Abstract
Background and purpose
Surface-guided deep inspiration breath-hold radiotherapy uses deep inspiration breath-hold CT for treatment planning, whereas a free-breathing body contour is needed for baseline surface registration. This requires an additional free-breathing CT acquisition. This study investigated the feasibility of generating a three-dimensional free-breathing body contour directly from routine biplanar CT scout images using deep learning.Material and methods
A total of 173 thoracic CT-simulation studies with paired coronal and sagittal CT scout images and corresponding free-breathing CT images were retrospectively collected. The CT-derived body mask was generated using an automated threshold- and morphology-based procedure and used as the reference contour. The dataset was divided into training, validation, and test cohorts of 121, 26, and 26 studies, respectively. Geometric performance was evaluated using the Dice coefficient, 95th percentile Hausdorff distance, and mean surface distance.Results
In the test cohort, the predicted contours demonstrated high geometric agreement with the reference contours, with a Dice coefficient of 0.979 ± 0.017, 95th percentile Hausdorff distance of 4.23 ± 2.87 mm, and mean surface distance of 1.05 ± 0.73 mm. Slice-wise analysis showed similar performance. Larger and more symmetric, geometrically regular torso shapes were associated with better performance.Conclusions
A three-dimensional free-breathing body contour was generated from routine biplanar CT scout images with high geometric agreement to the CT-derived reference contour. This approach may reduce the need for additional free-breathing CT acquisition for patient setup, but prospective validation in clinical setup, registration, and gating workflows is required.