2494 - Bayesian Model-Based Internal Target Volume Accounting for Interfractional Hilum Shifts during Radiotherapy for Locally Advanced Non-Small Cell Lung Cancer
Presenter(s)
K. Ueki1, M. Nakamura2, and N. Araki1; 1Department of Radiation Oncology, National Hospital Organization Kyoto Medical Center, Kyoto, Japan, 2Department of Advanced Medical Physics, Graduate School of Medicine, Kyoto University, Kyoto, Japan
Purpose/Objective(s): This study aimed to develop a prediction model-based internal target volume (ITVpred) for thoracic lymph node metastases, designed to account for interfractional hilum shifts during conventionally fractionated radiotherapy for locally advanced non-small cell lung cancer.
Materials/Methods: This exploratory study first examined interfractional shifts in thoracic lymph nodes using two landmarks, the carina (C) and the hilum closest to the primary tumor (H), in 23 patients. The cohort included patients who received 2 Gy per fraction for 30 or 35 fractions, with only the first 30 analyzed for all. After vertebrae-based registration of daily CBCT images, shifts were measured as the distance between the first CBCT (C1, H1) and each subsequent fraction (2–30). Based on these analyses, we selected five patients with larger hilum shifts for modeling analysis. The model predicted each patient’s vector H1Hmean (mean H position across sessions 2–30 relative to H1). This vector was formulated as a combination of unit vectors: H1C1 (H1 to C1), H1G (H1 to centroid of the total GTV, including primary and nodal tumor), and their cross product (H1C1 × H1G), scaled by coefficients as linear functions of GTVprox, where GTVprox was defined as the volume of any GTV within 2 cm of the proximal bronchial tree. H1Hmean = (a0 + a1·GTVprox) H1C1 + (b0 + b1·GTVprox) H1G + (c0 + c1·GTVprox) (H1C1 × H1G). A Bayesian framework with Markov chain Monte Carlo sampling estimated the coefficients. The observed H1Hmean shifts were fitted to a multivariate normal distribution, with convergence was assessed using the Gelman-Rubin statistic (R-hat = 1.01). The predictive performance of the model was evaluated using the mean absolute error (MAE) between the predicted and observed vectors. As the first step in generating ITVpred, the CTV node was created as a 5 mm isotropic expansion of the GTV node, adjusted for anatomy. A 95% confidence ellipsoid derived from the Bayesian model’s posterior covariance was applied to each CTV point with a scaling factor based on its spatial relationship to anatomical landmarks. Scaling factors varied continuously, approaching 1 near H1 and 0 near C1 and the vertebral body. The posterior samples, initially scattered in space, were mapped onto planning CT slices, and their outer boundaries defined the ITVpred.
Results: The five patients had the median 3-dimensional hilum shift distance of 6.5 mm (range: 5.3–15.4 mm) and the median carina shift distance of 4.6 mm (range: 2.3–7.8 mm). The MAE for the predicted shifts was 0.83 mm in distance and 27.2° in angle. The ITVpred was patient-specific and exhibited directional characteristics.
Conclusion: We proposed a data-driven approach using a Bayesian framework to generate an ITV that accounts for interfractional hilum shifts. Future research will explore geometric and dosimetric evaluation.