Deep learning-based automated detection of fetal corpus callosum abnormalities in prenatal ultrasound.

Objective: Prenatal detection of corpus callosum (CC) abnormalities is essential for assessing fetal neurodevelopment, yet conventional ultrasound diagnosis faces challenges from operator variability and suboptimal fetal positioning. Methods: We developed a novel deep learning framework CC-FocusNet that integrates automated region localization with an anatomy-aware dual-stream architecture for multi-view analysis. The model was trained on 496 cases and validated on an independent external cohort of 93 cases. We assessed both diagnostic performance and clinical interpretability through attention visualization. Results: Our framework achieved 97.36% accuracy on the external test set. Grad-CAM++ heatmaps revealed that model attention consistently focused on clinically relevant anatomical landmarks, demonstrating strong interpretability. When integrated into clinical workflows, the AI system enhanced diagnostic accuracy and efficiency, particularly reducing misdiagnosis rates in challenging cases. Conclusions: This interpretable AI system provides accurate and efficient prenatal detection of CC abnormalities, offering substantial potential to support clinical decision-making and enable timely intervention for at-risk pregnancies.