Fetal Ultrasound Monitoring Using MedSegDiff-V2–DUCKNet Ensemble Segmentation and Explainable Head Circumference Estimation
TL;DR
An AI-enhanced ultrasound monitoring system has been developed that automates the segmentation and analysis of fetal anatomical structures, and the proposed ensemble framework combines MedSegDiff-V2 and DUCK-Net through learnable and attention-based fusion mechanisms.
Abstract
Fetal health monitoring is a vital component of prenatal care. Ultrasonography is the primary imaging technique for assessing fetal development and detecting abnormalities. In this work, an AI-enhanced ultrasound monitoring system has been developed that automates the segmentation and analysis of fetal anatomical structures. We used the Radboudumc HC18 dataset comprising 1,335 fetal head ultrasound images with segmentation masks, pixel-size metadata, and ground-truth HC/AC measurements. We collected a novel dataset (Aalok dataset) comprising 400 2D ultrasound scans from a local hospital in Dhaka, Bangladesh. The scans were obtained using a Samsung WS80A Elite system, which covers gestational ages of 18–38 weeks and includes precise manual annotations by radiologists. It consists of imaging, biometric measurements, and clinical annotations for complete fetal estimation. We applied multiple segmentation architectures, including U-NeXt, Attention U-Net, TransUNet, DUCK-Net, and MedSegDiff-V2. Among them, MedSegDiff-V2 and DUCK-Net achieved the highest segmentation performance with Dice scores of 0.9852 and recall of 0.9864 with ~70 minutes training time, outperforming DUCK-Net, TransUNet and UNeXt. The proposed ensemble framework combines MedSegDiff-V2 and DUCK-Net through learnable and attention-based fusion mechanisms. The learnable ensemble achieved a Dice score of 0.9892 and a Recall of 0.9871. The proposed MedSegDiff-V2-Duck-Net ensemble technique demonstrated a strong correlation between predicted and actual HC values with a correlation of 0.9988. Grad-CAM++ explainable AI technique, has been applied to the ensemble model, visualizing the key ROI for clearer human understanding. The proposed system significantly reduces examination time and interobserver variability and increases diagnostic precision. The collected Aalok fetal head segmentation dataset is publicly available at the Aalok Dataset GitHub Repository