role of AI in this field. 13
The rapid evolution of DL architectures, coupled with improvements in accuracy and scalability, has enabled AI to keep pace with the expanding size and complexity of neuroimaging datasets as scanners and patient throughput continue to advance. Beyond image-level improvements, these methods now extend to higher-order neuroimaging tasks, such as mapping the structural and functional connectome, integrating multimodal data for diagnostic classification, and supporting applications in disease subtyping, prognosis and staging. In this review, we focus on recent and emerging AI methodologies, with particular emphasis on technical innovations that are reshaping neuroimaging analysis across modalities and tasks. Figure 1 provides an overview of the scope and organizational structure.
Schematic overview of the review scope and structure, organized into five major sections. The numbers shown in the figure correspond to the number of papers reviewed in each section.
The remainder of the paper is organized around the stages of a neuroimaging study, tracing the path from preprocessing raw data to clinical application. We first describe the literature search strategy and selection criteria. We then survey reconstruction and enhancement approaches, where AI has been shown to accelerate acquisition, recover missing measurements, and improve image quality. This is followed by a review of segmentation, parcellation, and registration methods, which provide the anatomical maps required for clinical diagnosis and group-level comparisons. We next turn to connectomics, covering both the structural and functional organization of the brain. Advances in diagnostic classification across neurodegenerative and developmental disorders are then discussed, followed by clinical applications including disease subtyping, prognosis, and staging. The review concludes with a discussion of persistent challenges, future directions, and overarching perspectives.
Search strategy and selection criteria
We conducted a PubMed search across major neuroscience and medical imaging journals using method-related terms (AI, DL, transformers, contrastive learning, self-supervised learning, graph neural networks (GNNs), convolutional neural network (CNN), diffusion models, generative adversarial networks (GANs), foundation models, large language models (LLM)) combined with domain terms (neuroimaging, brain imaging, functional MRI (fMRI), diffusion tensor imaging (DTI), connectivity, segmentation, registration, tractography). The Boolean query and search time window are provided in Supplemental Material S1. Given the large volume of AI-based neuroimaging research, we restricted the search to a 2-year period (August 2023–August 2025) and applied manual screening to prioritize methodological innovation and depth.
Our scope was limited to peer-reviewed, English-language original research articles applying AI methods to human neuroimaging data, published in established journals with broad readership in imaging, neuroscience, or biomedicine. We excluded non-peer-reviewed formats (conference abstracts, editorials, commentaries), studies not involving humans, non-imaging work (e.g. drug–target modeling), electroencephalography (EEG) or ultrasound studies, cardiac DTI, and publications in venues lacking methodological rigor. When multiple versions existed, journal articles were prioritized over preprints. Screening emphasized studies employing DL or clearly impactful machine learning (ML) approaches, with preference for those reporting external or multi-site validation and transparent datasets, code, or evaluation protocols. Articles categorized as having insufficient information support lacked sufficient methodological detail to assess model architecture, training procedure, or evaluation strategy, or did not report quantitative results relevant to the stated task. Borderline cases, such as studies using ML methods without substantial methodological contribution or studies where neuroimaging played a peripheral role, were resolved by retaining the article if it introduced a clearly novel modeling strategy, addressed an important neuroimaging task, or provided insight into generalization, interpretability, or clinical relevance. Preference was given to studies reporting external or multi-site validation, or those providing transparent descriptions of datasets, code, or evaluation protocols.
While the PubMed search served as the primary source of article identification and yielded most studies included in this review, additional studies were identified through targeted literature searches to improve coverage of rapidly evolving areas that were underrepresented in the primary search results. These studies were selected based on methodological novelty, relevance to neuroimaging AI, and their contribution to emerging areas of active research, including multimodal learning, large language models, advanced segmentation and registration methods, and quantitative imaging applications such as cerebral blood flow and perfusion estimation.
Figure 2 presents the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart summarizing the article selection process.
PRISMA flow diagram outlining the study selection procedure for this review, showing the stages of identification, screening, eligibility evaluation, and inclusion.
Reconstruction and enhancement
Conventional image reconstruction and processing approaches are fundamentally constrained by three factors: noisy and incomplete inputs, simplistic forward models that imperfectly approximate the physics of actual image acquisition, and regularization strategies that require manual hyperparameter selection. These limitations lead to noisy images, loss of clinically relevant detail, and reduced generalizability. Recent advances in AI are reshaping this landscape by shifting the focus from purely model-based approaches to hybrid approaches that learn mappings directly from data while embedding physics-based constraints into the forward model. Within this paradigm, methods can be grouped into four complementary directions in processing images: reconstruction (recovering images from raw acquisitions), enhancement (denoising, super-resolution, and other post-reconstruction improvement of image quality), synthesis (generating missing contrasts or translating across modalities), and harmonization (reducing cross-scanner and cross-site variability). Together, these innovations address long-standing bottlenecks in neuroimaging pipelines, providing higher-quality images for downstream use. Table 1 presents 18 representative studies organized by imaging target, modality, dataset size, and underlying architecture, while Figure 3 illustrates representative architectural commonly used in reconstruction and enhancement tasks, highlighting design elements discussed in this section.
Summary of advanced AI approach for reconstruction and enhancement.
An illustration of advanced AI frameworks for reconstruction and enhancement. The figure presents four widely used models: U-Net, GAN, transformer, and diffusion, which have been applied to image generation tasks such as denoising, deblurring, and reconstruction. Key characteristics of these models are highlighted (e.g. convolutional layers, skip connections, attention mechanisms, adversarial training, and iterative denoising). To improve image quality and fidelity, the training process incorporates tailored optimization strategies, loss functions, and evaluation criteria.
Reconstruction
Reconstruction in neuroimaging has traditionally been approached as an inverse problem based on a forward model of acquisition. However, imperfections in measurement often violate the assumptions of this model, causing noise and artifacts to propagate into the reconstructed image. As a result, conventional methods are limited by the need for long scan times, sophisticated quantitative corrections, and repeated acquisitions to obtain images for clinical utility. AI-based methods possess the potential to accelerate image reconstruction and directly estimate quantitative parameters.
