Roles and applications of artificial intelligence in fetal and placental MRI: a literature review Authors: Dandan Ding, Fangmei Zhu, Xue Tang, Zhongxiang Ding Published: 2026 ABSTRACT BackgroundArtificial intelligence (AI) has been increasingly integrated with fetal and placental magnetic resonance imaging (MRI) to enhance the detection of abnormalities and streamline diagnostic processes. MRI, known for its superior soft tissue contrast and multiplanar imaging capabilities, is a critical tool for evaluating complex fetal and placental conditions. However, the large volume of MRI data often poses challenges for clinicians in providing timely and accurate diagnoses. Fetal and placental MRI are subject to several inherent limitations, including fetal motion artifacts, low signal-to-noise ratio, and operator-dependent variability, which can reduce image quality and hinder accurate diagnosis. Artificial intelligence has shown significant potential in addressing these challenges. For example, AI-based methods have been applied to motion artifact reduction, automated organ segmentation, and disease classification. In addition, recent studies have demonstrated that AI can reduce MRI scan times by up to 60% without compromising image quality, thereby improving diagnostic accuracy and workflow efficiency.MethodsA systematic PubMed search was conducted on January 16, and May 20, 2024, using predefined terms related to fetal and placental MRI and AI. Peer-reviewed English studies were included, while irrelevant articles, reviews, and editorials were excluded. Disagreements during the review process were resolved by a third reviewer.ResultsAfter systematic screening, 74 studies on AI applications in fetal MRI and 32 studies on placental MRI were included. Key applications in fetal MRI included motion correction (17.6%), organ segmentation (48.6%), and disease classification (6.8%). For placental MRI, studies primarily focused on placental invasion assessment (53.1%) and segmentation (28.1%). Relevant studies published between 2016 and 2024 were categorized by application area and analyzed in detail.ConclusionsThis review synthesizes extensive research on AI applications in fetal and placental MRI, highlighting its potential to enhance imaging quality, automate tasks such as segmentation and motion correction, and improve diagnostic accuracy. However, challenges remain, including reliance on small, single-center datasets, limited demographic and pathological diversity, and a predominance of Two-Dimensional (2D) imaging techniques. Addressing these issues through the development of diverse, multi-center datasets and the exploration of advanced Three-Dimensional (3D) imaging methods is essential. By overcoming these barriers and integrating multimodal approaches, AI holds immense promise for revolutionizing prenatal diagnostics and advancing personalized care.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12884-026-09494-3. Introduction Fetal and placental magnetic resonance imaging (MRI) serves as a noninvasive diagnostic adjunct to ultrasound, offering detailed insights into fetal development, particularly in scenarios where abnormalities are suspected. This imaging modality is crucial in cases involving maternal obesity, oligohydramnios, or advanced gestational age [1], where ultrasound imaging may be limited.Despite its advantages, fetal and placental MRI faces inherent challenges, including fetal motion artifacts, reduced signal-to-noise ratio, partial volume effects, and operator-dependent variability, all of which may compromise image quality and hinder accurate diagnosis. In recent years, there has been a notable increase in the use of AI in the medical field. AI refers to computational systems designed to perform tasks that typically require human cognitive functions, such as pattern recognition, decision-making, and learning from data [2]. Machine learning (ML) and deep learning (DL), which are subsets of AI, further enhance this capability by “learning” and “adapting” patterns from data without overt instructions or external human intervention and without explicit programming. Deep learning models, particularly convolutional neural networks (CNNs), are widely used in medical image analysis due to their ability to automatically learn hierarchical features from imaging data. These models have been successfully applied to tasks such as segmentation [3, 4], classification [5], and image reconstruction [6], demonstrating their effectiveness in fetal MRI analysis. Researchers have increasingly integrated AI with fetal and placental MRI to develop more automated, efficient approaches for detecting fetal abnormalities and potential pregnancy complications. The implementation of AI tools has already shown promise in improving image quality, segmentation accuracy, and disease detection in fetal MRI [3–5,7]. For example, deep learning-based approaches have been applied to motion correction and super-resolution reconstruction [7], enabling improved image quality despite fetal movement. In addition, CNN-based segmentation models have demonstrated high accuracy in delineating fetal brain and body structures [3, 4], significantly reducing manual processing time. Furthermore, machine learning and deep learning algorithms have been used for disease classification and developmental assessment [5], showing strong diagnostic performance in identifying conditions such as cortical malformations. This review aims to (1) explore the applications of AI in fetal and placental imaging, focusing on the advantages and limitations of using AI in fetal and placental MRI, (2) highlight current and emerging trends in AI-enhanced fetal and placental MRI, and suggest future research opportunities, and (3) examine how AI-driven digital technologies might reshape clinical practices