This study was designed to evaluate the utility of artificial intelligence (AI)-based fetal MRI for the early screening of FGR and PI (Fig. 1). Specifically, the study aimed to validate the diagnostic performance of the built-in AIR™ Recon DL reconstruction algorithm, rather than to develop a new CNN model. The study population consisted of pregnant women undergoing fetal MRI examinations at the Obstetrics Department of West China Second University Hospital from September 2024 to May 2025. All participants carried singleton pregnancies. Based on the diagnostic criterion for FGR (EFW below the 10th percentile for gestational age), participants were assigned to an FGR group. This diagnostic threshold follows the clinical guidelines issued by the International Society of Ultrasound in Obstetrics and Gynecology (ISUOG), which define FGR as an EFW below the 10th percentile for gestational age [10]. Although all eligible participants were consecutively enrolled, to minimize the potential influence of gestational age on group comparisons and to maintain statistical balance, frequency matching by gestational age was applied during analysis. This approach resulted in equal numbers of participants in the FGR and non-FGR groups (n = 60 each), totaling 120 participants. The sample size was determined via power calculation ($\alpha$ = 0.05, power = 0.8) to detect a 15% difference in FGR diagnostic accuracy.

Fig. 1.

Flowchart of the study pipeline for AI-based fetal MRI using the AIR™ Recon DL algorithm. AI, artificial intelligence; MRI, magnetic resonance imaging; DL, deep learning.

This retrospective study was reviewed and approved by the Institutional Review Board (IRB) of the Medical Research Ethics Committee of the Second Hospital of West China, Sichuan University (Approval No. 313, 2024). The requirement for written informed consent was formally waived by the IRB due to the retrospective nature of the study. All procedures were conducted in accordance with the Declaration of Helsinki and applicable national regulations.

Inclusion criteria were: (1) gestational age of 24–36 weeks; (2) absence of severe complications (e.g., malignant tumors, or endocrine disorders); and (3) MRI examinations of acceptable diagnostic quality without severe motion artifacts. Only fetuses without major structural or chromosomal abnormalities were included to ensure dataset homogeneity. Exclusion criteria included: (1) severe fetal anomalies; (2) MRI contraindications (e.g., implanted metal devices); and (3) inability to tolerate scanning.

FGR diagnosis was defined as an EFW below the 10th percentile for gestational age based on the Hadlock growth chart, which was also used to establish percentile thresholds for postnatal birthweight confirmation [10]. Cases were retrospectively reviewed, and final FGR classification was supported by birthweight percentiles and available prenatal Doppler findings, including an umbilical artery pulsatility index $>$95th percentile or a cerebroplacental ratio (CPR) 5th percentile. Neonatal outcomes, such as LBW (2500 g) and Neonatal Intensive Care Unit (NICU) admission, were also recorded for reference [3, 5, 10]. PI was defined according to abnormal Doppler parameters, including an umbilical artery pulsatility index $>$95th percentile and/or a CPR 5th percentile for gestational age, consistent with ISUOG recommendations [10]. In cases where Doppler data were unavailable, PI diagnosis was supported by MRI findings, such as reduced placental volume or elevated placental T2 values, and by adverse perinatal outcomes, such as LBW or NICU admission [5].

During the study period, 156 pregnant women underwent fetal MRI examinations. Of these, 132 met the inclusion criteria, and 120 were ultimately analyzed after excluding cases with incomplete data or poor image quality.

