Ultrasound radiomics and deep learning for predicting antral follicle count and anti-Müllerian hormone

To overcome inter-observer variability in conventional antral follicle count (AFC) assessment and AMH testing limitations, we developed an AI-powered framework using routine 2D ultrasound to standardize ovarian-reserve evaluation in assisted reproductive technology (ART). This multicenter retrospective study analyzed 395 women with infertility from two affiliated hospitals. The cohort was divided into training (n = 210), internal-test (n = 91), and external-test (n = 94) cohorts. We established three prediction models: radiomics model, 674 IBSI-compliant features; deep-learning model, ResNet50-based feature extraction; fusion model, hybrid approach combining both modalities. Model performance was validated against the manual AFC and serum AMH levels. Sequential classification categorized patients into low, moderate, or high ovarian-response risk groups. Strong correlation and consistency existed between routine 2D ultrasound image AFCs and three-dimensional dynamic-scan AFCs. The deep learning-radiomics fusion model displayed superior AFC prediction (R²=0.743 internal/0.583 external), surpassing the performance of single-modality models (radiomics: 0.586/0.572; deep learning: 0.737/0.541). For AMH prediction, the fusion model maintained generalizability (external R²=0.509 vs. 0.420 radiomics and 0.352 deep learning, p < 0.05). In ovarian-response stratification, the fusion model achieved an AUC of 0.881 (95%CI: 0.828-0.925), which was 8.0% higher than that of individual models, with 69.1% sensitivity and 84.6% specificity for identifying high-risk patients requiring stimulation-protocol modifications. The developed AI framework enables standardized ovarian-reserve evaluation using routine 2D ultrasound, effectively bridging imaging limitations by synergizing radiomics and deep learning. Meanwhile, the model achieves clinical applicability by enabling personalized ovarian-stimulation protocol optimization, demonstrating particular value in resource-limited clinical environments without requiring advanced imaging infrastructure.