in fetal and placental MRI. Materials and methods This study was conducted as a narrative review to provide a structured overview of current applications of artificial intelligence in fetal and placental MRI. The databases of PubMed were searched using predefined research terms, including “fetus,” “fetal,” “placenta,” “placentas,” “placental,” “artificial intelligence,” “machine learning,” and “deep learning,” “MRI,” and their spelling variations, for relevant studies. The complete research terms and search strategy are provided in Tables 1 and 2, respectively. The initial literature search was conducted on January 16, 2024. To ensure the inclusion of the most recent studies, an updated search was performed on May 20, 2024, using the same strategy. Both searches were independently conducted and screened by two reviewers to ensure consistency and completeness of study selection.The first reviewer was a postgraduate student with training in medical imaging research, while the second reviewer was a board-certified radiologist with over 8 years of clinical experience, specializing in fetal MRI. The updated search involved a full re-screening of all retrieved records rather than limiting the review to newly published studies. No restriction on publication start date was applied during the search process. Both reviewers conducted an independent review of all titles and abstracts. The studies with no relevance (i.e., ultrasound and pathology imaging) were excluded. Table 1Complete research terms and search strategy used for studies on the application of AI in fetal MRI#Search termsResults1“fetal“[Title/Abstract] OR “fetus“[Title/Abstract] OR “embryo“[Title/Abstract] OR “embryonal“[Title/Abstract] OR “intrauterine“[Title/Abstract] OR “antepartum“[Title/Abstract]506,1952“artificial intelligence“[Title/Abstract] OR “AI“[Title/Abstract] OR “deep learning“[Title/Abstract] OR “machine learning“[Title/Abstract] OR “neural network“[Title/Abstract] OR “neural networks“[Title/Abstract]303,2933“MRI“[Title/Abstract] OR “imaging“[Title/Abstract] OR “radiology“[Title/Abstract]1,298,7734#1 AND #22,8435#4 AND #33796#5 AND English [language]377 Table 1Complete research terms and search strategy used for studies on the application of AI in fetal MRI#Search termsResults1“fetal“[Title/Abstract] OR “fetus“[Title/Abstract] OR “embryo“[Title/Abstract] OR “embryonal“[Title/Abstract] OR “intrauterine“[Title/Abstract] OR “antepartum“[Title/Abstract]506,1952“artificial intelligence“[Title/Abstract] OR “AI“[Title/Abstract] OR “deep learning“[Title/Abstract] OR “machine learning“[Title/Abstract] OR “neural network“[Title/Abstract] OR “neural networks“[Title/Abstract]303,2933“MRI“[Title/Abstract] OR “imaging“[Title/Abstract] OR “radiology“[Title/Abstract]1,298,7734#1 AND #22,8435#4 AND #33796#5 AND English [language]377 Complete research terms and search strategy used for studies on the application of AI in fetal MRI Table 2Complete research terms and search strategy used for studies on the application of AI in placental MRI#Search TermsResults1“placenta“[Title/Abstract] OR “placentas“[Title/Abstract] OR “placental“[Title/Abstract]118,8622“artificial intelligence“[Title/Abstract] OR “AI“[Title/Abstract] OR “deep learning“[Title/Abstract] OR “machine learning“[Title/Abstract] OR “neural network“[Title/Abstract] OR “neural networks“[Title/Abstract]239,3073“MRI“[Title/Abstract] OR “imaging“[Title/Abstract] OR “radiology“[Title/Abstract]1,298,7734#1 AND #24615#4 AND #3826#5 AND English[language]82 Table 2Complete research terms and search strategy used for studies on the application of AI in placental MRI#Search TermsResults1“placenta“[Title/Abstract] OR “placentas“[Title/Abstract] OR “placental“[Title/Abstract]118,8622“artificial intelligence“[Title/Abstract] OR “AI“[Title/Abstract] OR “deep learning“[Title/Abstract] OR “machine learning“[Title/Abstract] OR “neural network“[Title/Abstract] OR “neural networks“[Title/Abstract]239,3073“MRI“[Title/Abstract] OR “imaging“[Title/Abstract] OR “radiology“[Title/Abstract]1,298,7734#1 AND #24615#4 AND #3826#5 AND English[language]82 Complete research terms and search strategy used for studies on the application of AI in placental MRI After the initial screening, full texts of the studies meeting the initial screening criteria were acquired and reviewed by each reviewer. The inclusion criteria were peer-reviewed studies in English, whereas the exclusion criteria were studies not related to fetal or placental MRI imaging and AI. Additionally, review articles, commentaries, and editorials were also excluded. Discrepancies between reviewers were resolved through discussion with a third senior reviewer, a chief radiologist with over 20 years of clinical and research experience in fetal MRI and medical image analysis. Results The flow diagram of the screening process for the studies on the application of AI in fetal MRI is presented in Fig. 1. After the screening, 377 titles and abstracts were reviewed. Of these,275 studies that failed to meet the selection criteria were removed. The remaining 102 relevant studies were acquired for full-text analysis. Following the initial search, 377 records were identified. Titles and abstracts of all retrieved records were screened for relevance.Based on predefined inclusion and exclusion criteria, 275 studies were excluded at this stage.The remaining 102 studies were assessed through full-text review for eligibility.The screening process was conducted in accordance with the predefined selection criteria described in the Materials and Methods section. Subsequently, 28 studies were excluded for the following reasons: 9 articles were only reviews,8 were duplicates, and 11 were not relevant. A total of 74 studies on the application of AI in fetal MRI were included in this review. Fig. 1Flow diagram of the review Flow