All fetal MRI examinations were conducted using a 3.0-T MRI scanner (SIGNA™ Premier, GE Healthcare, Chicago, IL, USA) equipped with a 32-channel GE AIR™ Coil body coil for signal reception. The coil’s high sensitivity and uniform signal coverage optimized the SNR and contrast-to-noise ratio (CNR), while effectively minimizing fetal motion artifacts during imaging of the placenta and fetal abdomen. The standard fetal MRI protocol included axial T2-weighted fat-saturated (FS) single-shot fast spin-echo (SSFSE), axial T1-weighted gradient-echo, sagittal T2-weighted SSFSE, coronal T2-weighted SSFSE, and three-dimensional (3D) T2-weighted sequences to capture images of the placenta and fetal abdomen. Both two-dimensional (2D) and 3D acquisitions were performed to assess volumetric reconstruction and segmentation. FGR assessment relied on MRI-derived fetal abdominal volume combined with fetal femur length (FL) measured via ultrasound to calculate EFW. Ultrasound examinations were performed using a GE Voluson E10 ultrasound system (GE HealthCare, Zipf, Austria) by certified sonographers with at least 5 years of fetal ultrasound experience, following standardized biparietal diameter and FL measurement protocols, as per ISUOG guidelines [10]. MRI and ultrasound examinations were conducted within the same gestational week (24–36 weeks), with a maximum interval of 3 days to ensure data consistency. Doppler ultrasound was performed within three days of MRI acquisition to measure the umbilical artery pulsatility index, middle cerebral artery PI, and to calculate the CPR. All examinations were performed by certified obstetric sonographers following ISUOG standards for fetal Doppler assessment [10].

The standard MRI protocol had an average scanning time of 18 minutes and 45 seconds, while an accelerated protocol averaged 7 minutes and 30 seconds. Raw accelerated MRI images were transferred to post-processing software (Orchestra SDK, GE Healthcare) and reconstructed using the AIR™ Recon DL algorithm, with reconstructed images labeled as DL sequences. The AIR™ Recon DL pipeline incorporates a deep CNN with approximately 10,000 kernels and 4.4 million trainable parameters. Importantly, this CNN is an FDA (Food and Drug Administration)-cleared, commercially embedded algorithm within the GE scanner, rather than a newly developed research model. The CNN processes raw complex-valued imaging data to generate high-quality output images by: (1) reducing user-adjustable image noise, (2) minimizing truncation artifacts, and (3) enhancing edge sharpness. Integration into the scanner’s native inline reconstruction pipeline ensures access to raw, full-bit-depth data. The CNN was trained using supervised learning with image pairs consisting of near-perfect high-resolution images (minimal ringing, low noise) and conventional images synthesized to include higher noise and truncation artifacts. In our study, the use of AIR™ Recon DL under the same acquisition protocol reduced the mean scan time by approximately 60% while maintaining equivalent spatial resolution. The cohort-specific SNR increased by 92% $\pm$ 11% compared with non-DL sequences, consistent with vendor specifications. Motion artifact reduction was quantified using a 5-point artifact severity scale (1 = severe, 5 = minimal), showing an improvement from 3.1 $\pm$ 0.6 to 4.2 $\pm$ 0.5 (p 0.01).

This denoising and edge-enhancement capability further improved the diagnostic quality of MRI images that were acceptable but not entirely free of motion artifacts. To assess consistency and reproducibility, two independent radiologists evaluated 30 randomly selected MRI studies twice, separated by a 2-week interval, and inter-observer and intra-observer agreement were calculated using intraclass correlation coefficients (ICCs). Ultrasound data were stored in Digital Imaging and Communications in Medicine (DICOM) format and, together with MRI images, transferred to the Picture Archiving and Communication System (PACS; IMPAX 6, Agfa-Gevaert NV) for further analysis.

MRI image datasets were exported from the PACS and preprocessed using ITK-SNAP (version 3.8) for registration, denoising, and standardization. Segmentation of the placenta and fetal abdomen was performed using a 2D U-Net model, trained on an NVIDIA RTX 3090 GPU for 100 epochs with a cross-entropy loss function and a learning rate of 0.001. The Adam optimizer ($\beta$₁ = 0.9, $\beta$₂ = 0.999) was used, with a batch size of 8 and early stopping applied after 20 epochs without validation improvement. Data augmentation included random rotation ($\pm$10°), horizontal and vertical flipping, and intensity scaling ($\pm$10%) to improve model generalization.