One study demonstrated that DL methods can overcome the speed bottleneck by applying an unrolled variational network to accelerated 3 T FLAIR acquisitions in MRI, 14 that use the same acquisition parameters as conventional imaging but with higher acceleration factors for parallel imaging. Thus, with a shorter acquisition time, DL-accelerated FLAIR demonstrated improved sharpness and lesion detectability than conventional FLAIR. Another unrolled network employed a physics-driven, 15 self-supervised CNN for 10-fold accelerated multi-echo spiral fMRI acquisitions that are suitable for denser sampling but inherently slower to reconstruct. Despite aggressive undersampling, the method preserved blood oxygen level dependent (BOLD) signal fidelity, highlighting the benefit of embedding acquisition physics directly into the model.
The ability of AI extends to being able to estimate parameters directly from data. A residual network (ResNet) based network was trained to estimate arterial transit time and cerebral blood flow (CBF) from only one or two post-labeling delays (PLD), 16 eliminating the need for lengthy multi-delay arterial spin labeling (ASL) perfusion MRI. Similarly, a combined 1D CNN and a 2D U-Net was trained on 4D magnetic resonance (MR) perfusion time-series data to predict voxel-wise relative perfusion parameters relative cerebral blood volume (CBV) and relative CBF, 17 bypassing conventional time-consuming software pipelines used in dynamic susceptibility contrast (DSC) MR perfusion.
The benefits extend further to diffusion MRI (dMRI), where conventional acquisitions require numerous diffusion directions and b-values to capture microstructural information. A geometric DL method was introduced that reconstructs diffusion signals for arbitrary q-space samplings. 18 By learning directly in the sampling space rather than voxel space, the method generalizes across acquisition schemes, predicts unseen q-vectors, and preserves microstructural information. Diff-DTI, a diffusion model coupled with a transformer, 19 which recovers fractional anisotropy and mean diffusivity maps from as few as one b0 and three to six diffusion directions, compresses acquisition time by up to 10-fold while maintaining tensor contrast.
Enhancement
Neuroimaging data are often degraded by noise, motion, partial volume effects, and limited spatial resolution which can obscure subtle pathology and reduce diagnostic confidence. Conventional denoising and filtering methods suppress noise but risk blurring fine structures. AI-based enhancement techniques aim to recover high-quality images directly from low-resolution or noisy inputs while identifying and retaining finer structures.
For example, the image quality loss was addressed in radial-based silent fMRI, which is typically undersampled, by training 2D and 3D U-Nets on simulated “Looping Star” acquisitions to recover higher-quality ground truth. 20 In diffusion imaging, a multi-layer perceptron (MLP) was applied for spatial super-resolution of fiber orientation distribution functions (fODFs) represented in the spherical harmonic domain, enhancing fiber orientation estimates and improving the reliability of downstream tractography. 21 Another study demonstrated that a 1D CNN using BOLD signals and head motion parameters can reconstruct the respiratory-variation waveform directly from resting-state fMRI (rs-fMRI), enabling physiological corrections without external recordings. 22
Synthesis
In neuroimaging, synthesis refers to the generation of new image representations to fill missing information or complement existing data. This can take two main forms: within-modality synthesis, where methods enhance or reconstruct data within the same imaging sequence; and cross-modality synthesis, where one imaging modality is predicted from another. Trained DL models possess the ability to understand complex patterns of real neuroimaging data so that it can produce realistic and task-dependent synthetic brain images, both within-modality augmentation and cross-modal generation.
In structural imaging, a GAN with a U-Net generator synthesized 7 T-like MPRAGE images from 3 T inputs, producing high-field advantages for morphometry and lesion visualization when 7 T is unavailable. 23 Cross-modal synthesis was achieved using a 3D dense U-Net trained to synthesize tau-PET from MRI, fluorodeoxyglucose PET (FDG–PET), or amyloid-PET. 24 The creation of “virtual” tau maps across Mayo Clinic Study of Aging (MCSA) and Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohorts demonstrates direct clinical implications for Alzheimer’s disease (AD) research. At the decoding frontier, a dual-guided diffusion model reconstructed natural images from visual stimulus fMRI with improved semantic fidelity over prior GAN and variational autoencoder (VAE) based models. 25
At the population scale, the text-guided universal MRI synthesis (TUMSyn) framework was developed, 26 which both synthesizes missing MRI sequences and performs arbitrary-scale super-resolution. By combining routine MRI with imaging metadata and structured text prompts, this transformer-CNN hybrid was trained across 13 cohorts with more than 30,000 scans. Finally, diffusion microstructure estimation was addressed with a two-stage deep learning framework, coupling a CNN with a mixture density network for multi-fiber parameter estimation, trained on synthetic signals and validated on Human Connectome Project (HCP) data. 27
Harmonization
Variability across scanners, sites, and acquisition protocols remains a major barrier to the reproducibility and clinical translation of AI methods in neuroimaging. Data harmonization techniques offer flexible and data-driven solutions that improve the generalizability of models across cohorts.
At the feature level, DeepComBat was proposed to harmonize cortical thickness across sites using a conditional VAE with statistical batch-effect correction, 28 showing strong robustness to hyperparameter choices while preserving biological signals. At the image level, multi-scanner image harmonization via structure preserving embedding learning (MISPEL), 29 a supervised deep learning framework, harmonized multi-scanner structural MRI by mapping images into a structure-preserving middle-ground space. In molecular imaging, a surface-based diffusion model harmonized tau-PET standardized uptake value ratios (SUVRs) between ADNI and Health and Aging Brain Study–Health Disparities (HABS–HD) cohorts, 30 improving cross-study comparability while retaining individual variability. Within dMRI, a DL pipeline combining Attention U-Net for contrast normalization with residual encoder-decoder network (REDNet) for super-resolution reduced scanner- and field strength-related batch effects while maintaining lesion reproducibility. 31
Integrative comparison and implications
In reconstruction, recent progress has mainly come from using data-driven priors to reduce acquisition demands. Unrolled and physics-driven CNNs shorten scan time by incorporating the acquisition process into the model, enabling faster MRI and fMRI while preserving lesion contrast or BOLD signal. Other approaches bypass full image reconstruction and directly predict quantitative measures, such as CBF or CBV, simplifying both acquisition and post-processing. In diffusion MRI, two complementary ideas are being explored: methods that aim to remain flexible across acquisition schemes by learning diffusion signals in sampling space, and more aggressive approaches that recover diffusion metrics from very few measurements, with validation so far largely limited to research datasets.
Enhancement methods mainly address factors that affect downstream analysis, including low resolution, undersampling artifacts, and missing physiological signals. U-Net models are commonly used when spatial detail is critical, while simpler architectures such as MLPs or one-dimensional CNNs are better suited to signal-level features, improving tractography reliability or enabling physiological correction without external recordings.
