A Modular Deep-Learning Framework for Automated Fetal Biometry: Proof-of-Concept Evaluation of Head, Abdominal, and...

Background Accurate fetal biometry is pivotal for assessing fetal growth, well-being, and gestational age estimation. Despite its clinical importance, manual ultrasound measurements remain highly operator-dependent and subject to inter-observer variability. Artificial intelligence (AI)-based automation offers the potential to enhance consistency and efficiency; however, most existing systems are limited to single biometric parameters and lack scalability across multiple measurements. Objective The primary objective was to design and quantitatively evaluate a modular deep-learning framework for automated estimation of fetal head circumference (HC) using the HC18 dataset. Secondary objectives included developing and assessing the feasibility of prototype modules for abdominal circumference (AC) and femur length (FL) using de-identified institutional ultrasound images. Methods A retrospective, proof-of-concept study was conducted to develop an adaptable modular AI framework for fetal biometry. The HC module was trained and validated on the publicly available HC18 dataset (n = 1,334 images), while prototype modules for AC and FL were developed using de-identified still-frame ultrasound images from a tertiary care centre. Model outputs were evaluated using quantitative error metrics - mean absolute error (MAE) and mean squared error (MSE) - and qualitative analysis of anatomical accuracy and contour integrity. Results The HC module, based on a U-Net++ segmentation network with convex-hull contour fitting, achieved an MAE of 9.7 mm, an MSE of 113.3 mm², an accuracy of 0.90, a sensitivity of 0.97, and a specificity of 0.89 on the HC18 test subset (n = 110). The AC prototype, implemented with a hybrid Faster R-CNN + HRNet + Attention U-Net architecture, produced anatomically consistent contours with an average measurement difference of 5.5 ± 3.1 mm from manual references. The FL prototype demonstrated accurate femoral shaft segmentation and endpoint detection across all test images. The average inference time per image was 1.2 seconds, confirming computational feasibility for clinical integration. Conclusion This proof-of-concept study establishes the technical feasibility of a modular AI-driven framework for automated fetal biometry. The validated HC module and functional AC and FL prototypes collectively demonstrate a scalable, interpretable, and guideline-aligned approach for next-generation obstetric ultrasound automation. Further validation with multicentric and real-time ultrasound datasets is warranted before clinical deployment.