Developmental Brain Age Estimation From MRI Data: A Systematic Review of Deep Learning Approaches and Open Datasets

ABSTRACT

Brain age is an emerging concept that reflects complex, time‐dependent changes in brain structure, identifying departures from expected neurodevelopmental patterns. In the developing brain, accurate MRI‐based age estimation is a quantitative biomarker for detecting atypical neurodevelopment, facilitating early diagnosis, guiding clinical decision‐making, and potentially improving long‐term outcomes. Data‐driven models applied to neuroimaging have provided valuable insights into the pathogenesis of various congenital and acquired pediatric conditions. In particular, advanced deep learning approaches have recently gained prominence in a wide range of pediatric neuroimaging studies, offering state‐of‐the‐art performance in estimating developmental brain age. In this survey, we provide a comprehensive review of the current MRI applications of deep learning methodologies for developmental brain age (fetal stage—2 years) estimation. We provide details on both clinical and technical aspects, open‐access developmental MRI datasets, and compare the performance of these models utilizing evaluation metrics. Additionally, we discuss the applications of brain age estimation in clinical research contexts, highlighting its importance in understanding neurodevelopmental disorders. Finally, we address the challenges faced and propose future research directions to advance the field of brain age estimation. We aim to provide valuable insights for researchers and practitioners, facilitating advancements in both theoretical understanding and practical applications of MRI‐based deep learning brain age estimation of the developing brain.

Evidence Level: 3.

Technical Efficacy: Stage 2.

1. Introduction

Fetal and neonatal life are important periods for cerebral development: developmental processes establish most neural structures that are necessary for core motor, cognitive and behavioral functions [1] and thus for a healthy brain [2]. Disruptions during these stages can increase susceptibility to neurological and psychiatric diseases [3]. Despite its significance, early detection of neurological disturbances remains challenging, as formal neurodevelopmental or psychiatric diagnoses typically cannot be established until children reach at least 2 years of age, when standardized assessments become feasible [4].

Given this constraint, brain age estimation using brain MRI can offer a biologically informed measure of brain maturity without the need to wait until 2 years of age, capturing deviations from typical developmental trajectories through the calculation of the brain age gap. In the assessment of fetal and neonatal brain development, metrics such as gestational age (GA), postmenstrual age (PMA), and chronological age are commonly used [5]. Brain age reflects underlying physiological processes that may manifest as changes observable in neuroimaging. Brain age estimation could contribute to identifying and characterizing potential disorders [6], including neurodevelopmental disorders such as autism spectrum disorder [7], neuropsychiatric disorders such as early‐stage schizophrenia and bipolar disorder [8], and health conditions that affect brain development, such as preterm birth [9, 10] and congenital heart disease [11].

Among various modalities, brain MRI has been the primary source of data for brain age estimation across both pediatric and adult populations, as it provides sufficient spatial resolution and contrast to visualize age‐dependent structural and functional changes, without the risks associated with ionizing radiation. In fetuses, while ultrasound remains the primary clinical tool, fetal MRI is increasingly used in research settings to explore brain development. Structural MRI serves as the primary imaging modality for brain age estimation across pediatric and adult populations due to its sensitivity to age‐related changes in brain morphology, including cortical thickness, tissue volumes, and brain maturation patterns [12]. Compared to functional MRI, which requires specialized setups and is more susceptible to variability, structural MRI provides a more stable and accessible modality for brain age estimation studies [13].

Using MRI, simpler methods of machine learning (ML), supervised learning algorithms, such as support vector machines [14], random forests [15], and gradient boosting [16], are commonly trained on labeled datasets, where MRI‐derived features are linked to known chronological ages, for brain age estimation. However, compared to traditional ML algorithms, deep learning (DL) offers increased modeling capacity and often achieves superior performance [17]. In the context of brain age estimation, a typical DL framework involves training on a dataset of typically developing participants using supervised learning (e.g., regression algorithms) to model the relationship between brain features and the actual age of the participant. Several surveys have reviewed methodologies for brain age estimation [18, 19, 20, 21, 22]; however, these reviews have primarily focused on adult and aging populations, offering limited insight into the developing brain during the fetal period and early infancy (0–2 years). However, compared to traditional ML algorithms, DL offers increased modeling capacity and often achieves superior performance [22]. Developing generalizable DL models requires large and diverse datasets. Open‐access datasets addressing early brain development enable model training on varied populations, improve external validity, and support reproducibility across studies.

In this review, we present a systematic analysis of DL‐based brain age estimation for the early period from fetuses to infants using primarily structural MRI. Our key contributions include: (1) a focused review of DL techniques for estimating brain age during developmental stages (fetal stage to 2 years); (2) an overview of publicly available datasets containing structural brain MRIs for these age groups; and (3) a comprehensive discussion of the reviewed models, including their key features, clinical applications, and potential directions for future research. The graphical abstract is presented in Figure 1.

FIGURE 1.

Graphical abstract summarizing the review and clinical applications of brain age estimation in the developing brain.

2. Methods

2.1. Brain Age Estimation Literature Review

In May 2025, two reviewers systematically searched for eligible studies published since 2010 in Google Scholar and PubMed to ensure thoroughness and reduce bias. The decision to use two reviewers was based on the complexity and volume of the task, as well as the need for an unbiased and comprehensive search. One reviewer focused on searching for studies related to the keywords under the “Age group” category, whereas the second reviewer focused on the “Methodology” and “Modality” categories.

To ensure that no relevant papers were overlooked, the reviewers combined keywords from three categories: “Age group,” “Methodology,” and “Modality.” For the “Age group,” the keywords included: “Neonatal brain age,” “Fetal brain age,” “Gestational age,” “Neonates,” “Fetus,” and “Prenatal.” These were paired with “Methodology” keywords like “Convolutional neural networks,” “CNN,” “Recurrent neural networks,” “RNN,” “Transformer,” “Deep learning,” and “Ensemble learning,” as well as “Modality” keywords such as “MRI” and “Magnetic resonance imaging.”

After this search, studies were included if they (1) employed DL methods for age estimation; (2) focused on human subjects within the age range of fetal stage to 2 years of age; (3) utilized structural brain MRI as the imaging modality, including T1‐weighted imaging (T1WI) and/or T2‐weighted imaging (T2WI) sequences; and (4) reported relevant performance metrics and imaging details. Studies were excluded if they (1) used conventional image processing or shallow ML techniques; (2) involved populations outside the target developmental period (fetal stage to 2 years of age); (3) relied on non‐MRI modalities; or (4) lacked sufficient methodological detail to assess study quality. Only full‐text, peer‐reviewed journal articles written in English were considered. The review followed the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) [23] guidelines (Figure 2).

FIGURE 2.

PRISMA structure for literature review.

2.2. Open Dataset Search

To identify relevant and publicly available datasets focused on early brain development, a web‐based search was conducted. The search targeted datasets that included MRI data of the target population, with a specific focus on the fetal period up to 2 years of age. Datasets that primarily included subjects older than 2 years were excluded from the final dataset collection. Searches were performed using academic search engines including PubMed, IEEE Xplore, and Google Scholar, as well as general search platforms like Google. Institutional repositories and project‐specific portals were also used. The following keywords were used in various combinations: “open dataset,” “public dataset,” “neonatal neuroimaging,” “early childhood brain MRI,” “pediatric MRI,” “fetuses,” “prenatal,” and “longitudinal infant neuroimaging.” Only datasets that were openly accessible for research purposes were included.