Synthesis and harmonization target missing information and cohort heterogeneity. Generative models make it possible to impute missing contrasts or modalities at scale, but their outputs depend strongly on the training data and can be difficult to trust in clinical settings. In contrast, harmonization methods that explicitly address scanner and site effects are more directly focused on improving robustness across cohorts.
While these approaches demonstrate clear improvements in image quality, acquisition efficiency, or quantitative estimation, many have primarily been evaluated on research cohorts under controlled conditions. Broader validation across scanners, acquisition protocols, and clinical populations remains necessary to establish their robustness and generalizability.
Segmentation, parcellation, and registration
Accurate segmentation and alignment of brain images are critical component of workflows that diagnose neurological and neurodegenerative disorders or monitor tumors and lesions. Traditionally, these tasks have relied on manual annotation and alignment, which are labor-intensive, time-consuming, and prone to variability across observers and scans. AI-based approaches now offer automated, faster, and more reliable alternatives to these traditional methods. By learning from large datasets, they can delineate brain tissues such as gray matter (GM), white matter (WM), and cerebrospinal fluid with high accuracy, while also differentiating brain lesions from artifacts. In this context, segmentation generally refers to separating tissues and lesions, while parcellation denotes subdividing the brain into anatomical or functional regions. This section reviews the recent advances in AI methods for the segmentation, parcellation and registration, which together are critical for quantitative assessment of neuroimages. A summary of the studies covered in this section is shown in Table 2.
Studies included in advanced AI-based segmentation, parcellation, and registration.
Segmentation
Segmentation, the process of identifying and labeling regions of interest (ROI), is essential for tasks such as tumor volume estimation, lesion tracking, and tissue characterization in neurological and neurodegenerative disease. Recent DL approaches have advanced neuroimaging segmentation by improving accuracy, speed, and reproducibility, with applications spanning different model architecture (e.g. CNNs, transformers, hybrid frameworks), target structures (e.g. tumors, WM, acute infarcts), and imaging modality (single-sequence MRI, multi-sequence MRI, or dMRI).
Segmentation is especially important for small structures, where voxel-level precision is essential, and CNNs remain a strong baseline due to their ability to capture local detail. An edge-aware U-Net incorporating boundary guidance segmented glioma, 32 meningioma, and pituitary tumors from the Cheng 2015 dataset of 3064 contrast-enhanced T1 MRI. In parallel, many studies report results on the glioma focused brain tumor segmentation challenge (BraTS) benchmarks to enable standardized comparison across years. Capsule networks with Visual Geometry Group (VGG) layers were combined to develop a hybrid Caps-VGGNet for multi-grade tumor segmentation (glioma, meningioma, pituitary, and normal) using BraTS2020 for training and validation and BraTS2019 for testing, 33 achieving robust performance in glioma benchmarking.
Generative models have also been explored on this dataset. A multi-scale diffusion model with axial attention and multi-scale feature fusion (MSDMAT–BTS) was applied to glioma segmentation using multimodal BraTS MRI (T1, T1ce, T2, and FLAIR), 34 with experiments on BraTS2019 (n = 335) and BraTS2020 (n = 369) demonstrating improved robustness and accuracy over state-of-the-art baselines. Corrective diffusion (CorrDiff) introduced a corrective diffusion framework that refines U-Net outputs by addressing systematic segmentation errors. 35 This model was evaluated on BraTS2019, BraTS2020, and Cheng 2015 dataset (3,064 T1-CE tumor images), it showed improved segmentation of small tumors and boundary regions. At the frontier of foundation models, segment anything model (SAM) was evaluated for interactive glioma auto-segmentation on contrast-enhanced T1 MRI from BraTS2020 in the context of radiotherapy planning, 36 where it showed strong accuracy, underscoring the promise of foundation segmentation models in clinical neuroimaging. Foundation model pretraining is also emerging: BrainSegFounder pretrains a 3D transformer-U-Net framework on 41,400 UK Biobank (UKB) T1/T2-FLAIR scans and then adapts to BraTS tumors and ATLAS v2.0 stroke lesions, 37 showing superior performance with strong few-shot and modality flexibility. Finally, incremental architectural improvements focused on feature-fusion strategies introduced four feature-enhanced hybrid U-Net models (FE-HU-NETs), 38 which integrate image enhancement, customized U-Net layers, and CNN components with DeepLabv3 (a segmentation model using atrous convolutions for multi-scale context), reaching over 99% accuracy on two public datasets.
Beyond gliomas, segmentation methods are increasingly being applied to vascular and small-vessel diseases. In acute stroke a 3D U-Net model trained on 10,820 diffusion-weighted MRIs from 10 hospitals was evaluated on an internal cohort (n = 2159) and multicenter (n = 2777) cohorts, 39 along with additional datasets of atrial fibrillation (n = 50) and ISLES 2022 (n = 250), underscoring that both large-scale multi-site data and domain adaptation markedly improve infarct segmentation generalizability. For small-vessel disease, uncertainty quantification was incorporated into WM hyperintensity (WMH) segmentation using U-Net and no-new U-Net backbones. 40 Evaluations across ADNI, cerebrovascular disease (CVD), WMH segmentation challenge and multi-study small vessel disease segmentation (MSS3) datasets showed that calibrated probability and uncertainty maps improve Fazekas score prediction and support quality control (QC) in clinical setting.
Parcellation
Parcellation divides the brain into distinct regions (parcels), typically based on structural MRI (sMRI). However, diffusion-space parcellation suffers from mislabeling when projecting T1-based atlas onto dMRI. This challenge is addressed by deriving FreeSurfer Desikan–Killiany parcellations directly from DTI maps: fractional anisotropy (FA), mean diffusivity (MD), and tensor eigenvalues (E2, E3). 41 A CNN framework with multi-level fusion and tri-planar aggregation was trained and validated on HCP and evaluated for generalization on Consortium for Neuropsychiatric Phenomics (CNP), Parkinson’s Progression Markers Initiative (PPMI), and an in-house cohort. It improved both test-retest reproducibility and regional homogeneity compared to T1-registered baselines, with demonstrated generalizability across populations.
