Radiomics-based correlation analysis of fetal brain MRI features and children’s neurodevelopmental outcomes in monochorionic twins

Abstract

Objective

To characterize fetal brain MRI features in monochorionic twin pregnancies based on radiomics; and to investigate the correlation between these radiomic features and subsequent neurodevelopmental outcomes.

Methods

This retrospective cohort study analyzed 26 monochorionic twin pregnancies (36 fetus included) who underwent fetal brain MRI (Siemens Magnetom Skyra 3.0 T or Philips Ingenia 3.0 T). Neurodevelopmental assessment categorized 20 monochorionic twins into the good neurodevelopmental group and 16 into the moderate neurodevelopmental group. MRI textural features of different brain areas were quantified by composite radiomics score and individual radiomics-feature score, and the correlation between these scores and neurodevelopmental outcomes during postnatal follow-up was analyzed.

Results

Quantitative radiomic analysis revealed significantly higher radiomics score in the good neurodevelopmental group for the following regions: periventricular white matter (PWM), frontal, parietal and temporal lobes (all P < 0.05). Four specific radiomics-feature score demonstrated significant group differences in these regions: Gray Level Dependence Matrix (GLDM) in PWM, first-order statistical feature in frontal lobe, Gray Level Size Zone Matrix (GLSZM) in parietal lobe, and GLSZM in temporal lobe. For predictive modeling, we identified five high-discriminatory features representing distinct feature categories: shape features (Elongation), first-order statistical features (Kurtosis), and texture features (GLCM: Cluster Shade, GLRLM: Long Run High Gray Level Emphasis, GLSZM: Size Zone Non Uniformity). The logistic regression model with nested cross-validation incorporating these features achieved excellent predictive performance for neurodevelopmental outcomes [Mean of AUC = 0.8900 (± 0.1133)].

Conclusions

Radiomics scores were higher in good neurodevelopmental fetuses, and the selected radiomics features may be helpful for predicting the neurodevelopmental outcomes of monochorionic twins.

Introduction

Monochorionic twins comprise approximately 20% of multiple pregnancies and face elevated risks of neurodevelopmental impairment due to placental vascular anomalies that can disrupt cerebral blood flow, potentially leading to fetal brain injury [1, 2]. Such injuries may progress to long-term neurological deficits, including cognitive impairment and cerebral palsy, with significant implications for healthcare systems and families [3,4,5].

While MRI remains the gold standard for detecting structural brain injuries, radiomics,an emerging discipline combining medical imaging with machine learning,offers a powerful tool for extracting subvisual biomarkers from conventional scans [6], it enhances the precision of clinical decision support systems in diagnosis, prognosis, and prediction. Although most prenatal radiomic research has focused on placental and pulmonary assessment [7,8,9,10] some work suggests its potential in fetal neuroimaging. For instance, texture analysis has been used to differentiate microstructural patterns between growth-restricted and normally grown fetuses, with some features correlating with postnatal neurodevelopment [11].

Nevertheless, the application of radiomics to evaluate brain development in monochorionic twins remains largely unexplored. This study aims to bridge this gap by investigating early cerebral microstructural changes in monochorionic twins and assessing their association with neurodevelopmental outcomes.

Methods

Participants

This retrospective cohort study analyzed monochorionic twin pregnancies that underwent fetal brain MRI in the First Affiliated Hospital of Sun Yat-sen University between January 2016 and October 2022. Inclusion criteria: 1) at least one surviving fetus; and 2) complete clinical data(imaging and maternal demographics). Exclusion criteria: 1) significant motion artifacts on MRI; 2) major fetal anomalies or genetic disorders; 3) dual fetal demise; and 4) missing neurodevelopmental follow-up. Finally, a total of 26 monochorionic diamniotic (MCDA) twin pregnancies (36 fetuses included) were included in the study(Fig. 1).

[IMAGE OMITTED: SEE PDF]

Primary indications for MRI were: 1) Ultrasound-detected brain anomalies (e.g., ventricular dilation/septum pellucidum enlargement); 2) Complications including selective fetal growth restriction (sFGR) or twin-twin transfusion syndrome (TTTS) with hemodynamic compromise; and 3) Intrauterine fetal death (IUFD) of one twin.

The study was reviewed and approved by the Ethics Committee for Clinical Research of the First Affiliated Hospital of Sun Yat-Sen University, Guangzhou(Approval number: [2025]244).