3. Results

3.1. Summary of Brain Age Estimation Literature Review

The search identified a total of 2300 records (2080 from Google Scholar and 220 from PubMed). After removing 183 duplicates, 2117 unique records were retained for screening. Of these, 1826 were excluded for not aligning with the study focus: studies unrelated to medical imaging, brain development, or ML. The full texts of the remaining 291 records were then assessed for eligibility. An additional 265 articles were excluded for not meeting the previously described inclusion criteria or for meeting one or more of the exclusion criteria. Based on the final exclusion of the Reviewer's feedback 26 studies were included in the final review. In Table 1, we present an overview of the different DL methods. Table 2 provides a structured explanation of their strengths and limitations.

TABLE 1.

An overview of deep learning approaches for developmental brain age estimation.

Paper	Model	Modality|(2D/3D)	Dataset type a (no. of dataset and subject count)	Age range	MAE ± SD (weeks)	Coefficient (R 2)	MRI timing	Ratio train – test – validation|cross‐validation
Shen et al. (2019) [24]	Attention‐based ResNet (Multi‐view)	T2WI (2D)	Private (1; n1 = 741)	17.9−39 weeks, Gestational age	0.96 c  ± N	0.94	Fetal	70%–10%–20%, (single test set hold‐out)
Hong et al. (2021) [25]	2D single‐channel CNN (multi‐channel)	T1WI, T2WI (2D)	Two private (2; n1 = 220, n2 = 21)	15.9–38.7 weeks, Gestational age	0.3 ± 0.459	0.98	Fetal	(90%–10%) – 10%, (10 fold, independent validation set)
Kojita et al. (2021) [26]	CNN‐based (VGG, single plane)	T2WI (2D)	Private (7; n1 = 170, n2 = 6, n3 = 3, n4 = 2, n5 = 1, n6 = 1, n7 = 1)	14.0–41.4 weeks, gestational age	1.17 ± 1.47	0.964	Fetal	68.4%–15.8%–15.8%, (single test set hold‐out)
Shen et al. (2022) [27]	Attention‐guided ResNet‐50 (Multi‐view)	T2WI (2D)	Stanford fetal MRI dataset (1; n1 = 741)	19–39 weeks gestational age	0.95 ± N	0.81–0.90	Fetal	70%–10%–20%, (4 institute test set, 80% test)
Feng et al. (2024) [28]	PDFF‐CNN (Multi‐orientation)	T2WI (2D)	Private (1; n1 = 157)	22–34 weeks, gestational age	0.848 ± 0.037	0.904	Fetal	80%–20%–0, (single test set hold‐out, handle imbalance date)
Gangopadhyay et al. (2022) [29]	MTSE U‐Net, (Multi‐tasking, single encoder)	T1WI (2D)	FeTA 2.1 (1; n1 = 80)	20–35 weeks, gestational age	0.83 ± 1.88	N	Fetal	66.7%–33.3%–0, (3 fold)
Hasan et al. (2024) [30]	Xception + MHA, (Multi‐head, multi‐plane)	T2WI (2D)	Stanford Lucile Packard Children's Hospital (1; n1 = 741)	19–39 weeks, gestational age	0.52 ± N	0.980	Fetal	80%–20%–0, (5 fold)
Kwon et al. (2024) [31]	ResNet101V2, regression for feature, (cortical surface map)	T2WI (2D)	Private (1; n1 = 115)	19.9–38.7 weeks, gestational age	0.94 ± N	0.908	Fetal	80%–20%–0¸ (single test set hold‐out)
Liao et al. (2020) [32]	Deformable CNN with label distribution learning (multi‐branch)	T2WI (2D)	Private (1; n1 = 289)	21–36 weeks, gestational age	0.751 ± N	0.947	Fetal	80%–20%–0, (3 × 4 fold)
Vahedifard et al. (2024) [33]	Dynamic U‐Net‐based (Multi‐plane)	T2WI (2D)	Private (1; n1 = 52)	20–38 weeks, gestational age	0.46−2.26 ± N	0.91	Fetal	Not specify
Zhou et al. (2024) [34]	JoCoRank (Multi‐view, handle imbalance data)	T2WI (2D)	Private (1; n1 = 157)	22–34 weeks, gestational age	0.693 ± 0.064	0.930	Fetal	85%–15%–0, (10 random experiment)
Yun et al. (2025) [35]	CN (multi‐planner)	T2WI (2D)	Private (1; n1 = 755), 3 institution	15.9–38.7 weeks b , Gestational age	0.66–0.83 b (AUC)	0.88–0.94 b	*Fetal	*(90%–10%) – 10%, (10 fold, independent validation set)
Hu et al. (2019) [36]	DGFLDL (cortical morphometric)	T2WI (feature)	Private (1; n1 = 384)	5–121 weeks, chronological age	5.15 c  ± 0.32	0.963	Postnatal	10 fold cross validation
Shabanian et al. (2019) [37]	3D CNN (Multi‐modal, fuse inputs)	T1WI, T2WI, PDI (3D)	NIMH (1; n1 = 112)	3 weeks–3 years chronological age	N	NA	Postnatal	80%–20%–0, (single test set hold‐out)
He et al. (2020) [38]	2D‐ResNet18 + LSTM (sequence)	T1WI (2D)	NIH‐PD + MGHBCH (2; n1 = 1212, n2 = 428)	0–6 years chronological age	41.9–59.3 c  ± N	NA	Postnatal	80%–20%–0, (single test set hold‐out)
Liu et al. (2020) [39]	GCN (surface Graph)	T1WI (feature)	Private, UCSF + dHCP (1; n1 = 129, n2 = 407)	24–33, 29‐45 weeks, postmenstrual age	0.96 ± N	0.94	Postnatal	64%–20%–16%, (5‐fold, test set hold‐out)
Shabanian et al. 2020 [40]	3D CNN (Multi‐modal, fuse inputs)	T1WI, T2WI, PDI (3D)	NIMH (1; n1 = 112)	0–3 years chronological age	N	NA	Postnatal	80%–20%–0, (single test set hold‐out)
Vosylius et al. (2020) [41]	MeshCNN, PointNet++ GraphCNN Volumetric Benchmark	T2WI (feature)	dHCP (1; n1 = 650)	27–45 weeks, postmenstrual age	0.621 ± 0.4784	NA	Postnatal	65.7%–17.15%–17.15%, (single test set hold‐out)
Kawaguchi et al. (2021) [42]	CNN‐based (myelination pattern, non‐invasive)	T1WI (2D)	Private (1; n1 = 441)	0–2 years, chronological age	8.2 ± 28	0.94	Postnatal	60%–20%–20%, (4 fold, validation)
Wada et al. (2023) [43]	LSTM + PyramidNet (myelination stages)	T1WI, T2WI (2D)	Private (1; n1 = 119)	0–2 years, chronological age	4.3 ± 5.5|8.04 (external)	0.99|0.93 (external)	Postnatal	75%–25%–0, † (4 fold, external validation)
Chen et al. (2022) [44]	3D‐CNN (normal myelination)	T1WI, T2WI (3D)	Private (1; n1 = 518)	0–25 months, corrected gestational age	7.7 ± 1.7	0.93	Postnatal	70%–20%–10%, (independent test set)
Park et al. (2024) [45]	U‐Net++‐based XGBoost (Regional Brain)	T1WI, T2WI (features)	dHCP + HH (2; n1 = 163, n2 = 15) two private (2; n3 = 56, n4 = 13)	34–43 weeks, postmenstrual age	1.33−1.57 ± N	0.88	Postnatal	70%–30%–0, (2 external validation set)
Hu et al. (2023) [46]	ResNet‐18 (multiple input)	T1WI (3D)	Private (1; n1 = 658)	0–3 years, chronological age	9.67 ± N	0.91	Postnatal	80%–20%–0, (5 fold)
Tang et al. (2023) [47]	BAPNET (compare different modality)	T1WI, T2WI (2D, 3D)	Private (1; n1 = 281)	27−37 weeks, gestational age	1.15 ± N	0.96	Postnatal	75%–25%–0, (single test set hold‐out)
Zhao et al. (2024) [48]	HRINet (attention, inter‐hemispheric relation)	T2WI (2D)	dHCP + Private (3; n1 = 531, n2 = 47, n3 = 16)	34–45 weeks, postmenstrual age	0.53 ± 0.43	0.89	Postnatal	80%–10%–10%, (5 fold, evaluation on independent term and preterm set)
Zhao et al. (2024) [49]	Transformer‐based (multi‐modal fusion model)	T2WI, DWI (3D)	dHCP + Private (2; n1 = 478, n2 = 114)	37–44 weeks, postmenstrual age	0.5 ± N	0.89	Postnatal	80%–10%–10%, (5 fold)

