Cross-Guided Dual-View Pre-Training for Fracture-Healing Assessment on Orthogonal Radiographs

Radiographic fracture-healing assessment is subjective and time-consuming, and existing automated methods often underuse the complementary information in paired anteroposterior (AP) and lateral (LAT) radiographs. To address this limitation, we propose Dual-View Fracture Scoring (DV-FraS), a deep learning framework for automated cortex-level fracture-healing assessment from paired orthogonal radiographs. DV-FraS implements an automated end-to-end workflow that standardizes fracture localization, performs cross-guided orthogonal multi-view 2D representation learning, and predicts cortex-level modified radiographic union score for tibia (mRUST) from paired AP and LAT radiographs. In the representation-learning stage, a Dual-View Cross-Guided Masked Autoencoder (DC-MAE) reconstructs masked anatomical information across orthogonal views, encouraging view-consistent and complementary fracture representations. DV-FraS was evaluated using murine femur-fracture radiographs and a prospective human fracture cohort. On the murine held-out test set, it achieved a mean absolute error (MAE) of 0.042, Top-1 accuracy (Top-1 Acc.) of 97.07%, and an absolute-agreement intraclass correlation coefficient [ICC(A,1)] of 0.980. In human radiographs, DV-FraS achieved a mean absolute error of 0.192 and Top-1 Acc. of 89.47%, with statistically significant Top-1 Acc. gains over the leading dual-view competitors. Moreover, its predictions also preserved age-dependent longitudinal healing patterns. These findings suggest that DV-FraS provides an automated and clinically aligned framework that may improve the objectivity, consistency, and efficiency of radiographic fracture-healing assessment.