Adaptive Marker-Controlled Watershed Combined with Voxel Quantification for Automated Fetal Measurement

Abstract

Accurate and consistent fetal biometric measurement is essential for assessing fetal growth and gestational age in prenatal care. However, ultrasound (US) imaging presents several challenges, including speckle noise, shadowing artifacts, and low tissue contrast, which often degrade segmentation accuracy. Classical watershed algorithms, though effective for edge detection, tend to produce over-segmentation in such complex textures. The dataset used in this study consisted of 272 ultrasound images of patients from M. Djamil Hospital, Padang, West Sumatra. The dataset covers various phases of fetal development, from the first trimester to the third trimester. All images correspond exclusively to fetal ultrasound examinations and were used solely for automated fetal biometric analysis. To overcome these issues, this study introduced an Adaptive Marker-Controlled Watershed (AMCW) algorithm combined with Voxel Quantification (VQ) to achieve more reliable and automated fetal measurements. The proposed AMCW method integrates adaptive marker generation based on morphological gradient and local intensity statistics, enabling dynamic control of internal and external markers across varying fetal regions. After segmentation, spatially calibrated pixel-based quantification was employed to estimate the dimensional properties of segmented fetal structures. The method was applied exclusively to 2D B-mode ultrasound datasets across multiple gestational ages, targeting four key fetal parameters: Biparietal Diameter (BPD), Head Circumference (HC), Abdominal Circumference (AC), and Femur Length (FL). Although the present study is limited to 2D ultrasound images, the proposed framework may be extendable to 3D ultrasound data in future research. The combination of adaptive marker-controlled watershed segmentation and voxel-based quantification presents a robust, interpretable, and computationally efficient framework for automated fetal measurement. The CNN achieved a classification accuracy of 98.75% on the independent testing dataset, indicating that the extracted biometric features contain strong discriminative information for automated fetal condition assessment. This hybrid approach minimizes operator dependency and measurement variability aligning with clinical measurement trends.