MRI sequences

All MRI examinations were performed on Siemens Magnetom Skyra 3.0T or Philips Ingenia 3.0T scanner using an abdominal phased-array coil. Pregnant women were positioned in head-first supine or left lateral decubitus position for comfort and safety. T2-weighted imaging (T2WI) was finally acquired. Two board-certified radiologists (with ≥ 3-year experience in fetal MRI) independently interpreted the images. For each case, the images were initially reviewed by a junior radiologist and subsequently verified by an experienced senior radiologist. A consensus was reached through discussion for all cases with discrepant readings after the independent review process.

Assessment of neurodevelopmental outcomes

Neurodevelopmental outcomes were assessed using the Developmental scale for children aged 0–6 years (see Supplementary Table 1–2 for full protocol), which evaluates five attributes: (1) gross motor (head control, sitting, crawling, standing, walking, running, and jumping), (2) fine motor (manual dexterity), (3) language (receptive and expressive abilities), (4) adaptive behavior (environmental responsiveness), and (5) social-emotional development (interpersonal skills and self-care). Two standardized metrics were used: 1) Mental Age (MA): indicating cognitive ability age-equivalent; 2) Developmental Quotient (DQ): quantifying overall neurodevelopmental progress, and is calculated as follows:

$$\text{DQ}=\frac{MA}{chronological age}\times 100$$

(1)

Based on this scale classification: DQ ≥ 130: Advanced; 110–129: Good; 80–109: Moderate; 70–79: Borderline delay; < 70: Intellectual disability. Certified assessors conducted structured evaluations via telephone/video interview with parents.

Image preprocessing, segmentation and radiomic feature extraction

Due to the retrospective nature of this study, MRI acquisition parameters varied across scanners. To ensure radiomics reproducibility (following IBSI guidelines), all images underwent standardized preprocessing: Cubic B-spline interpolation was applied to normalize voxel sizes to 0.5mm *0.5mm *1.0mm3, optimizing feature extraction consistency. N4 bias field correction algorithm, the most widely used method was applied to remove the signal intensity inhomogeneities. For the subsequent calculation of texture features, the normalized image intensities were discretized using a fixed bin width method with a bin width of 25. Thereafter, all texture features (e.g., derived from the GLCM, etc.) were calculated from this discretized image. The regions of interest (ROIs) were delineated by two independent physicians, with inter-rater intraclass correlation coefficients (ICCs) ranging from 0.744 to 0.97.

Sagittal and coronal T2WI HASTE images, optimized for observing the brain, were retrieved for segmentation. A comprehensive segmentation protocol was implemented using 3D Slicer software (version 4.11) under radiologist supervision, encompassing six neuroanatomical regions: whole-brain parenchyma, PWM, bilateral frontal, parietal, occipital, and temporal cortices.

Radiomic features were extracted from segmented brain regions using PyRadiomics (v2.2.0). Features were categorized as: (1) Shape features(n = 14): Quantifying three-dimensional shape characteristics; (2) First-order statistics(n = 18): Describing voxel intensity distributions; (3) Texture features(n = 75): derived from texture matrices, including Grey-Level Co-ccurrence Matrix (GLCM), Grey-Level Run Length Matrix (GLRLM), Grey-Level Size Zone Matrix (GLSZM), Grey-Level Dependence Matrix (GLDM), Neighboring Gray Tone Difference Matrix (NGTDM). Feature definitions followed the PyRadiomics documentation is provided on the official website (https://pyradiomics.readthedocs.io/en/latest/features.html).

Feature selection

Two steps were followed to conduct feature selection to obtain the predictive, non-redundant, and reproducible radiomic features. Firstly, LASSO (least absolute shrinkage and selection operator) was applied for feature selection to extract the features that were statistically significantly associated with survival (P < 0.05). Secondly, the covariance analysis was applied to quantify the collinearity between features and to eliminate redundant features iteratively. Features with a high covariance (correlation coefficient |r|> 0.8, P < 0.05) were excluded.

Radiomics score was applied to quantify the contribution of different radiomics features. In the process of LASSO screening features, each feature was assigned a coefficient. First, the total radiomics score for the 3 categories of radiomics features was calculated with the following formula:

$$\begin{aligned}radiomics score&=\beta 0 +\beta 1 X1+\beta 2 X2\\&+\beta 3 X3+\beta 4 X4\\&+\beta 5 X5+\beta 6 X6\end{aligned}$$

(2)

To evaluate category-specific contributions, we computed separate radiomics score (named radiomics-feature score) for each feature category (shape/first-order/texture) using the same weighting method as the radiomics score (Equation ②). These category-specific scores were standardized (mean = 0, SD = 1) to enable cross-category comparison.