Registration
Neuroimage registration is a prerequisite for constructing population-based brain atlases, detecting group-level patterns, and comparing datasets across subjects, modalities, and timepoints. Conventional registration algorithms rely on iterative optimization process that are computationally intensive and further challenged by anatomical variability, motion artifacts, and the complexity of non-rigid transformations. These limitations can introduce errors like anatomical distortion and make large-scale studies difficult to manage. AI-based registration methods overcome many of these barriers by learning to align images directly, substantially reducing computation time from hours to seconds while preserving anatomical fidelity. These models are particularly effective at deformable MRI registration and cross-modal alignment, with the added capacity to retain clinically important local features such as lesions.
At the model level, Attention-enhanced Dual-Stream Registration Network (ADRNet) combines CNNs to extract shallow features, 42 a mixed-attention transformer for capturing self-attention, cross-attention, and local attention to capture correlations within and across images, and a gated adaptive fusion mechanism to enforce one-to-one matches across scales. U-Net based deep diffusion MRI template (DDTemplate) is a framework for groupwise dMRI registration that integrates whole-brain FA and tract orientation maps through a two-stage process of tract-specific and whole-brain template creation. 43 Trained on HCP young adult (HCP-YA) and tested on HCP-YA, adolescent brain cognitive development (ABCD), and PPMI, it shows strong cross-study generalization without requiring ground-truth deformation fields and enhances downstream analyses such as tract-based spatial statistics (TBSS) based sex difference detection. Moving to pipeline-level integration, DeepPrep is a scalable preprocessing pipeline that incorporates deep-learning modules for segmentation, 44 surface reconstruction, registration, and normalization. Evaluated on 55,000 scans across seven datasets, it demonstrated improved robustness and harmonization in real-world neuroimaging workflows.
Methodological comparison and practical considerations
In segmentation, most recent work still builds based on CNN models, particularly U-Net, because they handle voxel-level detail well and are relatively stable in practice. Recent work, however, extends beyond incremental architectural refinement. Diffusion-based segmentation introduces iterative or corrective mechanisms that improve boundary delineation and the detection of small lesions, while foundation models and large-scale pretraining, including transformer-based designs and interactive frameworks such as SAM, support improved transfer across tasks, modalities, and annotation regimes. Despite these advances, performance in clinical and multi-site settings is often strongly influenced by training data diversity, external validation strategies, and the availability of uncertainty estimates to support quality control.
Parcellation addresses a related but distinct challenge, focusing on the consistency of region definitions. Learning parcellations directly in diffusion space reduces error propagation associated with cross-modality atlas projection and has been shown to improve reproducibility. Although fewer studies exist in this area, parcellation plays a critical role in standardizing downstream analyses, and its impact is often indirect but substantial. At present, evidence remains limited to a relatively small number of studies and moderate cohort sizes.
Registration appears closer to routine use. Learning-based deformable registration substantially reduces computation time compared with traditional optimization-based approaches while maintaining acceptable anatomical consistency. Given that registration accuracy is already high with established methods, recent AI-based approaches primarily focus on improving efficiency and reducing manual intervention.
Taken together, these methods illustrate a trade-off between methodological sophistication and practical deployment. Foundation models and large-scale pretraining frameworks improve transferability and reduce annotation requirements but often require substantially greater computational resources and training data than conventional CNN-based approaches.
Structural and functional connectomics
Structural connectomics maps the anatomical wiring of the brain to provide a blueprint of the physical infrastructure, while functional connectomics captures how different brain regions communicate and coordinate in real-time. Together, they characterize the complete set of neural connections and link the organization of the nervous system to cognition and disease. Conventional approaches rely on handcrafted features and correlation-based measures, but these are often limited by noise, variability, and the high-dimensionality of neuroimaging data which make it hard to discern abstract patterns. This ties directly to the strength of DL frameworks, which can identify hidden features and patterns in high-dimensional data. This form of data-driven representation learning directly from imaging inputs. On the structural side, point cloud and graph formulations have enabled scalable tract-level representations and WM parcellations, while on the functional side, GNN and transformers now capture multiscale dependencies, sample-wise, spatial, and temporal, directly from BOLD time series. A concise overview of the representative studies discussed in this section is provided in Table 3.
Summary of advanced AI approach for structure and function connectomics.
Structural connectome (dMRI and tractography)
The structural connectome of the brain is typically obtained by reconstructing WM pathways from dMRI. Traditional tractography pipelines rely on clustering or atlas-based parcellations, which are sensitive to preprocessing steps and are unable to capture incomplete structures, limiting their scalability to multi-site cohorts. DL approaches are reshaping this landscape by learning tract representations directly from diffusion data. One such self-supervised implementation, deep fiber clustering (DFC), 45 introduces a point cloud framework that couples streamline geometry with GM parcellation, producing fast parcellations across three independent cohorts. Similarly, TractGraphFormer integrates graph CNN and transformer modules within anatomically informed graphs to capture both local geometry and global dependencies of white matter connections, 46 achieving strong age and sex prediction in ABCD and HCP-YA, underscoring that distributed WM connections encode demographic signals. Building on geometry to elevate shape as a first-class feature, TractShapeNet learned multiple shape descriptors directly from streamline point clouds in the HCP-YA cohort, 47 outperforming point clouds baseline methods in accuracy, running faster than conventional tools, and supporting downstream cognition prediction.
Besides full-brain approaches, robustness to incomplete acquisitions is addressed by TractCloud-FOV, 48 which augments training with synthetically cut tractograms and validates on real partial field of view (FOV) datasets. It improves tractography parcellation when cerebellar and brainstem coverage is limited, while also enhancing consistency on complete tractograms. New WM territories are brought into routine analysis as well; the superficial white matter analysis (SupWMA) framework employs a two-stage point cloud network with supervised contrastive learning to classify 198 superficial WM (SWM) clusters, 49 achieving fast and consistent parcellation across six independent datasets spanning different populations and acquisition protocols. Methods that embed spatial and anatomical constraints into streamline tracking reduce error propagation and show robust performance across public benchmarks such as TractoInferno and the ISMRM 2015 Challenge. 50 Finally, diffusion microstructure estimation in neonates and fetuses is addressed by a U-Net based model trained on the developing HCP (dHCP), 51 predicting fiber orientation distributions (FODs) from 6 or 12 dMRI measurements and validated against histology, linking tissue- and tract-level representations.
