Multitask learning for early treatment response and survival prediction in lung cancer radiotherapy using sequential CBCT imaging.

Background: Adaptive radiotherapy requires early prediction of treatment response and survival, yet the optimal use of longitudinal cone-beam computed tomography (CBCT) imaging and modeling strategy for small cohorts remains unclear. Purpose: To develop a multitask deep learning framework for simultaneous treatment response classification and progression-free survival (PFS) prediction using planning computed tomography (CT), dose, and sequential CBCT, and to evaluate the impact of including different numbers of early CBCT time points on prediction performance within an incremental analysis framework. Methods: A total of 142 lung cancer patients were retrospectively analyzed. A lightweight network with cross-modal attention fusion and dual task heads for treatment response classification and Cox-based survival prediction was trained end-to-end and benchmarked against 12 baseline methods via 5-fold cross-validation. Results: Using only the first on-treatment CBCT, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.858 ± 0.078 and a concordance index (C-index) of 0.672 ± 0.058, outperforming all baselines. Ablation analysis confirmed the contributions of attention fusion and multitask training. Additional CBCT scans degraded classification performance, likely due to increased dimensionality relative to the cohort size. Conclusions: The first on-treatment CBCT provided the strongest early predictive signal for treatment response classification in this cohort; adding subsequent scans degraded classification performance and produced only a marginal change in survival concordance under the current study setting.