diagram of the review Fig. 1 shows the flow diagram of the screening process of the studies on the application of AI in placental MRI. The application of AI to placental MRI was analyzed using the same method employed for fetal MRI. A total of 82 abstracts and titles were reviewed. However, 36 studies did not meet the selection criteria and were removed after initial screening. A full-text evaluation was conducted for the remaining 46 publications. Subsequently, 4 reviews, 2 editorials, 1 duplication and 7 studies were found to be irrelevant and were excluded. After applying the eligibility criteria, 32 studies relevant to the application of AI in placental MRI were included in our review. All studies were published between 2016 and 2024. The included studies were categorized based on their primary application in fetal and placental MRI. Each study was assigned to a single primary category according to its main objective, and in cases involving multiple applications, classification was based on the primary focus to avoid duplication. The categorization reflects observed patterns in the literature rather than a predefined framework. All included studies were carefully reviewed in full text by the research team. Based on the study content and their clinical relevance, a consensus was reached among the reviewers to categorize the studies into distinct application domains. The most clinically relevant and representative studies were selected and described in greater detail in the following sections (Fig. 2).As this review was designed as a narrative overview, no formal risk-of-bias assessment tool or grading rubric was applied. Instead, emphasis was placed on summarizing studies with high clinical relevance and methodological representativeness.Detailed characteristics and classification of all included studies are summarized in Supplementary Table S1. Fig. 2Different groups of AI in fetal and placental MRI. These pictures reproduced from Vollbrecht. et al. [8], 2024, Uus. et al. [3], 2024, Yehuda. et al. [5], 2023, Ji L. et al. [9], 2023, Zhu H. et al. [10], 2022 Different groups of AI in fetal and placental MRI. These pictures reproduced from Vollbrecht. et al. [8], 2024, Uus. et al. [3], 2024, Yehuda. et al. [5], 2023, Ji L. et al. [9], 2023, Zhu H. et al. [10], 2022 Application of AI in fetal MRI (n = 74) Image preprocessing of fetal MRI Fetal motion and artifact correction (n = 13, 17.6%) [8, 11–22].Fetal orientation (n = 1, 1.4%) [23]. Fetal motion and artifact correction (n = 13, 17.6%) [8, 11–22]. Fetal motion and artifact correction (n = 13, 17.6%) [8, 11–22]. Fetal orientation (n = 1, 1.4%) [23]. Fetal orientation (n = 1, 1.4%) [23]. Image postprocessing of fetal MRI Segmentation of fetal organs (n = 36, 48.6%) [3, 4, 7, 24–56].Automated fetal biometry measurement (n = 7, 9.5%) [57–63].Assessment of image quality (n = 2, 2.7%) [64, 65]. Segmentation of fetal organs (n = 36, 48.6%) [3, 4, 7, 24–56]. Segmentation of fetal organs (n = 36, 48.6%) [3, 4, 7, 24–56]. Automated fetal biometry measurement (n = 7, 9.5%) [57–63]. Automated fetal biometry measurement (n = 7, 9.5%) [57–63]. Assessment of image quality (n = 2, 2.7%) [64, 65]. Assessment of image quality (n = 2, 2.7%) [64, 65]. Data interpretation of fetal MRI Classification of disease (n = 5, 6.8%) [5, 66–69].Gestational age prediction (n = 5, 6.8%) [70–74].Fetal surgery (n = 1, 1.4%) [75]. Classification of disease (n = 5, 6.8%) [5, 66–69]. Classification of disease (n = 5, 6.8%) [5, 66–69]. Gestational age prediction (n = 5, 6.8%) [70–74]. Gestational age prediction (n = 5, 6.8%) [70–74]. Fetal surgery (n = 1, 1.4%) [75]. Fetal surgery (n = 1, 1.4%) [75]. Miscellaneous fetal MRI Generation of synthetic data (n = 2, 2.7%) [6, 76].Functional and dynamic MRI (n = 2, 2.7%) [9, 77]. Generation of synthetic data (n = 2, 2.7%) [6, 76]. Generation of synthetic data (n = 2, 2.7%) [6, 76]. Functional and dynamic MRI (n = 2, 2.7%) [9, 77]. Functional and dynamic MRI (n = 2, 2.7%) [9, 77]. Application of AI in placental MRI (n = 32) Segmentation of placenta (n = 9, 28.1%) [78–86].Assessment of placental invasion and placenta previa (n = 17, 53.1%) [10, 87–102].Placental phenotype and function in fetal congenital heart disease (n = 1, 3.1%) [103].Prediction of fetal growth restriction and postnatal outcome (n = 2, 6.3%) [104, 105].Spectroscopic analysis of quantitative MRI (n = 1, 3.1%) [106].Prediction of placental health (n = 2, 6.3%) [107, 108]. Segmentation of placenta (n = 9, 28.1%) [78–86]. Segmentation of placenta (n = 9, 28.1%) [78–86]. Assessment of placental invasion and placenta previa (n = 17, 53.1%) [10, 87–102]. Assessment of placental invasion and placenta previa (n = 17, 53.1%) [10, 87–102]. Placental phenotype and function in fetal congenital heart disease (n = 1, 3.1%) [103]. Placental phenotype and function in fetal congenital heart disease (n = 1, 3.1%) [103]. Prediction of fetal growth restriction and postnatal outcome (n = 2, 6.3%) [104, 105]. Prediction of fetal growth restriction and postnatal outcome (n = 2, 6.3%) [104, 105]. Spectroscopic analysis of quantitative MRI (n = 1, 3.1%) [106]. Spectroscopic analysis of quantitative MRI (n = 1, 3.1%) [106]. Prediction of placental health (n = 2, 6.3%) [107, 108]. Prediction of placental health (n = 2, 6.3%) [107, 108]. Overall, segmentation represented the most extensively studied application, while areas such as functional MRI and predictive modeling remained relatively underrepresented, suggesting potential directions for future research. This underrepresentation may be partly explained by technical and practical challenges. Functional MRI in the fetal setting is particularly sensitive to motion artifacts and requires prolonged acquisition times, making it difficult to obtain stable and reliable data. In addition, the complexity