Universal Zero-Shot Contour Propagation Via Prior-Guided Learning for Adaptive Radiotherapy
Abstract
Purpose
Adaptive radiotherapy (ART) requires accurate and efficient delineation of the planning target volume (PTV) on daily imaging. However, conventional automatic segmentation methods rely on large-scale, high-quality annotations, while registration-based contour propagation often loses accuracy under significant anatomical deformation. This study proposes a prior-guided zero-shot contour propagation method to automatically update PTVs from planning CT (pCT) to daily CT (dCT) without requiring daily PTV annotations.
Methods
This retrospective study analyzed four datasets from two clinical centers, including planning and daily CT scans of 409 patients (2295 CT scans for training). During training, only organs at risk (OARs) from head-and-neck, abdominal, and thoracic regions were used to guide the model in learning structural change relationships between pCT and dCT. During inference, the model transfers the image variation patterns learned from organs at risk during training to the PTV, enabling zero-shot PTV contour propagation from pCT to dCT. A multi-encoder network with a cross-attention fusion module was constructed to integrate multimodal features. Model performance was evaluated using quantitative metrics and qualitative visual analysis.
Results
On the internal test sets, high-accuracy zero-shot PTV contour propagation was achieved for nasopharyngeal carcinoma (Dice = 0.866) and colon cancer (Dice = 0.790). Without any fine-tuning, the model generalized well to external datasets, achieving Dice scores of 0.792 for external nasopharyngeal cancer and 0.713 for lung cancer. Overall, the proposed method consistently outperformed conventional segmentation-based approaches and registration-based contour propagation methods. Dosimetric evaluation further demonstrated that treatment plans generated using the predicted PTVs provided clinically acceptable target coverage and organ-at-risk sparing.
Conclusion
This study presents a prior-guided zero-shot contour propagation method that enables accurate and robust PTV propagation across disease sites. The proposed method has the potential to improve the efficiency and practicality of daily treatment planning and to facilitate the clinical implementation of ART.