Fetal ultrasound image segmentation is a crucial component of prenatal diagnosis, enabling the accurate measurement of key biometric parameters and the early detection of abnormalities. Manual segmentation is prone to variability due to factors such as operator skill, image quality, and acquisition conditions. This study compares conventional and deep learning-based segmentation methods, proposing an enhanced Multi-Feature Pyramid U-Net (MFP-Unet) to improve accuracy, robustness, and clinical applicability.

We utilised the HC18 Grand Challenge Dataset, which comprises 1,334 ultrasound images from 551 pregnancies without growth abnormalities. The proposed MFP-Unet integrates multi-scale feature pyramid fusion, dilated convolutions, and attention gates into the U-Net framework. Performance was compared with Dilated U-Net, U-Net, RU-Net, Attention U-Net, and R2U-Net using Dice Coefficient, Intersection over Union (IoU), and mean Average Precision (mAP). Preprocessing included Niblack’s adaptive thresholding and custom region extraction for femur length images.

MFP-Unet achieved the highest accuracy across all parameters: HC IoU = 90% (Dice = 0.95), FL IoU = 93% (Dice = 0.97), BPD IoU = 88% (Dice = 0.93), and overall mAP = 0.92. The architecture formed smoother anatomical boundaries, enhanced robustness to image variability, and compact measurement variations associated with other methods.

The MFP-Unet approach shows strong promise for use in clinical fetal ultrasound analysis, offering reliable and precise biometric measurements that can aid in timely interventions and enhance prenatal care outcomes. Future research should validate performance on multi-centre datasets and explore real-time deployment.