of fetal brain development and the lack of standardized acquisition protocols further limit the widespread application of functional magnetic resonance imaging (fMRI). Similarly, although three-dimensional MRI techniques offer improved spatial resolution and more comprehensive anatomical representation, their clinical use remains constrained by longer acquisition times and increased susceptibility to motion, especially in uncooperative fetal conditions. Nevertheless, with ongoing advances in motion correction, fast imaging sequences, and AI-based reconstruction techniques, both fMRI and 3D MRI are expected to play an increasingly important role in future fetal and placental imaging research. Image preprocessing of fetal MRI Motion artifacts result from a moving fetus or the mother. Excessive and unpredictable movements are most commonly seen with fetal heart and brain MRI. Usually, a highly trained technician is needed to adjust acquisition planes or reacquire sequences repeatedly.This can be relatively time-consuming and may vary in quality depending on the operator’s level of expertise. Preprocessing tools for motion correction are one of the potential solutions to this issue. Utilizing AI models, the automated correction for fetal and maternal motion can be theoretically less time-consuming and more accurate, thus accessing higher-quality images. However, models trained on high-quality datasets may face challenges in generalizability when applied to data acquired under different imaging conditions, raising concerns about potential overfitting and reduced robustness in real-world clinical settings, particularly when applied to data from different scanners or institutions.Therefore, external validation across multi-center and heterogeneous datasets is essential to ensure model robustness and clinical applicability. Vollbrecht et al. [8]conducted a study to assess the effectiveness of Speedsmart AI, a deep learning-based denoising technique, in enhancing the image quality of fetal cardiac cine MRI. The study included 25 fetuses with congenital heart disease (CHD) scanned using a 3T MRI system with balanced steady-state free precession (bSSFP) sequences and Doppler ultrasound gating. Compared to compressed sensing (CS), the deep learning method demonstrated significantly better image quality, as confirmed by a qualitative assessment using a 5-point Likert scale. The deep learning reconstructions notably increased both the apparent signal-to-noise ratio (aSNR: 13.4 ± 6.9 vs. 8.3 ± 3.6, P < 0.001) and contrast-to-noise ratio (aCNR: 26.6 ± 15.8 vs. 14.4 ± 6.8, P < 0.001). These improvements suggest that integrating Speedsmart AI into fetal cardiac MRI could enhance diagnostic confidence and overall image quality. Eisenstat et al. [23]developed a DL algorithm based on the novel convolutional neural networks (CNN) architecture called Fet-Net, which used fewer layers and parameters than 11 other prominently used CNN architectures, such as ResNet, VGG, Xception, and others. Despite the fewer architectural layers and parameters, Fet-Net still demonstrated improved classification accuracy regarding fetal orientation. Eisenstat et al. used 144 3D coronal T2-weighted imaging (T2WI) that demonstrated vertex, breech, oblique, and transverse fetal orientations and extracted 6120 2D MRI slices to train, validate, and test Fet-Net. After a 5-fold cross-validation experiment on the 6120 2D MRI slices, the Fet-Net yielded an average accuracy and F1 score of 97.68% and a loss of 0.06828, which outperformed the other 11 prominent architectures (P < 0.05). Even when noise was introduced to the images, Fet-Net still provided robust classification accuracy, outperforming 8 out of the 11 prominent architectures, and an ablation study was performed to prove each component’s statistical significance and contribution to improved performance. Fet-Net required less time for fetal MRI acquisition and allowed for expedited MRI acquisition, as well as a reduced number of required MRI technicians due to its capabilities for the automatic detection of fetal orientation. However, most studies, including this one, were conducted during early development stages and primarily focused on the feasibility of preprocessing tools rather than providing quantifiable results. While these studies often proposed potential pipelines for implementation, quantifiable outcomes regarding patient care and diagnostic accuracy were not thoroughly assessed. Future research should focus on testing these AI models for their impact on patient outcomes and diagnostic accuracy. In addition, motion correction and super-resolution reconstruction techniques have already demonstrated clinical utility, particularly in fetal brain MRI, where the relatively rigid anatomical structure allows more reliable reconstruction. These approaches can improve visualization of anatomical details in cases of persistent fetal motion, thereby supporting more accurate clinical interpretation. However, further validation across different fetal organs and imaging conditions is still required before widespread clinical adoption. Image postprocessing of fetal MRI Imaging postprocessing for fetal MRI mainly focuses on anatomical segmentation. For example, the manual delineation of 2D images is typically done to produce 3D reconstruction images. This technique is commonly done with fetal brain imaging by manually delineating all brain contours in all three orientational planes. This method is not only incredibly time-consuming but also unsuitable for large-data analyses. Usually, it takes approximately 30–40 min for a trained technician to complete a fetal brain segmentation. However, this previously labor-intensive manual delineation can be reduced to a few seconds by incorporating AI tools, such as the Brain Extraction Tool and other