Author: Steve B. Jiang, Ruiqi Li, Hua-Chieh Shao, Kenneth Westover, You Zhang, Tingliang Zhuang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab & Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose:
Respiratory motion is a long-standing challenge for lung SBRT, particularly for centrally-located lung tumors where increased toxicity demands more precise motion management during treatment. Current two-dimensional (2D) imaging approaches are insufficient for 3D tumor deformable motion tracking. This study aims to demonstrate the feasibility of transforming a surface imaging system into a 3D real-time imaging solution.
Methods:
We developed a Surface-derived Three-dimensional AI-driven Real-time (STAR) imaging system, by integrating two key components: 1) PMF-DREMM: a machine-learning sub-system for pre-treatment dynamic CBCT reconstruction and motion modeling; and 2) Surf2DefNet: a deep-learning model that correlates intra-treatment body surface images with internal 3D anatomy and motion fields, trained based on dynamic CBCT and motion model output of PMF-DREMM. Specifically, PMR-DREMM reconstructs a reference CBCT and solves a motion eigenvector-based motion model from a pre-treatment standard CBCT scan. By solving projection-specific weighting factors for motion eigenvectors, the motion model yields dynamic motion vector fields (MVFs) to deform the reference CBCT to a motion-resolved CBCT for each X-ray projection. Based on surface images extracted from PMF-DREMM-solved dynamic CBCTs, a Surf2DefNet was trained to predict the dynamic CBCTs’ corresponding motion eigen-vector weightings from surface images, enabling it to infer real-time CBCTs and MVFs using intra-treatment surface maps acquired later. We tested STAR using both a digital XCAT phantom and patient data. The relative error (RE) and tumor center-of-mass error (COME) metrics between STAR images and the ‘ground truth’ were evaluated.
Results:
The RE/COME for XCAT with regular and irregular motion are 0.15/0.74 mm and 0.15/0.96 mm respectively. For patients, using dynamic CBCT as ‘ground truth’, the RE/COME are 0.07/0.88 mm.
Conclusion:
The STAR imaging system can achieve accurate spatiotemporal reconstructions from surface images, yielding CBCTs and MVFs for intra-treatment real-time image-guidance, and has the potential to improve safety and efficacy of SBRT for lung cancer.