Abstract / Summary
Ventriculomegaly (VM) is one of the most common central nervous system (CNS) abnormalities, with important associations with other congenital defects, chromosomal abnormalities, and intrauterine infections. Accurate measurement of ventricular width is crucial for prenatal diagnosis and risk assessment. However, the operator-dependent nature of ultrasound introduces measurement variability, which can impact diagnostic accuracy. Although artificial intelligence (AI) has shown potential in ultrasound analysis, its performance in automated VM assessment and its agreement with radiologists with varying levels of expertise remain limitation. This study aims to develop a deep learning framework for automated measurement and diagnosis of fetal VM and evaluate its clinical applicability by comparing its performance with radiologists with varying levels of expertise. In this study, we collected ultrasound data from 1,018 fetuses across 4 hospitals to develop and evaluate a deep learning framework for automated fetal VM assessment. Images from normal fetuses and fetuses with VM were randomly assigned (ratio, 4:1) to training and internal validation datasets. An independent dataset was used for external validation and comparison with radiologists with different levels of expertise. We evaluated two automated strategies: landmark localization (AI-LandVM) and segmentation-based measurement (AI-SegVM), using a U-Net architecture to estimate lateral ventricular width. Diagnostic accuracy, sensitivity, specificity, measurement error, and Bland–Altman analysis were used to evaluate the performance of the AI framework and compare it with radiologists with varying levels of experience. The AI-SegVM framework achieved high diagnostic performance, with an accuracy of 98.20%, and demonstrated excellent agreement with expert annotations, achieving a mean absolute error (MAE) of $$0.53\pm1.17$$ mm and an intra-class correlation coefficient (ICC) of 93.02%. Compared with radiologists with different levels of expertise, AI-SegVM showed lower measurement variability and consistent diagnostic performance. In contrast, the AI-LandVM strategy exhibited inferior performance, likely due to its limited ability to capture comprehensive ventricular structural information. The AI-SegVM framework provides a reliable approach for automated fetal VM measurement and diagnosis, demonstrating potential to reduce operator-dependent variability and support standardized prenatal ultrasound assessment. These findings highlight the potential of AI-assisted approaches for improving the consistency and efficiency of fetal VM evaluation.