Note: N, Standard Deviation is not given; NA, not applicable; the imaging sequence was not specified.

Abbreviations: BAPNET, brain age prediction network; BCH, Boston Children's Hospital; BCP, baby connectome project; BrainNetCNN, brain network convolutional neural network; CNN, convolutional neural network; CNS, central nervous system; DGFLDL, deep granular feature‐label distribution learning; dHCP, developing human connectome project; DNN, deep neural network; DTI, diffusion tensor imaging; DWI, diffusion weighted imaging; GCN, graph convolutional network; HH, Hammersmith Hospital; HRINet, hemispheric relation inference network; JoCoRank, joint correlation ranking (with ranking similarity regularization); LSTM, long short‐term memory; MGHBCH, Massachusetts General Hospital and Boston Children's Hospital; MHA, multi‐head attention; MSN, multi‐scale network; MTSE, multi‐tasking single encoder U‐Net; NEOCIVET, cortical surface extraction and analysis; NIH‐PD, national institutes of health–pediatric data; NIMH, national institute of mental health; PDFF‐CNN, pose‐dependent feature fusion convolutional neural network; PDI, proton density imaging; PyramidNet, pyramid networks; SVM, support vector machine; T1WI, T1‐weighted imaging; T2WI, T2‐weighted imaging; TMC, Taipei Medical University; UCSF, University of California at San Francisco; Xception, extreme inception.

^a Dataset size refers to the number of patients for model development; the actual number of scans or images may be larger than the stated dataset size. Besides, Yun et al. [35] does not specify the number of cases from different institute.

^b Yun [35] mentioned they use exactly the same method as Hong [25] and achieve similar baseline results and then apply for clinical purposes in their paper but don't mention new MAE/SD though reported (AUC: 0.66–0.83).

^c MAE was originally reported in a different unit and has been converted to weeks for consistency. Converted from years to weeks (1 year = 52 weeks) when required.

TABLE 2.

Structured evaluation of deep learning studies for developmental brain age estimation (strengths and limitations).