Functional connectome (rs-fMRI and task-fMRI)
Conventional approaches to functional connectomics rely on correlation matrices or independent component analysis (ICA), which are limited in their ability to capture temporal dependencies and are challenging to scale to population-level datasets. AI methods address these limitations by learning representations that are temporally aware, robust across individuals, and transferable across cohorts. One such study using pure transformer modules to capture sample-wise, regional and temporal interdependencies in fMRI data, 52 and was evaluated on the Cambridge Centre for Aging and Neuroscience (Cam-CAN) and Nathan Kline Institute-Rockland Sample (NKI-RS) cohorts for age, IQ, and sex prediction, achieving superior performance over existing methods. Another innovation in the graph front is GraphCorr, 53 a plug-in GNN module that augments baseline classifiers by learning dynamic and lagged functional connectivity features via a transformer-based node embedder and lag filter. It was evaluated on HCP and ID1000 cohorts and demonstrated boosted sex classification and cognitive task decoding performance. Another study, BrainRGIN (brain ROI-aware graph isomorphism networks), 54 formulates functional connectivity graphs from rs-fMRI and, on ABCD and HCP cohorts, predicts fluid, crystallized, and total intelligence effectively. Using a dual-task fMRI design, an explainable extreme gradient boosting (XGBoost) with shapley additive explanations (SHAP) decoded the difficulty of noise-vocoded speech under divided attention, 55 highlighting a bilateral insular, superior frontal, and cingulate cortices as key regions for resources allocation between auditory and visual demands.
The recent generative approaches also find implementation in functional connectomics. One study introduced TransUNET–DDPM, a transformer-enhanced denoising diffusion probabilistic model (DDPM), which synthesizes subject-specific 3D intrinsic connectivity networks (ICNs) from rs-fMRI, 56 improving anatomical and functional fidelity and supporting applications such as schizophrenia (SZ) classification through data augmentation. Swin fMRI U-Net Transformer (SwiFUN) predicts task-evoked activation maps directly from rs-fMRI by capturing spatiotemporal representations with a transformer-U-Net architecture and contrastive loss. 57 It achieved accuracy comparable to test-retest reliability in UKB and exceeded it in ABCD, and outperforming both conventional and deep learning baseline methods. For clinical targets, an LLM driven, 58 task-oriented representation of functional networks improves amyloid-β status prediction across ADNI, open access series of imaging studies (OASIS), and hospital cohorts by aligning network features with the diagnostic objective. Adversarial graph contrastive learning (A-GCL) further strengthens diagnostic embeddings for neurodevelopmental disorders by pushing rs-fMRI graphs toward class-discriminative but augmentation-robust spaces. 59 A regression-assisted framework (RegAssist-cIVA) closes the loop by pairing scalable statistics with learned models to efficiently handle the large number of datasets encountered in multi-subject rs-fMRI, 60 while retaining interpretability and facilitating reliable cross-site comparisons in very large cohorts.
Multimodal learning
Multimodal learning seeks to align structural, functional, and anatomical information within a common representation space. By combining complementary signals from different imaging modalities, these approaches can bridge modality gaps. A bidirectional contrastive framework learns mappings between fMRI BOLD signal and dMRI derived brain networks, 61 preserving phenotype-relevant information across HCP and OASIS while reducing bias between one-way projections. Another cross-modal framework, FM-APP (foundation model for any phenotype prediction), leverages fMRI to sMRI knowledge transfer with a text memory bank and regressor synthesizer to enable zero-shot phenotype prediction across HCP and HCP-aging. 62 Functionally consistent tractography parcellation integrates fMRI signals into WM fiber clustering, 63 aligning structural units with functional organization and advancing toward unified connectome maps that respect both anatomy and activity.
Comparison across connectomic modeling approaches
Structure and functional connectomics both aim to characterize brain networks, but they rely on different imaging principles, which lead to distinct modeling approaches. In structural connectomics, many recent methods focus on how white matter pathways are represented, increasingly replacing atlas-based grouping with models that learn directly from streamline geometry. Point-cloud and graph-based approaches make it easier to handle large tractograms and incomplete coverage, which is useful in multi-site studies. At the same time, these methods take tractography outputs as their input, so choices made during tract reconstruction still impact the resulting connectome.
Functional connectomics addresses a different set of issues. The main challenge is how to model temporal variation and interactions between regions. Graph neural networks and transformer-based models are commonly used for this purpose and have been applied to a wide range of prediction and decoding tasks. However, results remain sensitive to preprocessing steps such as denoising, parcellation, and task design, which makes direct comparison across studies difficult.
Multimodal connectomic approaches attempt to combine structural and functional information within a shared framework. While these methods show potential for improving phenotype prediction and functional consistency, they also add complexity and introduce additional points of failure. In practice, connectomic models are more often evaluated in terms of stability, reproducibility, and behavior across datasets than by raw performance metrics. This reflects an important shift in the field, where methodological advances are increasingly assessed not only by predictive accuracy but also by their ability to produce stable and biologically meaningful representations across cohorts and imaging settings.
Diagnostic classification
Diagnostic classification in neuroimaging is relevant to a wide range of neurological conditions, including neurodegenerative disorders (e.g. predicting the conversion of mild cognitive impairment to AD), neurodevelopmental disorders (e.g. identifying patterns associated with autism spectrum disorder), and brain tumors/stroke (automatically detecting and segmenting lesions). Predicting disease trajectories can lead to earlier diagnosis for efficient personalized medicine. DL methods learn patterns from large neuroimaging datasets and can assist in diagnosing conditions, tracking disease progression, and predicting patient outcomes. Models that track temporal dependencies in fMRI, fuse complementary MRI and PET signals, or combine imaging with clinical and fluid biomarkers now support both disease-specific and cross-disease decisions. Table 4 provides an overview of the studies included, and Figure 4 presents a representative diagnostic classification approach using rs-fMRI connectomes, adapted from a specific study to illustrate how graph-based models are applied in practice.
Summary of advanced AI-based diagnostic classification.
Representative example of an AI approach for diagnostic classification in neuroimaging: (a) overall workflow of the multi-level GCN, which integrates spatiotemporal feature extraction, mutual-information-based graph construction, and node-edge embedding for classification and (b) classification accuracy across AD, MCI, and HC groups using different functional connectivity metrics. Together, these results demonstrate that a multi-level GCN trained on rs-fMRI functional connectomes can effectively distinguish diagnostic categories and highlight disease relevant regions.
