A rule-guided interpretable lightweight framework for fetal standard ultrasound plane capture and biometric measurement

Jintang Li, Zhan Gao, Chunlian Wang, Bin Pu, Kenli Li

Abstract: Automating prenatal ultrasound analysis is challenging due to the need for accurate detection of standard planes and biometric measurements. Existing methods typically rely on a multi-stage process, which suffers from errors due to sequential dependencies, limited anatomical perception, and over-confidence issues. To address these challenges, we propose a novel, rule-guided, and lightweight framework that directly captures standard planes and performs biometric measurements in a single, unified process. Our model incorporates a redundant feature selection module to handle ultrasound noise and capture rich anatomical representations. By designing separable dense blocks and controlling upsampling ratios, we optimize the framework for real-time processing on constrained devices. The expert knowledge-based scoring mechanism evaluates anatomical morphology, edge clarity, and model confidence, offering interpretability and ensuring high-quality plane selection. Additionally, a gradient refinement module fine-tunes measurement points to further enhance biometric accuracy. Comparative experiments demonstrate that our method significantly outperforms other state-of-the-art approaches in fetal anatomical structure segmentation, optimal plane capture, and biometric measurement. Its computational complexity and inference speed are also impressive, showcasing substantial clinical application potential.