DL-based segmentation tools. In addition, automated segmentation methods can reduce inter-observer variability by replacing manual delineation with standardized and reproducible algorithm-driven outputs, thereby minimizing subjective differences between observers and improving both inter- and intra-observer consistency in clinical assessment. Fidon et al. [25] introduced a Dempster-Shafer theory-driven approach to enhance the reliability of AI in fetal brain MRI segmentation. This method incorporates a fail-safe mechanism that seamlessly integrates expert knowledge, significantly boosting the robustness of deep learning models across 540 manually annotated 3D T2-weighted MRIs from 13 diverse centers. The framework not only refines the segmentation process but also aligns with emerging regulatory standards for trustworthy AI in medical imaging. Uus et al. [3, 61] developed an automated segmentation protocol for fetal body organs in 3D motion-corrected T2-weighted MRI, setting a new standard for fetal volumetry. This deep learning-based approach addresses the time-consuming nature of manual segmentation and reduces observer bias. The protocol has been validated across 91 normal control datasets, demonstrating robust performance and a significant reduction in the necessity for manual refinements. Covering 10 organ regions of interest (ROIs) enabled the creation of the first population-averaged T2WI MRI atlas, providing a standardized tool for fetal organ volumetry and growth assessment across gestational ages. This advancement enhances clinical diagnostics by enabling more efficient and accurate fetal development assessments, including detecting anomalies such as fetal growth restriction. Several studies [3, 4, 7, 24] have proposed segmentation and reconstruction of fetal MRI for growth monitoring, pathology identification, and surgical planning. Many automated segmentation methods compare images with gestation-specific reference atlases to classify structures of interest [3, 43], while other statistical AI algorithms can segment without requiring a reference atlas [41, 44]. Although most 3D segmentation and reconstruction efforts focus on the fetal brain due to its relatively firm structure and limited motion [7, 40, 41, 43], segmentation of other fetal structures, such as the thorax, has also been explored [24, 31, 39]. For instance, accurate fetal lung volume calculations can be made from 20 to 38 weeks of gestation through the segmentation and reconstruction of the thorax. Studies on fetal whole-body MRI extraction have also shown promising results [3, 4, 26]. Data interpretation of fetal MRI Detection, classification, and prognostication of diseases using fetal MRI are complex and require large datasets for training, which has limited number of dedicated studies. However, early detection of abnormalities is crucial for managing pregnancy complications. Attallah et al. [69] developed DL algorithms to detect embryonic neurodevelopmental disorders, evaluating three CNNs for accuracy, sensitivity, and specificity. Their algorithm demonstrated high classification accuracy in detecting embryonic neurodevelopmental abnormalities, potentially streamlining the diagnostic process for embryonic brain deficiencies. Yehuda et al. [5] an automated pipeline for fetal brain gyrification assessment based on routinely acquired 2D MRI, enabling quantitative evaluation of cortical development across gestation. The study included 162 fetuses, comprising 134 controls, 12 with lissencephaly, and 13 with polymicrogyria. Using machine learning classifiers, the method demonstrated strong performance in distinguishing abnormal from normal gyrification patterns. Specifically, an XGBoost-linear model achieved an area under the receiver operating characteristic curve (AUROC) of 0.90 with a recall of 0.83 for classifying lissencephaly versus controls (n = 32), while a random forest model achieved an AUROC of 0.84 with a recall of 0.62 for classifying polymicrogyria versus controls (n = 33). These results highlight the potential of AI-assisted quantitative analysis for improving the early detection of fetal cortical malformations. The most accurate method for predicting gestational age during the first trimester of pregnancy is ultrasound [71]. However, in many resource-limited settings, women do not seek prenatal care this early or are not identified during the first trimester to be enrolled in pregnancy cohorts. The AI platform can accurately predict gestational age based on fetal brain MRIs that are developmentally normal, which helps in managing pregnancy and fetal outcomes. Shen et al. [73] showed promising results of a prediction model, an end-to-end, attention-guided DL model, for estimating gestational age. The DL model was trained with T2WIs (741 training and 434 validation cases) for testing. The coefficient of determination (R2) was calculated based on the relationship between the model’s estimated gestational age and the reference standard. A high concordance correlation (R2 = 0.945) was obtained, indicating the high accuracy of the prediction model to guide real-time gestational age estimation after the first trimester. However, the applicability of such AI-based approaches may be limited in low-resource settings, where access to MRI remains restricted and prenatal imaging is less routinely performed.These constraints may reduce the feasibility of implementing MRI-based AI solutions in routine clinical practice in such environments. Furthermore, from a methodological perspective, the “black-box” nature of many deep learning models remains a major barrier to clinical adoption, as the lack of interpretability may limit clinician trust and hinder integration into routine clinical workflows. Emerging approaches such as explainable AI may help improve model transparency