Validation of an AI-Empowered Markerless Lung Tumor Tracking Method Using a Dynamic Chest Motion Phantom
Abstract
Purpose
Markerless lung tumor tracking has the potential to reduce target margins and improve organ-at-risk (OAR) sparing during radiotherapy. We previously proposed a deep learning–based target decomposition approach for real-time markerless lung tumor tracking. This study quantitatively validates the tracking accuracy of the proposed method using a dynamic chest phantom under multiple realistic respiratory motion patterns.
Methods
An AZ-M2 chest motion phantom, consisting of a mobile lung tumor and pulmonary vessels embedded within a static thorax (heart, spine, and ribs), was used to simulate respiratory motion. During CT simulation, free-breathing and 4DCT scans were acquired with the tumor moving longitudinally according to a sinusoidal waveform. A patient-specific deep learning model was trained using pairs of digitally reconstructed radiographs (DRRs) and decomposed target images (DTIs) to transform raw kV projection images into synthetic DTIs with enhanced tumor contrast and suppressed bony anatomy. During treatment delivery, the tumor motion in the phantom was programmed to follow: (1) a standard sinusoidal motion profile, and (2) three free-breathing (FB) and three deep inspiration breath-hold (DIBH) motion waveforms acquired from patients using a Real-time Position Management system. Tumor motion was tracked on synthetic DTIs using template matching, and the resulting trajectories were quantitatively compared with the phantom’s programmed ground-truth motion.
Results
The proposed framework achieved maximum tracking errors of 1.14 mm (with a Mean Absolute Error of 0.231±0.190 mm) for sinusoidal motion, 0.90 mm (0.306±0.196mm) for the FB waveforms, and 0.83 mm(0.268±0.169mm) for the DIBH waveforms. The end-to-end system latency was less than 100 ms, supporting feasibility for real-time clinical implementation.
Conclusion
Algorithm validation using the AZ-M2 chest motion phantom demonstrates that the proposed target-decomposition-based tracking framework achieves sub-millimeter accuracy across a range of realistic respiratory motion patterns. These results support its potential for real-time markerless lung tumor tracking during radiotherapy.