Real-time marker tracking for MV treatment beam imaging

A new approach to obtain real-time positions of fiducial markers in MV treatment beam images is proposed. The MV images are firstly preprocessed to enhance contrast in the treatment field. To deal with large variations in projected marker shape, we propose a learning-based approach to detect marker locations in the enhanced MV images. We also show that marker tracking can be accomplished much more efficiently and reliably by exploiting temporal correlation between consecutive MV images. Thus, the proposed framework can accurately localize multiple markers in low-contrast MV images while satisfying the real-time constraints imposed by the IGRT. Our method has been validated using manual marker annotations as ground-truth. The marker detection rate on patient images was at least 96% for the cases collected from multiple treatment fractions.