Neurodegenerative disorders
Neurodegenerative diseases such as AD and Parkinson’s disease (PD) are among the most pressing targets for neuroimaging-based diagnostic classification, given their high prevalence, progressive nature, and the clinical value of detecting them at the prodromal stage. In the classification of AD, a spatiotemporal graph transformer with adversarial training applied to rs-fMRI distinguished AD, 64 early and late stage mild cognitive impairment (MCI), and healthy controls (HC), demonstrating that coupling temporal attention with GNN layers captures disease dynamics. Another graph-based approach using multi-level graph convolutional network (GCN) and mutual-information-based connectome performed a similar classification across ADNI and OASIS-3, 65 emphasizing non-linear inter-regional dependencies as diagnostic signal. One study also performed multimodal classification on MRI and PET images, 66 using a CNN framework, neuroimaging based early detection of AD using deep learning (NEDA-DL), that uses a hybrid ResNet/AlexNet design to achieve multi-class staging. This lightweight approach argues for efficient diagnostic classification in resource-constrained settings. Finally, a pathology validated analysis confirmed distinct neuroimaging signatures for AD, 67 vascular dementia, and Lewy body disease, demonstrating clinical reliability anchored to post-mortem confirmation.
In the classification of PD, an rs-fMRI based PD-ARnet integrates amplitude of low frequency fluctuations (ALFF) and regional homogeneity (ReHO) with correlation-aware weighting and attention-enhanced fusion to classify PD, PPMI, and controls. 68 Another approach leverages knowledge-distillation from a 3D teacher to 2D students, 69 enables strong PD classification on dopamine transporter (DaT) SPECT (DaTscan) and generalizes to AV-133 PET, indicating a practical bridge between volumetric priors and efficient 2D classifiers. Several ML models applied to rs-fMRI metrics distinguish PD from controls and also separate freezing of gait (FOG) from non-FOG PD, 70 implicating extrapyramidal, visual, default-mode, and salience-motor circuitry as informative features. Multimodal MRI and plasma biomarkers with classical ML predicts PD with mild cognitive impairment versus cognitively normal PD, 71 underscoring value in combining structural and functional markers for prodromal cognitive status.
Psychiatric and developmental disorders
The classification of psychiatric and developmental disorders is more challenging than for neurodegenerative disorders, since they lack clear structural changes and are instead marked by heterogeneous connectome disruptions. AI methods are increasingly being used to classify these disorders based on subtle functional and connectivity patterns. For SZ, one multimodal multi-cohort approach implemented a vision transformer to fuse fMRI functional network connectivity and sMRI GM volume (GMV) to classify patients from controls, 72 using attention to localize diagnostic imaging cues. Another study instead made use of a graph framework, using a GNN to disentangle common and idiosyncratic connectivity patterns. 73 In major depressive disorder (MDD), a transformer was trained on REST-meta-MDD, 74 which supports classification of MDD from control and stratified recurrent MDD on a large scale with a multi-site cohort. Additionally, an attention-guided GCN that integrates plasma biomarkers with sMRI showed that fusing fluid and imaging markers can improve MDD classification even in modest-sized cohorts. 75
For neurodevelopmental conditions, spatiotemporal modeling is central. Spatio-temporal aggregation reorganization transformer (STARFormer) has been developed to classify autism spectrum disorder (ASD) and attention deficit hyperactivity disorder (ADHD), coupling temporal attention with downstream ML heads. 76 One approach involved federated population graph contrastive learning that supports multi-site classification of ASD from control without centralized data sharing, 77 underlining privacy-preserving training for cross-institution generalization. Diffusion-style augmentation and self-supervised graph contrastive learning further strengthen cross-site robustness for ASD and MDD detection, 78 pointing to pretraining and augmentation as practical routes to generalization. A diffusion-transformer with a vector quantized-VAE (VQ-VAE) encoder generates augmented functional connectivity to improves ASD classification across ABIDE-I/II, 79 blending generative and discriminative modeling. ADHD classification profits from modeling time-varying connectivity using dynamic rs-fMRI to distinguish ADHD from typically developing controls and explores ADHD subtypes. 80 Beyond the classic psychiatric spectrum, time-frequency analysis of rs-fMRI BOLD signals paired with conventional classifiers distinguishes opioid use disorder (OUD) from non-opioid controls, 81 showing that carefully engineered spectral features can remain competitive on small cohorts.
Cross-disease and methodological advances
General methods that transfer across diagnoses are increasingly influential. An adversarially trained GCN that uses persistent-homology features from brain connectomes classifies AD, ASD, and stroke, 82 suggesting that topological summaries offer disease-invariant signal for graph-based diagnosis. Analysis of dMRI with a CNN and a vision transformer enables biological sex classification and localization of discriminative WM signatures, 83 positioning dMRI as an “in vivo microscope” and offering a template for disease-agnostic classification pipelines. On FDG–PET, an explainable GNN performs differential diagnosis across parkinsonism (idiopathic PD vs multiple system atrophy (MSA) vs progressive supranuclear palsy (PSP)), 84 providing case-level interpretability on metabolic networks, a desirable property for clinical adoption. For dementia more broadly, a very large multimodal system spanning nine cohorts supports differential diagnosis across etiologies, 85 illustrating the scale at which transformer-style models can unify MRI, PET, and clinical data.
Performance context and generalizability
Diagnostic classification studies cover very different disease domains, but many of the modeling choices are similar. These trends are consistently observed across disease domains, including AD, brain tumors, stroke, and psychiatric disorders. Graph-based models and transformer architectures are commonly used to handle high-dimensional imaging-derived features and to capture relationships between spatially distant brain regions, while multimodal approaches often combine imaging with clinical or biological information to improve discrimination between diagnostic groups. Generative and contrastive approaches have also increasingly been adopted, mainly to augment data or stabilize learned representations. Alongside these developments, more conventional classifiers are still used in smaller cohorts or narrowly defined tasks, where they can perform competitively.
Reported classification performance is typically strong when evaluation is limited to well-curated research datasets. Many studies report accuracy or AUC values in the range of roughly 0.8–0.95, and in some binary settings even higher. When models are tested across sites or cohorts, however, performance is consistently lower. Cross-site accuracy is often closer to 0.65–0.75, which remains above chance but clearly reflects the impact of dataset shift and site-specific effects. This pattern appears across different model families rather than being tied to a specific architecture. Importantly, these trends are observed across neurodegenerative, psychiatric, developmental, and vascular disorders, indicating that many of the methodological challenges in diagnostic classification are shared across diseases rather than disease specific.