Paper	Bias/limitations	Strength
Shen et al. (2018) [24]	Single‐center (Stanford) data, leading to institutional and vendor bias (GE scanners). Dataset consists only of normal fetuses, limiting generalizability for anomaly detection. The 2D slice approach inherently loses full 3D volumetric context.	Relatively large fetal dataset for a single‐site study at the time. Novel use of an attention mechanism to automatically localize the brain, significantly reducing labor‐intensive preprocessing (no manual segmentation required). Model is computationally robust for noisy fetal MRI data.
Hong et al. (2021) [25]	The five‐fold increase in MAE during external validation (0.125 → 0.508) explicitly highlights the presence of significant scanner/site bias. The dedicated external test set is small (N = 21). The 3D model underperformed significantly, suggesting DL benefits are limited by motion and small fetal sample size.	Rigorous methodological comparison proving the superiority of the 2D multiplanar approach for dealing with motion. Strongest validation approach by including an external test set. Provides biological interpretability via saliency maps showing that the cortex and ventricles drive prediction.
Kojita et al. (2021) [26]	Smallest dataset in the comparison group (N = 184). Single‐center (Kobe University). Focused only on the second and third trimesters (Weeks 13–40).	Successfully demonstrated DL outperforming the traditional clinical method (BPD) as this utilized complex, multi‐planar information from 3D MRI volume. Simple, effective architecture (VGG) making adoption feasible.
Shen et al. (2022) [27]	Primary dataset limited to developmentally normal fetuses. Model shows underestimation bias due to data imbalance in late GA range. Performance variability across external sites is still observed.	Large, diverse dataset with validation on multiple external sites—the strongest test of generalization. High internal performance and validated effectiveness of the attention‐guided architecture.
Feng et al. (2024) [28]	Imbalanced GA distribution (though authors applied corrective loss). Moderate subject count (157) for inter subject variability. Only T2 MRI modality. Strong fetal motion may cause misalignment during feature fusion.	Innovative dynamic fusion of multi‐orientation features (attention, PSA modules) to cope with orientation uncertainty. Use of balanced MSE loss is methodologically strong for imbalance mitigation. Attention maps provide interpretability.
Gangopadhyay et al. (2022) [29]	Dataset size is relatively small (80 subjects) and lacks external validation. Exact slice count and split ratios are not reported. Possible bias toward specific FeTA scanner protocols. Joint multi‐tasking may compromise one task if one dominates training. Input is 2D slices; volumetric context may be lost. Unclear generalization across scanners/sites.	Multi‐task shared‐encoder design for joint Segmentation, GA Prediction, and Pathology Classification. High computational efficiency and robust cross‐validation performance.
Hasan et al. (2024) [30]	Relies on 2D slices (losing 3D context). Attention maps offer a degree of interpretability. The generalization to pathological/high‐risk fetuses is uncertain. Sensitivity to image quality, motion artifacts, and heterogeneity across scanners may affect model reliability.	Sophisticated architecture combining Xception and Multihead Attention (MHA) to capture cross‐plane dependencies. Large patient level sample enabling good statistical power Attention maps enable interpretability of which features/views contribute to gestational age estimation Experiments across different attention head/key/value dimensionalities provide robustness in model design
Kwon et al. (2024) [31]	Single‐center, proprietary dataset limits generalizability. Requires complex cortical registration pre‐processing step. Reliance on consistency of slice‐to‐volume reconstruction and surface mapping.	Novel approach focusing on cortical structural variance to improve age prediction. Balanced design with clear subject‐wise split. Use of test‐time augmentation reduces prediction noise. Analysis of cortical variance gives interpretability (which cortical features correlate with brain age).
Liao et al. (2020) [32]	The Multi‐Branch DCNN and Label Distribution Learning (LDL) add significant architectural complexity. Collaborative, but not publicly available dataset. Potential overfitting risk despite cross‐validation.	Introduces a robust Multi‐Branch Deformable Convolutional Neural Network (DCNN), which dynamically adapts to the high variance in fetal brain shape and size. Uses Label Distribution Learning (LDL) to model gestational age as a distribution rather than a single point, increasing prediction robustness.
Vahedifard et al. (2024) [33]	Biometric prediction relies heavily on the quality of automated segmentation (using Dynamic UNet/MONAIfbs). The biometrics‐based approach is sensitive to biological variability and relies on established (but potentially outdated) reference curves.	Demonstrates high accuracy using only common biometric measurements. Provides an accessible and interpretable AI method, suitable for clinical integration as it uses familiar clinical indices.
Zhou et al. (2024) [34]	The dataset is single center, potentially introducing institutional and scanner bias. Focuses on a relatively limited GA range (22–34 weeks).	Introduces JoCoRank, a novel loss function that uses Joint Correlation Learning with Ranking Similarity Regularization for multi‐view MRI and attention mechanism. Specifically designed to address the challenges of imbalanced data in fetal brain age regression.
Yun et al. (2025) [35]	The model's primary validation is its ability to distinguish VM fetuses from typical fetuses based on the predicted age difference (PAD/AAD). Using 2D slices may omit crucial 3D contextual information about structural deformities. Applicability is limited to fetal ventriculomegaly cases.	The strength lies in its clinical application: identifying fetuses with Cerebral Ventriculomegaly (CVM) in utero. Demonstrated that the Predicted Age Difference (PAD) and Absolute Predicted Age Difference (AAD) are significantly different in VM fetuses, acting as a potential imaging biomarker for associated CNS abnormalities.
Hu et al. (2019) [36]	Limited sample size for a deep learning model. Feature‐based approach relies on accurate preprocessing and may discard full image spatial context. Generalization to unseen cohorts is uncertain.	Proposes Deep Granular Feature‐Label Distribution Learning (DGFLDL) to handle small sample size and age label ambiguity by aggregating neighboring ages. Achieves high performance for the rapidly changing infant brain.
Shabanian et al. (2019) [37]	Very small dataset, major limitation for generalizability. The model performs classification, which is less granular and clinically useful than regression.	Unique Methodology: Using a deep learning model to perform categorical age classification.
He et al. (2020) [38]	Sequence representation may lose 3D spatial consistency. Performance may be sensitive to slice ordering.	Novel architecture combining 2D CNN + LSTM to model temporal (slice) dependencies across the 3D volume. Demonstrated superior generalization capability and robustness across two different, multi‐site datasets.
Liu et al. (2020) [39]	Focuses exclusively on preterm neonatal data (GA 30 to 45 weeks), limiting generalizability to the full infant/child population. Model performance is highly reliant on the accuracy of the initial cortical surface reconstruction.	Novel use of Graph Convolutional Networks (GCN) to incorporate topological and morphological relationships across the cortical mesh structure. Achieved high accuracy and demonstrated that predicted brain age relates to neurodevelopmental outcome, adding a strong clinical dimension.
Shabanian et al. 2020 [40]	Very small sample size. The primary task is classification (grouping age ranges) rather than continuous age regression, which is generally considered a less challenging task.	Utilized a 3D CNN architecture with multi‐modal fusion, effectively leveraging the full volumetric and complementary information from T1w, T2w, and PD sequences. Specifically targets the crucial infant period (0–2 years).
Vosylius et al. (2020) [41]	Focus limited to the cortical surface only (ignores subcortical and volumetric data). Data is exclusively the dHCP neonatal cohort (healthy and preterm), limiting generalization outside the neonatal period.	Pioneers the use of Geometric Deep Learning (GDL), leveraging mesh based, non‐Euclidean representation which is less influenced by image registration.
Kawaguchi et al. (2021) [42]	Uses single 2D slices, which inherently omits the full volumetric context of myelination. Small internal sample size.	Uses a simpler, clinically interpretable method (simulating a rule‐based assessment). Demonstrated utility in detecting maturational delay (myelination delay) by predicting a younger brain age for affected subjects.
Wada et al. (2023) [43]	Modest sample size. The model is limited to age. Generalization to pathology (SWS) showed some bias, especially in infants	Achieved high performance and demonstrated decent external generalization. Introduces an interpretable stacked deep learning model that mimics a clinician's established rule‐based myelination assessment.
Chen et al. (2022) [44]	Data is retrospective from a single academic center. Limited to subjects with “normal myelination,” which may cause poor generalization to cases of delayed or abnormal myelination.	Achieved strong performance for the T1 + T2 network. Demonstrated the power of 3D CNN regression to use the full volumetric context of myelination. Showed improved external test performance after partial fine‐tuning on the external data.
Park et al. (2024) [45]	Prediction is limited to neonates (up to 42 weeks GA). Relies entirely on the accuracy of the initial brain segmentation pipeline. Only uses structural‐normal MRI subjects (excludes known anomalies).	Derived a clinically useful Brain Maturation Index (BMI) alongside PMA prediction. Achieved excellent external generalization on the two external test sets.
Hu et al. (2023) [46]	Model is based solely on the T1 modality. The primary training cohort consists of healthy controls, which may limit direct generalization to diverse disease populations.	Achieved high internal correlation. Demonstrated clinical applicability by detecting significant delayed brain age in subjects with low birth weight and accelerated brain age in those with ADHD.
Tang et al. (2023) [47]	Moderate sample size (281 subjects). Only includes preterm infants (no term controls for comparison). Single‐site data.	Used transfer learning (from DeepBrainNet pretrained on ImageNet) to enhance performance. Demonstrated high performance both 2D and 3D model. Used heatmaps for model interpretability, highlighting attention on key developmental regions like the hindbrain.
Zhao et al. (2024) [48]	Model performance is dependent on accurate cortical surface reconstruction. Bias toward the specific characteristics of the dHCP population (scanners, demographics).	Proposed the Attention‐Based Hemispheric Relation Inference Network (HRINet), explicitly modeling the structural asymmetry and correlation between the two hemispheres to improve age prediction. Achieved superior performance over baseline models, demonstrating the value of incorporating topological relationship learning in the perinatal brain.
Zhao et al. (2024) [49]	Dataset is from a single cohort (dHCP), potentially limiting generalization to other scanner populations. Reliance on high quality multimodal MRI acquisition (T2w + dMRI), which is not always available clinically.	Uses a Transformer fusion architecture to effectively integrate structural and diffusion modalities. Achieved state‐of‐the‐art accuracy for term infants. Provides high classification accuracy for preterm and offers interpretability via saliency maps.

Based on the DL architecture, 9 studies were convolutional neural network (CNN) based, 7 studies were ResNet based, and 10 studies were from other DL models. A total of 11 studies utilized fetal brain MRI datasets, and 15 studies employed postnatal brain MRI datasets. Studies that used CNN‐based models extracted hierarchical spatial features from imaging data. Studies of ResNet‐based models utilized skip connections to enable deeper architectures and mitigate vanishing gradient problems of CNN models. Other DL models included leverage transfer learning to adapt features learned from datasets. A visualization of reviewed DL architectures subcategories by three age groups (fetus, neonate, and infant) is presented in Figure 3.

