Feature-enhanced respiratory motion estimation driven by spatial information compensation

The nonrigid motion of lung tissue caused by respiration represents a primary challenge impeding the safety of robotic interventional therapy. Understanding the intrinsic patterns of respiratory motion and achieving precise estimation are critical for optimizing interventional pathways in lung cancer treatment. Some methods relying on internal and external signals only provide localized descriptions of motion states within regions of interest, and their accuracy remains constrained. Traditional image-based approaches enable global motion modeling but are hindered by high computational complexity and extended processing times. Deep learning-based respiratory motion estimation offers a promising solution to reconcile the trade-off between accuracy and computational efficiency. To enhance estimation precision and address the challenges of large deformation motion, we present the Spatial Information Compensation and Enhancement Network (SICE). A novel spatial information compensation module additionally extracts global and local spatial features to compensate for information loss during estimation, thereby enhancing the global representational capacity of the model. Furthermore, a multi-scale feature fusion mechanism combining convolution and self-attention is established to facilitate effective interaction between features of different resolutions. Validation on 4DCT datasets demonstrates that the proposed framework achieves both high accuracy and computational efficiency.