The training dataset comprised images from 80 pregnant women (67%), with validation and test sets each including images from 20 pregnant women (16.7%). To ensure subject-level independence, all images from a single subject were assigned to only one subset. Segmentation masks were annotated by two radiologists (Kappa = 0.85). In addition, inter-rater reliability was further assessed using Dice overlap (mean = 0.87 $\pm$ 0.03) and Bland–Altman analysis, confirming strong agreement between observers. The 2D U-Net model utilized AIR™ Recon DL-reconstructed T2-weighted SSFSE images as input to generate masks for the placenta and fetal abdomen. This segmentation model was implemented exclusively for automated region delineation and not for diagnostic classification of abnormalities. When the Dice similarity coefficient (DSC) was 0.80, manual correction was performed by two radiologists according to predefined anatomical rules: (1) placental boundaries were delineated along the T2-weighted hypointense edge, excluding maternal myometrium; (2) fetal abdominal contours were refined to exclude umbilical cord and amniotic fluid. Disagreements were resolved through consensus review.

Fetal FL was measured via ultrasound using standardized ISUOG measurement protocols [10], and the data were input as numerical values into the subsequent EFW calculation model. FGR identification was based on EFW calculated from MRI-segmented fetal abdominal volume and ultrasound-measured FL, using the Hadlock formula, compared against birth weight, with EFW below the 10th percentile serving as the criterion for FGR prediction. Segmentation performance was evaluated using the Dice similarity coefficient (target $>$0.85), with manual correction applied as described above. EFW prediction error was assessed using MAE (target 150 g, along with accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUROC).

For the segmentation model, datasets were divided into training, validation, and test subsets (80, 20, and 20 subjects, respectively), as described in Section 2.3. For the subsequent regression model used for EFW prediction and FGR identification, data from 120 subjects were divided at the subject level into training (80%) and validation (20%) subsets, and five-fold cross-validation was applied to enhance model generalization.

Placental and fetal abdominal segmentation was performed using the 2D U-Net model, while the EFW prediction model integrated MRI-segmented abdominal volume and ultrasound-measured FL data using a regression framework based on the Hadlock formula. No additional CNN or classification networks were trained or fine-tuned in this study; all reconstructions were performed using the manufacturer-implemented AIR™ Recon DL algorithm. The regression model was trained to minimize prediction error between MRI-based EFW and actual birth weight. To prevent potential temporal bias, datasets were randomly partitioned according to acquisition date to ensure chronological independence between training and validation samples. Fold-wise performance metrics were averaged across five folds, and 95% confidence intervals (CIs) were calculated to assess model stability.

Evaluation metrics included the Dice similarity coefficient (target $>$0.85) for placental segmentation, MAE (target 150 g for EFW prediction, root mean square error (RMSE), coefficient of determination (R²), and AUROC for FGR identification. All reported performance values were obtained from the validation and independent test sets corresponding to each model. As this was a single-center study using one MRI scanner, potential institutional bias and limited generalizability should be acknowledged, and further external validation is warranted.

All statistical analyses were performed using SPSS software (version 26.0, IBM, Armonk, NY, USA). Image quality metrics (SNR, CNR, and edge sharpness scores) were compared between AIR™ Recon DL-reconstructed and non-reconstructed images using paired t-tests. For intergroup comparisons of continuous variables (e.g., per-case Dice Dcoefficients), independent-samples t-tests or Mann–Whitney U tests were applied according to the Shapiro–Wilk normality test. The “Consistency of placental volume measurement” reported in Table 3 represents a reliability coefficient (ICC). Because this metric is a statistical coefficient rather than a continuous variable, it was summarized descriptively (0.85 vs. 0.93). The p-value reflects a comparison of these coefficients using Fisher’s z-transformation rather than a t-test.

For Table 4, per-case continuous outcomes, such as placental-volume estimation error, were analyzed using paired t-tests or nonparametric alternatives. Summary performance metrics, such as accuracy and AUROC, were not treated as continuous variables; AUROC values were compared using DeLong’s test.