FIGURE 3.

Brain age estimation conceptual illustration of dataset, preprocessing and DL model architectures. BAPNET, brain age prediction network; BCP, baby connectome project; BOBs, baby open brains Repository; dHCP, developing human connectome project; DL, deep learning; FeTA, fetal tissue annotation challenge dataset; GCN, graph convolutional network; JoCoRank, joint correlation ranking (with ranking similarity regularization); LSTM, long short‐term memory; MTSE, multi‐tasking single encoder U‐Net; PDFF‐CNN, pose‐dependent fusion convolutional neural network; PyramidNet, pyramid networks.

Seven of the included 26 studies provided inference code (Table 3); however, none made their training code publicly available.

TABLE 3.

Available inference code for reference deep learning models.

Paper	Code language	Code link
Chen et al. (2022) [44]	Python	https://github.com/gunvantc/infant‐brain‐age‐public
Vosylius et al. (2020) [41]	Python	https://github.com/andwang1/BrainSurfaceTK/tree/master/models/MeshCNN
Liu et al. (2020) [39]	Python	https://github.com/bigting84/Brain‐Age‐Prediction
Zhao et al. (2024) [48]	Python	https://huggingface.co/uais‐zll/HRINet/tree/main
Hong et al. (2021) [25]	Python	https://github.com/jwhong1125/Fetal_BrainAge
Shen et al. (2022) [27]	Python	https://github.com/pytorch/vision/tree/main/torchvision/models
Gangopadhyay et al. (2022) [29]	Python	https://github.com/tg2001/MTSE‐U‐Net

3.1.1. CNN‐Based Models

CNNs are the predominant architecture employed in computer vision and have been widely adopted in medical imaging. They are composed of multiple convolutional layers, typically followed by pooling layers to progressively reduce spatial dimensions and extract hierarchical features. The pyramid squeeze attention guided dynamic feature fusion CNN (PDFF‐CNN; [28]) achieved a mean absolute error (MAE) of 0.85 weeks, whereas the multi‐branch deformable CNN [32] reported an MAE of 0.75 weeks. The PDFF‐CNN model was specifically designed to robustly predict GA from fetal brain MRI with variable brain localization and imbalanced age distribution. It comprises four key components: a transformation module, a feature extraction module, a dynamic feature fusion module, and a balanced mean square error (MSE) loss function. Gangopadhyay et al. [29] show that their single model MTSE U‐NET can perform multiple, but related, tasks, including segmentation, brain condition prediction (neurotypical/pathological), and brain age estimation simultaneously.

For neonatal brain imaging, geometric CNNs predicted PMA with an MAE of 0.621 weeks using T2WI [41]. Graph convolutional networks (GCNs) predicted PMA in preterm fetal capturing cortical maturation patterns using T1WI with an MAE of 0.96 weeks [39]. In fetal brain imaging, CNN‐based models have been developed to predict GA. For instance, Liao et al. introduced a multi‐branch deformable CNN that utilizes label distribution learning for fetal brain age prediction. The multi‐branch deformable CNN was designed to aggregate multi‐view information, and label distribution learning to deal with the small sample problem. Their model achieved an MAE of 0.751 weeks on T2‐weighted MRI, demonstrating its effectiveness in estimating fetal brain age from MRI scans [32].

3.1.2. ResNet Models

Residual network (ResNet) is a deep CNN that uses skip connections for training deep models. These models excel at learning residual features from MRI, handling anatomical variability, and overcoming challenges like class imbalance with limited data.

In fetal brain MRI, ResNet‐18 and ResNet‐50 models using T2WI predicted GA in fetuses with an MAE of 0.96 weeks [24]. The models had the ability to handle fetal brain structure variability. A refined 2D single‐channel CNN using T1WI and T2WI predicted GA and outperformed multi‐channel CNNs with an MAE of 0.3 weeks by leveraging multiplanar slices [25].

For neonatal brain imaging, a transfer learning model called BAPNET predicted brain age from T1WI and T2WI, achieving an MAE of 1.15 weeks [47]. Deep granular feature‐label distribution learning (DGFLDL), using T2WI, improved age estimation for small datasets, achieving an MAE of 5.15 weeks for infants by learning granular features [36]. By grouping adjacent labels into granules, granular label distribution (GLD) enables each MRI to inform not only its own age but also neighboring ages, preserving LDL's information augmentation while reducing label count. To further enhance small‐data learning, they propose granular feature distribution (GFD), which exploits within‐age image variability to significantly boost effectiveness.

In infants, ResNet‐18 predicted brain age from T1WI with an MAE of 9.67 weeks [46]. A hybrid model combining ResNet‐18 and long short‐term memory (LSTM) predicted brain age for children with an MAE of 41.9–59.3 weeks [38]. An LSTM‐Pyramid network (PyramidNet) approach trained on T1WI and T2WI achieved an MAE of 4.30 weeks for infants up to 2 years old [43].

3.1.3. Other DL Models

These models describe various models such as transfer learning, 3D analysis, attention mechanisms, and multi‐modal fusion. Transfer learning and pretrained models help with brain age estimation by enabling knowledge transfer and feature extraction, even with limited datasets.

In fetal brain MRIT, pretrained models predicted GA using Visual Geometry Group (VGG)16‐based CNN fine‐tuned with T2WI fetal MRI, outperforming traditional methods like biparietal diameter [26]. A joint correlation learning with ranking similarity regularization (JoCoRank) handled data imbalance in fetal brain age regression, improving prediction accuracy [34]. A 3D CNN based on myelination‐weighted structural MRI was proposed to estimate neonatal and infant brain age [34]. The model achieved an MAE of 0.96 weeks in neonates, highlighting its capacity to capture early myelination patterns across the developing brain [24]. U‐Net++ and eXtreme Gradient Boosting (XGBoost) predicted PMA, introducing a brain maturation index [27]. A 2D CNN model using multiplanar fetal MRI was applied to predict brain age and identify fetuses with cerebral ventriculomegaly [35].

For neonatal brain imaging, a transformer‐based multi‐modal MRI fusion model improved neonatal brain age estimation [46]. The model used a two‐stream dense network from structural and diffusion MRI of the brain individually with a transformer module for feature fusion, improving neonatal brain age estimation [49].

3.2. Summary of Open Dataset Search

Open‐access datasets that can be used for developmental brain age are summarized in (Table 4). These datasets provide multimodal brain MRI data, including structural, diffusion, and functional MRI. Datasets include the Developing Human Connectome Project (dHCP) [50], Baby Connectome Project (BCP) [51], and Fetal Tissue Annotation Challenge Dataset (FeTA) [52] which offer brain MRI data for fetuses, neonates, and infants. Additional datasets, like the Infant Brain MRI Segmentation Challenge 2019 (iSeg‐2019) [54] and Baby Open Brains (BOBs) Repository [58], provide manually segmented MRI scans.

TABLE 4.

Summary of open datasets of fetal, neonatal and infant brain development.

