Quality Assurance of Brachytherapy Treatment Planning Based on Template, Classification, and Segmentation Methods
Year: 2026
Type: preprint
Abstract: Background and purpose Accurate applicator reconstruction is essential for dose calculation in CT-based brachytherapy. Reconstruction errors—driven by limited CT slice thickness and metal artifacts—can produce clinically relevant dose deviations. We developed a fully automated quality assurance (QA) pipeline that combines deep learning classification and segmentation with high-resolution template registration to verify the geometric accuracy of planned source dwell positions. Materials and methods We retrospectively collected 3,702 treatment plans from 742 gynecological cancer patients. A lightweight 3D convolutional neural network classified applicators into five categories, and an nnU-Net (v2) model segmented the three channels of Fletcher applicators. High-resolution (0.6-mm) template centerlines were registered to 3-mm clinical CT images using a progressive weighted arc-segment Kabsch algorithm. Geometric consistency was quantified with point-to-curve Euclidean distances. Error-detection performance was assessed in a simulation study in which one randomly selected dwell position per plan was displaced by 1, 3, or 5 mm (50 plans per magnitude), with 50 error-free plans as controls. Results The classification model achieved 99.28% accuracy and a weighted F1 score of 0.99. The segmentation model yielded a mean Dice coefficient of 90.7% and a 95% Hausdorff distance of approximately 1 mm. Mean distances between auto-segmented centerlines and the registered template curve ranged from 0.39 to 0.70 mm; mean dwell-point-to-template distances were 0.34–0.51 mm. In the simulation study, specificity was 94.0%; sensitivity was 36.0% for 1-mm displacements and 100% for both 3-mm and 5-mm displacements. Conclusion The proposed pipeline enables automated, high-precision QA of brachytherapy dwell positions and reliably detects displacements of 3 mm or larger, providing an objective tool for clinical treatment-plan verification that mitigates thick-slice artifacts through high-resolution template registration.
Source: Research Square
Authors: Tianlong ji, Wenjun Tan, Dazhe zhao
Institution: Northeastern University
Language: English
Subfield: Radiation
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236864709