Predictive model construction, validation and evaluation

Two machine learning classifiers were implemented: logistic regression (LR) with L2 regularization and support vector machine (SVM) with a radial basis function (RBF) kernel. To ensure robust performance estimation, a nested cross-validation strategy was employed. The dataset was first partitioned into five outer stratified folds, each preserving the distribution of the outcome variable. In each iteration of the outer loop, one fold was held out as the validation set, while the remaining two folds constituted the training set. Within the training set, an inner cross-validation process was performed for hyperparameter tuning and model selection. The optimized model was then evaluated on the outer validation set to compute its performance. This procedure was repeated three times—once for each outer fold—and the results were averaged to yield a final, stable estimate of the generalization error for each classifier.

The predictive performance of the two models was assessed in terms of the area under curve (AUC), accuracy and F1 score.

Statistical methods

Statistical analyses were performed using SPSS (v20.0). Normality testing determined the application of parametric tests (t-test/ANOVA) or nonparametric Wilcoxon test for group comparisons of clinical and radiomics variables. Correlation between imaging biomarkers and neurodevelopmental outcomes was assessed using Pearson/Spearman coefficients based on data distribution. Statistical significance was defined as P < 0.05.

Results

The clinical characteristics of the participants

Based on neurodevelopmental assessment: 1 fetus (2.8%) scored ≥ 130 (advanced), 19 (52.8%) scored 110–129 (good), and 16 (44.4%) scored 90–109 (moderate). Fetuses were stratified into: Good Neurodevelopmental Group (GNG, n = 20; scored ≥ 110) and Moderate Neurodevelopmental Group (MNG, n = 16; scored 90–109). Delivery gestational age were later in the GNG than in the MNG (36.30 ± 1.57 vs 35.29 ± 1.23 weeks, P = 0.048). Birth weights of the GNG were larger than those of the MNG (2.54 ± 0.35 vs 2.26 ± 0.38 kg, P = 0.008). The incidence of sFGR in the GNG was lower than that of the MNG (15.0% vs 37.5%, P = 0.043). No significant differences were observed in the other complications (TTTS, IUFD), mode of delivery, and gestational week of MRI. Pearson correlation revealed significant differences between delivery week and neurodevelopmental outcome (r = 0.42, P = 0.012), also birth weight and neurodevelopmental outcome (r = 0.39, P = 0.018) (Table 1). Traditional MRI findings showed no significant group differences and no correlation with neurodevelopmental scores (Supplementary Table 3).

[IMAGE OMITTED: SEE PDF]

The radiomics features of the participants

From the segmented region of interest (ROI), 107 radiomics features were initially extracted (Supplementary Table 4.1–4.6). Correlations (P < 0.05) between radiomics score and neurodevelopmental scores for different ROI showed: PWM was correlated with gross motor (r = 0.47), adaptive (r = 0.51), and social behavioral (r = 0.36) score; frontal lobe was correlated with gross motor (r = 0.45) and adaptive (r = 0.44) score; parietal lobe was correlated with adaptive (r = 0.40), language (r = 0.53), and social behavioral (r = 0.44) score; temporal lobe was correlated with gross motor (r = 0.47), adaptive (r = 0.43), language (r = 0.39), and social behavioral (r = 0.42)score; no significant correlation were found between the radiomics score of cerebral/occipital and the neurodevelopmental scores (Fig. 2).

[IMAGE OMITTED: SEE PDF]

Comparison of radiomics score and radiomics-feature score between the GNG and MNG

Regional radiomic signature analysis revealed that the GNG showed significantly elevated radiomics score in PWM, frontal, parietal, and temporal lobes compared to the MNG (all P < 0.05), with no differences in cerebral and occipital regions (Fig. 3). Comparing of radiomics-feature score in different brain regions, we found significantly higher scores in the GNG, including GLDM scores of PWM, first-order statistical feature score of frontal lobe, GLSZM score of parietal lobe, and GLSZM score of temporal lobe. These findings indicate distinct gray-level intensity and spatial distribution patterns between groups (Fig. 4 A-F).