Dataset	Number of cases	Age range	Types of common scans	T1/T2 weighted images	Origin country	Hospital name
Developing Human Connectome Project (dHCP) [50]	40, 558, 783, 783 (1st to 4th release) a	26–44 weeks postmenstrual age	Structural MRI, Diffusion MRI, Functional MRI	Y/Y	UK	King's College London, Imperial College London
Baby Connectome Project (BCP) [51]	750	0–5 year	Structural MRI, Diffusion MRI, Functional MRI	Y/Y	USA	University of Minnesota (UMN) and University of North Carolina (UNC)
FeTA (Fetal Tissue Annotation Challenge Dataset) [52] b	120 (training), 180 (testing)	21–36 weeks gestational age	Structural MRI (T2‐weighted); Tissue segmentation masks	N/Y	Switzerland, Austria, USA	University Children's Hospital Zurich (Kispi), Medical University of Vienna, Lausanne University Hospital (CHUV), University of California San Francisco (UCSF)
Penn‐CHOP Infant Brain Atlases [53]	95	33 to 46 weeks postmenstrual age, 1‐year, 2‐year	Structural MRI, Diffusion Tensor Imaging (DTI)	Y/N	USA	Children's Hospital of Philadelphia, University of Pennsylvania
Infant Brain MRI Segmentation Challenge 2019 (iSeg‐2019) [54]	10 (training), 16 (validation)	6 months	Structural MRI with manual segmentations	Y/Y	USA	University of North Carolina, Stanford University, Emory University
ChildBrainAtlas [55]	100	0–7.5 years	Structural MRI, Diffusion MRI, Functional MRI	Y/Y	Singapore	National University of Singapore
Pediatric MRI Data Repository (PedsMRI)	73	0–18 years	Structural MRI, Diffusion MRI, Functional MRI	Y/Y	USA	National Institutes of Health
MRI Brain Template for Chinese Children [56]	180	1–6 years	Structural MRI templates with tissue probability maps	Y/Y	China	Huaxi MR Research Center (HMRRC)
NIH Pediatric MRI Data Repository (NIH‐PD) [57]	288	0–2 years	Structural MRI, Diffusion MRI	Y/Y	USA	Multiple U.S. Pediatric Centers (e.g., Children's Hospital Boston, UCLA)
MGHBCH Pediatric MRI Dataset [38]	428	0–6 years	Structural MRI	Y/Y	USA	Massachusetts General Hospital, Boston Children's Hospital (MGHBCH)
Baby Open Brains (BOBs) Repository [58]	71	1–9 months	Structural MRI with manually curated segmentations	Y/Y	USA	Masonic Institute for the Developing Brain (MIDB)
Stanford Fetal Brain MRI Dataset [27]	741	19–39 weeks (gestational age)	Structural MRI https://purl.stanford.edu/sf714wg0636	N/Y	USA	Stanford Lucile Packard Children's Hospital

^a dHCP Dataset has common cases for different release versions, N: sequence not present; Y: sequence present.

^b Only the Zürich dataset is openly available; other datasets are controlled access or hidden (test data) according to the Challenge design.

4. Discussion

4.1. Deep Learning Models for Developmental Brain Age Estimation

4.1.1. Performance Comparison and Strengths

A summary of MAE values of the reviewed articles is shown in Figure 4. Generally, models using data from younger subjects had lower MAEs: fetal imaging models had the lowest MAEs, and neonatal models had the highest. This trend can be attributed to smaller age ranges used for training in younger subjects [25], however, the brain undergoes more pronounced shape and size changes in the fetal period, allowing for a better estimation of age based on simpler morphometric features.

FIGURE 4.

Brain age estimation mean absolute error (MAE) across 23 studies from Table 1. Two papers without MAE values were not included (Shabanian [37] and Shabanian [40]). Also, Yun [35] was excluded, as it used the same 2D CNN method as Hong [25] and achieve same performance but reported AUC instead of MAE in their paper. MAE values are shown as dots, with whiskers indicating the standard deviation. Red represents MAE from the private dataset, and blue represents MAE from the public dataset.

In the fetal group, the lowest MAE of 0.3 weeks was achieved using a 2D single‐channel CNN on private data [25]. The MAE near 1 week came from a ResNet model, possibly due to 2D features and variable image quality [24]. A multi‐plane CNN with attention, trained on 741 images, achieved an MAE = 0.95 weeks [27] (Figure 5).

FIGURE 5.

Examples of heatmap generation and region of interest mask inference using attention‐guided deep learning for gestational age prediction using different planes of fetal brain MRI. Increasing activation values correspond to the color spectrum from violet to yellow. Figure from Shen et al. [27].

For neonates, multiplanar‐based fetal MRI yielded a higher AUC of about 0.8 in preterm neonates [35]. The highest accuracy was achieved with MAE = 0.50 weeks using a transformer‐based multi‐modal MRI fusion framework combining T2‐sMRI and dMRI data [49]. Higher MAEs of 7.7–9.8 weeks were observed using 3D CNNs [44].

In infants, MAE of 8.2 weeks was achieved using a custom CNN [42]. A hybrid LSTM PyramidNet model showed MAE of 4.3 months [43], whereas non‐neonatal architectures such as ResNet‐18 + LSTM produced much larger errors (MAE = 54.6 weeks; [38]).

4.1.2. Model Complexity, Scalability, and Computational Limitations

The reviewed DL models for neonatal brain age estimation include 2D and 3D CNNs, GCNs, and transformers. These show tradeoffs between accuracy, computation, and scalability. Hong et al. [25] used a 3D CNN on myelination with MAE 0.3 weeks but limited data. 3D models represent spatial relationships better but have higher computational demands [44] and are prone to motion artifacts [41]. Models employing ranking regularization to handle imbalanced data like JoCoRank [34] and transformer‐based approaches [49] have been developed, though generalization remains limited due to small, modality‐specific datasets and the high model complexity may increase susceptibility to overfitting. Attention mechanisms inherently raise complexity [28], a characteristic also evident in models such as [59]. ResNet with attention handled multi‐view MRI (MAE 0.96 weeks) but used proprietary data [24]. GCNs improved preterm brain age estimation performance but had high computation and real‐time limits [33]. A hybrid CNN–attention model improved prediction performance but increased complexity [30]. These show the need to balance accuracy, computation, and practical use. However, greater model complexity can be advantageous when it enables the network to learn richer spatial and contextual features that enhance predictive accuracy and generalization.

4.1.3. Important Brain Regions and Whole Brain Analysis

Accurate brain age estimation depends on both region‐specific and global neuroanatomical features and several studies relied on identifying neuroanatomical regions that are important for brain age estimation. For instance, GradCAM [60] highlighted the frontal and temporal cortices for age prediction. Liao et al. [32] also highlighted the frontal and temporal cortices as important for age prediction. In another study, subcortical regions, including the thalamus, were found to be critical for age estimation [31]. Additionally, saliency maps from transformer‐based models emphasized frontal, temporal, and subcortical regions as important for neonatal brain development analysis [49].

While these region‐specific findings may underscore the importance of localized brain structures, other studies have shifted focus toward global features derived from entire brain analyses. Integrative models have confirmed that holistic brain analysis outperforms regionally restricted approaches in fetal brain age estimation [32]. Entire brain analyses further emphasize that global features such as cortical thickness, cortical folding, myelination and volumetric changes contribute to robust age estimation [36, 47].