Moreover, in a substantial portion of the literature, diagnostic classification is not treated as an end goal in itself, but rather as a practical benchmark for evaluating representation learning, generalization across datasets, or model interpretability. In this context, diagnostic models that show stable behavior across datasets and disease subgroups tend to be more informative than those optimized for peak performance on a single cohort. Consequently, improvements in benchmark accuracy should be interpreted alongside evidence of external validation, robustness to dataset shift, and feasibility for deployment in real-world clinical environments.
Prognosis, subtyping, and progression
Recent advances in representation learning have moved prognosis away from conventional feature engineered scores toward patient-specific, biologically grounded predictions that integrate imaging with behavior, genetics, and clinical context. Across the studies reviewed here, DL models either discover latent disease subtypes, forecast trajectories of symptoms and treatment response, or infer biomarker stages directly from MRI and fMRI. Table 5 summarizes nine representative studies, while Figure 5 shows a representative longitudinal modeling approach for disease progression, adapted from a single study to illustrate a typical design.
Advanced AI strategies for prognosis, subtyping, and progression.
Representative example of disease progression prediction from longitudinal MRI: (a) framework of the Siamese ResNet model applied to serial MRI scans for predicting ADAS-Cog13 change between timepoints, (b) association between predicted aging-related cognitive decline slope and accelerated decline, stratified by clinical diagnosis, and (c) distribution of accelerated cognitive decline slopes across CN, MCI, and AD groups, with significant group-level differences. These results show that longitudinal MRI with a Siamese ResNet can predict cognitive trajectories and capture disease progression signals.
Imaging-based subtyping
A key theme is the use of multimodal MRI to identify subpopulations with distinct pathophysiology that are obscured by conventional case versus control analyses. In obsessive compulsive disorder (OCD), a semi-supervised framework that integrates GM volume with resting-state ALFF identifies two reproducible subtypes. 86 These subtypes show opposite relationships between structure and function and, within each group, a negative correlation between GMV and ALFF. The method combines similarity network fusion with heterogeneity by discriminant analysis (HYDRA) to distinguishes disease-related dimensions from nuisance variation, providing targets for intervention that are specific to each subtype.
Prognosis
AI methods are being applied to predict future disease risk and treatment response. One study on OASIS-3, a GCN model operating on functional connectivity classified MCI from controls and predicted future dementia risk before clinical diagnosis, 87 with contributions from default-mode, visual, ventral-attention, and somatomotor networks. Symptom and treatment forecasting has also advanced: task-fMRI with Gaussian-process regression predicted posttraumatic stress disorder (PTSD) symptom burden, 88 a 3D-CNN on ABCD reward-task fMRI predicted child irritability with region-wise salience in caudate nucleus, amygdala, hippocampus, and parahippocampal gyrus, 89 and a multicenter local-to-global GNN combining baseline rs-fMRI with clinical data predicted selective serotonin reuptake inhibitor (SSRI) remission with internal and external replication around the mid-70% accuracy range. 90 A hierarchical functional brain network (HFBN) based on rs-fMRI both diagnosed subclinical depression with sleep disorders and predicted non-pharmacological therapy success, highlighting altered coupling between sub-regions of default mode network (sub-DMN) and executive control network (ECN). 91
Progression and staging
Longitudinal analyses are being used to model disease trajectories. Dual-loss Siamese 3D ResNets trained on longitudinal T1w MRI separated normative from accelerated cognitive decline, 92 linking the latter to genetic pathways including a NELL1 genetic locus. In another study, rs-fMRI graphs were used to estimate individual amyloid-β PET grades and regional SUVRs, 93 pointing to limbic circuitry salience. At the population scale, multimodal analysis was performed on PET amyloid-β and tau status across seven cohorts, 94 supporting scalable staging and trial pre-screening.
Comparison of modeling approaches and objectives
The studies reviewed in this section address related but distinct problems. Subtyping work focuses on identifying latent disease patterns that are not captured by diagnostic labels. These studies are often weakly supervised or semi-supervised, using imaging and functional measures to separate disease-related variation from confounding factors. Their main challenge is stability, as subtype definitions can change with cohort composition and preprocessing, which is why models that integrate multimodal information or impose explicit constraints are commonly used.
Prognosis studies focus on predicting future outcomes, such as later cognitive decline or response to treatment. These tasks are more sensitive to noise in clinical measures and limited follow-up, and performance is typically moderate. Models that combine imaging with clinical variables or operate on network-level representations are well suited to this setting because they can incorporate distributed brain information together with non-imaging predictors.
Progression and staging studies focus on disease course over time, using either longitudinal data or cross-sectional indicators of disease stage. The key difficulty is sparse and irregular sampling, which makes temporal consistency more important than short-term prediction accuracy. Across all three settings, prediction commonly serves as a practical check on whether learned representations reflect meaningful disease structure. However, many of these approaches remain at an early stage of development, and further validation in larger longitudinal and multi-center cohorts will be important for establishing clinical utility and reproducibility.
Challenges and future directions
While advances in AI have substantially improved performance across multiple stages of the neuroimaging pipeline, they have also brought to light unresolved challenges at both the methodological and clinical translational levels. These challenges motivate important directions for future investigation.
Interpretability and clinician-facing trust
While interpretable frameworks are being developed to tackle the black-box nature of typical DL models with the aim of clarifying which factors in the training paradigm drive certain predictions, these models can be hard to assess or standardize. Many interpretability tools, while valuable for developers, are not end-user friendly. In the specific context of neuroimaging, explainability methods do not always align with or add to clinical or neuroscientific knowledge. Thus, despite existing efforts, this remains an active direction of research. One recent study analyzed fMRI data from subjects under challenging speech perception by engaging them in a concurrent visuomotor recognition task. Using SHAP to interpret the model, the study identified specific brain regions that contributed to degraded speech under divided attention. 55 This illustrates how accurate prediction and meaningful explanation can emerge together within a network. Similarly, generative models like the dual-guided brain diffusion model, 25 which reconstruct natural visual stimuli from fMRI signals, can help reveal what types of visual and semantic information the brain encodes. However, interpreting these reconstructions remains challenging, such as understanding subject-specific neural encoding patterns, dataset limitations, and avoiding artificially generated features that lack clear biological meaning.