[IMAGE OMITTED: SEE PDF]

[IMAGE OMITTED: SEE PDF]

Neurodevelopmental outcome prediction model construction and performance evaluation

Feature selection via LASSO regression (tenfold cross-validation, λ = 0.1) identified six discriminative radiomic features (Supplementary Table 5). These features exhibited low multicollinearity (all variance inflation factors < 3) (Supplementary Fig. 1). Univariable logistic regression was performed for each feature and ROC curves were generated (Supplementary Fig. 2). Five features, except for GLCM-Imc1 in the temporal lobe, demonstrated a high level of confidence in identification (AUC > 0.7).

The five selected features were used to construct predictive models via Logistic regression and SVM algorithms, respectively. After hyperparameter tunin(Supplementary Table 6.1- 6.2), both models achieved a perfect AUC of 1.00 in some configurations; however, the logistic regression model demonstrated more robust performance overall, with a higher mean AUC (0.89 ± 0.11) compared to the SVM (0.78 ± 0.20). Statistical comparison against a random classifier (AUC = 0.5) showed that the model's performance was highly significant (P < 0.001).(Fig. 5).

[IMAGE OMITTED: SEE PDF]

Discussion

The main innovations of this study compared to prior research include: 1) This is the first investigation to focus specifically on cerebral MRI radiomic features in monochorionic diamniotic (MCDA) twins. The unique shared placental circulation in MCDA pregnancies creates inherent hemodynamic vulnerabilities that may elevate the risk of adverse neurodevelopmental outcomes; 2) While previous studies in this population have primarily examined correlative relationships, our work represents the first attempt to develop a predictive model for neurodevelopmental outcomes, with the ultimate goal of enabling early identification of high-risk infants.

Brain texture features more adequately reflect brain development than traditional MRI features

Notably, while traditional MRI features showed no abnormalities in GNG fetuses with matched gestational age, both radiomics score and radiomics-feature score revealed significant intergroup differences (P < 0.01), suggesting the presence of subtle microstructural variations. Accumulating evidence has indicated that altered fetal cerebral hemodynamics may drive progressive microstructural reorganization via coupled metabolic-perfusion dysregulation during neurodevelopment [12, 13]. Our findings provide a potential explanation for the clinical paradox where children with normal prenatal MRI exhibit neurodevelopmental delays. By capturing nuanced texture variations (e.g., GLCM-Cluster Shade) and first order statistics (e.g., Kurtosis), radiomics may detect early signs of aberrant neurodevelopment that evade conventional visual assessment. However, the biological underpinnings of these radiomic signatures require further validation through histopathological correlation studies.

Brain texture features is correlated with neurodevelopmental attribute scores

Our study demonstrated significant correlations between radiomics score of different brain region and neurodevelopmental attribute scores, aligning with known functional neuroanatomy [14]. PWM textural features correlated with motor attribute, consistent with its role in corticospinal tract development. Temporal lobe textural features predicted language attribute, mirroring early auditory cortex maturation. Holli et al. [15] observed that Alzheimer’s disease-related radiomic textural features mirrored functional alterations seen in mild traumatic brain injury and mild cognitive impairment. The radiomic textural features were able to detect differences associated with disease progression, even in mildly affected areas. Accumulating evidence demonstrates that early-onset severe FGR disrupts fetal/infant brain development, manifesting as metabolic, sulcal, and microstructural abnormalities that are associated with adverse neurological outcomes [16, 17]. Sanz-Cortes M [18] identified distinct prenatal brain MRI texture features in term SGA compared to AGA fetuses, with these texture patterns demonstrating correlation with further childhood neurobehavioral outcomes. Similarly, our study found that the MRI radiomic textural features across different brain regions in twin fetuses were associated with neurodevelopment during infancy and childhood.

Selected radiomics features could predict the neurodevelopmental outcomes

The GNG exhibited elevated radiomics-feature score in key brain regions (PWM/temporal lobe), particularly in GLSZM features. Previous studies have documented neuroanatomical alterations in FGR, including reduction in hippocampal volumes, shallower sulcal gyrus and cortical thinning [19], Complementing these findings, Shin et al. [20] demonstrated that GLSZM features in preterm neonates predicted 12-month motor scores on the Bayley-III (p = 0.03), These observations collectively suggest that texture features may capture microstructural maturation patterns across developmental stages.