and facilitate clinical adoption. Miscellaneous fetal MRI Advances in synthetic data generation and functional and dynamic MRI have significantly improved fetal imaging by overcoming challenges such as data scarcity, motion artifacts, and the complexity of brain development [6, 9, 76, 77] .These techniques improve diagnostic accuracy and deepen our understanding of prenatal brain function, providing a solid foundation for early assessment and potential intervention in fetal brain development. Torrents-Barrena et al. [6] introduced a Generative Adversarial Network (GAN) with a balanced auto-encoder architecture to generate synthetic fetal brain MRI images in the axial, sagittal, and coronal planes. This method effectively mitigates data scarcity and privacy concerns by producing high-quality, anatomically accurate images that closely mimic actual scans, creating an invaluable resource for training machine learning models in fetal brain imaging. Ji et al. [9] utilized four-dimensional functional MRI (4D fMRI) to study fetal motor behavior and its relationship to brain network dynamics as gestation progresses. By analyzing motor activity in 120 fetuses, they observed a decrease in motor activity with age, likely due to spatial constraints, and identified significant associations between fetal motor activity, quantified using framewise displacement (a proxy for head motion), and network interactions in crucial brain areas, such as the supplementary motor area. This research highlights fMRI’s potential to reveal the maturation of fetal brain organization and its connection to postnatal motor development. Application of AI for placental MRI Fetal health and pregnancy outcomes are closely linked to the condition of the placenta, which changes throughout gestation in response to fetal demands. As the placenta ages, it becomes more lobulated, visible on low-intensity T2-weighted MRI. Accelerated placental aging is associated with increased risks of placental failure, fetal growth restriction, pre-eclampsia, and late stillbirth [109]. Conversely, delayed placental maturation has been associated with an increased risk of gestational diabetes and chromosomal abnormalities.The studies highlighted in this section were selected based on their clinical relevance, methodological representativeness, and their ability to reflect key application areas in placental MRI, following a comprehensive review of all included studies. Placental segmentation is widely used to detect pregnancy-related complications, such as placenta accreta or fetal growth restrictions. However, this review was limited by the relatively small size of the training datasets, which are essential for representing the diverse shapes and locations of the placenta. In addition, variability in MRI acquisition protocols across institutions—including differences in scanners, imaging sequences, and acquisition parameters—as well as variations in placental position (e.g., anterior, posterior, or lateral) may further influence model performance. These factors collectively limit the generalizability of current AI models. Variations in fetal size, appearance, and orientation at different gestational stages also impacted algorithm performance. While deep learning-based segmentation methods showed great potential for clinical applications in placental imaging, future research should focus on larger, multi-center datasets and improved standardization of imaging protocols to enhance model robustness and facilitate broader clinical applicability. Similarly, Ye et al. [94] proposed a combined radionics-clinical deep learning model for the prenatal detection of placenta accreta spectrum (PAS) disorder using preoperative T2-weighted images. Their study retrospectively recruited 407 pregnant patients from two institutions. Patients from the first institution were divided into a training cohort (298 patients) and a validation cohort (75 patients), while patients from the second institution formed the external test cohort (34 patients). The combined model, which integrates a clinical prediction model, radiomic features, and deep semantic features, achieved an Area under the curve (AUC) of 0.872 (95% CI, 0.843–0.908) in the validation cohort and 0.857 (95% CI, 0.808–0.894) in the external test cohort, demonstrating strong predictive performance for PAS. Zhu et al. [95] developed a ML model based on sagittal-view T2WI to improve diagnostic accuracy for PAS disorders, comparing the results with surgical and histological findings. Their study included 512 pregnant patients, with 358 in the training and 154 in the testing cohorts. The model, which combines U-net-based automatic segmentation with radiomic and clinical features, yielded an accuracy of 0.825, sensitivity of 0.830, specificity of 0.822, and an AUC of 0.800 (95% CI, 0.743–0.850) in the testing cohort. These results highlight the potential of ML models to enhance PAS diagnostic accuracy. Li et al. [104] developed a hybrid model that leverages automatic placental radiomics on T2-weighted preoperative MRI to assess fetal growth restriction (FGR). The model integrates deep learning for sensitive segmentation and a random forest algorithm for feature fusion, using data from 274 pregnant women, including both normal and FGR cases, retrospectively collected from two centers. This approach achieved an AUROC of 0.923 for the internal test set and 0.880 for the external validation set (accuracy: 0.771–0.825, sensitivity: 0.854, specificity: 0.750–0.822), highlighting its potential to enhance diagnostic precision in FGR. Discussion In this systematic literature review, we conducted a comprehensive search in the PubMed database using predefined keywords to identify studies on the applications of artificial intelligence (AI) in fetal and placental MRI. After an initial