A novel multi-task deep learning framework for classification and detection of intracranial structures in first-trimester fetal ultrasound images

INTRODUCTION: To develop a novel deep learning framework for automated classification and detection of fetal intracranial structures in first-trimester ultrasound images. METHODS: A total of 2569 sagittal ultrasound images from 2,560 pregnant women (11-14 weeks' gestation) were retrospectively collected from four clinical centers. All images were manually annotated for nine key fetal brain structures and underwent preprocessing (grayscale conversion, contrast enhancement, and augmentation). Two classification strategies were evaluated using EfficientNet and VGG16: end-to-end training and deep feature extraction. LASSO regression was used to select relevant features and generate a deep feature score (DFS), which was integrated with clinical predictors into a nomogram. YOLOv11 and the Swin Transformer were employed for multi-label detection of intracranial structures. Model performance was evaluated using accuracy, AUC, IoU, mean Average Precision (mAP), calibration curves, and decision curve analysis (DCA). RESULTS: Out of 2,569 images, 1,670 (65%) were classified as standard diagnostic frames. EfficientNet achieved the highest classification performance (AUC: 0.95 training, 0.93 testing), outperforming VGG16. Deep feature extraction with LASSO and nomogram analysis significantly improved classification accuracy. For detection, the Swin Transformer achieved superior results (mAP: 0.940, IoU: 0.923) compared to YOLOv11 (mAP: 0.825, IoU: 0.810). DCA demonstrated greater clinical benefit for the deep feature-based model. CONCLUSION: This is the first study to integrate deep feature selection, statistical modeling, and transformer-based detection for early fetal brain assessment. The proposed framework enhances classification accuracy, detection performance, and clinical interpretability, offering a robust tool for first-trimester prenatal screening.