Hierarchical Spatial and Temporal Aggregation As Progressive Denoising for Longitudinal CBCT Radiomics In Lung Cancer
Abstract
Purpose
Longitudinal CBCT radiomics acquired during radiotherapy suffers from temporal variability. We developed a novel cumulative spatial and temporal aggregation framework (CBCTc) to optimize both prediction performance and stability for lung cancer survival prediction.
Methods
We analyzed 225 radiotherapy courses from 189 lung cancer patients with weekly CBCT imaging over six treatment weeks. A total of 107 radiomics features were extracted at the lesion level, while 14 clinical variables were collected at the patient level. We developed a hierarchical aggregation pipeline to process multi-level and longitudinal data. The lesion-level radiomics were aggregated to the patient level using robust statistics (e.g., largest lesion, volume-weighted mean). We then implemented a cumulative temporal aggregation framework (CBCTc1 to CBCTc6). Survival prediction was modeled using Cox Proportional Hazards, Random Survival Forest (RSF), and Gradient Boosting Survival (GBS). Each cumulative model incorporated all prior weeks’ features using robust summary statistics. Models were evaluated using 5-fold patient-level cross-validation, with performance assessed by the Concordance Index (C-index) and stability by the Coefficient of Variation (CV).
Results
Cumulative temporal aggregation produced a monotonic reduction in variability, with CV decreasing from 8.47% at CBCTc1 to 1.97% at CBCTc6. The best configuration achieved C-index=0.719 with standard deviation 0.0141 (CBCTc6, GBS, mean aggregation). CBCTc6 provided 59% lower variance than CBCTc5 with equivalent performance (p=0.87). Multi-level aggregation improved signal-to-noise ratio from lesion-level C-index 0.64 to patient-level 0.72. Clinical feature fusion provided 11.4% improvement over radiomics-only. CBCTc6 outperformed clinical-only (17.0%), pre-treatment CT (8.8%), delta radiomics (3.4%), and deep learning features (5.6%).
Conclusion
Hierarchical cumulative aggregation of longitudinal CBCT radiomics enables progressive denoising across treatment, yielding robust survival prediction with low variability. By prioritizing stability (CV < 2%), end-of-treatment cumulative aggregation (CBCTc6) represents the optimal strategy for clinical deployment of CBCT-based prognostic models in lung cancer.