Author: Noriko Kishi, Takashi Mizowaki, Mitsuhiro Nakamura, Yukine Shimizu 👨🔬
Affiliation: Kyoto University, Kyoto Univercity, Department of Radiation Oncology and Image-applied Therapy, Kyoto University Graduate School of Medicine, Department of Radiation Oncology and Image-applied Therapy, Graduate School of Medicine, Kyoto University 🌍
Purpose: To predict tumor positions in markerless dynamic tumor tracking (ML-DTT)-VMAT by compensating for the asynchrony between the tumor and the diaphragm.
Methods: Rotational fluoroscopic X-ray images were acquired for 70 seconds from two orthogonal directions across 39 sessions from 21 cases with a moving tumor. For each case, DRRs were generated from planning 4D-CT images to simulate the fluoroscopic images, and diaphragm templates were constructed. The first 20 seconds of data were used for training, during which diaphragm positions in the fluoroscopic images were detected via template matching utilizing epipolar line geometry. The 3D diaphragm position was subsequently calculated using triangulation. Tumor position was estimated under two scenarios: Scenario A (Svar), in which a variable offset vector between the pseudo tumor (the fiducial marker centroid) and diaphragm, based on the respiratory phase, was applied; and Scenario B (Sno), in which the pseudo tumor was directly detected without the application of any offset vector. A prediction model was then developed incorporating tumor positions from each scenario along with IR marker signals. Prediction accuracy was evaluated using the remaining 50 seconds of data. The pseudo tumor was considered the ground truth, and the 95th percentile value of the tumor prediction errors (E95) was calculated for each scenario.
Results: The E95 values for LR, SI, and AP directions across all sessions were as follows; Svar: 2.7, 7.2, and 4.5 mm; and Sno: 2.8, 6.4, and 4.5 mm. In the Svar, five sessions exceeded the E95 across all sessions. In these sessions, the prediction accuracy in Svar was similar to that in Sno due to the irregularity of the IR marker signal.
Conclusion: The integration of diaphragm detection as a surrogate, along with a variable offset vector tailored to the respiratory phase, demonstrates the feasibility of achieving ML-DTT-VMAT.