Abstract
Preterm birth is a major cause of mortality and lifelong morbidity in childhood. Its complex and multifactorial origins limit the effectiveness of current clinical predictors and impede optimal care. In this study, a dual-branch deep learning architecture (PUUMA) was developed to predict gestational age (GA) at birth using T2* fetal MRI data from 295 pregnancies, encompassing a heterogeneous and imbalanced population. The model integrates both global whole-uterus and local placental features. Its performance was benchmarked against linear regression using cervical length measurements obtained by experienced clinicians from anatomical MRI and other Deep Learning architectures. The GA at birth predictions were assessed using mean absolute error. Accuracy, sensitivity, and specificity were used to assess preterm classification. Both the fully automated MRI-based pipeline and the cervical length regression achieved comparable mean absolute errors (3 weeks) and good sensitivity (0.67) for detecting preterm birth, despite pronounced class imbalance in the dataset. These results provide a proof of concept for automated prediction of GA at birth from functional MRI, and underscore the value of whole-uterus functional imaging in identifying at-risk pregnancies. Additionally, we demonstrate that manual, high-definition cervical length measurements derived from MRI, not currently routine in clinical practice, offer valuable predictive information. Future work will focus on expanding the cohort size and incorporating additional organ-specific imaging to improve generalisability and predictive performance.
Introduction
Preterm birth—delivery occurring before 37 weeks of gestation—is the leading cause of death for children under five. It is a complex syndrome driven by diverse factors including infection, inflammation, and vascular disease. Identifying high-risk pregnancies early is critical for clinical management, such as the administration of corticosteroids for fetal lung development or ensuring delivery occurs at a facility with specialized neonatal intensive care.
However, existing clinical tools like fetal fibronectin testing and transvaginal ultrasound for cervical length are often limited by operator dependency or invasive nature. This study introduces PUUMA, a novel deep learning architecture that leverages non-invasive T2* functional MRI to directly predict the exact Gestational Age (GA) at birth, providing a more granular risk assessment than simple binary classification.
Methods: The PUUMA Architecture
PUUMA is a dual-branch architecture based on the U-Mamba framework, which utilizes State Space Models to capture long-range dependencies in 3D medical images with linear computational complexity. This efficiency is vital for processing bulky volumetric MRI data.
As shown in the architecture diagram, the model processes data through two distinct paths:
- Global Branch: Analyzes a downsampled 3D whole-uterus volume to predict a placental segmentation mask and extract macro-level features.
- Local Branch: Focuses on high-resolution 3D patches of the placenta, sampled using the segmentation mask, to capture fine-grained microstructural details.
- Fusion Layer: The predictions from both branches are concatenated with the patient's GA at the time of the scan to produce the final regression of GA at birth and classification of preterm risk.
Results and Performance
The model was trained and validated on a heterogeneous dataset of 295 pregnancies. Performance was benchmarked against standard U-Net architectures and a clinical baseline consisting of manual cervical length measurements extracted from anatomical MRIs by experienced clinicians.
Key findings from the evaluation include:
- Competitive Accuracy: PUUMA achieved a Mean Absolute Error (MAE) of 3.05 ± 3.1 weeks, which is comparable to the manual cervical length baseline (2.94 ± 2.59 weeks).
- Sensitivity to Preterm Birth: The dual-branch approach proved essential; while a global-only U-Mamba model achieved only 0.33 sensitivity, the full PUUMA model reached 0.67, matching the clinical baseline.
- Detection of High-Risk Cases: PUUMA was the only deep learning model capable of correctly identifying "very preterm" subjects (born before 34 weeks), whereas other models tended to overestimate GA due to the inherent dataset imbalance favoring term births.
Discussion and Conclusion
This work provides a proof of concept for the fully automated prediction of gestational age at birth using functional MRI. A significant insight is that high-definition cervical length measurements derived from MRI—though not currently a clinical standard—hold immense predictive value, matching the performance of sophisticated deep learning pipelines.
While the results are promising, the study is limited by a small test set (n=26), which may affect generalizability. Future research will focus on expanding the cohort size and incorporating multi-organ imaging, such as fetal brain or lung MRI, to provide a more holistic view of fetal development and further improve predictive accuracy.
Summary updated