Abstract
Accurate classification of pubertal breast development is essential for evaluating pubertal timing and disorders. Breast ultrasound offers an objective and noninvasive approach that complements clinical staging but is limited by interobserver variability. This multicenter retrospective study included 2,576 ultrasound images from three hospitals to develop and validate a hybrid deep learning model, STransXNet, for automatic staging of pubertal breast development. The model was trained on internal data and tested on both internal and external cohorts. STransXNet achieved an accuracy of 88.0% on the internal test set, outperforming baseline models and resident radiologists, and showed robust generalizability across external datasets and imaging systems. Importantly, it demonstrated superior performance in differentiating intermediate stages, which are most challenging in clinical practice. These findings indicate that STransXNet can provide standardized and reproducible assessment of pubertal breast development and may support clinical decision-making in settings with limited radiological expertise.