Placental segmentation performance was assessed via the Dice similarity coefficient, and the diagnostic accuracy for FGR and PI was evaluated using ROC curve analysis with AUROC calculation (AUROC). For each key metric (AUROC, MAE, and Dice coefficient), 95% CIs were computed to assess statistical precision and robustness. Comparisons between the AUROC values of the DL system and manual assessment were performed using DeLong’s test. Multiple comparisons across subgroups were adjusted using the Benjamini–Hochberg false discovery rate (FDR) correction method to control type I error inflation. Correlations between MRI-derived placental parameters (e.g., placental volume, mean T2 value) and Doppler indices (umbilical artery PI, CPR) were assessed using Pearson or Spearman correlation coefficients, depending on data distribution.

EFW prediction error was evaluated using MAE, with a target MAE below 150 g. All statistical tests were two-sided, with significance thresholds adjusted for multiple comparisons (adjusted p 0.05). Results are presented as mean $\pm$ standard deviation (SD) or mean (95% CI), as appropriate.

A total of 120 pregnant women who underwent fetal MRI were included in the study and were grouped according to the presence of FGR: the FGR group (n = 60) and the non-FGR group (n = 60). The two groups showed no significant differences in baseline characteristics, including gestational age and maternal age (p $>$ 0.05). In the FGR group, placental volume, EFW, and T2 values were significantly lower than those in the control group, along with significant differences observed in the birth weight (Table 1).

High-precision segmentation of the placenta and fetal brain structures was achieved using the fetal MRI image analysis system based on DL. The Dice coefficient for placental segmentation was 0.892 $\pm$ 0.021, and the MAE for fetal weight estimation reached 148.2 g, indicating good predictive performance, with an R² value of 0.86 (Table 2).

All MRI images were collected through the GE SIGNA™ Premier 3.0-T system, with the AIR™ Recon DL image reconstruction algorithm applied. This algorithm enhances image SNR and CNR based on raw k-space data and attenuates Gibbs artifacts. Comparison of 20 cases before and after applying AIR™ Recon DL found noticeable enhancement in image quality and subsequent segmentation performance (Table 3).

The MRI system based on deep learning demonstrated significant advantages were observed in the FGR identification task. The system achieved an accuracy of 91.7%, substantially higher than the 84.5% of manual assessment. Additionally, the system showed a superior performance in predicting placental volume and identifying placental perfusion abnormalities compared with manual assessment. Combining the two methods further increased the accuracy to 92.2%. See Table 4.

We analyzed subgroups with different FGR and PI conditions. The system demonstrated superior performance in the FGR + PI group, showing the lowest EFW error (135.6 g) and an FGR detection accuracy of 93.4%. These results suggest that this system has greater application value in high-risk pregnant populations (Table 5).

Although this study demonstrated the potential of AIR™ Recon DL in placental segmentation and the assessment of FGR and PI, its clinical implementation still faces several challenges. Despite the improvement in image quality, fetal motion artifacts remain a persistent limitation in fetal MRI, especially during active fetal movement. Furthermore, the evaluation of PI relies not only on imaging quality but also on comprehensive integration with other clinical parameters, such as hemodynamic indices and laboratory data. The retrospective, single-center design of this study and the relatively small sample size may also limit the generalizability of the findings. In addition, the results may be influenced by scanner- and vendor-specific characteristics, as all MRI data were acquired on a single 3.0-T GE SIGNA™ Premier system using the AIR™ Recon DL algorithm. Therefore, caution should be exercised when extrapolating the results to other MRI platforms or reconstruction algorithms. Although ultrasound FL measurements followed ISUOG guidelines and were performed by experienced sonographers, operator- and equipment-related variability could not be completely eliminated, potentially affecting the accuracy of EFW. Future studies involving multiple vendors, standardized multicenter ultrasound protocols, and larger patient cohorts are warranted to further validate the robustness and general applicability of our findings.

With ongoing advancements in technology and imaging hardware, DL–based fetal MRI assessment is expected to further reduce manual intervention, improve efficiency, and enhance diagnostic precision. Such AI-assisted systems hold promise as valuable tools for the early detection and individualized management of high-risk pregnancies.