Interpretable automated cobb angle measurement in adolescent idiopathic scoliosis using a lightweight keypoint-to-geometry network: cross-center external validation

Adolescent idiopathic scoliosis (AIS) is routinely quantified on standing coronal radiographs using the Cobb angle; however, manual measurement is time-consuming and subject to intra-observer variability. We propose SpineNET, a lightweight, interpretable keypoint-based framework for automated scoliosis assessment on standing whole-spine posteroanterior (PA) radiographs. SpineNET is built on the YOLOv13 paradigm and enhances cross-scale structural modeling by replacing DS-C3k2 modules with Star-Blocks throughout the backbone and neck, while a lightweight shared-convolution detection head reduces redundancy across feature pyramid levels. The model is trained using only vertebral and pelvic bounding boxes and anatomical keypoints, without any Cobb angle supervision. During inference, Cobb angles as well as shoulder and pelvic coronal tilt angles are computed via a transparent keypoints-to-geometry pipeline with clinically auditable visual overlays. In a cross-center external validation (1,200 radiographs for training/validation and 150 independent external test radiographs), SpineNET achieved 94.2% mAP@50 and 85.7% mAP@50-95 with 1.7M parameters and 4.3B FLOPs. Compared with the consensus of three senior spine surgeons, SpineNET yielded a Cobb angle mean absolute error (MAE) of 1.44° and intraclass correlation coefficient (ICC) of 0.970, with minimal bias (0.03°) and narrow 95% limits of agreement. These results support SpineNET as an efficient and clinically interpretable measurement-assisted tool for high-throughput AIS screening and follow-up in resource-constrained settings.