Fetal ultrasound imaging plays a pivotal role in monitoring fetal growth and development during pregnancy. Accurate segmentation of fetal anatomical structures is crucial for precise measurement of biometric parameters, such as crown-rump length (CRL) and nuchal translucency thickness (NT), which are essential for assessing gestational age and screening for potential anomalies. However, manual segmentation is a time-consuming and laborious task, necessitating the development of automated segmentation techniques. This study evaluates the performance of deep learning models for automated fetal contour segmentation in sagittal view ultrasound scans of first trimester fetuses. Six deep learning models, including UNet, Multi-Attention Net, Trans-UNet, Attention UNet, MedSAM, and DeepLabV3+, were explored for automated fetal contour segmentation. A proprietary dataset of 10,500 annotated ultrasound images was used for training and evaluation. The models were compared on the basis of performance metrics such as Dice score, and Intersection over Union (IoU). AttentionUNet emerged as the top performer, achieving an impressive accuracy of 0.98, a Dice score of 0.95, and an IoU score of 0.92. This superior performance can probably be attributed to its effective spatial and channel attention mechanism, which helps it focus on relevant features. In contrast, transformer-based models like TransUNet and MedSAM struggled, showing diminished performance due to their high computational demands and insufficient fine-tuning. Meanwhile, convolutional architectures, despite their inherent strengths, have consistently faced challenges in precisely defining boundaries. This study highlights the potential application of deep learning techniques for automating fetal contour segmentation, which can streamline prenatal care and enhance the accuracy of fetal biometric measurements.
Sriraam et al. (Sat,) studied this question.