Multi-centre deep learning for placenta segmentation in obstetric ultrasound with multi-observer and cross-country generalization

Lisbeth Anita Andreasen, Aasa Feragen, Anders Nymark Christensen, Jonathan Kistrup Thybo, Morten Bo Søndergaard Svendsen, Kilian Zepf, Karim Lekadir, Martin G. TolsgaardScientific 2023 · DOI

Abstract

The placenta is crucial to fetal well-being and it plays a significant role in the pathogenesis of hypertensive pregnancy disorders. Moreover, a timely diagnosis of placenta previa may save lives. Ultrasound is the primary imaging modality in pregnancy, but high-quality imaging depends on the access to equipment and staff, which is not possible in all settings. Convolutional neural networks may help standardize the acquisition of images for fetal diagnostics. Our aim was to develop a deep learning based model for classification and segmentation of the placenta in ultrasound images. We trained a model based on manual annotations of 7,500 ultrasound images to identify and segment the placenta. The model's performance was compared to annotations made by 25 clinicians (experts, trainees, midwives). The overall image classification accuracy was 81%. The average intersection over union score (IoU) reached 0.78. The model's accuracy was lower than experts' and trainees', but it outperformed all clinicians at delineating the placenta, IoU = 0.75 vs 0.69, 0.66, 0.59. The model was cross validated on 100 2nd trimester images from Barcelona, yielding an accuracy of 76%, IoU 0.68. In conclusion, we developed a model for automatic classification and segmentation of the placenta with consistent performance across different patient populations. It may be used for automated detection of placenta previa and enable future deep learning research in placental dysfunction.

Introduction

The placenta is the fetus's life-support system, yet it remains one of the most difficult organs to evaluate systematically during pregnancy. Abnormalities in placental volume or position, such as placenta previa, are linked to hypertensive disorders, fetal growth restriction, and stillbirth. Currently, evaluating the placenta via ultrasound is highly operator-dependent. Manually tracing placental boundaries (segmentation) is time-consuming and rarely occurs in routine clinical practice. This paper addresses the challenge of creating a robust deep learning model that works on "messy," real-world data from multiple countries and clinical settings.

Methods

The researchers employed Mask R-CNN, a well-established architecture for instance segmentation. The model was initialized with weights from the ImageNet database and utilized extensive data augmentation—including shearing, rotation, and brightness shifts—to handle the inherent noise of ultrasound imaging. The model was designed for two primary tasks: classification (detecting the presence of a placenta) and segmentation (drawing precise boundaries). Performance was benchmarked against 25 clinicians, including fetal medicine experts, OB-Gyn residents, and midwives.

Datasets

The study utilized a multi-centre data strategy:

  1. Primary Dataset (Copenhagen, Denmark): 7,500 2D ultrasound images from routine screenings, reflecting heterogeneous real-world conditions.
  2. External Validation (Barcelona, Spain): 100 second-trimester images from a different patient population and equipment set to test cross-country generalization.
  3. Human Comparison Set: A subset of 24 images used to benchmark clinician performance.

Results and Key Insights

The model achieved an 81% image classification accuracy and an average IoU of 0.78 for segmentation. A key finding was the trade-off between human intuition and AI consistency: while experts were slightly better at identifying the presence of a placenta (84% vs 81%), the AI consistently outperformed all clinician groups at accurately tracing its boundaries. In the head-to-head segmentation task, the model achieved an IoU of 0.75, surpassing experts (0.69), residents (0.66), and midwives (0.59).

Performance metrics (Accuracy and IoU) across different trimesters, showing the model's robustness in later stages of pregnancy.

Performance varied by gestational age, with the model showing higher accuracy in the second and third trimesters compared to the first. Additionally, the model successfully distinguished between anterior and posterior placental locations with approximately 71% accuracy, a critical factor for diagnosing placenta previa.

Confusion matrix for placental positioning (Anterior vs. Posterior), demonstrating the model's classification performance.

Discussion and Conclusion

While the model demonstrated strong generalization across different patient populations, including the Spanish validation set (76% accuracy, 0.68 IoU), limitations remain. The model struggled more with first-trimester scans where the placenta is less distinct. Future work will focus on integrating this technology into real-time ultrasound video feeds. Ultimately, this research paves the way for automated screening tools that can identify high-risk pregnancies earlier and more reliably, particularly in settings with limited access to expert sonographers.

Summary updated

Abstract

The placenta is crucial to fetal well-being and it plays a significant role in the pathogenesis of hypertensive pregnancy disorders. Moreover, a timely diagnosis of placenta previa may save lives. Ultrasound is the primary imaging modality in pregnancy, but high-quality imaging depends on the access to equipment and staff, which is not possible in all settings. Convolutional neural networks may help standardize the acquisition of images for fetal diagnostics. Our aim was to develop a deep learning based model for classification and segmentation of the placenta in ultrasound images. We trained a model based on manual annotations of 7,500 ultrasound images to identify and segment the placenta. The model's performance was compared to annotations made by 25 clinicians (experts, trainees, midwives). The overall image classification accuracy was 81%. The average intersection over union score (IoU) reached 0.78. The model's accuracy was lower than experts' and trainees', but it outperformed all clinicians at delineating the placenta, IoU = 0.75 vs 0.69, 0.66, 0.59. The model was cross validated on 100 2nd trimester images from Barcelona, yielding an accuracy of 76%, IoU 0.68. In conclusion, we developed a model for automatic classification and segmentation of the placenta with consistent performance across different patient populations. It may be used for automated detection of placenta previa and enable future deep learning research in placental dysfunction.
BibTeX
@inproceedings{andreasen2023multi,
  author = {Lisbeth Anita Andreasen and Aasa Feragen and Anders Nymark Christensen and Jonathan Kistrup Thybo and Morten Bo Søndergaard Svendsen and Kilian Zepf and Karim Lekadir and Martin G. Tolsgaard},
  title = {{Multi-centre deep learning for placenta segmentation in obstetric ultrasound with multi-observer and cross-country generalization}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-29105-x},
}
Figure