screening, 74 studies on AI applications in fetal MRI and 32 studies on placental MRI were selected for review. In terms of performance, many segmentation studies reported high accuracy, with Dice similarity coefficients frequently exceeding 0.80, indicating reliable delineation of fetal structures. Similarly, classification models for disease detection demonstrated strong diagnostic performance, with reported AUC values often above 0.80. In image preprocessing, AI-based approaches have shown improvements in image quality, reflected by enhanced signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). Additionally, gestational age prediction models have demonstrated strong agreement with reference standards, indicating robust predictive capability.These quantitative findings further support the potential of AI to improve accuracy, efficiency, and consistency in fetal and placental MRI analysis. However, despite these promising results, several important challenges remain. The results revealed three main challenges: First, the datasets were generally small and mainly from single-center studies, limiting AI models’ generalisability and broader clinical applicability. Second, most research relied on deep learning techniques such as CNNs and GANs, focusing on two-dimensional imaging tasks such as segmentation and motion correction. There has been limited exploration of advanced three-dimensional approaches, such as volumetric reconstruction, three-dimensional segmentation, and spatiotemporal modeling, which are essential for capturing complex anatomical structures and dynamic fetal motion. Thirdly, although AI has demonstrated considerable potential in automating tasks such as organ segmentation and disease classification, its clinical application remains in the infancy and requires further validation before it can be fully implemented in clinical practice. Dataset Current studies on fetal and placental MRI often rely on small, institution-specific datasets, which limits the generalizability of AI models. The lack of diversity in patient demographics and pathology types further restricts their applicability in broader clinical settings. To address this, future work should prioritize more diverse and representative datasets and improved data sharing to enhance model robustness.Moreover, the limited availability of open-source datasets impedes collaboration and external validation efforts. To overcome these challenges, future research should focus on developing large, multi-center datasets that encompass a wide range of pathologies and demographic variations. This will allow AI models to be trained and validated on data that more accurately reflects real-world clinical environments, enhancing both their robustness and clinical relevance in prenatal diagnostics. Technique AI integration in fetal and placental MRI has predominantly employed deep learning models, particularly CNNs and GANs, for tasks like image segmentation, motion correction, and disease classification. While these methods show great potential, they require substantial computational resources and large, well-annotated datasets for training. Current research [5, 23, 60, 110] is mainly focused on 2D imaging.This is because 2D MRI techniques are more robust to fetal motion due to the use of fast single-shot acquisition sequences, which reduce motion artifacts and improve image reliability. In addition, 2D imaging requires shorter acquisition times and is more widely available in clinical practice, making it more feasible for routine use despite its limitations in spatial resolution.However, there is growing interest in transitioning to 3D imaging techniques, which can better capture complex anatomical structures and improve diagnostic precision. Future research should explore more advanced algorithms, such as transformer models, which have demonstrated the ability to handle complex multimodal data in other fields. Additionally, integrating genetic, environmental, and clinical data with imaging will likely enhance diagnostic accuracy and provide a more comprehensive understanding of fetal and placental health. In addition, standardized MRI acquisition protocols and data preprocessing pipelines are critical for the development of robust AI models. Variability in imaging parameters, scanner types, and data annotation practices can significantly affect model performance.Furthermore, careful data curation, including quality control and consistent annotation strategies, is essential to ensure the reliability of training datasets. Future research should emphasize data harmonization, quality control, and standardized preprocessing strategies to improve reproducibility and facilitate multi-center collaboration. Moreover, motion correction is not only essential for improving image quality but also represents a critical step in quantitative MRI pipelines [7]. In placental imaging, robust motion correction is particularly important for enabling reliable functional modeling, such as diffusion-based or perfusion-related analyses [111]. Without adequate motion compensation, these quantitative approaches may be significantly compromised, limiting their clinical applicability. Application AI has been applied to several key areas in fetal and placental MRI, demonstrating significant potential to improve diagnostic accuracy and streamline workflows. In fetal MRI, AI is mainly used for automatic fetal organ segmentation, motion artifact correction, and disease classification. In placental MRI, AI plays a vital role in placental segmentation, predicting conditions like PAS, and assessing placental function. These applications highlight AI’s capacity to minimize human error and reduce processing times, thus enhancing the efficiency of diagnostic workflows. However, despite these advancements, the clinical integration