Quantifying Primary and Lymph Node Tumor Volume Reduction and Mass Loss during Radiotherapy
Abstract
Purpose
Early assessment of radiotherapy response is essential for adaptive treatment planning and digital-twin development. Longitudinal quantification of tumor volume and mass changes depends on reliable deformable image registration (DIR), which remains challenging. This study compared an AI-based DIR model (ProRSeg) with ANTs for CT alignment in head and neck cancer (HNC) and evaluated deformation-derived quantitative features.
Methods
CT data from 104 HNC patients were analyzed, with scans acquired prior to radiotherapy and approximately three weeks after treatment initiation. Following rigid alignment, deformable registration was performed using ANTs and ProRSeg, a framework based on a 3D convolutional LSTM integrated with a U-Net encoder (Jiang et al., Med Phys), fine-tuned on head and neck anatomy using 62 patients. Registration accuracy was evaluated using Dice similarity coefficient (DSC) and 95th-percentile Hausdorff distance (HD95). Deformation vector fields were used to compute voxel-wise Jacobian maps, and summary Jacobian metrics within tumor regions were compared with percent volume and mass change per day derived from longitudinal segmentations.
Results
The fine-tuned ProRSeg model achieved higher registration accuracy than ANTs, with improved DSC (0.74 vs. 0.62) and lower HD95 (5.0 mm vs. 6.5 mm). Primary tumors showed mean volume(mass) reductions of 0.4%(0.5%)/day, respectively, while lymph nodes showed comparable reductions of 0.5%(0.5%)/day. Volume and mass loss were highly variable, with coefficients of variation of 1.4 (volume) and 1.3 (mass) for primary tumors, compared with 1.1 and 1.0 for lymph nodes. Volume and mass changes were strongly coupled. Jacobian mean/median correlated with lymph node volume and mass loss (Spearman ρ ≈ 0.5*, p < 0.001), but not with primary tumor changes.
Conclusion
ProRSeg applied to mid-treatment HNC CT images provided results consistent with ground-truth, resulting in quantitative estimates of tumor specific volume and mass loss. Future work will quantify the predictive value of these parameters, supporting digital-twin–based adaptive radiotherapy.