Feasibility of markerless real‐time tumor tracking volumetric modulated arc therapy using diaphragm‐based respiratory offset vectors

Abstract Background Real‐time tumor tracking (RTTT) is an effective strategy for managing respiratory motion during radiation therapy; however, it typically requires invasive fiducial markers. As a non‐invasive alternative, markerless RTTT (ML‐RTTT) has gained increasing interest, especially with the growing use of volumetric modulated arc therapy (VMAT), which requires efficient real‐time motion compensation. Purpose This study aimed to develop and evaluate a tumor position estimation algorithm for ML‐RTTT integrated with VMAT (ML‐RTTT‐VMAT), using the diaphragm dome projection as a surrogate, with an added method to compensate for phase‐dependent asynchrony between diaphragm and tumor motion. Methods The analysis included 43 sessions performed on 23 patients whose tumors exhibited motion greater than 10 mm and who had fiducial markers implanted near the tumors. Rotational kV x‐ray images were acquired from two orthogonal directions during free breathing. Infrared reflective (IR) markers were placed on the abdomen near or above the umbilicus to track respiratory motion. Diaphragm templates were generated from planning CT images, and diaphragm positions were identified by template matching with epipolar geometry. The 3D diaphragm dome position was computed via triangulation. The initial 20 s of each session were used to build a prediction model linking the IR marker signal to tumor position. Tumor position was approximated using the centroid of the implanted fiducial markers (pseudo tumor). Three scenarios were evaluated: (i) S no , a baseline scenario assuming direct availability of the pseudo‐tumor position without diaphragm detection or offset‐vector estimation; (ii) S con , a constant offset vector defined at an end‐expiratory phase and applied to all phases; and (iii) S var , a variable offset vector dependent on respiratory phase. The model was validated using the remaining 50 s of data, and the prediction accuracy of each model was assessed using the 90th percentile error ( E 90 ) in each direction. Results The E 90 values for left–right, superior–inferior, and anterior–posterior directions were as follows: S no : 2.4, 4.1, and 4.0 mm; S con : 2.7, 5.7, and 4.5 mm; and S var : 2.3, 4.5, and 3.7 mm. Large errors were observed in six sessions (14.0% of all S var sessions), mainly due to irregular respiratory patterns or pulsation artifacts in the IR marker signal. Conclusions The proposed ML‐RTTT‐VMAT approach using diaphragm‐based prediction with a phase‐dependent offset vector is feasible and holds promise for markerless real‐time motion management in radiation therapy.