A cascade deep learning network for simultaneous prediction of dose and linear energy transfer in proton therapy of prostate cancer

OBJECTIVE: Prediction of dose-averaged linear energy transfer (LETd) may facilitate LET-aware evaluation of proton therapy plans. This study developed a cascade deep learning framework for simultaneous three-dimensional prediction of dose and LETd distributions for conventional dose-based optimization (DoseOpt) and LETd-incorporated optimization (LETOpt) in prostate intensity-modulated proton therapy. APPROACH: A total of 150 patients with prostate cancer were included, each with DoseOpt and LETOpt plans. Planning CT images and anatomical structure masks were used as inputs. The cascade network transferred an intermediate latent representation from the dose prediction pathway to the LETd prediction pathway. Stepwise, multitask, and anatomy-only LETd (only-LET) networks served as comparators. Model performance was evaluated using five-fold cross-validation. MAIN RESULTS: All dose-predicting models showed high spatial agreement, with cascade Dice similarity coefficients of approximately 0.93, 0.96, and 0.95 for the 50%, 75%, and 95% prescription isodose volumes. No significant differences in structure-specific dose errors remained after false discovery rate correction. The cascade model achieved the lowest global LETd MAE for DoseOpt and LETOpt, 0.108 ± 0.025 and 0.111 ± 0.025 keV/μm, respectively, significantly lower than those of the multitask, stepwise, and only-LET models. Structure-specific LETd performance was endpoint dependent, with favorable cascade performance for several CTV and organ-at-risk endpoints. LETOpt primarily redistributed LETd while largely preserving dose, and the cascade model reproduced the principal plan-specific changes, including reductions in rectal and rectal-wall LETd. Mean inference time was 0.505 ± 0.020 s per patient. SIGNIFICANCE: The proposed framework enables rapid simultaneous prediction of dose and LETd for predefined planning strategies from patient anatomy without requiring a precomputed dose distribution, supporting LET-aware plan comparison and providing a basis for future LETd-guided biological optimization.