4.1.4. Clinical Application of DL Models

Potential clinical applications of brain age estimation in the developing brain are illustrated in Figure 1. The clinical applicability of a brain age estimation method depends on the target cohort. Reliable application requires sufficiently low error rates in healthy subjects (e.g., MAE) to detect subtle neurodevelopmental delays. For example, Zhao et al. [48] demonstrated that in preterm infants, the brain age gap becomes increasingly negative with greater degrees of prematurity: those born at 24–28 weeks showed a mean brain age gap of −1.18 weeks, those born at 28–32 weeks had a gap of −0.52 weeks, and those born at 32–36 weeks exhibited a near‐normal gap of +0.03 weeks. In comparison, the model achieved an MAE of 0.5 weeks in healthy term‐born neonates [49]. Only about half of the algorithms reviewed achieved an MAE below 1 week in neonatal cohorts, and none reached below 3 weeks for infants aged 0 to 24 months [25, 42, 43, 47], which might limit their usability in a clinical setting in preterm.

It should be noted that only a few of the studies in the reviewed literature have clearly examined clinical applications, which are summarized in Table 5. We found four studies that applied their methods to prematurity [45, 47, 48, 49]. One fetal brain study reported that fetuses with ventriculomegaly showed a larger MAE compared to typically developing fetuses [35], indicating a greater deviation from the normal trajectory.

TABLE 5.

Summary of the clinical applications of deep learning studies on developmental brain age estimation.

Paper	Model characteristics	Clinical application
Tang et al. (2023) [47]	Used preterm infant dataset as training and test set. Grad‐CAM heatmaps highlighted hindbrain and cortical folding areas.	Might be limited: the predicted brain age of preterm infants was unstable.
Zhao et al. (2024) [48]	Used term and preterm infant dataset. Mean curvature and sulcal depth were characterized for folding morphology.	The predicted brain age of preterm infants was younger than their chronological age, indicating delayed brain development in preterm infants.
Zhao et al. (2024) [49]	Used term and preterm infant dataset. Both structural and microstructural MRI data were used, with the frontal and temporal lobes as the main focus of classification.	The predicted brain age of preterm infants was younger than their chronological age, indicating delayed brain development in preterm infants.
Park et al. (2024) [45]	Used a multi‐site dataset of healthy term and preterm neonates. The most important predictive features were the right lateral ventricle, left caudate, and corpus callosum.	The predicted brain age of preterm infants was younger than their chronological age, indicating delayed brain development in preterm infants.
Yun et al. (2025) [35]	Used typical and vetriculomegaly fetal dataset.	The predicted brain age of fetuses with isolated ventriculomegaly was older than their chronological age, indicating atypical growth trajectory. The predicted brain age of fetuses with ventriculomegaly accompanied by other central nervous system abnormalities was younger than their chronological age, indicating delayed brain development.

Among studies on preterm infants, Zhao et al. [49] applied a multimodal fusion approach and showed deviations in the predicted age scale with prematurity, reflecting developmental delay. Its integration of structural and diffusion features yields biologically interpretable patterns, though dependence on both modalities may limit use in routine scans and it is unknown whether their approach is agnostic in other datasets from other hospitals. Tang et al.'s BAPNET models [47], trained and tested solely on preterm data, risk bias toward pathological norms, potentially masking the very brain‐age gaps they aim to detect. This highlights the need for training on healthy references to preserve sensitivity to developmental deviation. Zhao et al. [48] applied geometric DL. By leveraging cortical surface morphology and hemispheric asymmetry, it achieved high predictive performance with a MAE = 0.53 weeks [48]. The model effectively captured a slower maturational trajectory in preterm infants, with predicted brain age lagging behind chronological age. This framework connects structural folding features to neurodevelopmental timing.

In ventriculomegaly fetuses, Yun et al. [35] applied slice‐based 2D approach, showing increasing positive brain age gap from mild to severe ventriculomegaly fetuses, suggesting a neurodevelopmental advancement rather than a delay. This likely reflects morphological distortions associated with ventricular enlargement (e.g., cortical stretching) rather than true accelerated maturation. Nevertheless, the model captured disease severity and differentiation between isolated and syndromic ventriculomegaly.

Beyond preterm infants and fetuses, DL models can characterize atypical brain growth trajectories in infants and young children. Structural MRI‐based models have detected abnormal patterns associated with autism spectrum disorder, including children up to 5 years of age, beyond the 0–2‐year range of this review [61]. Hu et al. [46] also summarized DL studies across early childhood. An MRI‐based AI approach has focused on predicting neurodevelopmental outcomes rather than estimating brain age [37]. While most studies have focused on preterm or neonatal cohorts, conditions such as congenital heart disease remain underexplored.

Our proposed strategy for clinicians who intend to use DL‐based brain age estimation is to select methods that have sufficiently low MAE for their given cohort (e.g., < 1 week for fetal and neonatal cohorts or < 4 weeks for older infants), which are generalizable and rely on larger datasets (several hundred) for training, and which have documented training and inference code. Since several factors, such as imaging parameters or post‐processing can introduce bias, we recommend clinicians test any algorithm on images of typically developing controls acquired on the same imager as the cases they are studying.

Despite its potential, the clinical applicability of DL‐based brain age estimation remains limited by challenges such as model generalizability, dataset heterogeneity, and variations in MRI protocols. Case‐control designs may detect group differences, but they have limited utility for psychiatric disorders, which are diagnosed by symptom clusters rather than biomarkers. In addition, translating into clinical practice may be complicated by population variability or imaging differences, highlighting the need for multimodal approaches combining brain age with genetic and clinical data.

4.1.5. Diffusion and Functional MRI

While this review primarily focuses on structural MRI for DL–based brain age estimation, other modalities such as DWI and functional MRI (fMRI) have also been explored. Zhao et al. [49] applied a DenseNet model to neonatal DWI data from the dHCP, reporting an MAE of 0.50 compared to 0.53 with T2‐weighted MRI. When combining both modalities in a transformer framework, performance improved further to 0.51, highlighting the potential of multimodal integration. We did not find other DL estimators for our age range that focused on DWI and fMRI. Yet, several studies have examined DWI or fMRI without DL. Pruett et al. [62] used resting‐state fMRI with SVMs to classify infants at 6 versus 12 months, demonstrating marked reorganization of large‐scale functional networks during early development. In youth aged 8–22 years, resting‐state fMRI predicted chronological age with r = 0.60 and MAE = 2.43, and the brain age gap showed associations with psychopathology. DWI has also provided insights in neurodegeneration: DWI brain age appeared older than T1‐based estimates in individuals transitioning from cognitively normal to mild cognitive impairment, but younger in those with Alzheimer's disease. Together, these studies suggest that different modalities capture distinct biological processes of brain maturation and aging. Future multimodal approaches integrating structural, diffusion, and functional MRI, potentially alongside non‐MRI measures, may improve both accuracy and interpretability of brain age estimation.