Our study identified five discriminative radiomics features predictive of neurodevelopmental outcomes (P < 0.05), consisting of shape features (Elongation), first-order statistical features (Kurtosis), and texture features (GLCM: Cluster Shade, GLRLM: Long Run High Gray Level Emphasis, GLSZM: Size Zone Non Uniformity). The logistic regression-based model demonstrated superior predictive performance for neurodevelopmental outcomes compared to the SVM algorithm, achieving a higher mean AUC of 0.89. This model offers several advantages: (1) simplicity and interpretability, (2) reduced risk of overfitting particularly for small datasets with low complexity. While clinical factors (patient condition and obstetric history) are readily identifiable and ultrasound features remain operator-dependent, MRI provides a more comprehensive fetal brain assessment, Importantly, radiomics enables thorough quantification of brain features through multiple parameters. Consequently, the MRI-radiomics integrated model shows more stable performance in predicting fetal neurodevelopmental outcomes.

Due to the unique and dynamic developmental processes of the fetal brain, neurodevelopment is indeed influenced by a complex array of factors, including disruptions in developmental programming, hypoxia–ischemia, infections, alcohol exposure, and maternal nutrition. Although the radiomic parameters identified in our study demonstrated high predictive efficacy for neurodevelopmental outcomes, it is an oversimplification to attribute neurodevelopmental impairments in twins solely to hemodynamic changes associated with placental factors. This explanation, while relevant, likely accounts for only a significant portion of the observed effects. It is crucial to acknowledge that acute brain injuries can still occur during the perinatal period and for a considerable time postnatally.

Strengths, limitations and prospects

Our study revealed significant microstructural disparities in MCDA twin pregnancies affected by IUFD. These findings suggest divergent neurodevelopmental trajectories between IUFD survivors and unaffected twins. While preliminary, this evidence supports the integration of radiomics into current IUFD surveillance protocols, enabling earlier risk stratification and targeted neuroprotective interventions in high-risk cases. Validation in prospective cohorts is needed before clinical implementation. However, there are several limitations in this study. The study was retrospective design. Heterogeneous MRI protocols (scanner/SAR variations) may affect feature stability, particularly for gradient-based texture features (e.g., GLCM contrast). The neurobehavioral scores at visiting age may not fully predict later cognitive functions. Longitudinal follow-up until school age is needed, as neurodevelopmental trajectories often diverge post-infancy. Notably, our radiomic model attained clinically relevant accuracy (AUC =0.89) even with limited samples, potentially reflecting strong biological signals in monochorionic twin neurodevelopment. However, the wide confidence interval and lack of independent validation highlight this as a proof-of-concept study requiring: (1) expanded recruitment to reduce overfitting risk, and (2) prospective validation to assess real-world performance.

The heightened prevalence of cerebral injury and neurodevelopmental morbidity in monochorionic twin gestations represents a critical clinical challenge. Early identification of prenatal neurological compromise and prognostication of neurodevelopmental trajectories may enable targeted interventions to optimize neuroprotection and mitigate adverse perinatal outcomes. These findings underscore the imperative for multicenter prospective cohorts to expand sample sizes through longitudinal follow-up, establish standardized fetal neuroimaging protocols with quantitative biomarkers, and develop validated predictive algorithms integrating multimodal biomarkers for precision prenatal counseling frameworks.

Conclusion

This retrospective study established significant associations between fetal cerebral microstructural patterns on MRI radiomics and differential neurodevelopmental outcomes in monochorionic twins. The radiomics framework demonstrated robust differentiation between fetuses with favorable versus moderate neurodevelopmental trajectories, supporting its potential for prenatal prediction. These findings highlight the need for multicenter prospective studies to standardize gestational age-specific radiomics protocols, integrate placental pathophysiology with neural biomarkers, and translate predictive algorithms into precision neuroprotective interventions.

Data availability

The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author.

Abbreviations

AGA:

Appropriate for gestational age

FGR:

Fetal growth restriction

GLCM:

Gray Level Co-occurrence Matrix

GLDM:

Gray Level Dependence Matrix

GLRLM:

Gray Level Run Length Matrix

GLSZM:

Gray Level Size Zone Matrix

NGTDM:

Neighboring Gray Tone Difference Matrix

LASSO:

Least absolute shrinkage and selection operator

IUFD:

Intra-uterine fetal death

MCDA:

Monochorionic diamniotic

MRI:

Magnetic resonance imaging

PD:

Parkinson's disease

ROC:

Receiver operating characteristic curve

ROI:

Region of interest

sFGR:

Selective fetal growth restriction

SGA:

Small for gestational age

SVM:

Support vector machine

TTTS:

Twin-twin transfusion syndrome

TAPS:

Twin anemia polycythemia sequence

References

Khalil A, Liu B. Controversies in the management of twin pregnancy. Ultrasound Obstet Gynecol. 2021;57(6):888–902.

