FSU-YOLOv12 for vertebral abnormalities detection in fetal spinal ultrasound

Background and objectives Prenatal screening for fetal spinal abnormalities relies primarily on ultrasound, yet its accuracy is highly operator-dependent. Identifying subtle vertebral abnormalities is challenging, often leading to missed diagnoses. Artificial intelligence may improve automation and standardization, yet existing methods focus on gross fetal structures, leaving spinal pathology underexplored. This study proposes a dedicated real-time detector for vertebral abnormalities in fetal spinal ultrasound. Methods We present FSU-YOLOv12, a novel deep learning network specifically optimized for fetal spinal ultrasound that effectively alleviates the impact of small vertebrae, acoustic shadowing and artifacts, and low tissue contrast through targeted optimizations. The model was trained and validated on 968 images from 735 fetuses at Peking University Shenzhen Hospital, and independently tested on 221 images from 185 fetuses at Shenzhen Baoan Women’s and Children’s Hospital. Images included both normal spines and cases with vertebral abnormalities (hemivertebrae, butterfly vertebrae, block vertebrae). Results FSU‑YOLOv12 achieves competitive detection accuracy with 91.9% mAP 50 , 55.9% mAP 75 , and 52.3% mAP, while maintaining real‑time inference at 149 frames per second with only 2.6 M parameters and 6.0 GFLOPs. Its precision reaches 94.2%, recall 86.3%, and F1 score 90.1%. Under speckle noise and Gaussian blur, it demonstrates superior robustness, improving recall by 7.7% and 11.1% respectively over the baseline. Conclusion FSU‑YOLOv12, a real‑time detector tailored for vertebral abnormalities in fetal spinal ultrasound, combines high accuracy with low computational cost and robustness to common ultrasound artifacts, offering a practical solution to enhance prenatal screening, reduce missed diagnoses, and reduce operator dependence.