4.2. Open Datasets of the Developing Brain

4.2.1. Age Coverage

Publicly available perinatal and infant brain MRI datasets are valuable for clinical research as they provide reference data for method development and studying normal maturation; thus it is also important for brain age estimation methods. Datasets such as the dHCP Neonates Release and BOBs include perinatal brain MRI data (Table 3), which is valuable for investigating prematurity, birth complications, and early brain development. The dHCP Neonates Release further includes preterm infants, making it particularly useful for assessing brain age gaps in this population. iSeg‐2019 and Penn‐CHOP Infant Brain Atlases cover rapid brain changes from birth to 2 years. Datasets like BCP and ChildBrainAtlas provide older children's brain MRI up to 7.5 years with both longitudinal and cross‐sectional data. These data may be useful for identifying early deviations in children at risk for autism, ADHD, or intellectual disabilities. Stanford Fetal MRI provides a valuable resource of multi‐plane MRI scans to understand fetal brain growth and early brain maturation [27].

4.2.2. Regional Segmentation

Manually annotated datasets like iSeg‐2019 and BOBs are helpful for training and validating brain age estimation. Besides, NIH‐PD and ChildBrainAtlas support normative developmental modeling where FeTA offers manually segmented T2WI from multiple centers across GA of 18–36 weeks. In several neonatal and fetal brain age prediction frameworks, a segmentation‐first approach is adopted (Figure 6). Structural MRI scans are segmented into key brain tissues such as gray matter, white matter, and cerebrospinal fluid, and morphometric or volumetric features derived from these tissues are then used as input. Accurate segmentation is challenging during early development due to rapid anatomical changes, requiring specialized methods like InfantFreeSurfer [63]. Hasan et al. [30] applied segmentation for GA estimation, a technique also used by Vahedifard et al. [33].

FIGURE 6.

Example of a deep learning approach based on U‐Net++ architecture for segmenting 30 brain regions in neonatal brain MRI (A). Using regional brain volumes, the postmenstrual age of neonates was predicted, showing a positive correlation with the actual postmenstrual age (correlation coefficient = 0.875, p < 0.001) (B). Figures from Park et al. [45].

4.2.3. Ethnic and Geographic Characteristics of Open Datasets

Geographic imbalance of open datasets may limit the generalizability of neurodevelopmental findings, as brain development is influenced by genetic, environmental, and socio‐cultural factors that differ across populations. Most publicly available pediatric neuroimaging datasets originated from North America and Europe (Table 3). In Europe, the UK‐based dHCP is a leading resource for fetal and neonatal imaging. Asian representation remains limited, with Singapore and China. The FeTA dataset adds fetal MRI data from multiple European centers and the USA. In contrast, data from Africa, South Asia, Latin America are largely missing. Studies demonstrated that brain age estimation models trained on limited populations may be sensitive to cohort characteristics [42], emphasizing the need for diverse data sources [28]. To address these, efforts such as the multi‐site iSeg‐2019 challenge and the Chinese multimodal neuroimaging dataset [54], and the Chinese multimodal neuroimaging dataset [64] have been implemented to enhance representation.

5. Future Directions

The studies reviewed highlighted diverse approaches for brain age estimation, including multi‐modal data integration, attention mechanisms, and DL architectures like transformers. Despite advances, challenges such as data availability, computational efficiency, and model generalizability persisted. Future research could focus on hybrid models that combine attention mechanisms with lightweight 3D networks. Such integration could enhance the models' ability to capture both global contextual information (via attention) and fine‐grained spatial features (via 3D convolutions), leading to more accurate and robust representations of complex anatomical structures in medical images. In addition, synthetic data generation techniques hold promise for addressing data scarcity [65]. For instance, generating high‐quality synthetic fetal and neonatal MRIs can expand training datasets and improve model robustness, particularly in pathological cases such as ventriculomegaly. Other synthetic data generation approaches such as SynthSeg [61], which leverages domain randomization to produce highly varied synthetic data, enable robust generalization across diverse clinical settings and MRI sequences, and may hold potential for developmental brain age estimation. Improving model interpretability remains another important priority. Visualization tools such as Grad‐CAM might increase clinical applicability.

Several DL models now achieve high accuracy in brain age estimation (< 1 week MAE), often exceeding the sensitivity of radiologists' visual assessment. However, integration into routine radiological reporting has not yet been realized. The available algorithms are Python‐based research tools, which are not directly suitable for clinical use. A practical implementation would require a user‐friendly interface, calibration or adaptation to site‐specific data, computational efficiency, and safeguards such as fallback to conventional methods or ensembling multiple DL approaches to ensure robustness. For neonatologists and child development specialists, such a tool could offer a comprehensive way to detect infants at risk of neurodevelopmental delay, even in the absence of obvious structural brain abnormalities.

A major limitation of the algorithms included in this review is the absence of the adoption of open science practices, including sharing code, trained models, and complete analysis pipelines. Such practices would substantially improve reproducibility, transparency, and comparability across studies. None of the studies provided access to the code used for model training, and only approximately 30% made inference code accessible. Establishing standards for code and data sharing, documentation, and model evaluation can support the translation of AI tools into clinical practice. Moreover, there is a pressing need for publicly available benchmarks. This should include standardized dataset splits (e.g., training/validation/test partitions made publicly available), harmonized evaluation metrics such as MAE reported in weeks, and consistent reporting standards including dataset characteristics, preprocessing steps, and cohort demographics. A potential initiative could be a DL brain age estimation challenge, where multi‐center benchmarks could be used across a wide range of ages, from fetus to neonate and infant. Such a collaborative effort would not only drive methodological improvements but also provide a shared resource for training, evaluation, and reproducibility in developmental neuroimaging research.

A further limitation is, which is also the conclusion of our open dataset review in chapter 2.3, that the data used to train DL‐based brain age estimation models may have demographic and geographical biases, which could limit their generalizability. Because brain development is shaped by genetic, environmental, and socio‐cultural factors, models trained on homogeneous datasets may fail to generalize. Possible strategies to overcome this limitation would be the use of demographically more diverse datasets, federated learning (to enable multi‐site model training without data sharing), and advanced data augmentation and domain adaptation methods such as synthetic data generation techniques [65] to enhance diversity in underrepresented populations.

6. Conclusion

Brain age estimation serves as a valuable biomarker for supporting early diagnosis of neurodevelopmental disorders and for predicting outcomes associated with preterm birth. This survey provides a comprehensive review of DL approaches for brain age estimation from MRI data, focusing on the developmental period from the fetus to 2 years old and open access datasets. We systematically analyzed existing studies, focusing on modeling strategies, performance outcomes, and architectural choices. Each DL model addresses challenges such as spatial resolution, computational efficiency, and feature representation. Overall, DL models have demonstrated potential for brain age estimation in early life. Continued advancement will depend on overcoming current limitations through methodological innovation, collaborative data sharing, and integration with multimodal data. Open and well‐annotated datasets are essential, but current open neuroimaging resources vary in imaging protocols and lack geographic diversity; developing harmonized, globally representative datasets will improve generalizability and real‐world applicability.

Asma ull H., Kaandorp M. P. T., Jakab A., and Kim H. G., “Developmental Brain Age Estimation From MRI Data: A Systematic Review of Deep Learning Approaches and Open Datasets,” Journal of Magnetic Resonance Imaging 63, no. 3 (2026): 650–671, 10.1002/jmri.70180.