Beyond model development, interpretability for real-world deployment raises additional requirements that extend beyond developer-facing explanations. 95 In clinical settings, interpretability must support decision making, risk assessment, and trust at the point of care. This includes providing calibrated confidence or uncertainty estimates, identifying cases where predictions may be unreliable due to data quality issues or distributional shift, and presenting outputs in a form that clinicians can meaningfully integrate with existing diagnostic workflows. Clinically oriented studies typically require clinician-facing interpretability supported by rigorous and comprehensive validation. Many of the cutting-edge AI techniques reviewed here have not yet been evaluated in such contexts, representing an important challenge for clinician-facing interpretability and trust. At the same time, these methods outline clear directions for future clinically focused research.
Generalizability, domain shift, and harmonization
Models trained on one scanner, sequence, or site often fail to generalize to others because subtle acquisition differences are encoded implicitly as features in the model, and make the model inherently unusable in dissimilar clinical settings. 96 In practice, this means that the model may recognize the scanner instead of the disease. Harmonization methods that suppress domain-specific fingerprints while also learning task-related features provide a path forward. For example, recent work in dMRI demonstrates that DL methods combining contrast adjustment and super-resolution can effectively reduce variability caused by differences in scanners and imaging parameters. 31 This improves image consistency across multiple sites, while still clearly showing lesions, thus moving toward better clinical generalizability. 97 Looking ahead, robust generalizable solutions will likely combine pre-acquisition, physics-informed augmentation, and post-hoc harmonization, paired with uncertainty quantification to flag input samples that fall outside this trained distribution, consistent with challenges highlighted in prior disease-focused reviews of AD and neuro-oncology.98–100
Foundation models and LLM: Opportunities and risks
Foundation models and LLMs are beginning to deeply influence many medical domains and also have great promise in neuroimaging.101–103 Vision backbones pretrained at scale on non-neuroimaging datasets can be adapted for neuroimaging tasks with modest labeled data, enabling zero-shot or few-shot transfer learning. Text-conditioned vision models create a common interface between images and text and can bridge the gap between neuroimaging data and clinical reports. Recent multimodal LLM systems developed for neuroradiology have demonstrated the feasibility of generating high-quality radiology reports directly from 3D medical images, 104 achieving performance comparable to human experts. This development indicates opportunities for integrating automated reporting into clinical decision support.
However, the clinical adoption of foundation models and LLMs faces several challenges. In the near term, the most realistic roles are likely to remain supportive, such as feature extraction, report drafting, and decision support, where outputs can be viewed and validated by clinicians within current workflows. In the longer-term, the goal including fully automated interpretation or decision-making, remain difficult to justify in practice due to risks including hallucination, calibration of confidence estimates, and sensitivity to datasets. In addition to technical reliability, barriers related to data privacy, governance, and evaluation standards are also unresolved. Current benchmarks and validation protocols are often limited to assessing clinical robustness, especially across datacenter, populations, and imaging settings. Moreover, regulatory approval pathways still need to be defined, posing further challenges for safe and responsible clinical deployment.
Path to clinical translation
The rapid gains in accuracy and efficiency across various components of the neuroimaging pipeline, driven by AI, show strong potential for adoption into routine clinical practice; however, this translation continues to be challenged by concerns around trustworthiness, reliability, reproducibility, and generalizability. 5 Many of the studies reviewed in this paper are evaluated based on highly curated research cohorts, such as ADNI, HCP, and ABCD, or on single-center datasets with limited diversity in scanner, acquisition protocol, and patient population. As a result, strong benchmark performance may not reflect robustness under real-world clinical conditions, where data heterogeneity, motion, artifacts, and incomplete acquisitions are common. Limited prospective evaluation and inconsistent external validation further restrict the assessment of translational readiness. 105
Beyond data-related challenges, clinical translation also depends on broader practical considerations that typically fall outside of the scope of methodological studies and need to be addressed in subsequent application-driven work. These include integration into existing clinical workflows, acceptable computational cost and turnaround time, clarity regarding the clinical role of AI tools (e.g. decision support rather than standalone clinical decisions), and alignment with regulatory and reimbursement frameworks. Addressing these gaps is an important direction for future work toward bridging innovative AI methods and routine neuroimaging practice.
Failure modes and negative findings
Most of the studies reviewed in this paper report successful model performance under controlled research settings. In practice, however, applying these models to external datasets may lead to a noticeable performance drop.106,107 Models can be highly sensitive to data-related variations, including differences in acquisition protocols, scanners, parcellation choices, and preprocessing pipelines. For generative models, additional risks arise during synthesis and enhancement. Such models may generate structures that do not exist, distort pathological patterns, or oversmooth lesion regions. While the resulting images may appear visually convincing, they may lack clinical fidelity and potentially misrepresent underlying pathology. Although recent work increasingly relies on larger datasets and extensive experimentation to demonstrate performance, systematic characterization of where and how models fail remains limited. An important future direction is therefore to leverage these methods in broader and more clinically realistic scenarios to comprehensively evaluate feasibility, robustness, and limitations, including reporting of failure behavior and negative results.
Conclusion
Advancements in AI are rapidly reshaping various components of the neuroimaging pipeline, making scans faster to acquire and process, improving resolution, enabling more accurate alignment and segmentation, enhancing our understanding of brain networks, and supporting reliable clinical predictions. To ensure these advances translate into real-world diagnosis, prognosis, and patient care, it is important to recognize that performance gains under development settings may not fully address challenges in interpretability and generalizability.
For clinical use, models must be robust from a clinician-facing perspective, including multi-center datasets, comprehensive external validation, and consideration of practical translation factors such as cost, regulatory requirements, and workflow integration. Progress toward this goal could include a broader data-sharing agreement, open-source code that enables clinical researchers to evaluate emerging methods, and open acceptance of negative findings in the literature.
Supplemental Material
sj-docx-1-jcb-10.1177_0271678X261465843 – Supplemental material for Advances in artificial intelligence for neuroimaging
Supplemental material, sj-docx-1-jcb-10.1177_0271678X261465843 for Advances in artificial intelligence for neuroimaging by Fan Yang, Vibha Balaji, Ziyuan Zhou, Bowen Lei, Ziwei Liu, Tzu-An Song and Joyita Dutta in Journal of Cerebral Blood Flow & Metabolism
Footnotes
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported in part by the National Institutes of Health (NIH) under grant numbers R01AG072669, R21AG087392, and R03AG070750.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
