Image-guided radiation therapy (IGRT) is clinically appealing as it offers real-time patient specific image feedback and anatomical variation adaptation. Correcting non-linear misalignment of cone-beam CT (CBCT) to simulated 4DCT projection image intra-fractionally and inter-fractionally provides an opportunity to increase precision and accuracy of target localization. Taking advantage of on-board CBCT to perform deformable image registration (DIR), implemented on GPUs, allows for real-time and retrospective anatomical tracking. Increasing the spatial congruence of the on board imager to the simulation can allow for dose escalation using hypo-fractionated regimens in addition to dose reduction to organs at risk and improved target dose coverage, increasing local control without increasing toxicity. Robust respiratory signal extraction is crucial in pulmonary lesson localization. This work demonstrates the evaluation and application of DIR implemented on GPUs as a tool for IGRT. We first evaluate the performance of a realtime lung tumor tracking technique using DIR between simulated 4DCT and CBCT projections as a means to compensate for respiratory motion. Then we assess the performance of a respiratory signal extraction technique using DIR between consecutive CBCT projections. Finally, we test lower dimensional modeling using principal component analysis (PCA) as a means for 4D thoracic reconstruction and establish a correlation between respiratory modes and PCA eigenvalue decay rate.