Van Klink JM, Koopman HM, van Zwet EW, et al. Cerebral injury and neurodevelopmental impairment after amnioreduction versus laser surgery in twin-twin transfusion syndrome: a systematic review and meta-analysis. Fetal Diagn Ther. 2013;33:81–9.

Inklaar MJ, van Klink JM, Stolk TT, van Zwet EW, Oepkes D, Lopriore E. Cerebral injury in monochorionic twins with selective intrauterine growth restriction: a systematic review. Prenat Diagn. 2014;34(3):205–13.

Peng R, Xie HN, Lin MF, et al. Clinical outcomes after selective fetal reduction of complicated monochorionic twins with radiofrequency ablation and bipolar cord coagulation. Gynecol Obstet Invest. 2016;81(6):552–8.

Colmant C, Lapillonne A, Stirnemann J, et al. Impact of different prenatal management strategies in short- and long-term outcomes in monochorionic twin pregnancies with selective intrauterine growth restriction and abnormal flow velocity waveforms in the umbilical artery Doppler: a retrospective observational study of 108 cases. BJOG. 2021;128(2):401–9.

Mayerhoefer ME, Materka A, Langs G, et al. Introduction to radiomics. J Nucl Med. 2020;61(4):488–95.

Zamarian ACP, Caetano ACR, Grohmann RM, et al. Prediction of lung maturity in fetuses with growth restriction through quantitative ultrasound analysis. Ultrasound Med Biol. 2022;48(1):20–6.

Dahdouh S, Andescavage N, Yewale S, et al. In vivo placental MRI shape and textural features predict fetal growth restriction and postnatal outcome. J Magn Reson Imaging. 2018;47(2):449–58.

Wu Q, Yao K, Liu Z, et al. Radiomics analysis of placenta on T2WI facilitates prediction of postpartum haemorrhage: a multicentre study. EBioMedicine. 2019;50:355–65.

Shao Q, Xuan R, Wang Y, et al. Deep learning and radiomics analysis for prediction of placenta invasion based on T2WI. Math Biosci Eng. 2021;18(5):6198–215.

Sanz-Cortés M, Figueras F, Bonet-Carne E, et al. Fetal brain MRI texture analysis identifies different microstructural patterns in adequate and small for gestational age fetuses at term. Fetal Diagn Ther. 2013;33(2):122–9.

Wang WX, Wang T, Ma L, et al. New-onset lesions on MRI-DWI and cerebral blood flow changes on 3D-pCASL after carotid artery stenting. Sci Rep. 2021;11(1):8005.

Esteban FJ, Padilla N, Sanz-Cortes M, et al. Fractal-dimension analysis detects cerebral changes in preterm infants with and without intrauterine growth restriction. Neuroimage. 2010;53:1225–32.

Raichle ME, Mintun MA. Brain work and brain imaging. Annu Rev Neurosci. 2006;29:449–76.

Holli KK, Harrison L, Dastidar P, et al. Texture analysis of MR images of patients with mild traumatic brain injury. BMC Med Imaging. 2010;12(10):8.

Story L, Damodaram MS, Allsop JM, et al. Brain metabolism in fetal intrauterine growth restriction: a proton magnetic resonance spectroscopy study. Am J Obstet Gynecol. 2011;205(483):e481-488.

Padilla N, Falcon C, Sanz-Cortes M, et al. Differential effects of intrauterine growth restriction on brain structure and development in preterm infants: a magnetic resonance imaging study. Brain Res. 2011;1382:98–108.

Sanz-Cortes M, Ratta GA, Figueras F, et al. Automatic quantitative MRI texture analysis in small-for-gestational-age fetuses discriminates abnormal neonatal neurobehavior. PLoS ONE. 2013;8(7):e69595.

Sanz-Cortés M, Figueras F, Bargalló N, et al. Abnormal brain microstructure and metabolism in small-for-gestational-age term fetuses with normal umbilical artery Doppler. Ultrasound Obstet Gynecol. 2010;36(2):159–65.

Shin Y, Nam Y, Shin T, Choi JW, Lee JH, Jung DE, et al. Brain MRI radiomics analysis may predict poor psychomotor outcome in preterm neonates. Eur Radiol. 2021;31(8):6147–55.