Prediction of Fetal Growth Restriction Using Placental Image Features in BOLD MRI

• Published in: Journal of advanced computational intelligence and intelligent informatics Vol. 29; no. 3; pp. 508 - 518
• Main Authors: Nishida, Kentaro, Morita, Kento, Enomoto, Naosuke, Magawa, Shoichi, Nii, Masafumi, Wakabayashi, Tetsushi
• Format: Journal Article
• Language: English
• Published: Tokyo Fuji Technology Press Co. Ltd 20.05.2025
• Subjects: Annotations; Automation; Body weight; Classification; Deep learning; Feature extraction; Magnetic resonance imaging; Medical imaging; Oxygen; Oxygenation; Ultrasonic imaging
• ISSN: 1343-0130; 1883-8014
• DOI: 10.20965/jaciii.2025.p0508

Fetal growth restriction (FGR) is a disease during pregnancy that increases the risk of preterm birth and perinatal death. Currently, the diagnosis of FGR relies on ultrasonography-based estimated fetal body weight (EFBW). However, EFBW can only provide an indirect assessment of FGR because a low EFBW is only a result of growth restriction. Recent research has indicated that placental oxygenation function is a key indicator of fetal growth; however, its assessment through ultrasonography is impractical. Techniques other than ultrasonography for placental function have been investigated, and a significant difference in placental oxygenation function between FGR and non-FGR cases has been demonstrated using blood oxygen level-dependent magnetic resonance imaging (BOLD MRI). BOLD MRI can visualize oxygenation in vivo , and may be useful as a marker for the direct assessment of placental oxygenation function. However, visual assessment of placental oxygenation in BOLD MRI is challenging, even for experts, because of the complexity of analyzing the signal intensity on MRI. In this study, we proposed an automated method for predicting FGR by utilizing placental image features extracted from BOLD MRI during oxygen administration. In addition to the FGR/non-FGR classification method, we propose a placental region segmentation method to reduce the manual annotation burden. The proposed segmentation method achieved a Dice coefficient of 0.809, outperforming other deep learning methods. In the FGR/non-FGR classification experiments, conducted with four-fold cross-validation on 22 subjects, the highest performance was obtained using a pre-trained ResNet50 combined with a fully connected layer with transfer learning as a feature extractor (subject-wise accuracy of 0.908, ROC-AUC of 0.927, and F1 score of 0.922). These results demonstrate that placental image features extracted from BOLD MRI can effectively differentiate between FGR and non-FGR cases, suggesting the potential for a direct and automated approach to assess FGR through placental oxygenation function.