2285 - X-Ray Image-Based Real-Time Simultaneous Intrafractional Spine and Lung Tumor Tracking via Target Decomposition

Y. FU¹, P. Zhang², Q. Fan³, W. Cai¹, H. Pham², J. J. Cuaron¹, L. I. Cervino², J. M. Moran², T. Li³, and X. Li¹; ¹Memorial Sloan Kettering Cancer Center, New York, NY, ²Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, ³Memorial Sloan Kettering Cancer Center, NEW YORK, NY

Purpose/Objective(s): Given the inherent difficulties presented by X-ray projection images, including noise, low soft tissue contrast, and overlapping objects along the projection angle, accurate markerless tumor tracking during spine and lung radiotherapy poses a significant challenge. In addition, the patient’s respiratory motion and random body motion can be hard to distinguish from each other. To tackle these obstacles, we introduce an innovative deep-learning-based target decomposition technique to simultaneously track both spine and lung tumor motion in real-time. Materials/

Methods: A patient-specific model was developed to transform onboard kV projection images into decomposed target images (DTIs), enhancing target visibility. Digitally reconstructed radiographs (DRRs) from simulation CT scans were employed to simulate kV projection images obtained during treatment. The spine and lung tumor templates were derived from the planning CT through forward projection of a region of interest around them. Pairs of DRR-target template images were utilized to train a conditional generative adversarial network (cGAN). Structure-type encoding was employed to train a single model that can simultaneously decompose both the spine and lung tumors. Template matching between the target template and DTIs derived from (intrafraction motion review) IMR images was utilized for tumor tracking. The accuracy of tumor tracking was verified using a Lungman phantom with known spine and lung tumor shifts. The spine was manually shifted by up to 4mm in the superior-inferior, lateral, and vertical directions while the lung tumor was manually shifted by up to 9mm superiorly.
Results: Compared to the IMR, the DTI has increased the image correlation and the structural similarity index measure with the tumor template by approximately 83% and 75%, respectively. Our template-matching results demonstrated tumor tracking errors of 0.27±0.23mm for the spine and 0.1±0.3mm for the lung tumors. The latency for target tracking averaged approximately 100ms per frame.
Conclusion: Our target decomposition technique was able to simultaneously track the spine and lung tumors with high accuracy in a phantom. This approach significantly improved KV imaging quality, resulting in enhanced target visibility and more precise real-time markerless target tracking.