From Segmentation to Modeling Tumor Evolution: Trajectory-Aware Time-Series Learning (TRASE) for Adaptive Radiotherapy
Abstract
Purpose
Accurate delineation of the gross tumor volume of the primary lesion (GTVp) during radiotherapy is essential for ART, particularly in head-and-neck cancer where tumor regression and anatomical deformation occur during treatment. However, mid-treatment contouring remains challenging due to limited soft-tissue contrast in CBCT and the inability of conventional methods to model tumor evolution. This study proposes Trajectory-Aware Time-Series Learning (TRASE), a framework that formulates mid-treatment GTVp delineation as tumor evolution modeling rather than static segmentation.
Methods
A total of 101 patients with planning CT at week-0 and weekly CBCT during treatment were split into 65/15/20% training, validation, and test cohorts. TRASE models temporal tumor morphology evolution by integrating deformation-based shape propagation with longitudinal image-driven feature learning. Baseline GTVp geometry is propagated toward the target week to establish a warp-first morphological evolution backbone. Longitudinal CBCT image patches from weeks 1–3 are stacked as input channels and temporally integrated through patch-based convolutional representations to extract features capturing tumor shape and appearance changes, with inter-week intensity differences encoding evolution. Region-of-interest annotations at weeks 1–3 were used only during training. Performance was evaluated using the Dice similarity coefficient (DSC) and average surface distance (ASD) and compared with DIR, U-Net, and nnU-Net.
Results
At mid-treatment (week-4), TRASE achieved a DSC of 0.73 ± 0.03 and an ASD of 0.94 ± 0.08 mm, significantly outperforming DIR (DSC 0.44 ± 0.03, ASD 2.38 ± 0.11 mm), U-Net (DSC 0.34 ± 0.05, ASD 5.39 ± 0.68 mm), and nnU-Net (DSC 0.47 ± 0.03, ASD 5.75 ± 0.51 mm). Qualitative evaluation showed that TRASE produced evolution-consistent contours with stable boundaries, whereas conventional methods showed boundary degradation.
Conclusion
By explicitly modeling tumor evolution as a temporal trajectory, TRASE enables accurate and robust mid-treatment GTVp prediction without additional contouring workload, providing a physics-consistent and clinically practical foundation for ART.