of AI is still in its early stages. Further validation is necessary to transition these technologies from research to routine clinical practice. Future efforts should focus on testing AI tools in real-world clinical settings to ensure their robustness, generalizability, and clinical utility in prenatal diagnostics. Limitation This review has several limitations. While it provides a comprehensive overview of AI applications in fetal and placental MRI, a formal assessment of study quality or risk of bias was not performed, as this review was designed as a narrative overview, which may limit the ability to quantitatively compare findings across studies. First, it relies on studies indexed in major databases, which may have led to the omission of relevant research not covered by these sources. Second, the review focuses primarily on the medical implications of AI in fetal and placental MRI, neglecting non-medical research essential for translating AI advancements into practical clinical applications. Third, ethical considerations are critical for the clinical implementation of AI. Issues such as data privacy, patient consent, and potential bias in training datasets may disproportionately affect underrepresented populations. For example, if certain demographic groups or rare pathologies are underrepresented in training data, AI models may perform suboptimally when applied to these populations, thereby limiting their generalizability. In addition, inconsistencies in data acquisition, annotation practices, and institutional protocols across studies may limit the generalizability of AI models and hinder their transferability to real-world clinical settings. These variations can lead to performance degradation when models are applied to external datasets, highlighting the need for more standardized and representative data.The use of publicly available and multi-center datasets may partially alleviate these challenges by increasing data diversity and facilitating external validation; however, such datasets may still be limited in scale, quality, and population representation, and therefore cannot fully eliminate these concerns. Fourth, the high cost of MRI acquisition, infrastructure requirements, and the need for specialized expertise may limit the widespread clinical implementation of fetal and placental MRI, particularly in low-resource settings. These economic and logistical barriers may also restrict the deployment of AI-based tools that rely on high-quality MRI data. However, in some rapidly developing healthcare systems, the availability of fetal MRI has increased in recent years, particularly in tertiary care and specialized maternal–fetal centers, which may facilitate the future clinical translation of AI-based approaches. Future prospects Future efforts should focus on improving the robustness and clinical applicability of AI tools in real-world settings. Leveraging advanced AI techniques capable of handling multi-modal data will be crucial for improving diagnostic accuracy. In fetal AI, the focus should be on early diagnosis and prognosis of brain anomalies and neurodevelopmental disorders, with a comprehensive fetal brain atlas serving as a reference for detecting abnormalities across gestational ages. For placental AI, understanding the impact of placental abnormalities on fetal development, especially along the placenta-fetal heart/brain axis, is critical. Research should also explore how AI can predict and manage high-risk pregnancies like preeclampsia and gestational diabetes by analyzing placental function. In the future, research in these fields will continue to depend on the further development of public datasets and the collaborative efforts among obstetrical radiologists, perinatologists, and engineering technicians in this domain. Conclusions This comprehensive literature review synthesizes the extensive research on the applications of AI in fetal and placental MRI, underscoring its potential to enhance imaging quality, automate tasks such as segmentation and motion correction, and improve diagnostic accuracy. While AI has made significant strides, challenges remain, including the reliance on small, single-center datasets, limited demographic and pathological diversity, and a predominance of 2D imaging techniques. Addressing these issues is crucial through the development of diverse, multi-center datasets and the exploration of advanced 3D imaging methods. By overcoming these barriers and integrating multimodal approaches, AI holds immense promise for revolutionizing prenatal diagnostics and advancing personalized care. The integration of AI in fetal and placental MRI has the potential to extend to disease prevention, early diagnosis, treatment decision-making, and efficient patient management, offering transformative opportunities for prenatal and maternal-fetal healthcare. However, for AI to reach its full potential in clinical practice, further validation and real-world testing are necessary. Future research should focus on constructing large, multi-center datasets that reflect the breadth of pathologies and demographic variations, and on developing AI models capable of handling complex multimodal data. Collaborative efforts among obstetrical radiologists, perinatologists, and engineering technicians will be instrumental in translating these research advancements into clinical reality. In conclusion, AI’s impact on fetal and placental MRI is profound, offering a glimpse into a future where prenatal diagnostics are more accurate, efficient, and personalized. As the field progresses, the ethical considerations surrounding data privacy, patient consent, and the responsible use of AI must be addressed to ensure the technology’s successful and widespread adoption. Supplementary Information Supplementary Table S1. Supplementary Table S1.