An automated method for precise definition of fracture reduction targets

Anatomical reduction is essential for the treatment of long bone fractures, while well-defined reduction targets are critical for ensuring reliable outcomes. However, due to the subjective nature of physician experience and the structural deficiencies often found in imaging data, the efficient and high-quality definition of reduction targets remains a significant challenge. This paper proposes a feature dynamic fusion network model (FDF-Net) for target definition. FDF-Net adopts a dual-branch heterogeneous encoding-joint decoding architecture to construct an intelligent prediction system with bone structure perception capabilities. Through a feature fusion mechanism guided by skeletal repositioning prior knowledge, combined with geometry-enhanced point cloud feature encoding, and utilizing decoupled dual prediction heads for explicit modeling of rigid body transformation physical constraints, the model achieves precise mapping from raw point clouds to repositioning parameters. Comparative experiments against state-of-the-art neural network models demonstrated that FDF-Net achieved superior performance. On the simulated dataset, the model attained mean errors of 0.653 ± 0.387 mm in translation and 0.560 ± 0.301° in rotation. On the clinical dataset, the corresponding errors were 1.934 ± 1.146 mm and 0.829 ± 0.386°. FDF-Net enables precise and robust definition of reduction targets and demonstrates potential for translation into clinical practice.